{"cssClassNames":"page basicpage summit-page","templateName":"quickstart-page-template","canonicalLink":"https://www.snowflake.com/en/developers/guides/credit-risk-pd-model-snowflake-ml/","robotsTags":[],"allowedRenditionsWidth":["320","480","640","768","960","1200","1440","1920"],"language":"en","title":"From SAS to Snowflake: Productionizing Credit Risk Models on a Modern ML Platform","analyticsPageType":"developer","analyticsCategory":"developers","analyticsSubCategory":"guides","excludeFromAnalytics":false,":items":{"root":{"columnCount":12,"columnClassNames":{"markup_editor_1950346551":"aem-GridColumn aem-GridColumn--default--12","experiencefragment-banner":"aem-GridColumn aem-GridColumn--default--12","experiencefragment-header":"aem-GridColumn aem-GridColumn--default--12","responsivegrid":"aem-GridColumn aem-GridColumn--default--12","experiencefragment-footer":"aem-GridColumn aem-GridColumn--default--12","modal_container":"aem-GridColumn aem-GridColumn--default--12","markup_editor":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12",":items":{"experiencefragment-banner":{"id":"experiencefragment-ec986d1b8d","localizedFragmentVariationPath":"/content/experience-fragments/snowflake-site/language-masters/en/site/pushdown-banner/master/jcr:content","configured":true,":items":{"root":{"layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"pushdown_banner_copy_1875701434":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-0505492cbe",":items":{"pushdown_banner_copy_1875701434":{"id":"pushdown-banner-36a4200e1d","contentHeadline":"Expedition. Free, on-demand.","contentDescription":"Technical tracks for practitioners, outcomes for leaders.","contentJustifyContent":"center","linkStyle":"text-white","linkCTA":{"id":"link-cta","heapButtonClasses":["pushdown_banner"],"showOutboundIcon":false,"buttonLink":{"valid":true,"attributes":{"target":"_blank"},"url":"/en/expedition/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Watch now"},":type":"snowflake-site/components/pushdown-banner","appliedCssClassNames":"snowflake-pushdown-banner-text-white snowflake-pushdown-banner-background-black"}},":type":"snowflake-site/components/container",":itemsOrder":["pushdown_banner_copy_1875701434"]},"image":{":type":"nt:unstructured"},"cq:metadata":{":type":"nt:unstructured"}},":type":"snowflake-site/components/experiencefragment","classNames":"aem-xf",":itemsOrder":["root","image","cq:metadata"]},"experiencefragment-header":{"id":"experiencefragment-847ced72ad","localizedFragmentVariationPath":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/jcr:content","configured":true,":items":{"root":{"layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"mega_header":"aem-GridColumn aem-GridColumn--default--12","markup_editor":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-c93981dc62",":items":{"markup_editor":{"id":"markup-editor-d40cc833d7","title":" ","cssContent":".footer-nav__link-group .snowflake-button-container,.subnav__item--button,.snowflake-card-v2-advanced-button .snowflake-button-container{justify-content:flex-start}.mega-nav__sign-in.snowflake-button-container{display:none}@media screen and (min-width:768px){.mega-nav__sign-in.snowflake-button-container{display:inline-block;font-family:'Texta',sans-serif;font-weight:800 !important}}@media screen and (min-width:1024px) and (max-width:1199px){.snowflake-mega-nav-header-buttons-container .snowflake-button-blue .snowflake-button-container{font-size:13px !important}.snowflake-language-navigation .language-icon{width:18px !important;height:18px !important;margin-right:4px !important}}.mega-nav__sign-in svg{display:none}.nav-item__platform-parent-why-sf.snowflake-mega-nav-nav-item\u003Ea:hover,.nav-item__platform-parent.snowflake-mega-nav-nav-item\u003Ea:hover{background-color:transparent !important}.nav-platform-sidebar .snowflake-mega-nav-nav-item:hover.blue-icon .snowflake-mega-nav-nav-item-icon__inner{background-color:var(--ui-01) !important}@media screen and (min-width:1024px){.snowflake-mega-nav-navigation-dropdown{overflow:hidden}.meganav-platform-features{padding-left:64px}.meganav-platform-features::before{content:'';transform:translateX(-64px);display:block;z-index:0;width:100%;height:100%;position:absolute;top:0;background:#f7f9fa}.nav-item--si.snowflake-mega-nav-nav-item\u003Ea:hover{background-color:transparent}.nav-item--si{border-bottom:1px solid #ccc;padding-bottom:16px;margin-bottom:8px}.nav-item__platform-parent{border-bottom:1px solid #ccc;margin-bottom:8px;padding-bottom:16px}.nav-item__platform-parent-why-sf .snowflake-mega-nav-nav-item-description::after{content:'What Snowflake can do for you \u003E';display:block;color:var(--ui-01);margin-top:16px}.nav-item__platform-parent .snowflake-mega-nav-nav-item-description::after{content:'View the platform \u003E';display:block;color:var(--ui-01);margin-top:16px}}@media screen and (min-width:1367px){.snowflake-mega-nav-nav-item-description{font-size:13px !important;line-height:20px !important}.snowflake-mega-nav-nav-item-title-wrapper\u003E.snowflake-mega-nav-nav-item-title{font-size:17px !important}.nav-item__platform-parent-why-sf .snowflake-mega-nav-nav-item-title,.nav-item__platform-parent .snowflake-mega-nav-nav-item-title{font-size:24px !important;line-height:32px !important;margin-bottom:8px !important}.nav-item__platform-parent-why-sf .snowflake-mega-nav-nav-item-description,.nav-item__platform-parent .snowflake-mega-nav-nav-item-description{font-size:14px !important;line-height:20px !important}}html.wf-texta-n9-loading .display-1-v2{font-size:48px!important;line-height:50px!important;letter-spacing:-.5px!important;font-family:sans-serif!important}html.wf-texta-n9-loading .heading-4-v2{font-size:18px!important;line-height:24px!important;font-family:sans-serif!important}@media screen and (min-width:768px){html.wf-texta-n9-loading .display-2-v2{font-size:48px!important;line-height:50px!important;font-family:sans-serif!important}html.wf-texta-n9-loading .display-1-v2{font-size:55.5px!important;line-height:54px!important;letter-spacing:-.5px!important;font-family:sans-serif!important}html.wf-lato-n4-loading .body-2,html.wf-lato-n4-loading .heading-5-v2,html.wf-lato-n4-loading .snowflake-card-v2-advanced-text .snowflake-text p{font-size:15.5px!important;font-family:sans-serif!important}html.wf-texta-n9-loading .heading-2,html.wf-texta-n9-loading .heading-2-v2{font-size:34px!important;line-height:38px!important;letter-spacing:-.75px!important;font-family:sans-serif!important}html.wf-texta-n8-loading .heading-6-v2.snowflake-mega-nav-navigation-title{font-size:13.5px!important;font-family:sans-serif!important}html.wf-texta-n8-loading .heading-4,html.wf-texta-n8-loading .snowflake-button-container,html.wf-texta-n8-loading .snowflake-button-regular .snowflake-button-container{font-size:13px!important;line-height:20px!important;letter-spacing:.25px!important;font-family:sans-serif!important}}@media screen and (min-width:1024px){html.wf-lato-n4-loading .snowflake-mega-nav-nav-item-description{font-size:11.5px!important;font-family:sans-serif!important}html.wf-lato-n4-loading .body-2,html.wf-lato-n4-loading .text-size-regular .snowflake-text li,html.wf-lato-n4-loading .text-size-regular .snowflake-text p,html.wf-lato-n4-loading .text-size-regular .snowflake-text span[data-testid=text-content],html.wf-lato-n4-loading .text-size-regular.cq-Editable-dom li,html.wf-lato-n4-loading .text-size-regular.cq-Editable-dom p,html.wf-lato-n4-loading .text-size-regular.cq-Editable-dom span[data-testid=text-content]{font-size:13.5px!important;font-family:sans-serif!important}html.wf-texta-n8-loading .snowflake-button-compact .snowflake-button-container{font-size:12px!important;letter-spacing:0!important;line-height:18px!important}}@media screen and (min-width:1367px){html.wf-lato-n4-loading .hp-hero__eyebrow a\u003Eb:first-child{font-size:11px!important;font-family:sans-serif!important}html.wf-texta-n8-loading .hp-hero__eyebrow a{font-size:13px!important;font-family:sans-serif!important}html.wf-texta-n9-loading .display-2-v2{font-size:61px!important;line-height:60px!important;font-family:sans-serif!important}html.wf-texta-n9-loading .display-1-v2{font-size:74.5px!important;line-height:74px!important;letter-spacing:-.75px!important;font-family:sans-serif!important}html.wf-texta-n9-loading .heading-2,html.wf-texta-n9-loading .heading-2-v2{font-size:41px!important;letter-spacing:-.75px!important;font-family:sans-serif!important}html.wf-texta-n9-loading .heading-3-v2{font-family:sans-serif!important;letter-spacing:-.75px!important;font-size:33.75px!important}html.wf-texta-n9-loading .heading-4-v2{font-size:19.5px!important;line-height:26px!important;font-family:sans-serif!important}html.wf-texta-n8-loading .heading-6-v2{font-size:12px!important;font-family:sans-serif!important}html.wf-texta-n8-loading .heading-6-v2.snowflake-mega-nav-navigation-title{font-size:14px!important;font-family:sans-serif!important}html.wf-lato-n4-loading .body-1,html.wf-lato-n4-loading .cq-Editable-dom[data-cq-data-path*=text] ol\u003Eli,html.wf-lato-n4-loading .snowflake-text li,html.wf-lato-n4-loading .snowflake-text p,html.wf-lato-n4-loading .text-size-large .snowflake-text li,html.wf-lato-n4-loading .text-size-large .snowflake-text p,html.wf-lato-n4-loading .text-size-large .snowflake-text span[data-testid=text-content],html.wf-lato-n4-loading .text-size-large.cq-Editable-dom li,html.wf-lato-n4-loading .text-size-large.cq-Editable-dom p,html.wf-lato-n4-loading .text-size-large.cq-Editable-dom span[data-testid=text-content],html.wf-lato-n4-loading.cq-Editable-dom[data-cq-data-path*=text]\u003Ep,html.wf-lato-n4-loading.cq-Editable-dom[data-cq-data-path*=text]\u003Eul\u003Eli{font-size:17.5px!important;font-family:sans-serif!important}html.wf-lato-n4-loading .body-2,html.wf-lato-n4-loading .text-size-regular .snowflake-text li,html.wf-lato-n4-loading .text-size-regular .snowflake-text p,html.wf-lato-n4-loading .text-size-regular .snowflake-text span[data-testid=text-content],html.wf-lato-n4-loading .text-size-regular.cq-Editable-dom li,html.wf-lato-n4-loading .text-size-regular.cq-Editable-dom p,html.wf-lato-n4-loading .text-size-regular.cq-Editable-dom span[data-testid=text-content],html.wf-texta-n8-loading .snowflake-button-link .snowflake-button-container,html.wf-texta-n8-loading .snowflake-button-link-back .snowflake-button-container{font-size:15.5px!important;font-family:sans-serif!important}html.wf-lato-n4-loading .body-3,html.wf-lato-n4-loading .text-size-small .snowflake-text li,html.wf-lato-n4-loading .text-size-small .snowflake-text p,html.wf-lato-n4-loading .text-size-small .snowflake-text span[data-testid=text-content],html.wf-lato-n4-loading .text-size-small.cq-Editable-dom li,html.wf-lato-n4-loading .text-size-small.cq-Editable-dom p,html.wf-lato-n4-loading .text-size-small.cq-Editable-dom span[data-testid=text-content]{font-size:13.5px!important;font-family:sans-serif!important}}#industryPlatformSection,.sc-hero{background-position:top left;background-size:20% auto}.bwalignc,.bwalignr{list-style-position:inside}.snowflake-text p sup{font-size:10px}#industryPlatformSection .industry-platform__row .snowflake-flexible-column-container-items,.button-group-pair\u003E.container\u003E.cmp-container\u003E.aem-container,.snowflake-hero-system-content-container{gap:16px}.agenda-item p,.button-group-pair\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv,.partner-details p{margin:0!important}.button-group-pair\u003E.container\u003E.cmp-container\u003E.aem-container::after,.button-group-pair\u003E.container\u003E.cmp-container\u003E.aem-container::before,.hide-logo .snowflake-case-study-card-logo,.partner-page__powered-by-logo,.sc-hero div.code-toolbar\u003E.toolbar,.snowflake-card-v2-advanced.no-link .snowflake-card-v2-advanced-button,.snowflake-partner-hero-card-badge-container{display:none!important}.section--card-mobile-carousel .snowflake-flexible-column-container-items-with-carousel{max-width:100%!important}@media screen and (min-width:768px){.button-group-pair .snowflake-button-container.inline-button--desktop,.button-group-pair\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv{width:auto!important;display:inline-block!important}.button-group-pair\u003E.container\u003E.cmp-container\u003E.aem-container{align-items:center;justify-content:flex-start!important}.button-group-pair.center\u003E.container\u003E.cmp-container\u003E.aem-container{justify-content:center!important}.section--card-mobile-carousel{margin-left:var(--tablet-portrait-margin,48px)!important;margin-right:var(--tablet-portrait-margin,48px);width:calc(100% - 96px)!important;width:calc(100% - var(--tablet-portrait-margin) * 2)!important}}@media screen and (min-width:1024px){.section--card-mobile-carousel{margin-left:var(--tablet-horizontal-margin,48px)!important;margin-right:var(--tablet-horizontal-margin,48px);width:calc(100% - 96px)!important;width:calc(100% - var(--tablet-horizontal-margin) * 2)!important}.snowflake-mega-nav-header-mobile-icon{display:none!important}}@media screen and (min-width:1367px){.section--card-mobile-carousel{margin-left:var(--desktop-margin,6.5%)!important;margin-right:var(--desktop-margin,6.5%);width:87%!important;width:calc(100% - var(--desktop-margin) * 2)!important}.logo-container{min-width:143px}.sc-hero__headline .heading-1-v2{font-size:60px}.snowflake-mega-nav-navigation-title{font-size:17px}.snowflake-mega-nav-dropdown-footer-wrapper .snowflake-title-v2 .snowflake-title-v2-line:first-child{font-size:16px!important;line-height:24px!important}}.hero--home{overflow:hidden;background-color:var(--ui-01);z-index:2}.hp-hero__subheadline{width:90%}.hero--home .snowflake-button-container{transition:.3s}.hero--home .snowflake-button-primary a:hover,.hero--home .snowflake-button-secondary a:hover,.hero--home .snowflake-button-white a:hover{transition:.3s;background-color:var(--ui-02)!important;color:var(--ui-05)!important}.hero--home .snowflake-button-secondary a:hover{border-color:var(--ui-05)!important}.hero--home .snowflake-button-primary a:hover,.hero--home .snowflake-button-white a:hover{border-color:var(--ui-02)!important}.bwalignc,.hp-hero__eyebrow{text-align:center}.hp-hero__eyebrow a{display:inline-flex;flex-direction:column;justify-content:center;cursor:pointer;padding:8px;border-radius:var(--spacing-01);gap:8px;align-items:center;background-color:#45aee3;color:var(--ui-03);font-family:Texta,sans-serif;font-weight:800;font-size:16px;line-height:22px;transition:background-color .3s}.hp-hero__eyebrow a:hover{background-color:#7fc6ea;text-decoration:none;transition:background-color .3s}.hp-hero__eyebrow a\u003Eb:first-child{text-transform:uppercase;white-space:nowrap;display:inline-block;background-color:var(--ui-02);color:var(--ui-05);font-size:12px!important;line-height:16px!important;font-family:Lato,sans-serif;font-weight:500!important;padding:3px 6px;border-radius:2px;letter-spacing:1px}@media screen and (min-width:767px){.hp-hero__eyebrow{text-align:left}.hp-hero__eyebrow a{flex-direction:row;text-align:left}}.hero--home__inner .offset-video,.hero--home__inner .snowflake-experience-fragment,.offset-video__bg-image{max-height:200px;overflow:hidden}.hero--home__inner .offset-video .wistia-responsive-padding{padding-top:100%}.hero--home__inner .snowflake-experience-fragment,.offset-video__bg-image{position:absolute!important;top:0;left:0;width:100%}.offset-video__bg-image{z-index:-1}@media screen and (min-width:768px){.hero--home__inner .snowflake-experience-fragment,.offset-video,.offset-video__bg-image{position:absolute!important;max-height:none;top:0;left:0;width:250%;padding-bottom:250%;transform:translate(0,-50%);height:0}.workloads_7.unistore{max-width:317px}}.promo-banner--homepage{z-index:2}.homepage-banner-offset-container::after{content:\"\";display:block;position:absolute;bottom:0;z-index:1;left:0;width:100%;height:80%;background:#fff}.section--quicklinks .snowflake-button-full-width a{padding-left:24px!important;padding-right:24px!important;transition:box-shadow .25s cubic-bezier(.4,0,.2,1);text-align:left;display:flex;justify-content:center;align-items:center}.section--quicklinks .snowflake-button-full-width a:hover{box-shadow:0 16px 16px 0 rgb(0 0 0 / .16);transition:box-shadow .25s cubic-bezier(.4,0,.2,1)}.section--quicklinks .snowflake-button-container:focus-visible a::before,.section--quicklinks .snowflake-button-full-width a::before{content:\"\";width:23px;height:23px;flex-shrink:0;margin-right:12px;display:inline-block;background-size:cover;background-repeat:no-repeat;background-position:center}#industryPartnerSlider .snowflake-navigation-icon.swiper-button-disabled,#partnerResources .section--resource-hub a svg,.button-tabs span.snowflake-tabs-navigation-item:after,.customer-card--hide-cta .snowflake-case-study-card-button,.dot-tabs span.snowflake-tabs-navigation-item::after,.partner-sidebar__mobile-expand,html:not(.aem-AuthorLayer-initial):not(.aem-AuthorLayer-Edit) .tab-content:not(.is-active){display:none}.section--quicklinks .snowflake-button-full-width a.pricing::before{background-image:url(https://www.snowflake.com/content/dam/snowflake-site/general/icons/decorative-icons/pricing-icon.svg)}.section--quicklinks .snowflake-button-full-width a.snowflake_on_snowflake::before{background-image:url(https://www.snowflake.com/content/dam/snowflake-site/general/icons/navigation/nav-icon_snowflake-bug.svg)}.section--quicklinks .snowflake-button-full-width a.virtual_hands_on_labs::before{background-image:url(https://www.snowflake.com/content/dam/snowflake-site/general/icons/navigation/nav-icon__training.svg)}.section--quicklinks .snowflake-button-full-width a.weekly_demo::before{background-image:url(https://www.snowflake.com/content/dam/snowflake-site/general/icons/navigation/nav-icon__webinars.svg)}@media screen and (min-width:1024px){.hero--home__inner .snowflake-experience-fragment,.offset-video,.offset-video__bg-image{left:-50%}.section--quicklinks .snowflake-flexible-column-container-items{gap:24px}.snowflake-quote-item-inner{padding:32px 24px 24px!important}}#communitiesOuter_overflowBottomGray::after{max-height:100px}#caseStudyOuter_overflowBottomMidBlue::after{max-height:180px}#caseStudyInner .snowflake-case-study-card .snowflake-wistia-video{border-radius:0!important}#caseStudyInner .snowflake-case-study-card{box-shadow:none!important;border-radius:0}#caseStudyInner{max-width:1200px;margin:0 auto;box-shadow:rgb(152 162 179 / .1) 0 10px 20px 0,rgb(152 162 179 / .25) 0 2px 6px 0;border-radius:8px;overflow:hidden;position:relative;z-index:1}.case-study__logo-bar\u003E.snowflake-flexible-column-container-items{background:#f7f9fa;padding:32px 16px 40px}.case-study__logo-bar .cmp-image__image{width:90%;margin:0 auto;max-width:240px}.hp-platform__text-group\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:not(:first-child),.sc-sidebar__group .snowflake-button-link{margin-top:8px}.workloads_7.unistore{margin-left:auto;margin-right:auto}#homepageFootnotesInner .snowflake-simple-stat-disclaimer .snowflake-text p{color:#fff!important}.snowflake-simple-stat-disclaimer .snowflake-text p\u003Ea{border-bottom:1px solid var(--ui-03);color:var(--text-03)}.snowflake-card-v2-advanced{color:inherit}#workloadCardGridOuter .snowflake-card-v2-base-front{gap:0}.video-modal.snowflake-modal-window-open-inner{background-color:#fff0;padding:8px;border:none}.snowflake-container-arrow-dotted-faded .snowflake-container-arrow-dotted-faded-image{width:40%!important;max-width:420px;top:4%!important}.list--blue-bullets ul{margin:0!important;padding:0!important;list-style-type:none}.list--blue-bullets li{margin:0;padding:0 0 0 32px;position:relative}.list--blue-bullets li::before{content:\"\";display:block;border-radius:100%;background:#29b5e8;width:18px;height:18px;position:absolute;top:4px;left:0;border:5px solid #e5f2f7;box-sizing:border-box}.list--blue-bullets li:not(:last-child){margin-bottom:1rem}.logo-tabs .snowflake-navigation-container,.snowflake-simple-stat-content:empty,.summit-speaker-card .snowflake-card-v2-advanced-text{margin-bottom:0}#techResourceInner,#techResourceOuter,div.overflow-bottom--blue,div.overflow-bottom--gray,div.overflow-bottom--mid-blue,div.overflow-bottom--white,div.overflow-top--blue,div.overflow-top--gray,div.overflow-top--mid-blue,div.overflow-top--white,div[id$=overflowBottomGray],div[id$=overflowBottomMidBlue],div[id$=overflowTopBlue],div[id$=overflowTopGray]{position:relative}div.overflow-bottom--blue::after,div.overflow-bottom--gray::after,div.overflow-bottom--mid-blue::after,div.overflow-bottom--white::after,div.overflow-top--blue::after,div.overflow-top--gray::after,div.overflow-top--mid-blue::after,div.overflow-top--white::after,div[id$=overflowBottomGray]::after,div[id$=overflowBottomMidBlue]::after,div[id$=overflowBottomWhite]::after,div[id$=overflowTopBlue]::after,div[id$=overflowTopGray]::after,div[id$=overflowTopWhite]::after{content:\"\";display:block;position:absolute;left:0;width:100%;height:40%}div.overflow-top--blue::after,div.overflow-top--gray::after,div.overflow-top--mid-blue::after,div.overflow-top--white::after,div[id$=overflowTopBlue]::after,div[id$=overflowTopGray]::after,div[id$=overflowTopWhite]::after{top:0}div.overflow-bottom--blue::after,div.overflow-bottom--gray::after,div.overflow-bottom--mid-blue::after,div.overflow-bottom--white::after,div[id$=overflowBottomGray]::after,div[id$=overflowBottomMidBlue]::after,div[id$=overflowBottomWhite]::after{bottom:0}div.overflow-bottom--white::after,div.overflow-top--white::after,div[id$=overflowBottomWhite]::after,div[id$=overflowTopWhite]::after{background:#fff!important}div.overflow-bottom--gray::after,div.overflow-top--gray::after,div[id$=overflowBottomGray]::after,div[id$=overflowTopGray]::after{background:#f6f9fa!important}div.overflow-bottom--mid-blue::after,div.overflow-top--mid-blue::after,div[id$=overflowBottomMidBlue]::after,div[id$=overflowTopMidBlue]::after{background:#11567f!important}div.overflow-bottom--blue::after,div.overflow-top--blue::after,div[id$=overflowBottomBlue]::after,div[id$=overflowTopBlue]::after{background:#259edc!important}.snowflake-premium-content-banner.promo-banner--no-shadow{box-shadow:none!important}#industryPartnerSlider .cmp-image__image,#industryPartnerSlider .section--partner-tabs .snowflake-image-container .cmp-image__image,#partnerSidebar,.has-shadow .cmp-image__image{box-shadow:0 10px 20px 0 rgb(152 162 179 / .1),0 2px 6px 0 rgb(152 162 179 / .25)}.content-chip--has-desc{align-items:flex-start;padding:20px!important}.content-chip--has-desc .snowflake-content-chip-image{max-width:100px}.content-chip--has-desc .snowflake-content-chip-image__image{aspect-ratio:1}.content-chip--has-desc .snowflake-title-v2-line:first-child{font-size:18px!important}.content-chip--has-desc .snowflake-title-v2-line:nth-child(2){color:#000!important;font-weight:500!important;font-size:16px!important;line-height:22px!important;margin-top:2px!important}.content-chip--has-desc .snowflake-content-chip-button{margin-top:6px!important;font-size:18px!important;display:none}.square-image .snowflake-content-chip-image{aspect-ratio:1;max-width:120px}.section--logo-bar.smaller-logos .snowflake-image-container .cmp-image__image{max-width:200px;margin:0 auto}.snowflake-card-v2-advanced-tag,.snowflake-content-chip-tag{padding:3px 6px!important}.sc-overview__webinar-promo-banner .snowflake-content-chip-button,.snowflake-card-v2-advanced-title:first-child,.summit-pricing-block__aside ul{margin-top:0}.dot-tabs .snowflake-navigation-container .snowflake-tabs-navigation-item{width:40px;height:40px;display:flex;justify-content:center;align-items:center;margin:0!important}.dot-tabs .snowflake-navigation-container .snowflake-tabs-navigation-item p{width:12px;height:12px;background:var(--ui-12);border-radius:100%}.dot-tabs .snowflake-navigation-container .snowflake-tabs-navigation-item p,.logo-tabs .snowflake-navigation-container .snowflake-tabs-navigation-item p{font-size:0!important}.dot-tabs .snowflake-navigation-container .snowflake-tabs-navigation-item.active p{background:var(--ui-01)}.button-tabs .snowflake-navigation-container .swiper-wrapper{padding:8px 0}.button-tabs .snowflake-navigation-container .swiper-slide{margin:0 6px}.button-tabs .snowflake-navigation-container .snowflake-tabs-navigation-item{padding:8px 24px;background-color:#f6f9fa;border-radius:48px;margin:0}.button-tabs .snowflake-navigation-container .snowflake-tabs-navigation-item p{text-transform:uppercase;font-family:Texta,sans-serif;font-weight:700}.button-tabs .border-top{border-top:1px solid #ccc}.button-tabs .snowflake-navigation-container .snowflake-tabs-navigation-item.active{background-color:var(--ui-01);box-shadow:0 2px 6px 0 rgb(152 162 179 / .25),0 10px 20px 0 rgb(152 162 179 / .1)}.button-tabs .snowflake-navigation-container .snowflake-tabs-navigation-item.active p{color:#fff}.button-tabs.has-icons .snowflake-navigation-container .snowflake-tabs-navigation-item p::before{content:\"\";display:inline-block;width:20px;height:20px;background-size:contain;background-repeat:no-repeat;background-position:center center;margin-right:12px;vertical-align:middle;margin-top:-3px}.logo-tabs .snowflake-navigation-container .snowflake-tabs-navigation-item{width:220px;padding-bottom:50%;height:0;margin:0 8px!important;background-size:cover;background-repeat:no-repeat;opacity:.5;transition:opacity .3s}.logo-tabs .snowflake-navigation-container .snowflake-tabs-navigation-item:hover{opacity:.75;transition:opacity .3s}.logo-tabs .snowflake-navigation-container .snowflake-tabs-navigation-item.active{opacity:1;transition:opacity .3s}.dot-tabs .aem-container.cmp-tabs,.logo-tabs .aem-container.cmp-tabs{display:flex;flex-direction:column-reverse}.snowflake-icon.is-center{margin:0 auto;display:block}#industryPartnerSlider .snowflake-flexible-column-container-items,#partnerLogoSquare .snowflake-flexible-column-container-items{gap:24px}#techResourceOuter::after{content:\"\";display:block;position:absolute;top:0;left:0;width:100%;height:40%;background:#f6f9fa}#techResourceInner{z-index:1}.partner-tier-tag h6{display:inline-block!important;padding:2px 6px;border-radius:2px;color:#666}.partner-tier-tag.registered h6{background-color:#f6f9fa}.partner-tier-tag.elite h6{background-color:#11567f;color:#fff}.partner-tier-tag.premier h6{background-color:#b14c77;color:#fff}.partner-tier-tag.select h6{background-color:#5094a0;color:#fff}.partner-details\u003Espan{display:flex;gap:24px}.partner-details a{color:inherit!important;font-weight:400!important}.partner-details p::before{content:\"\";display:inline-block;vertical-align:middle;width:16px;height:16px;background-repeat:no-repeat;background-position:center;transform:translateY(-1px);background-size:auto 90%;margin-right:6px}.partner-details__location::before{background-image:url(\"data:image/svg+xml,%3Csvg width='13' height='18' viewBox='0 0 13 18' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M6.25 17.7531C6.4375 17.7531 6.6 17.6844 6.7375 17.5531C6.875 17.4219 6.95 17.2531 6.95 17.0531C6.95 16.8531 7.075 16.4281 7.3 15.7969C7.5875 15.0281 7.925 14.3156 8.30625 13.6406C8.8 12.7781 9.3125 12.1031 9.85 11.6094C10.75 10.7969 11.4125 9.96563 11.85 9.12188C12.2875 8.27813 12.5063 7.40313 12.5063 6.49063C12.5063 5.36563 12.2187 4.31563 11.6437 3.33438C11.0937 2.40313 10.3438 1.65938 9.4 1.10938C8.43125 .534376 7.375 .246876 6.24375 .246876C5.1125 .246876 4.06875 .534376 3.0875 1.10938C2.15625 1.65938 1.4125 2.40313 .862498 3.33438C.287498 4.31563 0 5.36563 0 6.49063C0 7.47188 .262499 8.42813 .787499 9.35938C1.14375 10.0031 1.65625 10.6656 2.3125 11.3344C2.75625 11.8031 3.24375 12.4781 3.78125 13.3656C4.225 14.0969 4.63125 14.8594 5 15.6656C5.35 16.3844 5.53125 16.8531 5.55625 17.0656C5.55625 17.2594 5.625 17.4156 5.7625 17.5531C5.9 17.6844 6.0625 17.7531 6.25 17.7531ZM6.16875 14.9156C5.775 14.0656 5.325 13.2469 4.825 12.4594C4.275 11.5594 3.7625 10.8719 3.28125 10.3969C2.625 9.71563 2.1375 9.05938 1.825 8.43438C1.5125 7.80313 1.35625 7.16563 1.35625 6.50313C1.35625 5.61563 1.575 4.80313 2.0125 4.05313C2.45 3.30313 3.04375 2.71563 3.7875 2.27813C4.5375 1.84063 5.35 1.62188 6.2375 1.62188C7.125 1.62188 7.9375 1.84063 8.6875 2.27813C9.4375 2.71563 10.0312 3.30313 10.475 4.04688C10.9187 4.80313 11.1375 5.62188 11.1375 6.50313C11.1375 7.90313 10.3937 9.26563 8.9125 10.5969C8.35 11.1094 7.8125 11.7906 7.3 12.6406C6.88125 13.3344 6.50625 14.0969 6.16875 14.9219V14.9156ZM6.26875 8.36563C6.65625 8.36563 7.01875 8.26563 7.35625 8.07188C7.69375 7.87813 7.95625 7.60938 8.14375 7.28438C8.3375 6.95313 8.43125 6.59063 8.43125 6.19688C8.43125 5.80313 8.33125 5.43438 8.1375 5.10313C7.9375 4.76563 7.675 4.50313 7.3375 4.31563C7 4.12813 6.6375 4.02813 6.24375 4.02813C5.85 4.02813 5.4875 4.12813 5.15625 4.32188C4.825 4.52188 4.56875 4.78438 4.375 5.12188C4.18125 5.45938 4.0875 5.82188 4.0875 6.20938C4.0875 6.59688 4.1875 6.95938 4.38125 7.29688C4.58125 7.63438 4.84375 7.89688 5.18125 8.08438C5.51875 8.27813 5.88125 8.37188 6.26875 8.37188V8.36563ZM6.24375 7.50313C5.8875 7.50313 5.575 7.37188 5.31875 7.11563C5.0625 6.85938 4.93125 6.55313 4.93125 6.19063C4.93125 5.82813 5.0625 5.52188 5.31875 5.26563C5.575 5.00938 5.88125 4.87813 6.24375 4.87813C6.60625 4.87813 6.9125 5.00938 7.16875 5.26563C7.425 5.52188 7.55625 5.82813 7.55625 6.19063C7.55625 6.55313 7.425 6.85938 7.16875 7.11563C6.9125 7.37188 6.60625 7.50313 6.24375 7.50313Z' fill='%2329B5E8'/%3E%3C/svg%3E%0A\")}.partner-details__website::before{background-image:url(\"data:image/svg+xml,%3Csvg width='18' height='16' viewBox='0 0 18 16' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M2.61587 2.96889C2.61587 2.75109 2.79633 2.57062 3.01413 2.57062C3.23192 2.57062 3.41238 2.75109 3.41238 2.96889C3.41238 3.18669 3.23192 3.36716 3.01413 3.36716C2.79633 3.36716 2.61587 3.18669 2.61587 2.96889ZM4.21512 2.96889C4.21512 2.75109 4.39558 2.57062 4.61338 2.57062C4.83117 2.57062 5.01163 2.75109 5.01163 2.96889C5.01163 3.18669 4.83117 3.36716 4.61338 3.36716C4.39558 3.36716 4.21512 3.18669 4.21512 2.96889ZM5.81438 2.96889C5.81438 2.75109 5.99484 2.57062 6.21264 2.57062C6.43043 2.57062 6.61089 2.75109 6.61089 2.96889C6.61089 3.18669 6.43043 3.36716 6.21264 3.36716C5.99484 3.36716 5.81438 3.18669 5.81438 2.96889ZM17.2518 .697559H1.19085C.811258 .697559 .506348 1.0025 .506348 1.38209V14.6179C.506348 14.9975 .811258 15.3024 1.19085 15.3024H17.2518C17.6314 15.3024 17.9363 14.9975 17.9363 14.6179V1.38209C17.9363 1.0025 17.6314 .697559 17.2518 .697559ZM16.5673 2.06035V3.90853H1.86914V2.06035H16.5673ZM1.86914 13.9334V4.78593H16.5673V13.9334H1.86914Z' fill='%2329B5E8'/%3E%3C/svg%3E%0A\")}#partnerSidebar{border-radius:4px;background-color:#fff;padding:24px 24px 32px;border-bottom:6px solid #29b5e8}#partnerSidebar h5,.newsletter-disclaimer p{font-size:14px!important}#partnerSidebar ul{margin-top:0;list-style-type:none;padding:0;display:flex;flex-wrap:wrap;gap:8px}#partnerSidebar li{border:1px solid;border-radius:2px;padding:0 4px!important;font-size:11px!important;letter-spacing:.25px;text-transform:uppercase}div.snowflake-partner-hero-card{width:100%;margin:0}.partner-details__logo{max-width:380px;margin:0 auto}@media screen and (max-width:767px){.left-alignment .hp-hero__subheadline{margin-left:auto;margin-right:auto}.left-alignment .hp-hero__headline .snowflake-title-v2-line,.left-alignment .hp-hero__subheadline .snowflake-title-v2-line{text-align:center}.hero--home__inner .snowflake-flexible-column-container-items-top-padding-large{padding-top:var(--spacing-02)}.section--logo-bar\u003E.snowflake-flexible-column-container-items{display:flex;flex-wrap:wrap;flex-direction:row;justify-content:center;gap:8px}.section--logo-bar\u003E.snowflake-flexible-column-container-items\u003Ediv{width:calc(33.33% - 8px)}.partner-sidebar__mobile-expand{display:inline-block;color:#249edc;border-color:#249edc!important}#partnerSidebar li:nth-child(n+6),.summit-nav__links .snowflake-button-tertiary{display:none}.sc-body__sidebar{background-color:#f6f9fa;padding:24px}.sc-body__content{padding:0 24px 24px}.summit-speaker-card .snowflake-card-v2-advanced-content{padding:24px}}#partnerResources h6,.snowflake-tabs-navigation-item p.body-1{font-size:16px!important}#partnerResources .section--resource-hub{padding:0 16px}#partnerResources .section--resource-hub a,.bwalignl{text-align:left}@media screen and (max-width:1023px){.hero--workload .snowflake-hero-system-media-container{width:100%}}.section--timely-content .snowflake-content-chip,.snowflake-mega-nav-dropdown-footer-wrapper{align-items:center}.section--timely-content .snowflake-content-chip-image{max-width:94px}.section--timely-content .snowflake-content-chip-image__inner{line-height:0}.section--timely-content .snowflake-content-chip-image__image{aspect-ratio:1;height:auto}.section--workload-overview .workload-overview__headline{max-width:280px;margin:0 auto}#industryPartnerSlider .swiper-slide{margin-top:0!important;padding:0 12px}#industryPartnerSlider .snowflake-tabs-navigation-item{margin-left:0!important;margin-right:0!important}#industryPartnerSlider .snowflake-premium-content-banner-background-grad-white .snowflake-premium-content-banner{box-shadow:none}#industryPartnerSlider .logo-slider__slide .aem-container{display:flex;padding:0 8px!important;flex-wrap:wrap;gap:16px!important;justify-content:center}#industryPartnerSlider .logo-slider__slide .aem-container\u003Ediv{width:48%;max-width:200px}#useCaseTabs{padding-top:24px;padding-bottom:24px;padding-right:24px}#useCaseTabs .tab-content.is-active{display:block}#useCaseTabs .vert-tab{border-bottom:1px solid #a0bbcc;padding-bottom:16px}#useCaseTabs .vert-tab p{display:inline-block}#useCaseTabs .vert-tab p:hover{cursor:pointer}#useCaseTabs .vert-tab p,#useCaseTabs .vert-tab.is-active p.not-active{color:#249edc}#useCaseTabs .vert-tab p.is-active,#useCaseTabs .vert-tab.is-active p{color:#000}#industryPlatformSection{background-image:url(/adobe/dynamicmedia/deliver/dm-aid--db074ad5-7122-4c51-87a3-76c3aa466182/double-arrow-bg%403x.png);background-repeat:no-repeat}.snowflake-text p.featured-quote__source{font-weight:900!important;text-transform:uppercase;font-size:16px!important;margin-top:2rem!important}.snowflake-text p.featured-quote__title{margin-top:0!important;font-size:16px!important}.snowflake-case-study-card-logo img{width:auto!important;height:100px!important;transform:translateX(-15%)}.snowflake-quote-item-quote-text{font-weight:600!important}#customerStoryStatsInner\u003E.container\u003E.cmp-container\u003E.aem-container{display:flex;flex-direction:row}#customerStoryStat1,#customerStoryStat2{max-width:240px}#storyHighlights{border-radius:4px;padding:1rem}.sc-overview__webinar-promo-banner .snowflake-content-chip-content .snowflake-title-v2-line,.summit-pricing-block__tile .black-blue-text-color .snowflake-title-v2-line{color:#000!important}.snowflake-youtube-embedded-wrapper{border-radius:var(--small-border-radius)}#arcticNavItem::before,#offset::before,#open-source::before{color:var(--text-05);font-family:Texta,sans-serif!important}#offset,.sc-architecture-caption{margin-top:16px}.hero--press .snowflake-title-v2-line{text-transform:none!important}@media screen and (min-width:768px){.subpage-timely-content__inner\u003E.snowflake-flexible-column-container-items{box-shadow:0 10px 20px 0 rgb(152 162 179 / .1),0 2px 6px 0 rgb(152 162 179 / .25);padding:var(--spacing-04);border-radius:4px;overflow:hidden}#partnerLogoSquare{padding:0 0 0 48px}.hero--workload .snowflake-container{max-width:1440px;margin:0 auto!important;align-items:center}#industryPartnerSlider.snowflake-flexible-column-container-2-column-40-60\u003E.snowflake-flexible-column-container-items{grid-template-columns:minmax(40%,4fr) minmax(0,6fr)}#industryPartnerSlider .swiper-slide{padding:0 24px}.sc-body{padding:48px}.sc-body\u003E.snowflake-flexible-column-container-items{grid-template-columns:7fr 3fr;gap:124px}}.snowflake-button-container.has-icon{display:inline-flex;justify-content:center;align-items:center;text-align:left}.snowflake-button-container.has-icon::before{content:\"\";display:inline-block;width:20px;height:20px;margin-right:12px;background-size:contain;background-repeat:no-repeat;background-position:center}.snowflake-button-container.is-video::before{background-image:url(\"data:image/svg+xml,%3Csvg width='18' height='18' viewBox='0 0 18 18' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M9 1.28663C13.2523 1.28663 16.7134 4.74768 16.7134 9C16.7134 13.2523 13.2523 16.7134 9 16.7134C4.74768 16.7198 1.28663 13.2588 1.28663 9C1.28663 4.74124 4.74768 1.28663 9 1.28663ZM9 0C4.0336 0 0 4.0336 0 9C0 13.9664 4.0336 18 9 18C13.9728 18 18 13.9664 18 9C18 4.0336 13.9728 0 9 0Z' fill='white'/%3E%3Cpath d='M7.75106 6.18211C7.42941 6.16925 7.16565 6.42658 7.16565 6.74823V11.2772C7.16565 11.7082 7.65457 11.9848 8.02126 11.7597L11.7975 9.4952C12.1578 9.27647 12.1578 8.74252 11.7975 8.52379L8.02126 6.25931C7.93763 6.21428 7.84756 6.18211 7.75106 6.18211Z' fill='white'/%3E%3C/svg%3E%0A\")}.snowflake-button-container.is-github::before{background-image:url(\"data:image/svg+xml,%3Csvg width='20' height='21' viewBox='0 0 20 21' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M10 .651794C4.475 .651794 0 5.12679 0 10.6518C0 15.0768 2.8625 18.8143 6.8375 20.1393C7.3375 20.2268 7.525 19.9268 7.525 19.6643C7.525 19.4268 7.5125 18.6393 7.5125 17.8018C5 18.2643 4.35 17.1893 4.15 16.6268C4.0375 16.3393 3.55 15.4518 3.125 15.2143C2.775 15.0268 2.275 14.5643 3.1125 14.5518C3.9 14.5393 4.4625 15.2768 4.65 15.5768C5.55 17.0893 6.9875 16.6643 7.5625 16.4018C7.65 15.7518 7.9125 15.3143 8.2 15.0643C5.975 14.8143 3.65 13.9518 3.65 10.1268C3.65 9.03929 4.0375 8.13929 4.675 7.43929C4.575 7.18929 4.225 6.16429 4.775 4.78929C4.775 4.78929 5.6125 4.52679 7.525 5.81429C8.325 5.58929 9.175 5.47679 10.025 5.47679C10.875 5.47679 11.725 5.58929 12.525 5.81429C14.4375 4.51429 15.275 4.78929 15.275 4.78929C15.825 6.16429 15.475 7.18929 15.375 7.43929C16.0125 8.13929 16.4 9.02679 16.4 10.1268C16.4 13.9643 14.0625 14.8143 11.8375 15.0643C12.2 15.3768 12.5125 15.9768 12.5125 16.9143C12.5125 18.2518 12.5 19.3268 12.5 19.6643C12.5 19.9268 12.6875 20.2393 13.1875 20.1393C17.1375 18.8143 20 15.0643 20 10.6518C20 5.12679 15.525 .651794 10 .651794Z' fill='%23249EDC'/%3E%3C/svg%3E%0A\")}.snowflake-button-container.is-quickstart::before{background-image:url(\"data:image/svg+xml,%3Csvg width='15' height='21' viewBox='0 0 15 21' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M13.8489 2.79368H11.6439V2.38493C11.6439 1.71368 11.1451 .967427 10.4251 .967427H8.94762C8.80887 .359927 8.37387 .299927 7.89762 .299927H7.23012C6.85512 .299927 6.26637 .299927 6.08637 .967427H4.68387C3.94887 .967427 3.35637 1.74368 3.35637 2.38493V2.79368H1.15137C.738867 2.79368 .401367 3.13118 .401367 3.54368V20.2537C.401367 20.6662 .738867 21.0037 1.15137 21.0037H13.8489C14.2614 21.0037 14.5989 20.6662 14.5989 20.2537V3.54368C14.5989 3.13118 14.2614 2.79368 13.8489 2.79368ZM4.29387 2.38493C4.29387 2.18243 4.54137 1.90493 4.68387 1.90493H6.50262C6.76137 1.90493 6.97137 1.69493 6.97137 1.43618C6.97137 1.33868 6.97887 1.27868 6.98637 1.24118C7.05012 1.23368 7.15512 1.23368 7.23387 1.23368H7.90137C7.95012 1.23368 8.00637 1.23368 8.05137 1.23368C8.05512 1.27868 8.05887 1.34243 8.05887 1.43243C8.05887 1.69118 8.26887 1.90118 8.52762 1.90118H10.4289C10.5301 1.90118 10.7101 2.14493 10.7101 2.38118V2.78993H4.29762V2.38118L4.29387 2.38493ZM13.0989 19.4999H1.90137V4.29368H13.0989V19.5037V19.4999Z' fill='%23249EDC'/%3E%3Cpath d='M3.82512 16.0424H11.1751C11.4339 16.0424 11.6439 15.8324 11.6439 15.5736V6.88486C11.6439 6.62611 11.4339 6.41611 11.1751 6.41611H3.82512C3.56637 6.41611 3.35637 6.62611 3.35637 6.88486V15.5736C3.35637 15.8324 3.56637 16.0424 3.82512 16.0424ZM4.29387 15.1049V13.3686H10.7064V15.1049H4.29387ZM10.7101 7.35361V12.4311H4.29762V7.35361H10.7101Z' fill='%23249EDC'/%3E%3Cpath d='M6.16512 9.35989H8.83887C9.09762 9.35989 9.30762 9.14989 9.30762 8.89114C9.30762 8.63239 9.09762 8.42239 8.83887 8.42239H6.16512C5.90637 8.42239 5.69637 8.63239 5.69637 8.89114C5.69637 9.14989 5.90637 9.35989 6.16512 9.35989Z' fill='%23249EDC'/%3E%3Cpath d='M6.16512 11.3624H8.83887C9.09762 11.3624 9.30762 11.1524 9.30762 10.8937C9.30762 10.6349 9.09762 10.4249 8.83887 10.4249H6.16512C5.90637 10.4249 5.69637 10.6349 5.69637 10.8937C5.69637 11.1524 5.90637 11.3624 6.16512 11.3624Z' fill='%23249EDC'/%3E%3C/svg%3E%0A\")}.snowflake-button-container.is-download::before{background-image:url(\"data:image/svg+xml,%3Csvg width='16' height='18' viewBox='0 0 16 18' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M15.2017 17.1637H.798265C.364425 17.1637 0 16.7993 0 16.3655V12.3568C0 11.923 .364425 11.5585 .798265 11.5585C1.2321 11.5585 1.59653 11.923 1.59653 12.3568V15.5498H14.4035V12.3568C14.4035 11.923 14.7679 11.5585 15.2017 11.5585C15.6356 11.5585 16 11.923 16 12.3568V16.3655C16 16.7993 15.6529 17.1637 15.2017 17.1637Z' fill='%23249EDC'/%3E%3Cpath d='M7.94793 12.9642C7.84381 12.9642 7.73969 12.9468 7.63557 12.8947C7.34056 12.7733 7.14967 12.4783 7.14967 12.1485L7.18437 .938127C7.18437 .504287 7.5488 .139862 7.98264 .139862C8.41648 .139862 8.7809 .504287 8.7809 .938127L8.7462 10.257L12.8416 6.33509C13.154 6.02273 13.6746 6.04008 13.9696 6.35244C14.282 6.66481 14.2646 7.18542 13.9523 7.48043L8.50325 12.7386C8.36442 12.8774 8.15617 12.9642 7.94793 12.9642Z' fill='%23249EDC'/%3E%3Cpath d='M7.94793 12.9642C7.73969 12.9642 7.54881 12.8947 7.39262 12.7386L2.03037 7.53249C1.718 7.22012 1.70065 6.71687 2.01301 6.40451C2.32538 6.09214 2.82863 6.07479 3.141 6.38715L8.50325 11.5932C8.81562 11.9056 8.83297 12.4088 8.52061 12.7212C8.36442 12.8774 8.15617 12.9642 7.94793 12.9642Z' fill='%23249EDC'/%3E%3C/svg%3E%0A\")}.snowflake-button-container.is-expand::before{background-image:url(\"data:image/svg+xml,%3Csvg width='18' height='18' viewBox='0 0 18 18' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M6.64375 10.9125C6.9375 11.2062 6.93125 11.6812 6.64375 11.9687L2.57502 16H3.79375C4.20625 16 4.54376 16.3375 4.54376 16.75C4.54376 17.1625 4.20625 17.5 3.79375 17.5H.756264C.556264 17.5 .36876 17.4187 .22501 17.2812C.22501 17.2812 .206248 17.25 .193748 17.2375C.143748 17.1812 .100004 17.1125 .0625038 17.0437C.0375038 16.9687 .0187492 16.8937 .0187492 16.8187C.0187492 16.8 .0062561 16.7813 .0062561 16.7625V13.725C.0187561 13.3125 .356257 12.9875 .768757 12.9937C1.16876 13 1.48752 13.325 1.50002 13.725V14.9688L5.5875 10.9187C5.88125 10.6312 6.35 10.6312 6.64375 10.9187V10.9125ZM17.5063 .743732C17.5063 .543732 17.425 .356235 17.2875 .218735C17.2875 .218735 17.2562 .199998 17.2437 .193748C17.1875 .137498 17.1188 .0937347 17.0438 .0624847C16.9688 .0374847 16.8938 .0187492 16.8188 .0187492C16.8 .0187492 16.7813 .00623703 16.7625 .00623703H13.725C13.3125 .00623703 12.975 .343745 12.975 .756245C12.975 1.16874 13.3125 1.50623 13.725 1.50623H14.9688L11.1312 5.37498C10.8437 5.67498 10.8563 6.14999 11.1563 6.43124C11.45 6.71249 11.9063 6.70624 12.1938 6.43124L16.0125 2.575V3.79375C16.0125 4.20625 16.35 4.54372 16.7625 4.54372C17.175 4.54372 17.5125 4.20625 17.5125 3.79375V.756245L17.5063 .743732ZM16.7562 12.9688C16.3437 12.9688 16.0063 13.3063 16.0063 13.7188V14.8937L12.1938 10.925C11.9063 10.625 11.4375 10.6188 11.1375 10.9063C10.8375 11.1938 10.8313 11.6625 11.1188 11.9625L15 16.0062H13.7188C13.3063 16.0062 12.9688 16.3437 12.9688 16.7562C12.9688 17.1687 13.3063 17.5063 13.7188 17.5063H16.7562C16.85 17.5063 16.95 17.4875 17.0375 17.45C17.0875 17.425 17.1313 17.3937 17.175 17.3625C17.2063 17.3437 17.2438 17.325 17.275 17.3C17.3313 17.2375 17.375 17.1687 17.4125 17.1C17.4188 17.0875 17.4375 17.075 17.4438 17.0562C17.45 17.025 17.4563 16.9938 17.4625 16.9625C17.4813 16.9 17.5 16.8375 17.5 16.7687V13.725C17.5 13.3125 17.1687 12.975 16.7562 12.975V12.9688ZM.750008 4.53125C1.16251 4.53125 1.50002 4.19374 1.50002 3.78124V2.5L5.59376 6.43124C5.89376 6.71874 6.36251 6.70626 6.65001 6.41251C6.93751 6.11876 6.92501 5.64375 6.63126 5.35625L2.61251 1.49998H3.7875C4.2 1.49998 4.53751 1.16249 4.53751 .749989C4.53751 .337489 4.2 0 3.7875 0H.743752C.668752 0 .600004 .0187355 .531254 .0437355C.506254 .0499855 .481263 .0437477 .462513 .0562477C.443763 .0687477 .425015 .0812462 .406265 .0937462C.337515 .124996 .275004 .168741 .218754 .224991H.212498C.212498 .224991 .175 .28125 .15625 .3125C.11875 .3625 .0812477 .4125 .0562477 .46875C.0374977 .525 .0249992 .587499 .0187492 .643749C.0124992 .674999 0 .712482 0 .743732V3.78124C0 4.19374 .337508 4.53125 .750008 4.53125Z' fill='white'/%3E%3C/svg%3E%0A\")}@keyframes slow-scroll{100%{transform:translateY(-50%)}}.sc-hero{overflow:hidden;background-color:#212d35;background-repeat:repeat-y;background-image:url(\"data:image/svg+xml,%3Csvg width='389' height='17' viewBox='0 0 389 17' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M.638672 7.80824L.638672 9.2566C.638672 9.52364 .85538 9.74024 1.12262 9.74024H2.57204C2.83928 9.74024 3.05598 9.52364 3.05598 9.2566V7.80824C3.05598 7.54119 2.83928 7.32472 2.57204 7.32472L1.12262 7.32472C.85538 7.32472 .638672 7.54119 .638672 7.80824Z' fill='url(%23paint0_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M10.9639 7.80824V9.2566C10.9639 9.52364 11.1806 9.74024 11.4478 9.74024L12.8972 9.74024C13.1645 9.74024 13.3812 9.52364 13.3812 9.2566V7.80824C13.3812 7.54119 13.1645 7.32471 12.8972 7.32471L11.4478 7.32471C11.1806 7.32471 10.9639 7.54119 10.9639 7.80824Z' fill='url(%23paint1_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M21.2891 7.80823V9.2566C21.2891 9.52364 21.5058 9.74024 21.773 9.74024L23.2224 9.74024C23.4897 9.74024 23.7064 9.52364 23.7064 9.2566V7.80823C23.7064 7.54119 23.4897 7.32471 23.2224 7.32471L21.773 7.32471C21.5058 7.32471 21.2891 7.54119 21.2891 7.80823Z' fill='url(%23paint2_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M31.6143 7.80823V9.2566C31.6143 9.52364 31.831 9.74024 32.0982 9.74024H33.5476C33.8149 9.74024 34.0316 9.52364 34.0316 9.2566V7.80823C34.0316 7.54119 33.8149 7.32471 33.5476 7.32471L32.0982 7.32471C31.831 7.32471 31.6143 7.54119 31.6143 7.80823Z' fill='url(%23paint3_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M41.9395 7.80823V9.2566C41.9395 9.52364 42.1562 9.74024 42.4234 9.74024H43.8728C44.1401 9.74024 44.3568 9.52364 44.3568 9.2566V7.80823C44.3568 7.54119 44.1401 7.32471 43.8728 7.32471L42.4234 7.32471C42.1562 7.32471 41.9395 7.54119 41.9395 7.80823Z' fill='url(%23paint4_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M52.5076 7.80823V9.2566C52.5076 9.52364 52.7243 9.74024 52.9916 9.74024H54.441C54.7082 9.74024 54.9249 9.52364 54.9249 9.2566V7.80823C54.9249 7.54119 54.7082 7.32471 54.441 7.32471L52.9916 7.32471C52.7243 7.32471 52.5076 7.54119 52.5076 7.80823Z' fill='url(%23paint5_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M62.8331 7.80823V9.2566C62.8331 9.52364 63.0493 9.74024 63.3165 9.74024H64.7664C65.0332 9.74024 65.2504 9.52364 65.2504 9.2566V7.80823C65.2504 7.54119 65.0332 7.32471 64.7664 7.32471L63.3165 7.32471C63.0493 7.32471 62.8331 7.54119 62.8331 7.80823Z' fill='url(%23paint6_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M73.1583 7.80823V9.2566C73.1583 9.52364 73.3745 9.74024 73.6417 9.74024H75.0916C75.3584 9.74024 75.5756 9.52364 75.5756 9.2566V7.80823C75.5756 7.54119 75.3584 7.32471 75.0916 7.32471L73.6417 7.32471C73.3745 7.32471 73.1583 7.54119 73.1583 7.80823Z' fill='url(%23paint7_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M83.4835 7.80823V9.2566C83.4835 9.52364 83.6997 9.74024 83.9669 9.74024H85.4168C85.6836 9.74024 85.9008 9.52364 85.9008 9.2566V7.80823C85.9008 7.54119 85.6836 7.32471 85.4168 7.32471L83.9669 7.32471C83.6997 7.32471 83.4835 7.54119 83.4835 7.80823Z' fill='url(%23paint8_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M93.8087 7.80823V9.2566C93.8087 9.52364 94.0249 9.74024 94.2921 9.74024H95.742C96.0088 9.74024 96.226 9.52364 96.226 9.2566V7.80823C96.226 7.54119 96.0088 7.32471 95.742 7.32471L94.2921 7.32471C94.0249 7.32471 93.8087 7.54119 93.8087 7.80823Z' fill='url(%23paint9_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M104.134 7.80823V9.2566C104.134 9.52364 104.35 9.74024 104.617 9.74024H106.067C106.334 9.74024 106.551 9.52364 106.551 9.2566V7.80823C106.551 7.54119 106.334 7.32471 106.067 7.32471L104.617 7.32471C104.35 7.32471 104.134 7.54119 104.134 7.80823Z' fill='url(%23paint10_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M114.702 7.80823V9.2566C114.702 9.52364 114.918 9.74024 115.185 9.74024L116.635 9.74024C116.902 9.74024 117.119 9.52364 117.119 9.25659V7.80823C117.119 7.54119 116.902 7.32471 116.635 7.32471L115.185 7.32471C114.918 7.32471 114.702 7.54119 114.702 7.80823Z' fill='url(%23paint11_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M125.027 7.80823V9.25659C125.027 9.52364 125.243 9.74024 125.511 9.74024L126.961 9.74024C127.227 9.74024 127.445 9.52364 127.445 9.25659V7.80823C127.445 7.54119 127.227 7.32471 126.961 7.32471L125.511 7.32471C125.243 7.32471 125.027 7.54119 125.027 7.80823Z' fill='url(%23paint12_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M135.352 7.80823V9.25659C135.352 9.52364 135.569 9.74024 135.836 9.74024H137.286C137.553 9.74024 137.77 9.52364 137.77 9.25659V7.80823C137.77 7.54119 137.553 7.32471 137.286 7.32471L135.836 7.32471C135.569 7.32471 135.352 7.54119 135.352 7.80823Z' fill='url(%23paint13_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M145.678 7.80823V9.25659C145.678 9.52364 145.894 9.74024 146.161 9.74024H147.611C147.878 9.74024 148.095 9.52364 148.095 9.25659V7.80823C148.095 7.54119 147.878 7.32471 147.611 7.32471L146.161 7.32471C145.894 7.32471 145.678 7.54119 145.678 7.80823Z' fill='url(%23paint14_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M156.003 7.80823V9.25659C156.003 9.52364 156.219 9.74024 156.486 9.74024H157.936C158.203 9.74024 158.42 9.52364 158.42 9.25659V7.80823C158.42 7.54119 158.203 7.32471 157.936 7.32471L156.486 7.32471C156.219 7.32471 156.003 7.54119 156.003 7.80823Z' fill='url(%23paint15_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M166.328 7.80823V9.25659C166.328 9.52363 166.544 9.74024 166.811 9.74024H168.261C168.528 9.74024 168.745 9.52363 168.745 9.25659V7.80823C168.745 7.54119 168.528 7.32471 168.261 7.32471L166.811 7.32471C166.544 7.32471 166.328 7.54119 166.328 7.80823Z' fill='url(%23paint16_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M176.896 7.80823V9.25659C176.896 9.52363 177.112 9.74023 177.38 9.74023H178.83C179.096 9.74023 179.313 9.52363 179.313 9.25659V7.80823C179.313 7.54119 179.096 7.32471 178.83 7.32471L177.38 7.32471C177.112 7.32471 176.896 7.54119 176.896 7.80823Z' fill='url(%23paint17_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M187.221 7.80823V9.25659C187.221 9.52363 187.438 9.74023 187.705 9.74023H189.155C189.421 9.74023 189.639 9.52363 189.639 9.25659V7.80823C189.639 7.54119 189.421 7.32471 189.155 7.32471L187.705 7.32471C187.438 7.32471 187.221 7.54119 187.221 7.80823Z' fill='url(%23paint18_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M199.639 7.80824V9.2566C199.639 9.52364 199.855 9.74024 200.123 9.74024H201.572C201.839 9.74024 202.056 9.52364 202.056 9.2566V7.80824C202.056 7.54119 201.839 7.32472 201.572 7.32472L200.123 7.32472C199.855 7.32472 199.639 7.54119 199.639 7.80824Z' fill='url(%23paint19_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M209.964 7.80824V9.2566C209.964 9.52364 210.181 9.74024 210.448 9.74024L211.897 9.74024C212.164 9.74024 212.381 9.52364 212.381 9.2566V7.80824C212.381 7.54119 212.164 7.32471 211.897 7.32471L210.448 7.32471C210.181 7.32471 209.964 7.54119 209.964 7.80824Z' fill='url(%23paint20_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M220.289 7.80823V9.2566C220.289 9.52364 220.506 9.74024 220.773 9.74024L222.222 9.74024C222.49 9.74024 222.706 9.52364 222.706 9.2566V7.80823C222.706 7.54119 222.49 7.32471 222.222 7.32471L220.773 7.32471C220.506 7.32471 220.289 7.54119 220.289 7.80823Z' fill='url(%23paint21_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M230.614 7.80823V9.2566C230.614 9.52364 230.831 9.74024 231.098 9.74024H232.548C232.815 9.74024 233.032 9.52364 233.032 9.2566V7.80823C233.032 7.54119 232.815 7.32471 232.548 7.32471L231.098 7.32471C230.831 7.32471 230.614 7.54119 230.614 7.80823Z' fill='url(%23paint22_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M240.939 7.80823V9.2566C240.939 9.52364 241.156 9.74024 241.423 9.74024H242.873C243.14 9.74024 243.357 9.52364 243.357 9.2566V7.80823C243.357 7.54119 243.14 7.32471 242.873 7.32471L241.423 7.32471C241.156 7.32471 240.939 7.54119 240.939 7.80823Z' fill='url(%23paint23_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M251.508 7.80823V9.2566C251.508 9.52364 251.724 9.74024 251.992 9.74024H253.441C253.708 9.74024 253.925 9.52364 253.925 9.2566V7.80823C253.925 7.54119 253.708 7.32471 253.441 7.32471L251.992 7.32471C251.724 7.32471 251.508 7.54119 251.508 7.80823Z' fill='url(%23paint24_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M261.833 7.80823V9.2566C261.833 9.52364 262.049 9.74024 262.317 9.74024H263.766C264.033 9.74024 264.25 9.52364 264.25 9.2566V7.80823C264.25 7.54119 264.033 7.32471 263.766 7.32471L262.317 7.32471C262.049 7.32471 261.833 7.54119 261.833 7.80823Z' fill='url(%23paint25_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M272.158 7.80823V9.2566C272.158 9.52364 272.374 9.74024 272.642 9.74024H274.092C274.358 9.74024 274.576 9.52364 274.576 9.2566L274.576 7.80823C274.576 7.54119 274.358 7.32471 274.092 7.32471L272.642 7.32471C272.374 7.32471 272.158 7.54119 272.158 7.80823Z' fill='url(%23paint26_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M282.483 7.80823V9.2566C282.483 9.52364 282.7 9.74024 282.967 9.74024H284.417C284.684 9.74024 284.901 9.52364 284.901 9.2566V7.80823C284.901 7.54119 284.684 7.32471 284.417 7.32471L282.967 7.32471C282.7 7.32471 282.483 7.54119 282.483 7.80823Z' fill='url(%23paint27_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M292.809 7.80823L292.809 9.2566C292.809 9.52364 293.025 9.74024 293.292 9.74024H294.742C295.009 9.74024 295.226 9.52364 295.226 9.2566V7.80823C295.226 7.54119 295.009 7.32471 294.742 7.32471L293.292 7.32471C293.025 7.32471 292.809 7.54119 292.809 7.80823Z' fill='url(%23paint28_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M303.134 7.80823L303.134 9.2566C303.134 9.52364 303.35 9.74024 303.617 9.74024H305.067C305.334 9.74024 305.551 9.52364 305.551 9.2566V7.80823C305.551 7.54119 305.334 7.32471 305.067 7.32471L303.617 7.32471C303.35 7.32471 303.134 7.54119 303.134 7.80823Z' fill='url(%23paint29_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M313.702 7.80823L313.702 9.2566C313.702 9.52364 313.918 9.74024 314.185 9.74024L315.635 9.74024C315.902 9.74024 316.119 9.52364 316.119 9.25659V7.80823C316.119 7.54119 315.902 7.32471 315.635 7.32471L314.185 7.32471C313.918 7.32471 313.702 7.54119 313.702 7.80823Z' fill='url(%23paint30_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M324.027 7.80823V9.25659C324.027 9.52364 324.243 9.74024 324.511 9.74024L325.961 9.74024C326.227 9.74024 326.445 9.52364 326.445 9.25659V7.80823C326.445 7.54119 326.227 7.32471 325.961 7.32471L324.511 7.32471C324.243 7.32471 324.027 7.54119 324.027 7.80823Z' fill='url(%23paint31_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M334.352 7.80823V9.25659C334.352 9.52364 334.569 9.74024 334.836 9.74024H336.286C336.553 9.74024 336.77 9.52364 336.77 9.25659L336.77 7.80823C336.77 7.54119 336.553 7.32471 336.286 7.32471L334.836 7.32471C334.569 7.32471 334.352 7.54119 334.352 7.80823Z' fill='url(%23paint32_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M344.678 7.80823V9.25659C344.678 9.52364 344.894 9.74024 345.161 9.74024H346.611C346.878 9.74024 347.095 9.52364 347.095 9.25659L347.095 7.80823C347.095 7.54119 346.878 7.32471 346.611 7.32471L345.161 7.32471C344.894 7.32471 344.678 7.54119 344.678 7.80823Z' fill='url(%23paint33_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M355.003 7.80823V9.25659C355.003 9.52364 355.219 9.74024 355.486 9.74024H356.936C357.203 9.74024 357.42 9.52364 357.42 9.25659L357.42 7.80823C357.42 7.54119 357.203 7.32471 356.936 7.32471L355.486 7.32471C355.219 7.32471 355.003 7.54119 355.003 7.80823Z' fill='url(%23paint34_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M365.328 7.80823V9.25659C365.328 9.52363 365.544 9.74024 365.811 9.74024H367.261C367.528 9.74024 367.745 9.52363 367.745 9.25659V7.80823C367.745 7.54119 367.528 7.32471 367.261 7.32471L365.811 7.32471C365.544 7.32471 365.328 7.54119 365.328 7.80823Z' fill='url(%23paint35_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M375.896 7.80823V9.25659C375.896 9.52363 376.112 9.74023 376.38 9.74023H377.83C378.096 9.74023 378.313 9.52363 378.313 9.25659V7.80823C378.313 7.54119 378.096 7.32471 377.829 7.32471L376.38 7.32471C376.112 7.32471 375.896 7.54119 375.896 7.80823Z' fill='url(%23paint36_linear_8295_70635)'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M386.221 7.80823V9.25659C386.221 9.52363 386.438 9.74023 386.705 9.74023H388.155C388.421 9.74023 388.639 9.52363 388.639 9.25659V7.80823C388.639 7.54119 388.421 7.32471 388.155 7.32471L386.705 7.32471C386.438 7.32471 386.221 7.54119 386.221 7.80823Z' fill='url(%23paint37_linear_8295_70635)'/%3E%3Cdefs%3E%3ClinearGradient id='paint0_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint1_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint2_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint3_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint4_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint5_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint6_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint7_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint8_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint9_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint10_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint11_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint12_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint13_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint14_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint15_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint16_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint17_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint18_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint19_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint20_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint21_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint22_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint23_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint24_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint25_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint26_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint27_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint28_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint29_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint30_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint31_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint32_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint33_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint34_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint35_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint36_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3ClinearGradient id='paint37_linear_8295_70635' x1='-47.5' y1='8.99989' x2='332' y2='8.99989' gradientUnits='userSpaceOnUse'%3E%3Cstop stop-color='%2329B5E8' stop-opacity='.8'/%3E%3Cstop offset='1' stop-color='%2329B5E8' stop-opacity='0'/%3E%3C/linearGradient%3E%3C/defs%3E%3C/svg%3E%0A\")}.sc-hero__inner\u003E.snowflake-flexible-column-container-items\u003Ediv:first-child{position:relative;z-index:3}.sc-hero__inner\u003E.snowflake-flexible-column-container-items\u003Ediv:last-child{position:absolute;height:100%;width:100%;top:0;left:-24px}.sc-hero__inner\u003E.snowflake-flexible-column-container-items\u003Ediv:last-child::before{content:\"\";display:block;z-index:1;position:absolute;top:-64px;left:0;width:150%;height:calc(100% + 160px);background-color:rgb(32 44 53 / .9)}.sc-body__content .heading-3-v2,.sc-hero__headline .heading-1-v2{text-transform:none}.sc-body__content span.snowflake-image-caption{display:block!important;font-style:italic}.sc-body__content .snowflake-text p+ul{margin-top:24px!important;padding-left:16px!important}.white-blue-text-color .snowflake-title-v2.solution-center-hero__certification .snowflake-typographyv2\u003Espan.snowflake-title-v2-line{color:#e9eaeb!important;font-size:16px}.white-blue-text-color .snowflake-title-v2.solution-center-hero__certification.is-large .snowflake-typographyv2\u003Espan.snowflake-title-v2-line{color:#fff!important;font-size:18px}.solution-center-hero__certification\u003E.snowflake-title-v2-line\u003Espan:first-child{display:flex;justify-content:flex-start;align-items:center;gap:8px}.solution-center-hero__certification\u003E.snowflake-title-v2-line\u003Espan:first-child::before{content:\"\";display:inline-block;width:16px;height:16px;background-image:url(\"data:image/svg+xml,%3Csvg width='16' height='16' viewBox='0 0 16 16' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M8 0C3.58146 0 0 3.58146 0 8C0 12.4185 3.58146 16 8 16C12.4185 16 16 12.4185 16 8C16 3.58146 12.4185 0 8 0ZM12.7184 5.91984L7.33471 11.3026C7.31293 11.3244 7.31293 11.3454 7.29198 11.3454L7.20653 11.4308C6.94933 11.688 6.54132 11.7525 6.21962 11.6235C6.11238 11.5808 6.00514 11.5163 5.9197 11.4308L5.83425 11.3454C5.83425 11.3454 5.83425 11.3236 5.81246 11.3236L3.28149 8.79347C2.93799 8.44997 2.93799 7.87107 3.28149 7.50664L3.36694 7.42119C3.71044 7.07769 4.28934 7.07769 4.65377 7.42119L6.58401 9.35143L11.3877 4.5477C11.7312 4.2042 12.3101 4.2042 12.6746 4.5477L12.76 4.63315C13.0826 4.99758 13.0828 5.55541 12.7184 5.91984Z' fill='%230E8A16'/%3E%3C/svg%3E%0A\");background-size:contain;background-repeat:no-repeat;background-color:#fff;border-radius:100%}.sc-hero__byline{padding-top:8px}.sc-hero__byline p{color:#e2e2e2;margin-top:0!important}.sc-hero pre[class*=language-]{overflow:visible}.snowflake-code-snippet,.snowflake-code-snippet code,.snowflake-code-snippet pre{font-size:16px}.sc-hero__code-snippet:not(pre)\u003Ecode[class*=language-],.sc-hero__code-snippet pre[class*=language-]{background:0 0}.sc-hero__code-snippet{opacity:.8;background-color:transparent!important;position:absolute;top:0;right:0;width:100%;animation:240s linear 1s forwards slow-scroll}.sc-hero__button-container .snowflake-flexible-column-container-items{padding:0 0 24px;margin-top:-8px;margin-left:24px}.sc-sidebar__partner-logo{width:100%;max-width:140px;margin-top:8px}.sc-sidebar__partner-logo .cmp-image__image{border-radius:0}.sc-tag-cluster.snowflake-text ul{list-style-type:none;padding:0;display:flex;flex-wrap:wrap;gap:8px;margin:0}.sc-tag-cluster.snowflake-text li{color:#373f41;border-radius:4px;display:inline-block;padding:6px;text-transform:uppercase;letter-spacing:1px;font-size:12px!important;line-height:12px!important;margin:0!important;background-color:#f3f3f3}.sc-body .share-icon svg{height:24px;cursor:pointer}.sc-body .share-icon svg:hover path{fill:var(--ui-02)}.sc-overview__webinar-promo-banner{align-items:center;border:1px solid #ccc;padding:var(--spacing-02)}.sc-overview__webinar-promo-banner .snowflake-content-chip-image{max-width:32px;margin-right:var(--spacing-02);line-height:0}.sc-overview__webinar-promo-banner .snowflake-content-chip-image__image,.summit-speaker-card .snowflake-card-v2-advanced-image__image{aspect-ratio:1}.sc-overview__webinar-promo-banner .snowflake-content-chip-content .heading-5-v2{font-size:14px;font-family:Lato,sans-serif}.sc-overview__webinar-promo-banner .snowflake-content-chip-content .snowflake-title-v2-line:not(:first-child){font-weight:400}.sc-overview__webinar-promo-banner .snowflake-content-chip-button .snowflake-button-container{font-size:14px!important}.diagram-group__button{position:absolute;bottom:24px;right:24px;background-color:#212c35!important}.section--mountains-bottom,.summit-hp-hero{position:relative}.sc-cert-banner{background-color:#212d35;border-radius:8px;padding:24px;overflow:hidden}.sc-cert-banner\u003E.container\u003E.cmp-container\u003E.aem-container{display:flex;flex-direction:row;align-items:center}:root{--text-secondary:#706f6f;--summit-bg-ltblue:#eaf8fd;--summit-bg-blue:#249edc;--summit-border:#d2d1d4;--summit-border-radius:8px;--summit-card-padding:32px;--summit-card-padding-sm:28px}.section--mountains-bottom::after,.section--mountains-bottom::before{content:\"\";display:block;position:absolute;bottom:-1px;max-width:400px;background-size:100% auto;height:100%;width:30%;line-height:0;background-repeat:no-repeat}.button-group\u003E.container\u003E.cmp-container\u003E.aem-container{justify-content:center;align-items:center}.button-group\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv{width:auto!important;margin:0 8px!important}.button-group .snowflake-button-container{font-family:Texta,sans-serif}.section--summit-bg-ltblue{background-color:var(--summit-bg-ltblue)}.section--summit-bg-blue,.summit-hero-secondary{background-color:var(--summit-bg-blue)}.section--mountains-bottom::before{left:0;background-image:url(\"data:image/svg+xml,%3Csvg width='402' height='309' viewBox='0 0 402 309' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M401.523 308.761H0V0L181.63 182.431L228.479 135.531L401.523 308.761Z' fill='%23249EDC'/%3E%3C/svg%3E%0A\");background-position:bottom left}.section--mountains-bottom::after{right:0;background-image:url(\"data:image/svg+xml,%3Csvg width='402' height='309' viewBox='0 0 402 309' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M0 308.761H401.523V0L219.893 182.431L173.044 135.531L0 308.761Z' fill='%23249EDC'/%3E%3C/svg%3E%0A\");background-position:bottom right}.summit-hp-hero{overflow:hidden}.summit-hero__bg-video{position:absolute;top:50%;left:50%;width:120%;height:100%;opacity:.3;transform:translate(-50%,-50%)}.summit-hero__bg-svg,.summit-prefooter__bg-image,.summit-secondary-hero__bg-image{position:absolute;bottom:0;left:0;width:100%}.summit-hp-promo-banner__headline .heading-4-v2{font-weight:900}.summit-hero-secondary .hero-lottie__left{position:absolute;bottom:0;left:0;width:30%;line-height:0}.summit-timeline__card::after,.summit-timeline__card::before{bottom:0;left:50%;position:absolute;display:block;background-color:var(--ui-01);content:\"\"}.summit-hero-secondary .snowflake-text p{font-size:24px!important;line-height:32px!important;max-width:720px;margin:0 auto}.summit-stat-container\u003E.container\u003E.cmp-container\u003E.aem-container{display:flex;justify-content:center}.summit-stat-container\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv{width:auto!important;max-width:25%}.summit-stat-container\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:not(:last-child){border-right:1px solid #fff}.summit-timeline__card{border:1px solid var(--summit-border);border-radius:var(--summit-border-radius);padding:var(--summit-card-padding);position:relative;background-color:#fff}.summit-timeline__card::before{width:20px;height:20px;border-radius:100%;transform:translate(-50%,50%)}.summit-timeline__card::after{width:3px;height:50px;transform:translate(-50%,100%)}.summit-timeline-card__icon{width:48px;height:48px}.summit-timeline-card__headline .heading-3-v2{font-size:32px}.faq-group{border:1px solid var(--ui-12);border-radius:4px;background-color:#fff}.faq-group__question{padding:24px}.faq-group__question:hover{color:var(--ui-01);cursor:pointer}.faq-group__question .heading-4-v2,.faq-group__question .heading-5-v2{position:relative;padding-right:64px}.faq-group__question .heading-4-v2::after,.faq-group__question .heading-5-v2::after{content:\"\";display:block;width:32px;height:32px;background-image:url(\"data:image/svg+xml,%3Csvg width='29' height='16' viewBox='0 0 29 16' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M14.16 14.6807C14.2537 14.7957 14.3719 14.8884 14.506 14.952C14.64 15.0157 14.7866 15.0487 14.935 15.0487C15.0834 15.0487 15.2299 15.0157 15.3639 14.952C15.498 14.8884 15.6162 14.7957 15.71 14.6807V14.6807L28.51 2.00068C29.07 1.43068 29.07 .92068 28.51 .44068C27.95 -.0393204 27.43 -.11932 26.96 .44068L14.94 12.0007L2.99996 .45068C2.90725 .322624 2.7855 .218374 2.6447 .146483C2.50389 .0745926 2.34805 .0371094 2.18996 .0371094C2.03187 .0371094 1.87603 .0745926 1.73522 .146483C1.59442 .218374 1.47267 .322624 1.37996 .45068C.819961 .93068 .819961 1.45068 1.37996 2.01068L14.16 14.6807Z' fill='black'/%3E%3C/svg%3E%0A\");background-size:80% auto;background-repeat:no-repeat;background-position:center;position:absolute;top:-2px;right:0;transition:.3s 150ms}.faq-group__question .heading-5-v2::after{top:-4px}.faq-group__answer{max-height:0;overflow:hidden;width:95%;padding:0 24px;transition:.5s}.faq-group__answer\u003Espan{display:block;padding-bottom:24px}.is-open .faq-group__answer{max-height:600px;transition:1s}.is-open .faq-group__question .heading-4-v2::after,.is-open .faq-group__question .heading-5-v2::after{transform:rotate(180deg);transition:.3s}.summit-agenda{box-shadow:2px 4px 10px 0 rgb(156 156 156 / .52);border-radius:8px;background-color:#fff;max-width:980px;margin-left:auto;margin-right:auto;padding:40px;width:90%}.agenda-item{border-radius:8px;background-color:#d4f0fa;padding:16px;border-left:4px solid var(--ui-01);position:relative}.summit-pricing-block__tile.is-past,.summit-pricing-block__tile.is-upcoming{pointer-events:none;border-color:#d2d1d4}p.agenda-item__time{width:25%;font-family:Texta!important;font-size:32px!important;font-weight:900!important;text-transform:uppercase!important;max-width:140px}@media screen and (max-width:991px){#partnerResources .section--resource-hub .snowflake-button-link .snowflake-button-container{font-size:14px!important;line-height:20px!important;margin-top:4px}#industryPartnerSlider\u003E.snowflake-flexible-column-container-items{display:flex;flex-direction:column}#industryPartnerSlider\u003E.snowflake-flexible-column-container-items\u003Ediv{width:100%}.sc-cert-banner__left{text-align:center}.sc-cert-banner__left .solution-center-hero__certification .snowflake-title-v2-line{justify-content:center}.summit-hero__bg-video{width:200%}.summit-leadership-grid .snowflake-flexible-column-container-items{grid-template-columns:repeat(2,1fr)}.summit-stat-container\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv{width:50%!important;max-width:50%!important}.summit-stat-container\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:not(:last-child){border-right:none!important}.summit-agenda{padding:24px}p.agenda-item__time{font-size:24px!important;width:auto;white-space:nowrap;padding-right:24px}}.agenda-item\u003Espan{display:flex;align-items:center}.summit-add-on-block,.summit-pricing-block{border:1px solid #d2d1d4;border-radius:8px;overflow:hidden;box-shadow:2px 4px 10px 0 rgb(156 156 156 / .52);background-color:#fff}.summit-add-on-block__content,.summit-pricing-block__content{padding:0 20px 20px}.summit-pricing-block__tile{padding:24px 20px;border-radius:4px;background:#fff;border:1px solid var(--ui-01);position:relative;transition:background-color .3s}.summit-pricing-block__tile:hover{background-color:var(--ui-01);transition:background-color .3s}.summit-pricing-block__tile.is-past{background-color:#d4f0fa}.summit-pricing-block__tile:hover .black-blue-text-color .snowflake-title-v2-line{color:#fff!important;transition:color .3s}.partner-card__logo-grid\u003E.container\u003E.cmp-container\u003E.aem-container::after,.partner-card__logo-grid\u003E.container\u003E.cmp-container\u003E.aem-container::before,.summit-add-on-block__content\u003E.container\u003E.cmp-container\u003E.aem-container::after,.summit-add-on-block__content\u003E.container\u003E.cmp-container\u003E.aem-container::before,.summit-pricing-block__tile.is-past .snowflake-content-chip-button,.summit-pricing-block__tile.is-upcoming .snowflake-content-chip-button,.summit-speaker-card .snowflake-card-v2-advanced-tag-indicator{display:none}.summit-pricing-block__tile.is-past .black-blue-text-color .snowflake-title-v2-line{color:#7cc7eb!important}.summit-pricing-block__tile.is-upcoming .black-blue-text-color .snowflake-title-v2-line{color:#8c8c8c!important}.summit-pricing-block__aside{background-color:#d4f0fa;border:1px solid #d2d1d4;border-radius:8px;padding:24px;width:100%}.summit-pricing-block__aside li::marker{color:var(--ui-01)}.summit-pricing-block__aside-headline .heading-5-v2{font-weight:900;margin-bottom:12px}.summit-pricing-block__header{background:#000;padding:24px 40px}.summit-pricing-block__header .heading-4-v2{font-weight:900;letter-spacing:.5px}.bwwidth100,.snowflake-mega-nav-dropdown-footer-content,.summit-pricing-block__tile .black-blue-text-color{width:100%}.summit-pricing-block__tile .heading-5-v2{position:static}.summit-pricing-block__tile .heading-5-v2 span.snowflake-title-v2-line:first-child{text-transform:uppercase;font-weight:900!important;letter-spacing:.25px;font-size:24px!important}.summit-pricing-block__tile .heading-5-v2 span.snowflake-title-v2-line:nth-child(2){margin-top:8px;font-family:Lato,sans-serif;font-size:14px;font-style:normal;font-weight:400;line-height:16px}.summit-pricing-block__tile .heading-5-v2 span.snowflake-title-v2-line:last-child{font-weight:900!important;font-size:40px!important}.snowflake-mega-nav-nav-item\u003Ea:hover .snowflake-mega-nav-nav-item-title-wrapper\u003E.snowflake-mega-nav-nav-item-title,.summit-pricing-block__tile:not(.is-upcoming):not(.is-past) .heading-5-v2 span.snowflake-title-v2-line:last-child{color:var(--ui-01)!important}.summit-pricing-block__tile:hover:not(.is-upcoming):not(.is-past) .heading-5-v2 span.snowflake-title-v2-line:last-child{color:#fff!important}.summit-pricing-block__tile.is-past .heading-5-v2 span.snowflake-title-v2-line:last-child{text-decoration:line-through}.summit-pricing-block__tile .snowflake-content-chip-button{margin-top:0;white-space:nowrap;display:none}.snowflake-card-v2-advanced.no-link{pointer-events:none!important}.snowpro-card{border:1px solid var(--summit-border);border-radius:var(--summit-border-radius);padding:var(--summit-card-padding-sm);display:flex;height:100%}.snowpro-card__headline{margin:24px 0 12px}.snowpro-card__pricing{margin-top:48px}.snowpro-card .snowflake-text .snowpro-card__price{color:var(--ui-01);font-weight:900;font-size:40px!important;font-family:Texta,sans-serif}.summit-stat-container\u003E.container\u003E.cmp-container\u003E.aem-container{display:flex;flex-direction:row;flex-wrap:wrap}.summit-stat-container\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:not(:last-child){border-right:1px solid var(--summit-border)}.summit-stat-card{padding:0 40px}.summit-stat .heading-2-v2 .snowflake-title-v2-line:first-child{font-size:64px;line-height:52px;margin-bottom:8px}.summit-stat .heading-2-v2 .snowflake-title-v2-line:last-child{font-size:32px;line-height:30px;margin-bottom:16px}.summit-speaker-card .snowflake-card-v2-advanced-title{margin-bottom:var(--spacing-01)}.summit-add-on-card{padding:24px;border:1px solid #d2d1d4;border-radius:8px}.summit-add-on__subhead{padding-left:40px;padding-right:40px}.partner-card__logo-grid,.partner-card__logo-single{padding:40px}.partner-card__logo-grid .snowflake-image-container .cmp-image__image,.partner-card__logo-single .snowflake-image-container .cmp-image__image{border-radius:0;max-width:240px;margin:0 auto}.partner-card\u003E.container,.partner-card\u003E.container\u003E.aem-container,.partner-card\u003E.container\u003E.cmp-container{height:100%}.summit-add-on-block__content\u003E.container\u003E.cmp-container\u003E.aem-container{display:flex;flex-direction:row;gap:24px;align-items:stretch}.partner-card__logo-grid\u003E.container\u003E.cmp-container\u003E.aem-container{display:flex;flex-direction:row;flex-wrap:wrap;gap:40px 24px;justify-content:center;align-items:center}.partner-card__logo-grid\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv{width:calc(33.3333% - 24px);margin:0!important}.partner-card{border-radius:8px;border:1px solid #d2d1d4;overflow:hidden;height:100%;background-color:#fff}.partner-card__header{padding:16px 24px;border-bottom:1px solid #d2d1d4}.partner-card__header.is-purple{background-color:#7d44cf}.partner-card__header h4{display:flex;flex-direction:row!important;align-items:center;gap:12px}.partner-card__header h4::before{vertical-align:middle;content:\"\";display:inline-block;width:20px;height:20px;background-size:contain;background-repeat:no-repeat;background-image:url(\"data:image/svg+xml,%3Csvg width='21' height='23' viewBox='0 0 21 23' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M20.0375 12.8374C20.1644 12.439 20.2172 12.0289 20.2077 11.6237C20.193 11.3305 20.1548 11.0373 20.0712 10.7441C19.8196 9.83306 19.223 9.01989 18.3294 8.50724L5.61817 1.2017C3.82388 .173815 1.53618 .784335 .506483 2.56804C-.533615 4.34915 .0797871 6.62351 1.87408 7.65398L8.97715 11.7427L1.87408 15.8201C.0797871 16.8527 -.531016 19.1271 .506483 20.9156C1.53618 22.6941 3.82388 23.302 5.61817 22.2746L18.3294 14.9643C19.1871 14.4728 19.7693 13.7027 20.0375 12.8374Z' fill='black'/%3E%3C/svg%3E%0A\")}.partner-card__header.is-purple h4::before{background-image:url(\"data:image/svg+xml,%3Csvg width='21' height='23' viewBox='0 0 21 23' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M20.0375 12.8374C20.1644 12.439 20.2172 12.0289 20.2077 11.6237C20.193 11.3305 20.1548 11.0373 20.0712 10.7441C19.8196 9.83306 19.223 9.01989 18.3294 8.50724L5.61817 1.2017C3.82388 .173815 1.53618 .784335 .506483 2.56804C-.533615 4.34915 .0797871 6.62351 1.87408 7.65398L8.97715 11.7427L1.87408 15.8201C.0797871 16.8527 -.531016 19.1271 .506483 20.9156C1.53618 22.6941 3.82388 23.302 5.61817 22.2746L18.3294 14.9643C19.1871 14.4728 19.7693 13.7027 20.0375 12.8374Z' fill='white'/%3E%3C/svg%3E%0A\")}.sf-blue-mountains{background-size:90% auto;background-repeat:no-repeat;background-position:center bottom;background-image:url(\"data:image/svg+xml,%3Csvg width='1361' height='410' viewBox='0 0 1361 410' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M1360.25 410L1065.53 114.309L976.256 203.875L773.049 0L364.393 410H1360.25Z' fill='%233AA8DF'/%3E%3Cpath d='M274.778 410L137.467 272.238L.15625 410H274.778Z' fill='%233AA8DF'/%3E%3C/svg%3E%0A\")}.bwalignr,.main-pr-body .bwalignr{text-align:right}.bwblockalignl{margin-left:0;margin-right:auto}.bwcellpmargin{margin-top:0;margin-bottom:0}.bwlistdisc{list-style-type:disc}.bwpadb3{padding-bottom:4px}.bwpadb4{padding-bottom:5px}.bwpadl0{padding-left:0}.bwpadl3{padding-left:15px}.bwpadl6{padding-left:30px}.bwpadl9{padding-left:45px}.bwpadl12{padding-left:60px}.bwpadr0{padding-right:0}.bwtablemarginb{margin-bottom:10px}.bwvertalignb{vertical-align:bottom}.bwvertalignt{vertical-align:top}.bwsinglebottom{border-bottom:1pt solid #000}.bwdoublebottom{border-bottom:2.25pt double #000}.bwwidth1{width:1%}.bwwidth2{width:2%}.bwwidth6{width:6%}.bwwidth7{width:7%}.bwwidth8{width:8%}.bwwidth10{width:10%}.bwwidth12{width:12%}.bwwidth32{width:32%}.bwwidth44{width:44%}.bwwidth72{width:72%}.bwwidth97{width:97%}.main-pr-body{font-size:18px;line-height:26px}.main-pr-body img{display:block;width:100%;height:auto!important;border-radius:var(--small-border-radius)}.main-pr-body table{width:100%;display:block}.main-pr-body tbody{background-color:#f7f7f7}.main-pr-body .bwsinglebottom{border-bottom:1pt solid #000!important}.main-pr-body td.bwwidth44{padding-right:40px}.main-pr-body .bw-release-story{font-family:Lato,sans-serif}.main-pr-body .bw-release-story sup,.snowflake-mega-nav-dropdown-header-content-right a{white-space:nowrap}.main-pr-body .bw-release-story\u003E*,.main-pr-body\u003Espan\u003E*{margin-bottom:2rem!important}.snowflake-text.main-pr-body tbody,.snowflake-text.main-pr-body tbody p{font-size:14px!important;line-height:20px!important;width:100%;display:block}.press-body .snowflake-flexible-column-container-items{gap:var(--spacing-08)}.about-snowflake{border:1px solid #ccc;background-color:var(--ui-background-05);padding:24px;border-radius:8px;margin-top:0}.about-snowflake__logo{max-width:140px;margin-top:16px}.hero--press .snowflake-hero-system-inner{max-width:1408px;margin:0 auto!important}#arcticNavItem{flex-direction:column}#arcticNavItem::before{content:\"Featured Open Source Technologies\";display:block;margin-top:48px;margin-bottom:24px;font-size:16px!important;line-height:16px!important;font-weight:800!important;text-transform:uppercase}@media screen and (min-width:768px){.sc-hero__inner\u003E.snowflake-flexible-column-container-items\u003Ediv:last-child{position:relative;height:100%;top:auto;left:auto;width:auto}.sc-hero__inner\u003E.snowflake-flexible-column-container-items\u003Ediv:last-child::before{background:linear-gradient(180deg,#202c35 -7.5%,#fff0 51.25%,#202c35 107.69%)}.sc-hero__byline\u003Espan{display:flex;flex-wrap:wrap}.sc-hero__byline p:not(:last-child)::after{content:\"|\";margin:0 12px;opacity:.5}.sc-hero__button-container .snowflake-flexible-column-container-items{position:absolute;bottom:0;padding:0;margin:0 24px 0 0}.sc-hero__button-container .hero-watch-the-demo{padding:12px 16px!important;float:right;margin-bottom:48px;background-color:rgb(35 45 54 / .8)}.summit-overview-stat{padding:0 40px}.summit-timeline{border-bottom:3px solid var(--ui-01);margin-bottom:64px}.summit-add-on-block__content,.summit-pricing-block__content{padding:0 40px 40px}#arcticNavItem::before{font-size:12px!important;margin-bottom:8px;margin-top:16px}.snowflake-mega-nav-nav-item-title-wrapper\u003E.snowflake-mega-nav-nav-item-title{line-height:20px!important}.snowflake-card .heading-2.snowflake-title-line{font-size:24px!important;line-height:28px!important}}@media screen and (min-width:992px){.hp-hero__eyebrow a{gap:12px;margin-left:0;margin-right:0}.hp-hero__eyebrow a::after{content:\"\";background-image:url(\"data:image/svg+xml,%3Csvg width='6' height='11' viewBox='0 0 6 11' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M5.49134 5.79438C5.53447 5.75922 5.56923 5.71489 5.5931 5.66463C5.61697 5.61436 5.62935 5.55941 5.62935 5.50376C5.62935 5.44811 5.61697 5.39316 5.5931 5.34289C5.56923 5.29263 5.53447 5.2483 5.49134 5.21314L.736339 .413136C.522589 .203135 .331339 .203135 .151339 .413136C-.0286612 .623135 -.0586612 .818135 .151339 .994386L4.48634 5.50188L.155089 9.97938C.107068 10.0142 .0679743 10.0598 .0410153 10.1126C.0140562 10.1654 0 10.2238 0 10.2831C0 10.3424 .0140562 10.4009 .0410153 10.4537C.0679743 10.5065 .107068 10.5521 .155089 10.5869C.335089 10.7969 .530089 10.7969 .740089 10.5869L5.49134 5.79438Z' fill='black'/%3E%3C/svg%3E%0A\");display:inline-block;width:12px;height:12px;background-repeat:no-repeat;background-size:auto 100%;background-position:left center}.promo-banner--homepage{padding-top:32px}.homepage-banner-offset-container::after{height:50%}#storyHighlights{padding:2rem}.body-display-v2.snowflake-quote-item-quote-text{line-height:28px!important}.snowflake-hero-system-headline .heading-1-v2{line-height:48px;font-size:54px!important}.sc-overview__webinar-promo-banner .snowflake-content-chip-content{flex-direction:row;justify-content:space-between;align-items:center;width:100%}.sc-overview__webinar-promo-banner .snowflake-content-chip-content .heading-5-v2{flex-direction:row}.sc-overview__webinar-promo-banner .snowflake-content-chip-content .snowflake-title-v2-line:not(:first-child)::before{content:\"|\";margin:0 6px}.sc-cert-banner{padding:40px}.sc-cert-banner\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv{margin:0!important;width:50%}.sc-cert-banner\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:first-child{flex-grow:1;padding-right:24px}.sc-cert-banner\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:last-child{max-width:240px}.summit-pricing-block__content\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:last-child{width:70%;padding-left:40px}.summit-pricing-block__content\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:first-child{width:30%}.summit-add-on-block__content\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv{width:calc(33.3333% - 24px);margin:0!important;display:flex}.summit-pricing-block__tile .snowflake-content-chip-content{display:flex;flex-direction:row;align-items:center;width:calc(100% - 200px)}.summit-pricing-block__tile .heading-5-v2 span.snowflake-title-v2-line:last-child{position:absolute;top:50%;transform:translate(0,-50%);right:40px}.press-body\u003E.snowflake-flexible-column-container-items\u003Ediv:last-child{position:sticky;top:120px}.snowflake-mega-nav-navigation-title:hover{color:var(--ui-01)}}@media screen and (min-width:1024px){.about-snowflake{padding:28px}.about-snowflake__logo{max-width:none;padding:0 0 0 48px;margin-bottom:0}.hero--press .snowflake-hero-system-layout-70-30 .snowflake-hero-system-content-container{width:85%}.snowflake-hero-system{padding-bottom:var(--spacing-04);padding-top:var(--spacing-07)}.hero--press .display-2-v2{font-size:64px;line-height:56px}.about-snowflake\u003E.container\u003E.cmp-container\u003E.aem-container{flex-direction:row;flex-wrap:nowrap;align-items:center}.about-snowflake\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:last-child{max-width:280px}.about-snowflake\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:first-child{flex-grow:1;margin-bottom:0!important}#polarisNavItem{margin-top:40px}.snowflake-mega-nav-nav-item-description{line-height:18px!important}.snowflake-mega-nav-column-items{gap:var(--spacing-01);grid-gap:var(--spacing-01)}.snowflake-mega-nav-navigation-title{text-transform:none}}div[id*=blueIcon] .snowflake-mega-nav-nav-item-icon__inner{background:var(--ui-01);padding:8px}div[id*=blueIcon]:hover .snowflake-mega-nav-nav-item-icon__inner{background:var(--ui-01)!important}.snowflake-mega-nav-nav-item-icon__inner{border-radius:4px;background:var(--ui-background-05);padding:6px}.snowflake-mega-nav-nav-item:hover .snowflake-mega-nav-nav-item-icon__inner{background:#fff!important}.snowflake-mega-nav-nav-item-icon.snowflake-image-container{height:40px;width:40px}.snowflake-mega-nav-dropdown-footer-links\u003E.snowflake-button-link\u003E.snowflake-button-container{font-size:16px!important;font-family:Texta!important;font-weight:800!important}.snowflake-mega-nav-dropdown-footer-icon.snowflake-image-container{margin-right:8px;width:40px!important;height:40px!important}#viewAllCapabilities a:hover{background:0 0!important}#platformFooter .snowflake-title-v2 .snowflake-title-v2-line:last-child{font-family:Lato;font-size:14px;font-weight:500}#platformFooter .snowflake-mega-nav-dropdown-footer-links{flex-grow:1;justify-content:flex-end;align-items:center}#platformFooter .snowflake-mega-nav-dropdown-footer-content{flex-direction:row}#offset,#open-source{flex-direction:column;border-top:1px solid #ccc}#offset::before,#open-source::before{content:\" \";display:block;width:100%;font-weight:800!important;font-size:12px!important;line-height:14px;text-transform:uppercase;white-space:nowrap;margin-top:16px;margin-bottom:8px}#open-source::before{content:\"Open Source Technologies\"}.snowflake-mega-nav-dropdown-menu-close-button{margin:var(--spacing-04) 0 var(--spacing-03)}.snowflake-mega-nav-column{gap:var(--spacing-02)!important}.snowflake-mega-nav-nav-item\u003Ea{width:100%;margin-left:-8px;padding:8px;border-radius:4px}.snowflake-mega-nav-nav-item\u003Ea:hover{background-color:var(--ui-background-05)}.snowflake-mega-nav-nav-item-description{margin-top:2px;display:block}#promobanner_overflowBottomDarkBlue::before{content:'';display:block;position:absolute;bottom:0;left:0;width:100%;height:50%;background:#212d35}#promobanner_overflowTopDarkBlue::before{content:'';display:block;position:absolute;top:0;left:0;width:100%;height:50%;background:#212d35}.overview-card\u003Ediv{box-shadow:0 0 14px 0 rgba(0,0,0,.10);background-color:#fff;border-radius:16px;overflow:hidden}.overview-card-text{padding:40px}.overview-card-image img{border-radius:0 !important}.overview-card-text h3,.overview-card-text .heading-3-v2{font-size:18px;line-height:1.1;margin-top:0}",":type":"snowflake-site/components/markup-editor","isGSAPEnabled":false},"mega_header":{"additionalClasses":"heap-nav-header","layout":"SIMPLE","id":"container-28ddda232e",":items":{"nav_mega":{"activeItem":"item_1719963657751_c_663444255","id":"tabs-a94705570a",":type":"snowflake-site/components/nav/nav-mega",":items":{"item_1719963657751_c_663444255":{"id":"nav-dropdown-menu-6a5550f2ce","enableDropdown":true,"nav_column_container":{"layout":"SIMPLE","id":"container-887710fabe",":items":{"nav_column":{"additionalClasses":"nav-platform-sidebar","numberOfSubColumns":"one-column","minWidth":"230","maxWidth":"350","layout":"SIMPLE","id":"container-120c1e56b0",":items":{"nav_item_copy_copy_2_793631646":{"id":"nav-item-3be07e9027","additionalClasses":"nav-item__platform-parent is-platform","linkDescription":"Develop AI products, apps and more on a fully managed platform that securely connects businesses globally — across any type or scale of data.","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/product/platform/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"The Snowflake Platform"},":type":"snowflake-site/components/nav/nav-item"},"nav_item":{"id":"nav-item-f7fde9bd5c","additionalClasses":"nav-item nav-item--si is-si","linkDescription":"All your knowledge. One trusted enterprise agent.","flag":"NOW GA","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/product/snowflake-cowork/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Snowflake CoWork"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_copy_2_836345186":{"id":"nav-item-5202ee09b7","additionalClasses":"blue-icon is-analytics","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/product/analytics/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Analytics"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_copy_2":{"id":"nav-item-42ad1f4416","additionalClasses":"blue-icon is-ai","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/product/ai/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"AI"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_copy_2_1314771042":{"id":"nav-item-9e7b88899d","additionalClasses":"blue-icon is-data-eng","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/product/data-engineering/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Data Engineering"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_144634":{"id":"nav-item-9ff6b840e3","additionalClasses":"blue-icon is-apps-collab","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/product/applications-and-collaboration/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Applications & Collaboration"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_144634_2013333117":{"id":"nav-item-3fcf2a2bff","additionalClasses":"blue-icon is-transactions","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/product/transactions/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Transactions"},":type":"snowflake-site/components/nav/nav-item"}},":type":"snowflake-site/components/nav/nav-column",":itemsOrder":["nav_item_copy_copy_2_793631646","nav_item","nav_item_copy_copy_2_836345186","nav_item_copy_copy_2","nav_item_copy_copy_2_1314771042","nav_item_copy_144634","nav_item_copy_144634_2013333117"]},"nav_column_copy_copy":{"additionalClasses":"meganav-platform-features","navColumnTitle":"Featured Capabilities","numberOfSubColumns":"one-column","layout":"SIMPLE","id":"container-8aa016eeda",":items":{"nav_item_copy_212715":{"id":"nav-item-430fcb7398","additionalClasses":"is-cortex-code","linkDescription":"Snowflake-native AI coding agent ","flag":"Now GA","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/product/snowflake-coco/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Snowflake CoCo"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_511717659_c":{"id":"nav-item-21d4c63fe9","additionalClasses":"is-observe","linkDescription":"AI-powered observability for faster production troubleshooting","flag":"New","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/product/observe/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Observe"},":type":"snowflake-site/components/nav/nav-item"},"nav_item":{"id":"nav-item-876bef2da0","additionalClasses":"is-cortex-ai","linkDescription":"Instant access to industry-leading LLMs","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"https://www.snowflake.com/en/product/features/cortex/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_EXTERNAL","text":"Cortex AI"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_660590635":{"id":"nav-item-b75afe39e0","additionalClasses":"is-marketplace","linkDescription":"Third-party data sources connected within minutes","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"https://www.snowflake.com/en/product/features/marketplace/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_EXTERNAL","text":"Marketplace"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_660590":{"id":"nav-item-35b9fbf0da","additionalClasses":"is-snowpark","linkDescription":"Libraries and code execution environments that run Python and more","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"https://www.snowflake.com/en/product/features/snowpark/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_EXTERNAL","text":"Snowpark"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_660590_1664056128":{"id":"nav-item-5eb24238e0","additionalClasses":"is-snowpark","linkDescription":"Compute that right-sizes itself for every workload","flag":"Now GA","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/product/features/adaptive-compute/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Adaptive Compute"},":type":"snowflake-site/components/nav/nav-item"}},":type":"snowflake-site/components/nav/nav-column",":itemsOrder":["nav_item_copy_212715","nav_item_511717659_c","nav_item","nav_item_copy_660590635","nav_item_copy_660590","nav_item_copy_660590_1664056128"]},"nav_column_692142673":{"navColumnTitle":" ","numberOfSubColumns":"one-column","layout":"SIMPLE","id":"container-1592d85d9b",":items":{"nav_item_copy_660590_1739526127":{"id":"nav-item-20411f9702","additionalClasses":"is-postgres","linkDescription":"Fully compatible open source Postgres running on Snowflake","flag":"Now GA","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/product/features/postgres/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Postgres"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_660590_2011265650":{"id":"nav-item-26d617dd20","additionalClasses":"is-streamlit","linkDescription":"Framework for transforming Python scripts into web apps","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"https://www.snowflake.com/en/product/features/streamlit-in-snowflake/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_EXTERNAL","text":"Streamlit"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_185565":{"id":"nav-item-e6befa3a83","additionalClasses":"is-dcr","linkDescription":"Streamlined model development and MLOps from a centralized UI","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/product/features/end-to-end-ml-workflows/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Snowflake ML"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_212715":{"id":"nav-item-5c48dbf5cd","additionalClasses":"is-openflow","linkDescription":"Effortless data movement for integrations","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"https://www.snowflake.com/en/product/features/openflow/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_EXTERNAL","text":"Openflow"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_660590":{"id":"nav-item-aab9fc89bd","additionalClasses":"is-notebooks","linkDescription":"Interactive dev environment for data and AI teams","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"https://www.snowflake.com/en/product/features/notebooks/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_EXTERNAL","text":"Notebooks"},":type":"snowflake-site/components/nav/nav-item"}},":type":"snowflake-site/components/nav/nav-column",":itemsOrder":["nav_item_copy_660590_1739526127","nav_item_copy_660590_2011265650","nav_item_copy_185565","nav_item_copy_212715","nav_item_copy_660590"]},"nav_column_782221091":{"navColumnTitle":" ","numberOfSubColumns":"one-column","layout":"SIMPLE","id":"container-fff05e03bc",":items":{"nav_item_258535199":{"id":"nav-item-d9d554700d","propertiesId":"workload-nav-1","additionalClasses":"is-native-apps","linkDescription":"End-to-end, Snowflake-native app creation and distribution","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"https://www.snowflake.com/en/product/features/native-apps/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_EXTERNAL","text":"Native Apps"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy":{"id":"nav-item-f4df54609e","additionalClasses":"is-light-gray-icon is-horizon-catalog","linkDescription":"Universal AI catalog","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/product/features/horizon/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Horizon Catalog"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_660590_1293798742":{"id":"nav-item-ad41859397","additionalClasses":"is-snowflake-ml","linkDescription":"Governed context layer that keeps AI, BI and data apps working from one truth","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/product/features/horizon-context/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Horizon Context"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_511717659_c":{"id":"nav-item-2d26845462","additionalClasses":"is-unistore","linkDescription":"Unify transactional and analytical workloads in Snowflake for enhanced simplicity","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/product/features/unistore/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Unistore"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_511717659_c_1006104884":{"id":"nav-item-253cf2a492","additionalClasses":"is-observe","linkDescription":"Use any engine on a single governed data copy","flag":"Now GA","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/product/use-cases/interoperable-lakehouse/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Interoperable Lakehouse"},":type":"snowflake-site/components/nav/nav-item"}},":type":"snowflake-site/components/nav/nav-column",":itemsOrder":["nav_item_258535199","nav_item_copy","nav_item_copy_660590_1293798742","nav_item_511717659_c","nav_item_511717659_c_1006104884"]}},":type":"snowflake-site/components/nav/nav-column/nav-column-container",":itemsOrder":["nav_column","nav_column_copy_copy","nav_column_692142673","nav_column_782221091"]},":type":"snowflake-site/components/nav/nav-dropdown-menu","cq:panelTitle":"Product"},"nav_dropdown_menu_2":{"id":"nav-dropdown-menu-d26f95c13b","enableDropdown":true,"nav_column_container":{"layout":"SIMPLE","id":"container-0619ad7aea",":items":{"nav_column":{"navColumnTitle":"INDUSTRIES","numberOfSubColumns":"one-column","minWidth":"280","layout":"SIMPLE","id":"container-df5ad8fa86",":items":{"nav_item_copy_361384_2056203141":{"id":"nav-item-6695b0e96d","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/solutions/industries/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"All Industries"},":type":"snowflake-site/components/nav/nav-item"},"nav_item":{"id":"nav-item-325bcfea73","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/solutions/industries/advertising-media-entertainment/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Advertising, Media & Entertainment"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy":{"id":"nav-item-2ace2f7d02","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/solutions/industries/financial-services/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Financial Services"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_1970515619":{"id":"nav-item-1d6d28bec9","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/solutions/industries/healthcare-and-life-sciences/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Healthcare & Life Sciences"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_1533429516":{"id":"nav-item-b446e128f6","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/solutions/industries/manufacturing/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Manufacturing"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_1444458226":{"id":"nav-item-85bc8dce09","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/solutions/industries/public-sector/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Public Sector"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_1149488919":{"id":"nav-item-784e51af14","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/solutions/industries/retail-consumer-goods/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Retail & Consumer Goods"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_57417040":{"id":"nav-item-68e0f746a0","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/solutions/industries/technology/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Technology"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_361384674":{"id":"nav-item-c150ed9440","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/solutions/industries/telecom/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Telecom"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_361384":{"id":"nav-item-769db15910","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/solutions/industries/travel-hospitality/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Travel & Hospitality"},":type":"snowflake-site/components/nav/nav-item"}},":type":"snowflake-site/components/nav/nav-column",":itemsOrder":["nav_item_copy_361384_2056203141","nav_item","nav_item_copy","nav_item_copy_1970515619","nav_item_copy_1533429516","nav_item_copy_1444458226","nav_item_copy_1149488919","nav_item_copy_57417040","nav_item_copy_361384674","nav_item_copy_361384"],"appliedCssClassNames":"snowflake-responsive-container-inner-padding-extra-small"},"nav_column_copy":{"navColumnTitle":"DEPARTMENTS","numberOfSubColumns":"one-column","minWidth":"160","layout":"SIMPLE","id":"container-714fa0a3af",":items":{"nav_item":{"id":"nav-item-828c348166","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/solutions/departments/finance/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Finance"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_1970515619":{"id":"nav-item-71c9e49db4","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"https://www.snowflake.com/en/solutions/departments/marketing/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_EXTERNAL","text":"Marketing"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_197051":{"id":"nav-item-d0e1396be2","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/solutions/departments/sales/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Sales"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy":{"id":"nav-item-73d3033c10","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/solutions/departments/information-technology/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"IT"},":type":"snowflake-site/components/nav/nav-item"}},":type":"snowflake-site/components/nav/nav-column",":itemsOrder":["nav_item","nav_item_copy_1970515619","nav_item_copy_197051","nav_item_copy"]},"nav_column_833417450":{"navColumnTitle":"Enablement Solutions","numberOfSubColumns":"one-column","layout":"SIMPLE","id":"container-76be9ce16b",":items":{"nav_item_copy_107772":{"id":"nav-item-8f246c2327","linkDescription":"Confident migration to a unified platform","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/migrate-to-the-cloud/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Migrate to the AI Data Cloud"},"icon":{"id":"icon","height":"64","alt":"Cloud icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/nav_dropdown_menu_2/nav_column_container/nav_column_833417450/nav_item_copy_107772/icon.coreimg.svg/1723828484100/nav-icon-cloud.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_copy":{"id":"nav-item-a3defc7fbd","linkDescription":"Snowflake experts to help you accelerate and achieve business goals","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/solutions/services-delivery/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Services Delivery"},"icon":{"id":"icon","height":"64","alt":"Migrate icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/nav_dropdown_menu_2/nav_column_container/nav_column_833417450/nav_item_copy_copy/icon.coreimg.svg/1768354429188/nav-icon--migrate.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"}},":type":"snowflake-site/components/nav/nav-column",":itemsOrder":["nav_item_copy_107772","nav_item_copy_copy"]},"nav_column_copy_copy":{"navColumnTitle":"PARTNER SOLUTIONS","numberOfSubColumns":"one-column","layout":"SIMPLE","id":"container-adf1f4c076",":items":{"nav_item":{"id":"nav-item-ba3b2d012f","linkDescription":"Programs with product, solutions and cloud partners","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/why-snowflake/partners/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Snowflake Partner Network"},"icon":{"id":"icon","height":"64","alt":"Partner Network icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/nav_dropdown_menu_2/nav_column_container/nav_column_copy_copy/nav_item/icon.coreimg.svg/1723828498700/nav-icon--partner-network.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy":{"id":"nav-item-0cf85f0993","linkDescription":"Partners, apps and solutions for enhanced deployment","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/why-snowflake/partners/all-partners/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Partner Finder"},"icon":{"id":"icon","height":"64","alt":"Partner Finder icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/nav_dropdown_menu_2/nav_column_container/nav_column_copy_copy/nav_item_copy/icon.coreimg.svg/1726173927645/nav-icon--partner-finder.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_1970515619":{"id":"nav-item-6525cb01ef","linkDescription":"Live and virtual events","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/why-snowflake/partners/event-partnership-opportunities/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Event Partnership Opportunities"},"icon":{"id":"icon","height":"64","alt":"Calendar icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/nav_dropdown_menu_2/nav_column_container/nav_column_copy_copy/nav_item_copy_1970515619/icon.coreimg.svg/1726173935655/nav-icon--events.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"}},":type":"snowflake-site/components/nav/nav-column",":itemsOrder":["nav_item","nav_item_copy","nav_item_copy_1970515619"]}},":type":"snowflake-site/components/nav/nav-column/nav-column-container",":itemsOrder":["nav_column","nav_column_copy","nav_column_833417450","nav_column_copy_copy"]},":type":"snowflake-site/components/nav/nav-dropdown-menu","cq:panelTitle":"Solutions"},"item_1719963657751_c":{"id":"nav-dropdown-menu-c27d4e5024","enableDropdown":true,"nav_column_container":{"layout":"SIMPLE","id":"container-69c0dccee6",":items":{"nav_column":{"numberOfSubColumns":"one-column","minWidth":"230","maxWidth":"350","layout":"SIMPLE","id":"container-7f0a4d7214",":items":{"nav_item_copy_copy_2_793631646":{"id":"nav-item-19ac3df140","additionalClasses":"nav-item__platform-parent-why-sf","linkDescription":"Collaborate locally and globally to reveal new insights, create previously unforeseen business opportunities, and identify your customers with seamless experiences.","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/why-snowflake/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Why Snowflake"},":type":"snowflake-site/components/nav/nav-item"}},":type":"snowflake-site/components/nav/nav-column",":itemsOrder":["nav_item_copy_copy_2_793631646"]},"nav_column_copy_copy":{"additionalClasses":"meganav-platform-features","numberOfSubColumns":"two-columns","maxWidth":"1200","layout":"SIMPLE","id":"container-6542bd9da3",":items":{"nav_item":{"id":"nav-item-12d64d4c20","propertiesId":"testID","linkDescription":"Case studies and videos showcasing how global organizations use Snowflake","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/customers/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Customers"},"icon":{"id":"icon","height":"64","alt":"Customer icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719963657751_c/nav_column_container/nav_column_copy_copy/nav_item/icon.coreimg.svg/1739839279367/nav-icon--partner-network.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_258535199":{"id":"nav-item-9567af40d4","propertiesId":"workload-nav-1","linkDescription":"Learn how to connect, share and integrate the data and apps on the AI Data Cloud","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/why-snowflake/what-is-data-cloud/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"The AI Data Cloud Explained"},"icon":{"id":"icon","height":"64","alt":"Cloud icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719963657751_c/nav_column_container/nav_column_copy_copy/nav_item_258535199/icon.coreimg.svg/1739840490955/nav-icon-cloud.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_185565":{"id":"nav-item-0c4c6154d6","linkDescription":"Comprehensive security through built-in features, robust cloud infrastructure protection, and more","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/why-snowflake/snowflake-security-hub/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Security Hub"},"icon":{"id":"icon","height":"64","alt":"User with security lock icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719963657751_c/nav_column_container/nav_column_copy_copy/nav_item_copy_185565/icon.coreimg.svg/1758909528089/user-security-admins-ciso-icon.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy":{"id":"nav-item-64695433df","additionalClasses":"is-light-gray-icon","linkDescription":"Reduce compute costs across queries, warehouses and storage","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/pricing-options/cost-and-performance-optimization/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Cost Optimization"},"icon":{"id":"icon","height":"64","alt":"Cost Optimization icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719963657751_c/nav_column_container/nav_column_copy_copy/nav_item_copy/icon.coreimg.svg/1791996291063/nav-icon-cost-optimization-performance.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_185565_903555964":{"id":"nav-item-47579a15b6","linkDescription":"Startups building applications in the AI Data Cloud","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/why-snowflake/startup-program/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Snowflake for Startups"},"icon":{"id":"icon","height":"64","alt":"Launch","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719963657751_c/nav_column_container/nav_column_copy_copy/nav_item_copy_185565_903555964/icon.coreimg.svg/1758732224323/launch.svg","lazyEnabled":true,"isLcpImage":false,"width":"65",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_copy":{"id":"nav-item-c502727653","additionalClasses":"is-light-gray-icon","linkDescription":"Keep queries fast as workloads grow with automatic tuning ","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/why-snowflake/workload-optimization/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Workload Optimization"},"icon":{"id":"icon","height":"64","alt":"Management Optimizer","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719963657751_c/nav_column_container/nav_column_copy_copy/nav_item_copy_copy/icon.coreimg.svg/1791996347807/management-optimizer.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"}},":type":"snowflake-site/components/nav/nav-column",":itemsOrder":["nav_item","nav_item_258535199","nav_item_copy_185565","nav_item_copy","nav_item_copy_185565_903555964","nav_item_copy_copy"]}},":type":"snowflake-site/components/nav/nav-column/nav-column-container",":itemsOrder":["nav_column","nav_column_copy_copy"]},":type":"snowflake-site/components/nav/nav-dropdown-menu","cq:panelTitle":"Why Snowflake"},"item_1719961362824":{"id":"nav-dropdown-menu-3f3c291291","enableDropdown":true,"nav_column_container":{"layout":"SIMPLE","id":"container-7ba9508aef",":items":{"nav_column_copy":{"navColumnTitle":"Connect","numberOfSubColumns":"one-column","minWidth":"124","layout":"SIMPLE","id":"container-081d1e1212",":items":{"nav_item":{"id":"nav-item-4baee8f3bf","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/blog/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Blog"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_180298689":{"id":"nav-item-2384e1952f","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/events/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Events"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_1639361946":{"id":"nav-item-489c9f3a63","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"https://www.snowflake.com/en/support/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_EXTERNAL","text":"Support"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_680912746":{"id":"nav-item-bc8c6878ef","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"https://www.snowflake.com/en/contact/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_EXTERNAL","text":"Contact us"},":type":"snowflake-site/components/nav/nav-item"}},":type":"snowflake-site/components/nav/nav-column",":itemsOrder":["nav_item","nav_item_180298689","nav_item_1639361946","nav_item_680912746"]},"nav_column_44600420__826130542":{"navColumnTitle":"Learn","numberOfSubColumns":"two-columns","layout":"SIMPLE","id":"container-fd51d9c3eb",":items":{"nav_item_copy":{"id":"nav-item-a66cdfb4e7","linkDescription":"Ebooks, videos, white papers and more","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/resources/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Resource Library"},"icon":{"id":"icon","height":"64","alt":"Notebooks icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719961362824/nav_column_container/nav_column_44600420__826130542/nav_item_copy/icon.coreimg.svg/1736877128196/nav-icon--notebooks.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item":{"id":"nav-item-1748ccbb73","linkDescription":"Overview of Snowflake's educational offerings","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"https://www.snowflake.com/en/resources/learn/training/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_EXTERNAL","text":"Training"},"icon":{"id":"icon","height":"64","alt":"Training icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719961362824/nav_column_container/nav_column_44600420__826130542/nav_item/icon.coreimg.svg/1722385094416/nav-icon--training.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_144634_1984107859":{"id":"nav-item-62ff5aab27","linkDescription":"Expert-led discussions and demos across industries and use cases","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/webinars/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Webinars"},"icon":{"id":"icon","height":"64","alt":"Webinars icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719961362824/nav_column_container/nav_column_44600420__826130542/nav_item_copy_144634_1984107859/icon.coreimg.svg/1759424691990/nav-icon--webinars.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_1438098918":{"id":"nav-item-dbe6b20b3a","linkDescription":"Snowflake's technical industry professional certifications","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"https://www.snowflake.com/en/resources/learn/certifications/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_EXTERNAL","text":"Certifications"},"icon":{"id":"icon","height":"64","alt":"Certification icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719961362824/nav_column_container/nav_column_44600420__826130542/nav_item_copy_1438098918/icon.coreimg.svg/1722382780833/nav-icon--cert.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_143809":{"id":"nav-item-5d8d0a552a","linkDescription":"Weekly product demos showcasing key features and live Q&A ","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/webinars/demo/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Live Demos"},"icon":{"id":"icon","height":"64","alt":"Live Demo icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719961362824/nav_column_container/nav_column_44600420__826130542/nav_item_copy_143809/icon.coreimg.svg/1759424359543/nav-icon--live-demo.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_333890638":{"id":"nav-item-01f03ccd1e","linkDescription":"Training courses for all levels, on-demand or instructor-led","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"attributes":{"target":"_blank"},"url":"https://learn.snowflake.com/en/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_EXTERNAL","text":"Snowflake University"},"icon":{"id":"icon","height":"64","alt":"Education icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719961362824/nav_column_container/nav_column_44600420__826130542/nav_item_copy_333890638/icon.coreimg.svg/1722382769808/nav-icon--education.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_189945":{"id":"nav-item-d621cb0a61","linkDescription":"Instructor-led virtual workshops for exploring key Snowflake features","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/webinars/virtual-hands-on-lab/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Hands-On Labs"},"icon":{"id":"icon","height":"64","alt":"Hands-on Labs icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719961362824/nav_column_container/nav_column_44600420__826130542/nav_item_copy_189945/icon.coreimg.svg/1759388182903/nav-icon--labs.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_333890":{"id":"nav-item-39ea203d74","linkDescription":"Academic papers written by Snowflake researchers","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/resources/publications/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Snowflake Research Publications"},"icon":{"id":"icon","height":"64","alt":"Copy","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719961362824/nav_column_container/nav_column_44600420__826130542/nav_item_copy_333890/icon.coreimg.svg/1756326371387/copy.svg","lazyEnabled":true,"isLcpImage":false,"width":"65",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_333890_930852828":{"id":"nav-item-81f93bba67","linkDescription":"Informative articles about AI and data topics","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/fundamentals/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Fundamentals"},"icon":{"id":"icon","height":"64","alt":"Document with list","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719961362824/nav_column_container/nav_column_44600420__826130542/nav_item_copy_333890_930852828/icon.coreimg.svg/1756853637155/data-sheet.svg","lazyEnabled":true,"isLcpImage":false,"width":"65",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"}},":type":"snowflake-site/components/nav/nav-column",":itemsOrder":["nav_item_copy","nav_item","nav_item_copy_144634_1984107859","nav_item_copy_1438098918","nav_item_copy_143809","nav_item_copy_333890638","nav_item_copy_189945","nav_item_copy_333890","nav_item_copy_333890_930852828"]}},":type":"snowflake-site/components/nav/nav-column/nav-column-container",":itemsOrder":["nav_column_copy","nav_column_44600420__826130542"]},"nav_promo_section":{"id":"nav-promo-section-db282bac86","experience_fragment_1":{"id":"experiencefragment-a7b9d7e744","localizedFragmentVariationPath":"/content/experience-fragments/snowflake-site/language-masters/en/site/nav-promo-card/master1/jcr:content","configured":true,":items":{"root":{"layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"nav_promo_card":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-6626185ec6",":items":{"nav_promo_card":{"id":"nav-promo-card-3ca9c5d5c2","openInNewWindow":true,"layout":"horizontal","headline":"The Modern Marketing Data Stack 5th Edition","description":"AI agents are changing the marketing stack. See how to govern the shift. ","linkTitle":"Learn more","linkUrl":"/en/the-modern-marketing-data-stack-report/","image":{"id":"image","height":"540","alt":"MMDS report 5th edition","src":"/adobe/dynamicmedia/deliver/dm-aid--b3030d24-fd50-45e6-bfe6-9520d3eb46d8/web-inside-the-mmds-5th-960x540.png?preferwebp=true&quality=85","lazyEnabled":true,"isLcpImage":false,"width":"960",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-promo-card"}},":type":"snowflake-site/components/container",":itemsOrder":["nav_promo_card"]},"cq:metadata":{":type":"nt:unstructured"}},":type":"snowflake-site/components/experiencefragment","classNames":"aem-xf",":itemsOrder":["root","cq:metadata"]},"experience_fragment_2":{"id":"experiencefragment-e17fad6077","localizedFragmentVariationPath":"/content/experience-fragments/snowflake-site/language-masters/en/site/nav-promo-card/navigation-promo-card-2/jcr:content","configured":true,":items":{"root":{"layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"nav_promo_card":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-8831137806",":items":{"nav_promo_card":{"id":"nav-promo-card-2ee88e6767","openInNewWindow":true,"layout":"horizontal","headline":"The ROI of Gen AI and Agents 2026","description":"Discover how 92% of early adopters are achieving positive ROI with gen AI.","linkTitle":"Learn More","linkUrl":"/en/lp/radical-roi-generative-ai/","image":{"id":"image","height":"540","alt":"roi of gen ai and agents","src":"/adobe/dynamicmedia/deliver/dm-aid--0c15edae-1a97-4739-8b16-c7f3941a6d9e/web-roi-of-gen-ai-and-agents-2026-r02-960x540.png?preferwebp=true&quality=85","lazyEnabled":true,"isLcpImage":false,"width":"960",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-promo-card"}},":type":"snowflake-site/components/container",":itemsOrder":["nav_promo_card"]},"cq:metadata":{":type":"nt:unstructured"}},":type":"snowflake-site/components/experiencefragment","classNames":"aem-xf",":itemsOrder":["root","cq:metadata"]},"experience_fragment_3":{"id":"experiencefragment-d2943df820","localizedFragmentVariationPath":"/content/experience-fragments/snowflake-site/language-masters/en/site/nav-promo-card/navigation-promo-card-3/jcr:content","configured":true,":items":{"root":{"layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"nav_promo_card":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-a33098e4bb",":items":{"nav_promo_card":{"id":"nav-promo-card-a7d3b48cb8","openInNewWindow":true,"layout":"horizontal","headline":"EXPEDITION 2026 | ON-DEMAND","description":"Embark on your agentic transformation with deep technical workshops and curated executive sessions.","linkTitle":"Watch now","linkUrl":"/en/expedition/","image":{"id":"image","height":"540","alt":"expedition","src":"/adobe/dynamicmedia/deliver/dm-aid--b02be612-1a04-464a-bc67-97ff0446e009/expedition-card-thumb.jpg?preferwebp=true&quality=85","lazyEnabled":true,"isLcpImage":false,"width":"960",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-promo-card"}},":type":"snowflake-site/components/container",":itemsOrder":["nav_promo_card"]},"cq:metadata":{":type":"nt:unstructured"}},":type":"snowflake-site/components/experiencefragment","classNames":"aem-xf",":itemsOrder":["root","cq:metadata"]},":type":"snowflake-site/components/nav/nav-promo-section"},":type":"snowflake-site/components/nav/nav-dropdown-menu","cq:panelTitle":"Resources"},"item_1719963657751":{"id":"nav-dropdown-menu-7919dea5b8","enableDropdown":true,"nav_column_container":{"layout":"SIMPLE","id":"container-9619f13465",":items":{"nav_column_copy_copy":{"navColumnTitle":"Build","numberOfSubColumns":"one-column","layout":"SIMPLE","id":"container-3565eda27e",":items":{"nav_item":{"id":"nav-item-e884a9a9a9","propertiesId":"testID","linkDescription":"Overview of the dev resources you need to build and scale","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/developers/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Snowflake for Developers"},"icon":{"id":"icon","height":"64","alt":"Developers icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719963657751/nav_column_container/nav_column_copy_copy/nav_item/icon.coreimg.svg/1731362494574/nav-icon--devs.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_1855651246":{"id":"nav-item-10a7b7264f","linkDescription":"Reference architectures, use cases and best practices","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/developers/guides/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Developer Guides"},"icon":{"id":"icon","height":"64","alt":"Solution Center icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719963657751/nav_column_container/nav_column_copy_copy/nav_item_copy_1855651246/icon.coreimg.svg/1761677891705/nav-icon--solution-center.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy":{"id":"nav-item-289677f2e9","additionalClasses":"is-light-gray-icon","linkDescription":"The latest software versions, drivers, libraries and relevant docs","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/developers/downloads/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Downloads"},"icon":{"id":"icon","height":"28","alt":"Download icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719963657751/nav_column_container/nav_column_copy_copy/nav_item_copy/icon.coreimg.svg/1731362660050/nav-icon-download.svg","lazyEnabled":true,"isLcpImage":false,"width":"28",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"}},":type":"snowflake-site/components/nav/nav-column",":itemsOrder":["nav_item","nav_item_copy_1855651246","nav_item_copy"]},"nav_column_copy_copy_1367930678":{"navColumnTitle":"Learn","numberOfSubColumns":"one-column","layout":"SIMPLE","id":"container-7f6b3150a6",":items":{"nav_item":{"id":"nav-item-ee79191b03","propertiesId":"testID","linkDescription":"Reference docs, guides, tutorials and announcements","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"attributes":{"target":"_blank"},"url":"https://docs.snowflake.com/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_EXTERNAL","text":"Documentation"},"icon":{"id":"icon","height":"64","alt":"Docs icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719963657751/nav_column_container/nav_column_copy_copy_1367930678/nav_item/icon.coreimg.svg/1731361950527/nav-icon--docs.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy":{"id":"nav-item-c7de769cc3","additionalClasses":"is-light-gray-icon","linkDescription":"Key projects Snowflake engineers maintain and support","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/developers/open-source/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Open Source"},"icon":{"id":"icon","height":"32","alt":"Open Source icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719963657751/nav_column_container/nav_column_copy_copy_1367930678/nav_item_copy/icon.coreimg.svg/1731365437016/nav-icon-open-source.svg","lazyEnabled":true,"isLcpImage":false,"width":"32",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_copy":{"id":"nav-item-624ca4411b","additionalClasses":"is-light-gray-icon","linkDescription":"Online and in-person classes and workshops to upskill on Snowflake","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/developers/northstar/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Builder Education"},"icon":{"id":"icon","height":"32","alt":"Northstar logo","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719963657751/nav_column_container/nav_column_copy_copy_1367930678/nav_item_copy_copy/icon.coreimg.svg/1731362475640/nav-icon--northstar.svg","lazyEnabled":true,"isLcpImage":false,"width":"32",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"}},":type":"snowflake-site/components/nav/nav-column",":itemsOrder":["nav_item","nav_item_copy","nav_item_copy_copy"]},"nav_column_copy_copy_1101894776":{"navColumnTitle":"Connect","numberOfSubColumns":"one-column","layout":"SIMPLE","id":"container-ce55c76797",":items":{"nav_item":{"id":"nav-item-a0114c1a1a","propertiesId":"testID","linkDescription":"Snowflake’s technical leaders on what, why and how they build features","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"https://www.snowflake.com/engineering-blog/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_EXTERNAL","text":"Engineering Blog"},"icon":{"id":"icon","height":"64","alt":"software engineers icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719963657751/nav_column_container/nav_column_copy_copy_1101894776/nav_item/icon.coreimg.svg/1789759077904/software-engineers.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_892639794":{"id":"nav-item-5bc364b999","linkDescription":"Ideas, solutions and tools for developers of data-intensive applications","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"url":"/en/developers/blog/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_INTERNAL","text":"Developer Blog"},"icon":{"id":"icon","height":"32","alt":"developer icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719963657751/nav_column_container/nav_column_copy_copy_1101894776/nav_item_892639794/icon.coreimg.svg/1789759047695/nav-icon--developer-center.svg","lazyEnabled":true,"isLcpImage":false,"width":"32",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"},"nav_item_copy_1855651246":{"id":"nav-item-99b474285d","linkDescription":"Tips, tricks and discussion with fellow Snowflake developers","button":{"id":"button","showOutboundIcon":false,"buttonLink":{"valid":true,"attributes":{"target":"_blank"},"url":"https://community.snowflake.com/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_EXTERNAL","text":"Community"},"icon":{"id":"icon","height":"64","alt":"Partner Network icon","src":"/content/experience-fragments/snowflake-site/language-masters/en/site/mega-nav-header/master/_jcr_content/root/mega_header/nav_mega/item_1719963657751/nav_column_container/nav_column_copy_copy_1101894776/nav_item_copy_1855651246/icon.coreimg.svg/1731362644348/nav-icon--partner-network.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-item"}},":type":"snowflake-site/components/nav/nav-column",":itemsOrder":["nav_item","nav_item_892639794","nav_item_copy_1855651246"]}},":type":"snowflake-site/components/nav/nav-column/nav-column-container",":itemsOrder":["nav_column_copy_copy","nav_column_copy_copy_1367930678","nav_column_copy_copy_1101894776"]},"nav_promo_section":{"id":"nav-promo-section-eeb8536cec","experience_fragment_1":{"id":"experiencefragment-cf44e38543","localizedFragmentVariationPath":"/content/experience-fragments/snowflake-site/language-masters/en/site/nav-promo-card/nav-promo-5/jcr:content","configured":true,":items":{"root":{"layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"nav_promo_card":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-87d062c2ac",":items":{"nav_promo_card":{"id":"nav-promo-card-405c13c8e0","openInNewWindow":false,"layout":"horizontal","headline":"Get started with your first Snowflake Notebook","description":"Write and execute code, visualize results, and tell the story of your analysis all in one place.","linkTitle":"Learn More","linkUrl":"/en/developers/solutions-center/getting-started-with-your-first-snowflake-notebook-project/","image":{"id":"image","height":"210","alt":"alt","src":"/adobe/dynamicmedia/deliver/dm-aid--dc7e334a-c38b-4283-b1de-fcf829952eef/nav-promo-first-notebook.jpg?preferwebp=true&quality=85","lazyEnabled":true,"isLcpImage":false,"width":"415",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-promo-card"}},":type":"snowflake-site/components/container",":itemsOrder":["nav_promo_card"]},"cq:LiveSyncConfig":{":type":"cq:LiveCopy","cq:isDeep":true,"cq:rolloutConfigs":[],"cq:master":"/content/experience-fragments/snowflake-site/language-masters/en/site/nav-promo-card/nav-promo-card-4"}},":type":"snowflake-site/components/experiencefragment","classNames":"aem-xf",":itemsOrder":["root","cq:LiveSyncConfig"]},"experience_fragment_2":{"id":"experiencefragment-ac4cb35100","localizedFragmentVariationPath":"/content/experience-fragments/snowflake-site/language-masters/en/site/nav-promo-card/nav-promo-card-4/jcr:content","configured":true,":items":{"root":{"layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"nav_promo_card":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-74a636112e",":items":{"nav_promo_card":{"id":"nav-promo-card-3cdc49fe0b","openInNewWindow":true,"layout":"horizontal","headline":"Northstar Builder Workshops","description":"Join other developers as you roll up your sleeves and explore the possibilities of Snowflake.","linkTitle":"Learn More","linkUrl":"/en/nav-promos/northstar-builders-workshop/","image":{"id":"image","height":"700","alt":"Snowflake Northstar logo","src":"/adobe/dynamicmedia/deliver/dm-aid--14341ced-bc5e-4a29-9762-b7857f6cadfc/nav-promo-northstar.jpg?preferwebp=true&quality=85","lazyEnabled":true,"isLcpImage":false,"width":"1440",":type":"snowflake-site/components/image"},":type":"snowflake-site/components/nav/nav-promo-card"}},":type":"snowflake-site/components/container",":itemsOrder":["nav_promo_card"]},"cq:LiveSyncConfig":{":type":"cq:LiveCopy","cq:isDeep":true,"cq:rolloutConfigs":[],"cq:master":"/content/experience-fragments/snowflake-site/language-masters/en/site/nav-promo-card/master"}},":type":"snowflake-site/components/experiencefragment","classNames":"aem-xf",":itemsOrder":["root","cq:LiveSyncConfig"]},":type":"snowflake-site/components/nav/nav-promo-section"},":type":"snowflake-site/components/nav/nav-dropdown-menu","cq:panelTitle":"Developers"},"item_1718247180324":{"id":"nav-dropdown-menu-a2ba81924c","enableDropdown":false,"link_url":"/en/pricing-options/",":type":"snowflake-site/components/nav/nav-dropdown-menu","cq:panelTitle":"Pricing"}},":itemsOrder":["item_1719963657751_c_663444255","nav_dropdown_menu_2","item_1719963657751_c","item_1719961362824","item_1719963657751","item_1718247180324"]},"languagenavigation":{"id":"language-navigation-9bb0f0cb96","languageNavItems":[{"title":"English","path":"/en/developers/guides/credit-risk-pd-model-snowflake-ml/","locale":"en","active":true},{"title":"日本語","path":"/ja/","locale":"ja","active":false},{"title":"한국어","path":"/ko/","locale":"ko","active":false},{"title":"中文（简体）","path":"/zh_cn/","locale":"zh-cn","active":false},{"title":"Português","path":"/pt_br/","locale":"pt-br","active":false},{"title":"Deutsch","path":"/de/","locale":"de","active":false},{"title":"Français","path":"/fr/","locale":"fr","active":false},{"title":"Español","path":"/es/","locale":"es","active":false},{"title":"Italiano","path":"/it/","locale":"it","active":false}],":type":"snowflake-site/components/nav/language-navigation"},"button_1177328691":{"id":"button-78d7babae8","heapButtonClasses":["mega-nav__sign-in"],"showOutboundIcon":false,"buttonLink":{"valid":true,"attributes":{"target":"_blank"},"url":"https://app.snowflake.com/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","appliedCssClassNames":"snowflake-button-link snowflake-button-black snowflake-button-compact","linkType":"SNOWFLAKE_EXTERNAL","text":"Sign in"},"button":{"id":"button-d0be5e1405","heapButtonClasses":["contact_nav","heap-nav-contact"],"showOutboundIcon":true,"buttonLink":{"valid":true,"url":"/en/contact-sales/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","appliedCssClassNames":"snowflake-button-secondary snowflake-button-blue snowflake-button-compact","linkType":"SNOWFLAKE_INTERNAL","text":"CONTACT SALES"},"button_288358396":{"id":"button-2c2fab3bd2","heapButtonClasses":["start_for_free_nav","heap-nav-start-for-free"],"showOutboundIcon":false,"buttonLink":{"valid":true,"url":"https://signup.snowflake.com/"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","appliedCssClassNames":"snowflake-button-primary snowflake-button-blue snowflake-button-compact","linkType":"SNOWFLAKE_EXTERNAL","text":"start for free"}},":type":"snowflake-site/components/mega-header",":itemsOrder":["nav_mega","languagenavigation","button_1177328691","button","button_288358396"],"appliedCssClassNames":"snowflake-header-container white"}},":type":"snowflake-site/components/container",":itemsOrder":["markup_editor","mega_header"]},"image":{":type":"nt:unstructured"},"cq:targetMetadata":{":type":"nt:unstructured","cq:targetStatus":"OUT_OF_SYNC","cq:exportTime":1793653056655,"cq:targetOfferId":860250}},":type":"snowflake-site/components/experiencefragment","classNames":"aem-xf",":itemsOrder":["root","image","cq:targetMetadata"]},"markup_editor_1950346551":{"id":"markup-editor-99c30f9d8d","title":" ","cssUrl":"/conf/snowflake-site/settings/wcm/templates/quickstart-page-template/structure/_jcr_content/root/markup_editor_1950346551.e08bb1bc93839db7.css",":type":"snowflake-site/components/markup-editor","isGSAPEnabled":false},"responsivegrid":{"columnCount":12,"columnClassNames":{"quickstart_hero":"aem-GridColumn aem-GridColumn--default--12","flexible_column_cont":"aem-GridColumn aem-GridColumn--default--12","markup_editor":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12",":items":{"quickstart_hero":{"id":"quickstart-hero-7eb93269f6",":type":"snowflake-site/components/quickstart/quickstart-hero","fragmentPath":"/content/dam/snowflake-site/en/content-fragments/quickstarts/credit-risk-pd-model-snowflake-ml","quickstartHeroTitle":{"lines":["From SAS to Snowflake: Productionizing Credit Risk Models on a Modern ML Platform"],"type":"heading2",":type":"snowflake-site/components/title-v2"},"quickstartHeroAuthor":"Gayatri Ghanakota","quickstartHeroFirstSnowflakeFeatureTag":{"tagText":"Model Development","tagColor":"#29B5E8","tagPath":"/content/cq:tags/snowflake-site/taxonomy/snowflake-feature/model-development","tagIcon":""},"quickstartHeroForkRepoLink":{"id":"button-5e23262426","showOutboundIcon":false,"buttonLink":{"valid":true,"attributes":{"target":"_blank"},"url":"https://github.com/Snowflake-Labs/sfquickstarts/tree/master/site/sfguides/src/credit-risk-pd-model-snowflake-ml"},"linkTargetContentType":"GENERIC",":type":"snowflake-site/components/button","linkType":"SNOWFLAKE_EXTERNAL","text":"Fork Repo"},"quickstartHeroBreadcrumbs":[{"title":"From SAS to Snowflake: Productionizing Credit Risk Models on a Modern ML Platform","url":"https://www.snowflake.com/content/snowflake-site/global/en/developers/guides/credit-risk-pd-model-snowflake-ml","currentPage":true},{"title":"Guides","url":"https://www.snowflake.com/content/snowflake-site/global/en/developers/guides","currentPage":false},{"title":"Snowflake for Developers","url":"https://www.snowflake.com/content/snowflake-site/global/en/developers","currentPage":false}],"isDeveloperGuidesPage":false},"flexible_column_cont":{"id":"flexible-column-container-ce9c19be21","propertiesId":"quickstart-template-main-flexible-container","type":"2-column-75-25","alignColumns":"top","containerMaxWidth":"extra-large","topPadding":"none","bottomPadding":"none","spaceBetween":"small","reverseOnMobile":false,"carouselOnMobile":false,"backgroundImageOption":"none","flexible_column_content_container_1":{"layout":"SIMPLE","id":"container-d950ad3f2f",":items":{"contentfragment":{"id":"contentfragment-464094341a","paragraphs":["<!-- ------------------------ -->\n","\u003Ch2\u003EOverview\u003C/h2\u003E\n","\u003Cp\u003EBanks have built Probability of Default (PD), Loss Given Default (LGD) and Exposure at Default (EAD) models in SAS for decades. The models work, but the infrastructure around them (data extracts to SAS Grid, macro-generated datasets, cron-driven batch jobs, spreadsheet model inventories and email sign-offs) makes reproducibility, lineage and monitoring expensive to prove to regulators.\u003C/p\u003E\n","\u003Cp\u003EIn this Quickstart you migrate a SAS \u003Ccode\u003EPROC GRADBOOST\u003C/code\u003E PD model to \u003Cstrong\u003EXGBoost on Snowflake ML\u003C/strong\u003E and run the whole model lifecycle in one governed platform: features, training data, experiments, registry, explainability, batch scoring, scenario-weighted Expected Credit Loss (ECL) scoring, monitoring, human-in-the-loop promotion and scheduled pipelines.\u003C/p\u003E\n","\u003Cp\u003E\u003Cimg src=\"https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/solution-architecture.png?v=54b7b1dd\" alt=\"Solution architecture\"\u003E\u003C/p\u003E\n","\u003Cp\u003EThe complete notebook is in the Quickstart assets on GitHub: \u003Cstrong\u003E\u003Ca href=\"https://github.com/Snowflake-Labs/sfquickstarts/blob/master/site/sfguides/src/credit-risk-pd-model-snowflake-ml/assets/credit_risk_probability_of_default_model.ipynb\"\u003Ecredit_risk_probability_of_default_model.ipynb\u003C/a\u003E\u003C/strong\u003E. Each step below shows the notebook code with an explanation, so you can follow along cell by cell.\u003C/p\u003E\n","\u003Ch3\u003EWhat You'll Learn\u003C/h3\u003E\n\u003Cul\u003E\u003Cli\u003EHow to convert a SAS \u003Ccode\u003EPROC GRADBOOST\u003C/code\u003E PD model to Python with CoCo, and map its hyperparameters to \u003Ccode\u003EXGBClassifier\u003C/code\u003E\u003C/li\u003E\u003Cli\u003EHow to build Feature Views (backed by Dynamic Tables) for behavioral, macroeconomic and forward-looking scenario features\u003C/li\u003E\u003Cli\u003EHow to generate immutable, point-in-time training Datasets\u003C/li\u003E\u003Cli\u003EHow to log runs with Experiment Tracking and register models on Snowflake Model Registry\u003C/li\u003E\u003Cli\u003EHow to explain predictions with the Model Registry's built-in SHAP \u003Ccode\u003Eexplain\u003C/code\u003E function and trace ML Lineage\u003C/li\u003E\u003Cli\u003EHow to batch score with \u003Ccode\u003Erun()\u003C/code\u003E and compute probability-weighted ECL across macro scenarios\u003C/li\u003E\u003Cli\u003EHow to attach Model Monitors for drift and performance\u003C/li\u003E\u003Cli\u003EHow to automate monthly retraining and weekly scoring with Task DAGs, including an email-driven, role-gated model approval step\u003C/li\u003E\u003C/ul\u003E\n","\u003Ch3\u003EWhat You'll Build\u003C/h3\u003E\n\u003Cul\u003E\u003Cli\u003EA Feature Store with three Feature Views: \u003Ccode\u003ELOAN_BEHAVIORAL_FEATURES\u003C/code\u003E, \u003Ccode\u003ELOAN_MACRO_FEATURES\u003C/code\u003E and \u003Ccode\u003EMACRO_SCENARIO_FEATURES\u003C/code\u003E\u003C/li\u003E\u003Cli\u003EA versioned training Dataset, \u003Ccode\u003EPD_TRAINING_DS\u003C/code\u003E\u003C/li\u003E\u003Cli\u003EA registered model, \u003Ccode\u003EPD_XGB_MODEL\u003C/code\u003E, with champion and challenger versions, SHAP explainability and lineage\u003C/li\u003E\u003Cli\u003EModel Monitors for the champion and challenger models\u003C/li\u003E\u003Cli\u003EA monthly \u003Cstrong\u003Etraining pipeline\u003C/strong\u003E Task DAG that retrains both models, scores them against production, gates them, and emails a reviewer for approval\u003C/li\u003E\u003Cli\u003EA weekly \u003Cstrong\u003Escoring pipeline\u003C/strong\u003E Task DAG that scores the portfolio with the approved models, computes scenario-weighted ECL, keeps history, and emails a summary\u003C/li\u003E\u003C/ul\u003E\n","\u003Ch3\u003EWhat You'll Need\u003C/h3\u003E\n\u003Cul\u003E\u003Cli\u003EA Snowflake account with access to \u003Cstrong\u003ESnowflake Notebooks\u003C/strong\u003E and a role that can create databases, warehouses, roles and notification integrations (the notebook was built with \u003Ccode\u003EACCOUNTADMIN\u003C/code\u003E)\u003C/li\u003E\u003Cli\u003EA verified email address on your Snowflake user, for the review and scoring notification emails\u003C/li\u003E\u003Cli\u003ETwo free Snowflake Marketplace listings (installed in the next step):\n\u003Cul\u003E\u003Cli\u003E\u003Ca href=\"https://app.snowflake.com/marketplace/listing/GZT0ZOSU8UC\"\u003EEquifax Analytic Dataset (Demo/Sample)\u003C/a\u003E\u003C/li\u003E\u003Cli\u003E\u003Ca href=\"https://app.snowflake.com/marketplace/listing/GZTYZ40XYNE\"\u003EFRED - Federal Reserve Economic Data\u003C/a\u003E\u003C/li\u003E\u003C/ul\u003E\n\u003C/li\u003E\u003Cli\u003EFamiliarity with Python and SQL. Credit risk knowledge helps but is not required.\u003C/li\u003E\u003C/ul\u003E\n<!-- ------------------------ -->\n","\u003Ch2\u003EGet Data and Notebook\u003C/h2\u003E\n","\u003Ch3\u003EInstall the Marketplace data\u003C/h3\u003E\n\u003Col\u003E\u003Cli\u003EIn Snowsight, open \u003Cstrong\u003EData Products > Marketplace\u003C/strong\u003E.\u003C/li\u003E\u003Cli\u003ESearch for \u003Cstrong\u003EEquifax Analytic Dataset (Demo/Sample)\u003C/strong\u003E and select \u003Cstrong\u003EGet\u003C/strong\u003E. Keep the database name \u003Ccode\u003EANALYTIC_DATASET_DEMOSAMPLE\u003C/code\u003E.\u003C/li\u003E\u003Cli\u003ESearch for \u003Cstrong\u003EFRED - Federal Reserve Economic Data\u003C/strong\u003E (provided by Fruit Stand) and select \u003Cstrong\u003EGet\u003C/strong\u003E. Keep the database name \u003Ccode\u003EFRED_FEDERAL_RESERVE_ECONOMIC_DATA_SERIES_OBSERVATIONS\u003C/code\u003E.\u003C/li\u003E\u003C/ol\u003E\n","\u003Cp\u003EThe notebook reads \u003Ccode\u003EANALYTIC_DATASET_DEMOSAMPLE.DATASET.ANALYTICSSAMPLE_LATEST\u003C/code\u003E and \u003Ccode\u003EFRED_FEDERAL_RESERVE_ECONOMIC_DATA_SERIES_OBSERVATIONS.FRED.FRED_SERIES_OBSERVATIONS_VW\u003C/code\u003E. If you choose different database names, update those references.\u003C/p\u003E\n","\u003Ch3\u003EImport the notebook\u003C/h3\u003E\n\u003Col\u003E\u003Cli\u003EDownload \u003Ccode\u003Ecredit_risk_probability_of_default_model.ipynb\u003C/code\u003E from the \u003Ca href=\"https://github.com/Snowflake-Labs/sfquickstarts/blob/master/site/sfguides/src/credit-risk-pd-model-snowflake-ml/assets/credit_risk_probability_of_default_model.ipynb\"\u003EQuickstart assets\u003C/a\u003E.\u003C/li\u003E\u003Cli\u003EIn Snowsight, go to \u003Cstrong\u003EProjects > Notebooks\u003C/strong\u003E, choose \u003Cstrong\u003EImport .ipynb file\u003C/strong\u003E, and select the downloaded notebook.\u003C/li\u003E\u003Cli\u003EMake sure these packages are available to the notebook: \u003Ccode\u003Esnowflake-ml-python\u003C/code\u003E, \u003Ccode\u003Exgboost\u003C/code\u003E, \u003Ccode\u003Escikit-learn\u003C/code\u003E, \u003Ccode\u003Escipy\u003C/code\u003E, \u003Ccode\u003Ematplotlib\u003C/code\u003E, \u003Ccode\u003Ealtair\u003C/code\u003E.\u003C/li\u003E\u003Cli\u003EUpload \u003Ca href=\"https://github.com/Snowflake-Labs/sfquickstarts/blob/master/site/sfguides/src/credit-risk-pd-model-snowflake-ml/assets/python_pd_model.py\"\u003E\u003Ccode\u003Epython_pd_model.py\u003C/code\u003E\u003C/a\u003E from the Quickstart assets to the same notebook, next to the \u003Ccode\u003E.ipynb\u003C/code\u003E file. The notebook and the training pipeline import the model from it (see the next step).\u003C/li\u003E\u003C/ol\u003E\n","\u003Cp\u003EThe Quickstart assets also contain \u003Ca href=\"https://github.com/Snowflake-Labs/sfquickstarts/blob/master/site/sfguides/src/credit-risk-pd-model-snowflake-ml/assets/sas_pd_model.sas\"\u003E\u003Ccode\u003Esas_pd_model.sas\u003C/code\u003E\u003C/a\u003E (the original SAS model) and \u003Ca href=\"https://github.com/Snowflake-Labs/sfquickstarts/blob/master/site/sfguides/src/credit-risk-pd-model-snowflake-ml/assets/python_pd_model.py\"\u003E\u003Ccode\u003Epython_pd_model.py\u003C/code\u003E\u003C/a\u003E (the Python conversion produced with CoCo). The next step shows how that conversion is done.\u003C/p\u003E\n<!-- ------------------------ -->\n","\u003Ch2\u003EConvert SAS with CoCo\u003C/h2\u003E\n","\u003Cp\u003EThe starting point is a SAS PD model, \u003Ccode\u003Esas_pd_model.sas\u003C/code\u003E. It follows a pattern found in many banks:\u003C/p\u003E\n\u003Col\u003E\u003Cli\u003E\u003Ccode\u003EPROC SQL\u003C/code\u003E joins Equifax loan data to FRED macro indicators, with \u003Ccode\u003E%let\u003C/code\u003E macro variables for libraries and the observation window.\u003C/li\u003E\u003Cli\u003EA \u003Ccode\u003EDATA\u003C/code\u003E step imputes missing values with 0.\u003C/li\u003E\u003Cli\u003E\u003Ccode\u003EPROC SURVEYSELECT\u003C/code\u003E makes an 80/20 split.\u003C/li\u003E\u003Cli\u003E\u003Ccode\u003EPROC GRADBOOST\u003C/code\u003E trains a gradient boosting model.\u003C/li\u003E\u003Cli\u003E\u003Ccode\u003EPROC ASTORE\u003C/code\u003E scores the test set from the saved analytic store.\u003C/li\u003E\u003Cli\u003E\u003Ccode\u003EPROC LOGISTIC\u003C/code\u003E computes AUC, and the metrics go into an Excel tracker by hand.\u003C/li\u003E\u003C/ol\u003E\n","\u003Cp\u003EThe core of the model is the \u003Ccode\u003EPROC GRADBOOST\u003C/code\u003E step. Because it is gradient boosting, like XGBoost, its options map closely onto \u003Ccode\u003EXGBClassifier\u003C/code\u003E:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-sas\"\u003Eproc gradboost data=&OUT_LIB..PD_TRAIN\n    ntrees=200\n    learningrate=0.1\n    maxdepth=6\n    vars_to_try=5\n    minleafsize=20\n    seed=42;\n\n    target DEFAULT_FLAG / level=nominal;\n\n    input VANTAGE_SCORE_3 NUMBER_OF_ACCOUNTS NUMBER_ACCOUNTS_PAST_DUE\n          NUMBER_ACCOUNTS_MAJOR_DEROGATORY HIGH_UTIL_ACCTS\n          AGE_OLDEST_ACCOUNT TOTAL_PAST_DUE_AMOUNT BALANCE\n          / level=interval;\n    input UNEMPLOYMENT_RATE FED_FUNDS_RATE CPI_INDEX / level=interval;\n\n    savestate rstore=&OUT_LIB..PD_GRADBOOST_STORE;\n    ods output VariableImportance = &OUT_LIB..PD_IMPORTANCE;\n    ods output FitStatistics      = &OUT_LIB..PD_FITSTATS;\nrun;\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Ch3\u003EPrompt CoCo\u003C/h3\u003E\n","\u003Cp\u003E\u003Cstrong\u003E\u003Ca href=\"https://www.snowflake.com/en/product/snowflake-coco/\"\u003ECoCo\u003C/a\u003E\u003C/strong\u003E is Snowflake's AI coding assistant. It's available in Snowsight and as a desktop app. Open the repository folder in CoCo (or attach \u003Ccode\u003Esas_pd_model.sas\u003C/code\u003E to the chat) and ask for the conversion in plain language. Name the target frameworks and the Snowflake features you want each SAS construct to become:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-text\"\u003EConvert @sas_pd_model.sas into a Python script, python_pd_model.py, that runs on Snowflake.\n\n- Replace PROC GRADBOOST with xgboost.XGBClassifier. Map every GRADBOOST option to its closest\n  XGBoost hyperparameter and explain each mapping in a comment.\n- Replace the %let macros and PROC SQL join with Snowflake Feature Store Feature Views\n  (one for the Equifax behavioral features, one for the FRED macro features), keyed on TRADE_ID\n  with ARCHIVE_DATE as the timestamp.\n- Replace PROC SURVEYSELECT and the manual dataset freeze with a versioned Dataset from\n  fs.generate_dataset, then a stratified 80/20 split with seed 42.\n- Keep the DATA step behavior: impute missing numeric values with 0.\n- Replace PROC LOGISTIC ROC with AUC, Gini and KS computed in Python.\n- Replace the saved analytic store with Snowflake Model Registry registration and a "champion" alias.\n- Replace PROC ASTORE with model.run() batch scoring on a warehouse.\nUse the Snowpark session from get_active_session(); data is in CREDIT_RISK_ML.PD_MODEL.\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003ECoCo reads the SAS file and returns \u003Ca href=\"https://github.com/Snowflake-Labs/sfquickstarts/blob/master/site/sfguides/src/credit-risk-pd-model-snowflake-ml/assets/python_pd_model.py\"\u003E\u003Ccode\u003Epython_pd_model.py\u003C/code\u003E\u003C/a\u003E (in the Quickstart assets). Ask it to structure the result as a reusable module, not just a one-off script: the feature lists, the SAS-mapped hyperparameters, and a small function for each SAS step. The notebook and the training pipeline then import the same code, so the converted model is the one that is actually trained in production:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# XGBoost hyperparameters mapped from SAS PROC GRADBOOST:\n#   ntrees=200        →  n_estimators=200\n#   learningrate=0.1  →  learning_rate=0.1\n#   maxdepth=6        →  max_depth=6\n#   vars_to_try=5     →  colsample_bynode = 5/11 (features tried per split)\n#   minleafsize=20    →  min_child_weight (approx)\n#   seed=42           →  random_state=42\nSAS_MAPPED_PARAMS = {\n    "n_estimators": 200,\n    "max_depth": 6,\n    "learning_rate": 0.1,\n    "colsample_bynode": 5 / len(FEATURE_COLS),\n    "min_child_weight": 20,\n    "objective": "binary:logistic",\n    "eval_metric": "auc",\n    "random_state": 42,\n}\n\n\ndef prepare_features(pdf):\n    """Feature matrix and label from a pandas frame. Replaces the SAS DATA step (missing → 0)."""\n    return pdf[FEATURE_COLS].fillna(0), pdf[TARGET_COL]\n\n\ndef split_train_test(X, y):\n    """Stratified 80/20 split with seed 42. Replaces PROC SURVEYSELECT samprate=0.8 seed=42."""\n    return train_test_split(X, y, test_size=0.2, random_state=42, stratify=y)\n\n\ndef train_pd_model(X, y, params=None):\n    """Fit the PD model. Replaces PROC GRADBOOST."""\n    return XGBClassifier(**(params or SAS_MAPPED_PARAMS)).fit(X, y)\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003EThe module also defines \u003Ccode\u003EBEHAVIORAL_COLS\u003C/code\u003E, \u003Ccode\u003EMACRO_COLS\u003C/code\u003E, \u003Ccode\u003EFEATURE_COLS\u003C/code\u003E, \u003Ccode\u003ETARGET_COL\u003C/code\u003E, \u003Ccode\u003ECHALLENGER_PARAMS\u003C/code\u003E and \u003Ccode\u003Epd_metrics()\u003C/code\u003E (AUC, Gini, KS and Brier, replacing \u003Ccode\u003EPROC LOGISTIC\u003C/code\u003E). Running the file directly (\u003Ccode\u003Emain(session)\u003C/code\u003E) still performs the converted SAS workflow end to end.\u003C/p\u003E\n","\u003Cp\u003ETo get this shape, add a line like this to the prompt:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-text\"\u003EStructure python_pd_model.py as an importable module: constants for the feature lists and the\nSAS-mapped hyperparameters, one function per SAS step (prepare, split, train, evaluate), and the\nend-to-end workflow in main(session) under if __name__ == "__main__".\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003EThe notebook in this guide imports this module and builds the rest of the lifecycle around it: Experiment Tracking, a challenger model, SHAP explainability, monitoring and orchestration. The remaining steps walk through it cell by cell.\u003C/p\u003E\n<!-- ------------------------ -->\n","\u003Ch2\u003ESet Up the Environment\u003C/h2\u003E\n","\u003Cp\u003EEdit the configuration block at the top of the first cell before running anything:\u003C/p\u003E\n\u003Cul\u003E\u003Cli\u003E\u003Ccode\u003EREVIEWER_USERS\u003C/code\u003E: Snowflake users who will receive the \u003Ccode\u003EMODEL_REVIEWER\u003C/code\u003E role and approve or reject models.\u003C/li\u003E\u003Cli\u003E\u003Ccode\u003EREVIEWER_EMAILS\u003C/code\u003E: verified email addresses of users in your account. They become the allowed recipients of the email notification integration.\u003C/li\u003E\u003Cli\u003E\u003Ccode\u003ESCORING_NOTIFY_EMAILS\u003C/code\u003E: recipients of the weekly scoring summary (must also be in \u003Ccode\u003EREVIEWER_EMAILS\u003C/code\u003E).\u003C/li\u003E\u003Cli\u003E\u003Ccode\u003EREVIEW_EMAIL_INTEGRATION\u003C/code\u003E: name of the email notification integration the notebook creates.\u003C/li\u003E\u003C/ul\u003E\n","\u003Cp\u003EThe cell then creates the \u003Ccode\u003ECREDIT_RISK_ML\u003C/code\u003E database, the \u003Ccode\u003EPD_MODEL\u003C/code\u003E schema and an extra-small \u003Ccode\u003ECREDIT_RISK_WH\u003C/code\u003E warehouse.\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# ---------------------------------------------------------------------------\n# Configuration — edit these before running the notebook\n# ---------------------------------------------------------------------------\n# Snowflake users who review and approve models (granted the MODEL_REVIEWER role in section 10)\nREVIEWER_USERS = ["<YOUR_SNOWFLAKE_USERNAME>"]\n# Email addresses for model review requests (section 10). Each must be the verified email\n# of a user in this Snowflake account; they become the email integration's ALLOWED_RECIPIENTS.\nREVIEWER_EMAILS = ["<your.email@example.com>"]\n# Recipients of the weekly scoring summary (section 10b) — must also appear in REVIEWER_EMAILS\nSCORING_NOTIFY_EMAILS = REVIEWER_EMAILS\n# Email notification integration used for both emails (created in section 10)\nREVIEW_EMAIL_INTEGRATION = "PD_MODEL_REVIEW_EMAIL"\n# ---------------------------------------------------------------------------\n\nfrom snowflake.snowpark.context import get_active_session\n\nsession = get_active_session()\n\nsession.sql("CREATE DATABASE IF NOT EXISTS CREDIT_RISK_ML").collect()\nsession.sql("CREATE SCHEMA IF NOT EXISTS CREDIT_RISK_ML.PD_MODEL").collect()\nsession.sql("""\n    CREATE WAREHOUSE IF NOT EXISTS CREDIT_RISK_WH\n    WAREHOUSE_SIZE = 'XSMALL' AUTO_SUSPEND = 120 AUTO_RESUME = TRUE\n""").collect()\nsession.sql("USE SCHEMA CREDIT_RISK_ML.PD_MODEL").collect()\nsession.sql("USE WAREHOUSE CREDIT_RISK_WH").collect()\nprint("Setup complete: CREDIT_RISK_ML.PD_MODEL")\n\u003C/code\u003E\u003C/pre\u003E\n<!-- ------------------------ -->\n","\u003Ch2\u003EExplore the Data\u003C/h2\u003E\n","\u003Cp\u003EStart by profiling the two Marketplace sources. The Equifax demo sample contains 500K loan-level records across product categories such as auto, credit card, mortgage, student and personal loans.\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003Eexplore_equifax = session.sql("""\n    SELECT COUNT(*) AS total_rows,\n           COUNT(DISTINCT CONSUMER_ID) AS unique_consumers,\n           COUNT(DISTINCT TRADE_ID) AS unique_trades,\n           MIN(ARCHIVE_DATE) AS min_date,\n           MAX(ARCHIVE_DATE) AS max_date,\n           COUNT(DISTINCT PRODUCT_CATEGORY) AS product_categories\n    FROM ANALYTIC_DATASET_DEMOSAMPLE.DATASET.ANALYTICSSAMPLE_LATEST\n""")\nexplore_equifax.show()\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003EThe target is \u003Cstrong\u003Edefault\u003C/strong\u003E, defined as 90+ days past due (\u003Ccode\u003ESTATUS_CATEGORY >= 4\u003C/code\u003E). This is the standard PD definition and it mirrors the SAS model. Check the class balance:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Target variable distribution\n# STATUS_CATEGORY: 1=Current, 2=30DPD, 3=60DPD, 4=90DPD, 5=120+DPD\n# Default flag: STATUS_CATEGORY >= 4 (90+ days past due)\ntarget_dist = session.sql("""\n    SELECT CASE WHEN STATUS_CATEGORY >= 4 THEN 1 ELSE 0 END AS DEFAULT_FLAG,\n           COUNT(*) AS count,\n           ROUND(COUNT(*) * 100.0 / SUM(COUNT(*)) OVER(), 1) AS pct\n    FROM ANALYTIC_DATASET_DEMOSAMPLE.DATASET.ANALYTICSSAMPLE_LATEST\n    GROUP BY DEFAULT_FLAG ORDER BY DEFAULT_FLAG\n""")\ntarget_dist.show()\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003EFrom FRED, the model uses three monthly macro series: unemployment rate (\u003Ccode\u003EUNRATE\u003C/code\u003E), the federal funds rate (\u003Ccode\u003EFEDFUNDS\u003C/code\u003E) and CPI (\u003Ccode\u003ECPIAUCSL\u003C/code\u003E).\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# FRED macro series available for this model\nexplore_fred = session.sql("""\n    SELECT SERIES_ID, SERIES_NAME, FREQUENCY, UNITS,\n           MIN(DATE) AS min_date, MAX(DATE) AS max_date, COUNT(*) AS obs_count\n    FROM FRED_FEDERAL_RESERVE_ECONOMIC_DATA_SERIES_OBSERVATIONS.FRED.FRED_SERIES_OBSERVATIONS_VW\n    WHERE SERIES_ID IN ('UNRATE', 'FEDFUNDS', 'CPIAUCSL')\n    GROUP BY SERIES_ID, SERIES_NAME, FREQUENCY, UNITS\n    ORDER BY SERIES_ID\n""")\nexplore_fred.show()\n\u003C/code\u003E\u003C/pre\u003E\n<!-- ------------------------ -->\n","\u003Ch2\u003EPrepare the Data\u003C/h2\u003E\n","\u003Cp\u003EManaged Feature Views are backed by Dynamic Tables, which can only read a shared Marketplace view if the provider has enabled change tracking on its underlying tables. Consumers can't enable change tracking on shared objects. To avoid depending on the provider, first materialize the data into the project schema with change tracking enabled.\u003C/p\u003E\n","\u003Cp\u003E\u003Ccode\u003ELOAN_DATA\u003C/code\u003E keeps the loan key, observation date, the derived \u003Ccode\u003EDEFAULT_FLAG\u003C/code\u003E and the behavioral attributes. \u003Ccode\u003ECHANGE_TRACKING = TRUE\u003C/code\u003E keeps the downstream Dynamic Tables refreshable after the table is re-created.\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Materialize Equifax loan data\n# CHANGE_TRACKING keeps the Feature View dynamic tables refreshable after this table is re-created\nsession.sql("""\n    CREATE OR REPLACE TABLE CREDIT_RISK_ML.PD_MODEL.LOAN_DATA CHANGE_TRACKING = TRUE AS\n    SELECT \n        TRADE_ID, CONSUMER_ID, ARCHIVE_DATE,\n        CASE WHEN STATUS_CATEGORY >= 4 THEN 1 ELSE 0 END AS DEFAULT_FLAG,\n        VANTAGE_SCORE_3, NUMBER_OF_ACCOUNTS, NUMBER_ACCOUNTS_PAST_DUE,\n        NUMBER_ACCOUNTS_MAJOR_DEROGATORY,\n        BANKCARD_ACCTS_OVER_75_PERCENT_UTILIZATION AS HIGH_UTIL_ACCTS,\n        AGE_OLDEST_ACCOUNT, TOTAL_PAST_DUE_AMOUNT, BALANCE,\n        PRODUCT_CATEGORY, STATUS_CATEGORY\n    FROM ANALYTIC_DATASET_DEMOSAMPLE.DATASET.ANALYTICSSAMPLE_LATEST\n""").collect()\nprint(f"LOAN_DATA: {session.table('CREDIT_RISK_ML.PD_MODEL.LOAN_DATA').count():,} rows")\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003E\u003Ccode\u003EMACRO_INDICATORS\u003C/code\u003E pivots the three FRED series from long to wide, with one row per month:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Pivot FRED monthly series into a wide table\nsession.sql("""\n    CREATE OR REPLACE TABLE CREDIT_RISK_ML.PD_MODEL.MACRO_INDICATORS CHANGE_TRACKING = TRUE AS\n    SELECT DATE AS OBSERVATION_MONTH,\n           MAX(CASE WHEN SERIES_ID = 'UNRATE' THEN VALUE END) AS UNEMPLOYMENT_RATE,\n           MAX(CASE WHEN SERIES_ID = 'FEDFUNDS' THEN VALUE END) AS FED_FUNDS_RATE,\n           MAX(CASE WHEN SERIES_ID = 'CPIAUCSL' THEN VALUE END) AS CPI_INDEX\n    FROM FRED_FEDERAL_RESERVE_ECONOMIC_DATA_SERIES_OBSERVATIONS.FRED.FRED_SERIES_OBSERVATIONS_VW\n    WHERE SERIES_ID IN ('UNRATE', 'FEDFUNDS', 'CPIAUCSL') AND DATE >= '2022-01-01'\n    GROUP BY DATE HAVING UNEMPLOYMENT_RATE IS NOT NULL\n    ORDER BY DATE\n""").collect()\nprint(f"MACRO_INDICATORS: {session.table('CREDIT_RISK_ML.PD_MODEL.MACRO_INDICATORS').count()} rows")\nsession.table("CREDIT_RISK_ML.PD_MODEL.MACRO_INDICATORS").show(5)\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Ch3\u003EFeature exploration\u003C/h3\u003E\n","\u003Cp\u003EBefore engineering features, run a few standard credit risk checks. Each aggregation runs in Snowflake, and only small summaries come back to pandas for plotting.\u003C/p\u003E\n","\u003Cp\u003EClass balance and default rate by product:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Class balance + default rate by product category\nimport matplotlib.pyplot as plt\n\nclass_pdf = session.sql("""\n    SELECT DEFAULT_FLAG, COUNT(*) AS N\n    FROM CREDIT_RISK_ML.PD_MODEL.LOAN_DATA GROUP BY DEFAULT_FLAG ORDER BY DEFAULT_FLAG\n""").to_pandas()\n\nproduct_pdf = session.sql("""\n    SELECT PRODUCT_CATEGORY, AVG(DEFAULT_FLAG) AS DEFAULT_RATE, COUNT(*) AS N\n    FROM CREDIT_RISK_ML.PD_MODEL.LOAN_DATA\n    GROUP BY PRODUCT_CATEGORY ORDER BY DEFAULT_RATE DESC\n""").to_pandas()\n\nfig, axes = plt.subplots(1, 2, figsize=(14, 5))\naxes[0].bar(["Non-default (0)", "Default (1)"], class_pdf["N"], color=["#4C9BE8", "#E8684C"])\naxes[0].set_title("Class Balance")\naxes[0].set_ylabel("Loans")\n\naxes[1].bar(product_pdf["PRODUCT_CATEGORY"], product_pdf["DEFAULT_RATE"], color="#7A6FD6")\naxes[1].set_title("Default Rate by Product Category")\naxes[1].set_ylabel("Default rate")\naxes[1].tick_params(axis="x", rotation=45)\nplt.tight_layout()\nplt.show()\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003E\u003Cimg src=\"https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/class-balance-default-rate-by-product.png?v=54b7b1dd\" alt=\"Class balance and default rate by product category\"\u003E\u003C/p\u003E\n","\u003Cp\u003EMacro trends against the portfolio default rate over time:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Macro trends vs portfolio default rate over time\ntrend_pdf = session.sql("""\n    SELECT m.OBSERVATION_MONTH, m.UNEMPLOYMENT_RATE, m.FED_FUNDS_RATE, m.CPI_INDEX,\n           AVG(l.DEFAULT_FLAG) AS DEFAULT_RATE\n    FROM CREDIT_RISK_ML.PD_MODEL.MACRO_INDICATORS m\n    LEFT JOIN CREDIT_RISK_ML.PD_MODEL.LOAN_DATA l\n        ON DATE_TRUNC('MONTH', l.ARCHIVE_DATE) = m.OBSERVATION_MONTH\n    GROUP BY 1, 2, 3, 4 ORDER BY 1\n""").to_pandas()\n\nfig, axes = plt.subplots(2, 2, figsize=(14, 8), sharex=True)\nseries = [("UNEMPLOYMENT_RATE", "Unemployment rate (%)"),\n          ("FED_FUNDS_RATE", "Fed funds rate (%)"),\n          ("CPI_INDEX", "CPI index"),\n          ("DEFAULT_RATE", "Portfolio default rate")]\nfor ax, (c, title) in zip(axes.flat, series):\n    ax.plot(trend_pdf["OBSERVATION_MONTH"], trend_pdf[c], marker=".")\n    ax.set_title(title)\n    ax.tick_params(axis="x", rotation=45)\nplt.suptitle("FRED Macro Indicators and Portfolio Default Rate (Monthly)")\nplt.tight_layout()\nplt.show()\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003E\u003Cimg src=\"https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/macro-trends-vs-default-rate.png?v=54b7b1dd\" alt=\"FRED macro indicators and portfolio default rate\"\u003E\u003C/p\u003E\n","\u003Cp\u003EA correlation matrix on a 20K-row sample, taken in Snowflake before collecting:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Correlation heatmap on a 20K-row sample (sampled in Snowflake before collecting)\nimport numpy as np\n\ncorr_cols = behavioral_cols + ["UNEMPLOYMENT_RATE", "FED_FUNDS_RATE", "CPI_INDEX", "DEFAULT_FLAG"]\nsample_pdf = session.sql(f"""\n    SELECT * FROM (\n        SELECT {", ".join("l." + c for c in behavioral_cols + ["DEFAULT_FLAG"])},\n               m.UNEMPLOYMENT_RATE, m.FED_FUNDS_RATE, m.CPI_INDEX\n        FROM CREDIT_RISK_ML.PD_MODEL.LOAN_DATA l\n        JOIN CREDIT_RISK_ML.PD_MODEL.MACRO_INDICATORS m\n            ON DATE_TRUNC('MONTH', l.ARCHIVE_DATE) = m.OBSERVATION_MONTH\n    ) SAMPLE (20000 ROWS)\n""").to_pandas()\n\ncorr = sample_pdf[corr_cols].corr()\nfig, ax = plt.subplots(figsize=(11, 9))\nim = ax.imshow(corr, cmap="RdBu_r", vmin=-1, vmax=1)\nax.set_xticks(range(len(corr_cols)))\nax.set_xticklabels(corr_cols, rotation=90)\nax.set_yticks(range(len(corr_cols)))\nax.set_yticklabels(corr_cols)\nfor i in range(len(corr_cols)):\n    for j in range(len(corr_cols)):\n        ax.text(j, i, f"{corr.iloc[i, j]:.2f}", ha="center", va="center", fontsize=7)\nfig.colorbar(im, ax=ax, shrink=0.8)\nax.set_title("Feature Correlation Matrix (incl. DEFAULT_FLAG)")\nplt.tight_layout()\nplt.show()\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003E\u003Cimg src=\"https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/correlation-heatmap.png?v=54b7b1dd\" alt=\"Feature correlation matrix\"\u003E\u003C/p\u003E\n<!-- ------------------------ -->\n","\u003Ch2\u003EBuild the Feature Store\u003C/h2\u003E\n","\u003Cp\u003EIn SAS, features usually live in macro-generated datasets refreshed on custom schedules. Here they become managed \u003Cstrong\u003EFeature Views\u003C/strong\u003E: you declare the transformation once, and Snowflake handles refresh, versioning and point-in-time correct retrieval.\u003C/p\u003E\n","\u003Cp\u003ECreate the Feature Store in the project schema and register the \u003Ccode\u003ELOAN\u003C/code\u003E entity, keyed on \u003Ccode\u003ETRADE_ID\u003C/code\u003E:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003Efrom snowflake.ml.feature_store import FeatureStore, Entity, FeatureView, CreationMode\n\nfs = FeatureStore(\n    session=session,\n    database="CREDIT_RISK_ML",\n    name="PD_MODEL",\n    default_warehouse="CREDIT_RISK_WH",\n    creation_mode=CreationMode.CREATE_IF_NOT_EXIST\n)\n\n# Entity: one row per loan trade\nloan_entity = Entity(name="LOAN", join_keys=["TRADE_ID"])\nfs.register_entity(loan_entity)\nprint(f"Registered entity: {loan_entity.name}")\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003E\u003Cstrong\u003EBehavioral Feature View.\u003C/strong\u003E Loan-level credit attributes from Equifax, time-stamped by \u003Ccode\u003EARCHIVE_DATE\u003C/code\u003E so later joins are point-in-time correct:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Behavioral Feature View — loan-level attributes from Equifax\nbehavioral_source = session.sql("""\n    SELECT TRADE_ID, ARCHIVE_DATE,\n           VANTAGE_SCORE_3, NUMBER_OF_ACCOUNTS, NUMBER_ACCOUNTS_PAST_DUE,\n           NUMBER_ACCOUNTS_MAJOR_DEROGATORY, HIGH_UTIL_ACCTS,\n           AGE_OLDEST_ACCOUNT, TOTAL_PAST_DUE_AMOUNT, BALANCE\n    FROM CREDIT_RISK_ML.PD_MODEL.LOAN_DATA\n""")\n\nbehavioral_fv = FeatureView(\n    name="LOAN_BEHAVIORAL_FEATURES",\n    entities=[loan_entity],\n    feature_df=behavioral_source,\n    timestamp_col="ARCHIVE_DATE",\n    refresh_freq="1 day",\n    desc="Loan-level behavioral features from Equifax"\n)\nbehavioral_fv = fs.register_feature_view(feature_view=behavioral_fv, version="V1")\nprint(f"Registered: {behavioral_fv.name} v{behavioral_fv.version}")\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003E\u003Cstrong\u003EMacro Feature View.\u003C/strong\u003E FRED indicators joined to each loan's observation month. These historical values are used for training:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Macro Feature View — FRED economic indicators joined to loan population\nmacro_source = session.sql("""\n    SELECT l.TRADE_ID, l.ARCHIVE_DATE,\n           m.UNEMPLOYMENT_RATE, m.FED_FUNDS_RATE, m.CPI_INDEX\n    FROM CREDIT_RISK_ML.PD_MODEL.LOAN_DATA l\n    JOIN CREDIT_RISK_ML.PD_MODEL.MACRO_INDICATORS m\n        ON DATE_TRUNC('MONTH', l.ARCHIVE_DATE) = m.OBSERVATION_MONTH\n""")\n\nmacro_fv = FeatureView(\n    name="LOAN_MACRO_FEATURES",\n    entities=[loan_entity],\n    feature_df=macro_source,\n    timestamp_col="ARCHIVE_DATE",\n    refresh_freq="1 day",\n    desc="Macroeconomic indicators from FRED joined to loan population"\n)\nmacro_fv = fs.register_feature_view(feature_view=macro_fv, version="V1")\nprint(f"Registered: {macro_fv.name} v{macro_fv.version}")\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003E\u003Cstrong\u003EMacro Scenario Feature View.\u003C/strong\u003E IFRS 9 and CECL require forward-looking, scenario-weighted loss estimates. This view holds four illustrative scenarios (baseline, upside, downside and severe downside), each a shock applied to the latest FRED observation with a probability weight. In production these would come from your economics team, regulator-prescribed scenarios or a Marketplace provider. The view is keyed by \u003Ccode\u003ESCENARIO_ID\u003C/code\u003E rather than by loan, because each loan is scored once per scenario at scoring time.\u003C/p\u003E\n\u003Ctable\u003E\u003Cthead\u003E\u003Ctr\u003E\u003Cth colspan=\"1\" rowspan=\"1\"\u003EScenario\u003C/th\u003E\u003Cth colspan=\"1\" rowspan=\"1\"\u003EWeight\u003C/th\u003E\u003Cth colspan=\"1\" rowspan=\"1\"\u003EUnemployment shift\u003C/th\u003E\u003Cth colspan=\"1\" rowspan=\"1\"\u003EFed funds shift\u003C/th\u003E\u003Cth colspan=\"1\" rowspan=\"1\"\u003ECPI factor\u003C/th\u003E\u003C/tr\u003E\u003C/thead\u003E\u003Ctbody\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003EBASELINE\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E0.50\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E0.0\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E0.00\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E1.000\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003EUPSIDE\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E0.15\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E-0.5\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E+0.25\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E1.010\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003EDOWNSIDE\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E0.30\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E+2.0\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E-1.00\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E0.990\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003ESEVERE_DOWNSIDE\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E0.05\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E+4.5\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E-2.00\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E0.970\u003C/td\u003E\u003C/tr\u003E\u003C/tbody\u003E\u003C/table\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Macro Scenario Feature View — forward-looking scenarios for IFRS 9 / CECL scenario-weighted ECL\n# Illustrative shocks applied to the latest FRED observation; in production these come from the\n# economics team, regulator-prescribed scenarios or a Marketplace provider.\nfrom python_pd_model import MACRO_COLS  # shared with the training module\n\nsession.sql("""\n    CREATE OR REPLACE TABLE CREDIT_RISK_ML.PD_MODEL.MACRO_SCENARIOS CHANGE_TRACKING = TRUE AS\n    WITH latest AS (\n        SELECT * FROM CREDIT_RISK_ML.PD_MODEL.MACRO_INDICATORS\n        QUALIFY ROW_NUMBER() OVER (ORDER BY OBSERVATION_MONTH DESC) = 1\n    ),\n    shocks (SCENARIO_ID, SCENARIO_WEIGHT, UNRATE_SHIFT, FEDFUNDS_SHIFT, CPI_FACTOR) AS (\n        SELECT * FROM VALUES\n            ('BASELINE',        0.50,  0.0,   0.00, 1.000),\n            ('UPSIDE',          0.15, -0.5,   0.25, 1.010),\n            ('DOWNSIDE',        0.30,  2.0,  -1.00, 0.990),\n            ('SEVERE_DOWNSIDE', 0.05,  4.5,  -2.00, 0.970)\n    )\n    SELECT s.SCENARIO_ID, s.SCENARIO_WEIGHT,\n           l.OBSERVATION_MONTH AS SCENARIO_ANCHOR_MONTH,\n           l.UNEMPLOYMENT_RATE + s.UNRATE_SHIFT               AS UNEMPLOYMENT_RATE,\n           GREATEST(l.FED_FUNDS_RATE + s.FEDFUNDS_SHIFT, 0)   AS FED_FUNDS_RATE,\n           l.CPI_INDEX * s.CPI_FACTOR                         AS CPI_INDEX\n    FROM shocks s CROSS JOIN latest l\n""").collect()\n\nscenario_entity = Entity(name="MACRO_SCENARIO", join_keys=["SCENARIO_ID"])\nfs.register_entity(scenario_entity)\n\nscenario_fv = FeatureView(\n    name="MACRO_SCENARIO_FEATURES",\n    entities=[scenario_entity],\n    feature_df=session.table("CREDIT_RISK_ML.PD_MODEL.MACRO_SCENARIOS").select(\n        "SCENARIO_ID", "SCENARIO_WEIGHT", *MACRO_COLS),\n    refresh_freq="1 day",  # in practice quarterly, or ad hoc when economics revises forecasts\n    desc="Forward-looking macro scenarios (baseline/upside/downside/severe) with probability weights",\n)\nscenario_fv = fs.register_feature_view(feature_view=scenario_fv, version="V1")\nprint(f"Registered: {scenario_fv.name} v{scenario_fv.version}")\nfs.read_feature_view(scenario_fv).show()\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003EList the registered Feature Views:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# List all registered feature views\nfv_list = fs.list_feature_views().to_pandas()\nprint(f"Feature Views registered: {len(fv_list)}")\nfv_list[["NAME", "VERSION", "DESC"]]\n\u003C/code\u003E\u003C/pre\u003E\n<!-- ------------------------ -->\n","\u003Ch2\u003EGenerate a Versioned Training Dataset\u003C/h2\u003E\n","\u003Cp\u003ERegulators ask: \u003Cem\u003E"What data trained this model?"\u003C/em\u003E A Snowflake \u003Cstrong\u003EDataset\u003C/strong\u003E answers that with a single immutable, versioned reference.\u003C/p\u003E\n","\u003Cp\u003EBuild a spine of loan IDs, observation dates and labels:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003Efrom snowflake.snowpark.functions import col\n\n# Build spine: loan IDs + observation dates + labels\nspine_df = session.table("CREDIT_RISK_ML.PD_MODEL.LOAN_DATA").select(\n    col("TRADE_ID"), col("ARCHIVE_DATE"), col("DEFAULT_FLAG")\n)\nprint(f"Spine rows: {spine_df.count():,}")\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003EThen let the Feature Store join the behavioral and macro features as of each loan's \u003Ccode\u003EARCHIVE_DATE\u003C/code\u003E, so no feature value from the future leaks into training:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Generate training dataset with point-in-time joins (no look-ahead bias)\ndataset = fs.generate_dataset(\n    name="PD_TRAINING_DS",\n    spine_df=spine_df,\n    features=[behavioral_fv, macro_fv],\n    spine_timestamp_col="ARCHIVE_DATE",\n    spine_label_cols=["DEFAULT_FLAG"],\n    output_type="dataset"\n)\nprint(f"Dataset: {dataset.fully_qualified_name}")\nprint(f"Version: {dataset.selected_version.name}")\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003EInspect the result:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Inspect the versioned dataset\nds_df = dataset.read.to_snowpark_dataframe()\nprint(f"Rows: {ds_df.count():,}")\nprint(f"Columns: {ds_df.columns}")\nds_df.show(5)\n\u003C/code\u003E\u003C/pre\u003E\n<!-- ------------------------ -->\n","\u003Ch2\u003ETrain with Experiment Tracking\u003C/h2\u003E\n","\u003Cp\u003EThe PD model is an \u003Ccode\u003EXGBClassifier\u003C/code\u003E, the Python equivalent of the SAS \u003Ccode\u003EPROC GRADBOOST\u003C/code\u003E model (\u003Ccode\u003Esas_pd_model.sas\u003C/code\u003E). It was converted with \u003Cstrong\u003ECoCo\u003C/strong\u003E. The hyperparameters map directly from the SAS source:\u003C/p\u003E\n\u003Ctable\u003E\u003Cthead\u003E\u003Ctr\u003E\u003Cth colspan=\"1\" rowspan=\"1\"\u003ESAS PROC GRADBOOST\u003C/th\u003E\u003Cth colspan=\"1\" rowspan=\"1\"\u003EXGBClassifier\u003C/th\u003E\u003Cth colspan=\"1\" rowspan=\"1\"\u003EValue\u003C/th\u003E\u003C/tr\u003E\u003C/thead\u003E\u003Ctbody\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003Entrees=200\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003En_estimators\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E200\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003Elearningrate=0.1\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003Elearning_rate\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E0.1\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003Emaxdepth=6\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003Emax_depth\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E6\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003Evars_to_try=5\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003Ecolsample_bynode\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E5/11 (~0.45), per split\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003Eminleafsize=20\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003Emin_child_weight\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E20 (approximate)\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003Eseed=42\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003Erandom_state\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E42\u003C/td\u003E\u003C/tr\u003E\u003C/tbody\u003E\u003C/table\u003E\n","\u003Cp\u003EImport the model definition from \u003Ccode\u003Epython_pd_model.py\u003C/code\u003E, then load the Dataset and create the stratified 80/20 split over the 11 features (8 behavioral, 3 macro) with the module's \u003Ccode\u003Eprepare_features\u003C/code\u003E and \u003Ccode\u003Esplit_train_test\u003C/code\u003E:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# The model definition lives in python_pd_model.py (converted from SAS by Cortex Code).\n# The notebook and the training pipeline both import it, so they train exactly the same model.\nimport numpy as np\nfrom python_pd_model import (FEATURE_COLS, TARGET_COL, SAS_MAPPED_PARAMS, CHALLENGER_PARAMS,\n                             prepare_features, split_train_test, train_pd_model, pd_metrics)\n\n# Load dataset to pandas\ntrain_pdf = ds_df.to_pandas()\n\nX, y = prepare_features(train_pdf)                           # SAS DATA step: missing -> 0\nX_train, X_test, y_train, y_test = split_train_test(X, y)    # SAS PROC SURVEYSELECT: 80/20, seed 42\nprint(f"Train: {X_train.shape[0]:,} rows | Test: {X_test.shape[0]:,} rows")\nprint(f"Default rate — Train: {y_train.mean():.3f} | Test: {y_test.mean():.3f}")\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003ETrain two candidates with the module's \u003Ccode\u003Etrain_pd_model\u003C/code\u003E and log both to the \u003Ccode\u003EPD_MODEL_DEVELOPMENT\u003C/code\u003E experiment: the SAS-mapped baseline, \u003Ccode\u003ESAS_MAPPED_PARAMS\u003C/code\u003E (the future \u003Cstrong\u003Echampion\u003C/strong\u003E), and a tuned configuration, \u003Ccode\u003ECHALLENGER_PARAMS\u003C/code\u003E (the \u003Cstrong\u003Echallenger\u003C/strong\u003E). Each run records its hyperparameters, the exact Dataset version it consumed, and the standard credit risk metrics:\u003C/p\u003E\n\u003Cul\u003E\u003Cli\u003E\u003Cstrong\u003EGini\u003C/strong\u003E (2 × AUC − 1): discrimination\u003C/li\u003E\u003Cli\u003E\u003Cstrong\u003EKS\u003C/strong\u003E: the maximum separation between the score distributions of defaulters and non-defaulters\u003C/li\u003E\u003Cli\u003E\u003Cstrong\u003EBrier score\u003C/strong\u003E: calibration\u003C/li\u003E\u003C/ul\u003E\n","\u003Cp\u003EThis replaces the Word or Excel tracker in which modelers record what they tried.\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Train XGBoost — converted from SAS PROC GRADBOOST (see sas_pd_model.sas) — with Experiment Tracking\n# Every run (params, dataset version, Gini/KS/Brier) is logged to the PD_MODEL_DEVELOPMENT experiment\nfrom snowflake.ml.experiment import ExperimentTracking\n\nexp = ExperimentTracking(session=session, database_name="CREDIT_RISK_ML", schema_name="PD_MODEL")\nexp.set_experiment("PD_MODEL_DEVELOPMENT")\n\n# Champion candidate: hyperparameters mapped from SAS PROC GRADBOOST (see SAS_MAPPED_PARAMS in python_pd_model.py)\nparams = SAS_MAPPED_PARAMS\n# Challenger candidate: CHALLENGER_PARAMS (slower learning rate, more trees, row subsampling)\n\nruns = {}\nfor run_name, run_params in [("sas_mapped_baseline", params), ("challenger_lr005_d5", CHALLENGER_PARAMS)]:\n    with exp.start_run(run_name):\n        exp.log_params({**run_params, "dataset": dataset.fully_qualified_name,\n                        "dataset_version": dataset.selected_version.name})\n        clf = train_pd_model(X_train, y_train, run_params)\n        run_metrics = pd_metrics(y_test.values, clf.predict_proba(X_test)[:, 1])\n        exp.log_metrics(run_metrics)\n    runs[run_name] = (clf, run_metrics)\n    print(f"{run_name:<22} AUC={run_metrics['auc']:.4f}  Gini={run_metrics['gini']:.4f}  "\n          f"KS={run_metrics['ks']:.4f}  Brier={run_metrics['brier']:.4f}")\n\nmodel, champion_metrics = runs["sas_mapped_baseline"]\nchallenger_model, challenger_metrics = runs["challenger_lr005_d5"]\ny_pred_proba = model.predict_proba(X_test)[:, 1]\nprint("\\nCompare runs in Snowsight: AI & ML > Experiments > PD_MODEL_DEVELOPMENT")\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003EPlot the champion's feature importance:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Feature importance\nimport matplotlib.pyplot as plt\n\nimportances = model.feature_importances_\nsorted_idx = np.argsort(importances)\n\nfig, ax = plt.subplots(figsize=(10, 6))\nax.barh(range(len(sorted_idx)), importances[sorted_idx])\nax.set_yticks(range(len(sorted_idx)))\nax.set_yticklabels([FEATURE_COLS[i] for i in sorted_idx])\nax.set_xlabel("Feature Importance (Gain)")\nax.set_title("PD Model — XGBoost Feature Importance")\nplt.tight_layout()\nplt.show()\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003E\u003Cimg src=\"https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/xgboost-feature-importance.png?v=54b7b1dd\" alt=\"XGBoost feature importance\"\u003E\u003C/p\u003E\n","\u003Cp\u003EYou can compare the runs in Snowsight under \u003Cstrong\u003EAI & ML > Experiments > PD_MODEL_DEVELOPMENT\u003C/strong\u003E.\u003C/p\u003E\n","\u003Ch3\u003EOptional: scale out with ML Jobs and Many Model Training\u003C/h3\u003E\n","\u003Cp\u003EThe demo portfolio trains comfortably in the notebook kernel. For larger portfolios, the same code runs as an \u003Cstrong\u003EML Job\u003C/strong\u003E on a CPU compute pool. \u003Cstrong\u003EMany Model Training\u003C/strong\u003E fits one PD model per segment (here \u003Ccode\u003EPRODUCT_CATEGORY\u003C/code\u003E) in a single distributed run. Both cells are gated by \u003Ccode\u003ERUN_ON_COMPUTE_POOL = False\u003C/code\u003E, so the notebook runs end to end without a compute pool. Set it to \u003Ccode\u003ETrue\u003C/code\u003E to try them.\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# ML Job: run the challenger training on a compute pool instead of the notebook kernel\nRUN_ON_COMPUTE_POOL = False\nCOMPUTE_POOL = "SYSTEM_COMPUTE_POOL_CPU"\n\nsession.sql("CREATE STAGE IF NOT EXISTS CREDIT_RISK_ML.PD_MODEL.ML_ARTIFACTS").collect()\n\nif RUN_ON_COMPUTE_POOL:\n    from snowflake.ml.jobs import remote\n\n    @remote(COMPUTE_POOL, stage_name="CREDIT_RISK_ML.PD_MODEL.ML_ARTIFACTS",\n            pip_requirements=["xgboost", "scikit-learn"], session=session)\n    def train_pd_on_pool(dataset_name: str, dataset_version: str, run_params: dict, feature_cols: list):\n        from snowflake.snowpark import Session\n        from snowflake.ml.dataset import load_dataset\n        from xgboost import XGBClassifier\n        from sklearn.metrics import roc_auc_score\n\n        s = Session.builder.getOrCreate()\n        pdf = load_dataset(s, dataset_name, dataset_version).read.to_pandas()\n        X_all, y_all = pdf[feature_cols].fillna(0), pdf["DEFAULT_FLAG"]\n        clf = XGBClassifier(**run_params).fit(X_all, y_all)\n        print(f"In-sample Gini: {2 * roc_auc_score(y_all, clf.predict_proba(X_all)[:, 1]) - 1:.4f}")\n        return clf\n\n    job = train_pd_on_pool(dataset.fully_qualified_name, dataset.selected_version.name,\n                           CHALLENGER_PARAMS, FEATURE_COLS)\n    print(f"ML Job {job.id} submitted to {COMPUTE_POOL}")\n    job.wait()\n    pool_model = job.result()\n    print(job.get_logs())\nelse:\n    print("Skipped ML Job — set RUN_ON_COMPUTE_POOL = True to train on a compute pool")\n\u003C/code\u003E\u003C/pre\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Many Model Training: one PD model per product segment, trained in parallel on the compute pool\nif RUN_ON_COMPUTE_POOL:\n    from snowflake.ml.modeling.distributors.many_model import ManyModelTraining\n\n    segment_train_df = ds_df.join(\n        session.table("CREDIT_RISK_ML.PD_MODEL.LOAN_DATA").select("TRADE_ID", "PRODUCT_CATEGORY"),\n        on="TRADE_ID",\n    )\n\n    def train_segment_model(data_connector, context):\n        from xgboost import XGBClassifier\n        seg = data_connector.to_pandas()\n        print(f"Training PD model for segment {context.partition_id}: {len(seg):,} loans")\n        return XGBClassifier(**params).fit(seg[FEATURE_COLS].fillna(0), seg[TARGET_COL])\n\n    mmt = ManyModelTraining(train_segment_model, "CREDIT_RISK_ML.PD_MODEL.ML_ARTIFACTS")\n    mmt_run = mmt.run(\n        partition_by="PRODUCT_CATEGORY",\n        snowpark_dataframe=segment_train_df,\n        run_id="pd_by_product_category",\n    )\n    print(f"Many Model Training finished: {mmt_run.wait()}")\nelse:\n    print("Skipped Many Model Training — set RUN_ON_COMPUTE_POOL = True to train per-segment models")\n\u003C/code\u003E\u003C/pre\u003E\n\u003Cblockquote\u003E\n","\u003Cp\u003Easide negative\nThe per-segment models are a demonstration only. They are not registered or used by the pipelines later in this guide.\u003C/p\u003E\n\u003C/blockquote\u003E\n<!-- ------------------------ -->\n","\u003Ch2\u003ERegister and Explain Model\u003C/h2\u003E\n","\u003Cp\u003ERegister both models in the \u003Cstrong\u003EModel Registry\u003C/strong\u003E as versions of \u003Ccode\u003EPD_XGB_MODEL\u003C/code\u003E: \u003Cstrong\u003EV1\u003C/strong\u003E gets the \u003Ccode\u003Echampion\u003C/code\u003E alias (production) and \u003Cstrong\u003EV2\u003C/strong\u003E gets the \u003Ccode\u003Echallenger\u003C/code\u003E alias (candidate). Logging each model with a Snowpark sample from the versioned Dataset does two things:\u003C/p\u003E\n\u003Cul\u003E\u003Cli\u003Eit records \u003Cstrong\u003EML Lineage\u003C/strong\u003E from the model back to the Dataset and Feature Views;\u003C/li\u003E\u003Cli\u003Ewith \u003Ccode\u003Etarget_platforms=["WAREHOUSE"]\u003C/code\u003E, it enables the Registry's built-in SHAP \u003Ccode\u003Eexplain\u003C/code\u003E function.\u003C/li\u003E\u003C/ul\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003Efrom snowflake.ml.registry import Registry\n\nreg = Registry(session=session, database_name="CREDIT_RISK_ML", schema_name="PD_MODEL")\n\n# Drop existing model if re-running so V1/V2 can be re-registered\ntry:\n    reg.delete_model("PD_XGB_MODEL")\n    print("Dropped existing PD_XGB_MODEL for re-registration")\nexcept Exception:\n    pass\n\n# Snowpark sample from the versioned Dataset -> records lineage Dataset -> Model\nsample_input = ds_df.select(FEATURE_COLS).limit(5)\n\nregistered = {}\nfor version_name, alias, mdl, mtr, desc in [\n    ("V1", "champion", model, champion_metrics, "SAS-mapped hyperparameters"),\n    ("V2", "challenger", challenger_model, challenger_metrics, "tuned candidate"),\n]:\n    mv_i = reg.log_model(\n        model=mdl,\n        model_name="PD_XGB_MODEL",\n        version_name=version_name,\n        sample_input_data=sample_input,\n        comment=f"{alias.upper()} — PD model, XGBoost migrated from SAS PROC GRADBOOST ({desc})",\n        metrics=mtr,\n        target_platforms=["WAREHOUSE"],  # warehouse target + sample input enables the explain function\n    )\n    mv_i.set_alias(alias)\n    registered[alias] = mv_i\n    print(f"Registered {version_name} as '{alias}': Gini={mtr['gini']:.4f}, KS={mtr['ks']:.4f}, Brier={mtr['brier']:.4f}")\n\nmv = registered["champion"]\npd_model = reg.get_model("PD_XGB_MODEL")\npd_model.default = mv\npd_model.show_versions()\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Ch3\u003ESHAP explainability\u003C/h3\u003E\n","\u003Cp\u003ECredit decisions must be explainable, for audit purposes. The \u003Ccode\u003Eexplain\u003C/code\u003E function runs on the warehouse next to the data, with no external SHAP tooling. Start with global importance:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# SHAP Explainability from the Registry — critical for adverse action notices and SR 11-7 compliance\n# The explain function runs on the warehouse next to the data; no local SHAP tooling required\nimport altair as alt\nfrom snowflake.ml.monitoring.explain_visualize import plot_violin\n\nexplain_input = ds_df.select("TRADE_ID", *FEATURE_COLS).limit(200)\nexplanations = mv.run(explain_input, function_name="explain").to_pandas()\n\nshap_cols = [f"{c}_explanation" for c in FEATURE_COLS]\nshap_df = explanations[shap_cols].rename(columns=lambda c: c.replace("_explanation", ""))\nfeature_df = explanations[FEATURE_COLS]\n\n# Global importance: mean |SHAP| per feature\nprint(shap_df.abs().mean().sort_values(ascending=False).round(4).to_string())\n\nalt.renderers.enable("mimetype")\nplot_violin(shap_df=shap_df, feature_df=feature_df)\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003E\u003Cimg src=\"https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/shap-explainability.png?v=54b7b1dd\" alt=\"SHAP values per feature\"\u003E\u003C/p\u003E\n","\u003Cp\u003EThen explain a single prediction. This is the input to an adverse action notice ("why was this applicant declined?"):\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Per-prediction explanation (adverse action reporting) from the Registry's explain output\nsample_idx = 0\nsample_trade = explanations["TRADE_ID"].iloc[sample_idx].item()  # numpy scalar -> native Python for Snowpark literal\nsample_pd = mv.run(\n    explain_input.filter(col("TRADE_ID") == sample_trade), function_name="predict_proba"\n).to_pandas()["output_feature_1"].iloc[0]\n\nprint(f"Loan {sample_trade}: PD = {sample_pd:.4f}\\n")\nprint("Top adverse action factors:")\nfactors = sorted(\n    zip(FEATURE_COLS, shap_df.iloc[sample_idx].values, feature_df.iloc[sample_idx].values),\n    key=lambda x: abs(x[1]), reverse=True,\n)\nfor feat, shap_val, feat_val in factors[:5]:\n    direction = "increases" if shap_val > 0 else "decreases"\n    print(f"  {feat} = {feat_val:.1f} ({direction} PD by {abs(shap_val):.4f})")\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Ch3\u003EML Lineage\u003C/h3\u003E\n","\u003Cp\u003EAnswer "what data trained this model?" with a query instead of a document hunt. Walk upstream from the champion to its Dataset, Feature Views and source tables:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# ML Lineage — trace the champion back to its Dataset, Feature Views and source tables\n# Answers the validator question "what data trained this model?" with a query, not a document hunt\nfor node in mv.lineage(direction="upstream"):\n    print(f"model {mv.version_name} <- {node}")\n    if hasattr(node, "lineage"):\n        for parent in node.lineage(direction="upstream"):\n            print(f"    <- {parent}")\n\u003C/code\u003E\u003C/pre\u003E\n<!-- ------------------------ -->\n","\u003Ch2\u003EBatch Score and ECL\u003C/h2\u003E\n","\u003Cp\u003ECredit risk models run in batch, for monthly provisioning and quarterly stress tests. Features come straight from the Feature Store through \u003Ccode\u003Eretrieve_feature_values\u003C/code\u003E, so scoring uses exactly the same feature definitions as training.\u003C/p\u003E\n","\u003Cp\u003EScore the same portfolio with the champion and the challenger. This is \u003Cstrong\u003Eoffline champion/challenger\u003C/strong\u003E: both models see the identical population, and nothing reaches production decisions until a model is explicitly promoted.\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Score the same portfolio with champion and challenger — features served by the Feature Store\nfrom snowflake.snowpark.types import TimestampType\n\npd_model = reg.get_model("PD_XGB_MODEL")\n\nscoring_spine = session.table("CREDIT_RISK_ML.PD_MODEL.LOAN_DATA").select(\n    "TRADE_ID", "ARCHIVE_DATE", "DEFAULT_FLAG", "PRODUCT_CATEGORY"\n).limit(100000)\n\nscoring_df = fs.retrieve_feature_values(\n    spine_df=scoring_spine,\n    features=[behavioral_fv, macro_fv],\n    spine_timestamp_col="ARCHIVE_DATE",\n)\n\nfor alias in ["champion", "challenger"]:\n    mv_alias = pd_model.version(alias)\n    preds = mv_alias.run(scoring_df, function_name="predict_proba")\n    # Model monitors require a TIMESTAMP_NTZ timestamp column\n    preds = preds.with_column("ARCHIVE_DATE", col("ARCHIVE_DATE").cast(TimestampType()))\n    preds.write.save_as_table(f"CREDIT_RISK_ML.PD_MODEL.PD_SCORES_{alias.upper()}", mode="overwrite")\n    print(f"{alias:<10} ({mv_alias.version_name}): scored {preds.count():,} loans "\n          f"-> PD_SCORES_{alias.upper()}")\n\nsession.table("CREDIT_RISK_ML.PD_MODEL.PD_SCORES_CHAMPION").show(5)\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003ECompare discrimination and calibration on that population:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Champion vs challenger on the identical population: discrimination (Gini, KS) and calibration (Brier)\nfrom sklearn.metrics import brier_score_loss\n\ncomparison = []\nfor alias in ["CHAMPION", "CHALLENGER"]:\n    s = session.table(f"CREDIT_RISK_ML.PD_MODEL.PD_SCORES_{alias}").select(\n        "DEFAULT_FLAG", '"output_feature_1"').to_pandas()\n    y_s, p_s = s["DEFAULT_FLAG"].values, s["output_feature_1"].values\n    a = roc_auc_score(y_s, p_s)\n    comparison.append({\n        "model": alias, "auc": a, "gini": 2 * a - 1,\n        "ks": ks_2samp(p_s[y_s == 0], p_s[y_s == 1]).statistic,\n        "brier": brier_score_loss(y_s, p_s),\n        "mean_pd": p_s.mean(), "observed_dr": y_s.mean(),\n    })\n\nimport pandas as pd\npd.DataFrame(comparison).set_index("model").round(4)\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003EFor very large portfolios, \u003Ccode\u003Erun_batch()\u003C/code\u003E distributes inference across a compute pool. This step is optional and gated by the same flag:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Large portfolios: job-based batch inference with run_batch() on an SPCS compute pool\n# (warehouse run() above is enough for this demo portfolio; RUN_ON_COMPUTE_POOL is set in section 6b)\nif RUN_ON_COMPUTE_POOL:\n    from snowflake.ml.model.batch import JobSpec, OutputSpec, SaveMode\n\n    batch_job = pd_model.version("champion").run_batch(\n        X=scoring_df,\n        compute_pool=COMPUTE_POOL,\n        output_spec=OutputSpec(\n            stage_location="@CREDIT_RISK_ML.PD_MODEL.ML_ARTIFACTS/batch_scores/",\n            mode=SaveMode.OVERWRITE,\n        ),\n        job_spec=JobSpec(function_name="predict_proba"),\n    )\n    batch_job.wait()\n    print(f"run_batch job {batch_job.id}: {batch_job.status}")\nelse:\n    print("Skipped run_batch() — set RUN_ON_COMPUTE_POOL = True to score on a compute pool")\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Ch3\u003EScenario-weighted Expected Credit Loss\u003C/h3\u003E\n","\u003Cp\u003EECL = PD × LGD × EAD, weighted across macro scenarios:\nScenario-conditioned scoring: score each loan once per macro scenario, then probability-weight\nLGD is a flat illustrative assumption and BALANCE is used as the EAD proxy — this guide models PD only\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003ELGD_ASSUMPTION = 0.45\n\nscenarios_df = fs.read_feature_view(scenario_fv)\nbehavioral_cols = [c for c in FEATURE_COLS if c not in MACRO_COLS]\n\n# Replace each loan's historical macro values with the scenario's forward-looking values\nscenario_scoring_df = scoring_df.select("TRADE_ID", "ARCHIVE_DATE", *behavioral_cols).cross_join(scenarios_df)\nscenario_preds = pd_model.version("champion").run(\n    scenario_scoring_df.select("TRADE_ID", "SCENARIO_ID", "SCENARIO_WEIGHT", *FEATURE_COLS),\n    function_name="predict_proba",\n)\nscenario_preds.write.save_as_table("CREDIT_RISK_ML.PD_MODEL.PD_SCORES_SCENARIO", mode="overwrite")\n\necl_by_scenario = session.sql(f"""\n    SELECT SCENARIO_ID, ANY_VALUE(SCENARIO_WEIGHT) AS SCENARIO_WEIGHT,\n           AVG("output_feature_1") AS MEAN_PD,\n           SUM("output_feature_1" * BALANCE * {LGD_ASSUMPTION}) AS ECL\n    FROM CREDIT_RISK_ML.PD_MODEL.PD_SCORES_SCENARIO\n    GROUP BY SCENARIO_ID\n    ORDER BY MEAN_PD\n""").to_pandas()\nweighted_ecl = (ecl_by_scenario["SCENARIO_WEIGHT"] * ecl_by_scenario["ECL"]).sum()\n\n# Loan-level probability-weighted PD for provisioning\nsession.sql("""\n    CREATE OR REPLACE TABLE CREDIT_RISK_ML.PD_MODEL.PD_SCORES_WEIGHTED AS\n    SELECT TRADE_ID, SUM(SCENARIO_WEIGHT * "output_feature_1") AS WEIGHTED_PD\n    FROM CREDIT_RISK_ML.PD_MODEL.PD_SCORES_SCENARIO\n    GROUP BY TRADE_ID\n""").collect()\n\nprint(ecl_by_scenario.round(4).to_string(index=False))\nprint(f"\\nProbability-weighted ECL: {weighted_ecl:,.0f}  (LGD = {LGD_ASSUMPTION}, EAD = BALANCE)")\n\u003C/code\u003E\u003C/pre\u003E\n<!-- ------------------------ -->\n","\u003Ch2\u003EMonitor the Models\u003C/h2\u003E\n","\u003Cp\u003EAttach a \u003Cstrong\u003EModel Version Monitor\u003C/strong\u003E to the champion and to the challenger. Each monitor:\u003C/p\u003E\n\u003Cul\u003E\u003Cli\u003Ereads its model's scoring output (\u003Ccode\u003EPD_SCORES_<ROLE>\u003C/code\u003E);\u003C/li\u003E\u003Cli\u003Ecompares it with a baseline, which is that version's scores on its training Dataset;\u003C/li\u003E\u003Cli\u003Etracks data drift and prediction drift against that baseline;\u003C/li\u003E\u003Cli\u003Etracks performance against realized defaults (\u003Ccode\u003EDEFAULT_FLAG\u003C/code\u003E);\u003C/li\u003E\u003Cli\u003Ebreaks every metric out by \u003Ccode\u003EPRODUCT_CATEGORY\u003C/code\u003E.\u003C/li\u003E\u003C/ul\u003E\n","\u003Cp\u003EThis replaces quarterly manual monitoring reports.\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Model Version Monitors — one per version, each on its own scoring output (champion vs challenger)\n# predict_proba returns lowercase columns: "output_feature_1" is P(default), so it is the score column\n# Binary classification: ground truth (0/1 DEFAULT_FLAG) goes in ACTUAL_CLASS_COLUMNS, not ACTUAL_SCORE_COLUMNS\nfor alias, version in [("CHAMPION", "V1"), ("CHALLENGER", "V2")]:\n    # Baseline = this version's scores on its training Dataset; drift (PSI, JS divergence) is measured against it\n    baseline_input = ds_df.join(\n        session.table("CREDIT_RISK_ML.PD_MODEL.LOAN_DATA").select("TRADE_ID", "PRODUCT_CATEGORY"), on="TRADE_ID"\n    )\n    pd_model.version(version).run(baseline_input, function_name="predict_proba").with_column(\n        "ARCHIVE_DATE", col("ARCHIVE_DATE").cast(TimestampType())\n    ).write.save_as_table(f"CREDIT_RISK_ML.PD_MODEL.PD_BASELINE_{alias}", mode="overwrite")\n\n    session.sql(f"""\n        CREATE OR REPLACE MODEL MONITOR CREDIT_RISK_ML.PD_MODEL.PD_MODEL_MONITOR_{alias}\n        WITH\n            MODEL = CREDIT_RISK_ML.PD_MODEL.PD_XGB_MODEL\n            VERSION = '{version}'\n            FUNCTION = 'predict_proba'\n            SOURCE = CREDIT_RISK_ML.PD_MODEL.PD_SCORES_{alias}\n            BASELINE = CREDIT_RISK_ML.PD_MODEL.PD_BASELINE_{alias}\n            TIMESTAMP_COLUMN = ARCHIVE_DATE\n            PREDICTION_SCORE_COLUMNS = ('output_feature_1')\n            ACTUAL_CLASS_COLUMNS = (DEFAULT_FLAG)\n            ID_COLUMNS = (TRADE_ID)\n            SEGMENT_COLUMNS = (PRODUCT_CATEGORY)\n            WAREHOUSE = CREDIT_RISK_WH\n            AGGREGATION_WINDOW = '1 DAY'\n            REFRESH_INTERVAL = '1 DAY'\n    """).collect()\n    print(f"Model Monitor created: PD_MODEL_MONITOR_{alias} (version {version})")\n\n# Drop the single-version monitor from earlier runs of this notebook\nsession.sql("DROP MODEL MONITOR IF EXISTS CREDIT_RISK_ML.PD_MODEL.PD_MODEL_MONITOR").collect()\nprint("Tracking per version: data drift vs training baseline, prediction drift, Gini/KS/calibration vs realized defaults")\nprint("View in Snowsight: AI & ML > Models > PD_XGB_MODEL > Monitors")\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003EView the monitors in Snowsight under \u003Cstrong\u003EAI & ML > Models > PD_XGB_MODEL > Monitors\u003C/strong\u003E. When the training pipeline promotes a new version, it re-points that role's monitor automatically.\u003C/p\u003E\n<!-- ------------------------ -->\n","\u003Ch2\u003EAutomate Monthly Retraining\u003C/h2\u003E\n","\u003Cp\u003ENow automate the lifecycle with \u003Cstrong\u003ETask DAGs\u003C/strong\u003E, replacing the cron and shell scripts common in SAS environments. Each task calls a Python stored procedure.\u003C/p\u003E\n","\u003Cp\u003E\u003Cimg src=\"https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/training-pipeline.png?v=54b7b1dd\" alt=\"Training pipeline\"\u003E\u003C/p\u003E\n","\u003Cp\u003EEvery month the training pipeline produces \u003Cstrong\u003Etwo candidates\u003C/strong\u003E on a fresh Dataset:\u003C/p\u003E\n\u003Cul\u003E\u003Cli\u003Ea \u003Cstrong\u003Echampion candidate\u003C/strong\u003E: the production champion's hyperparameters, refit;\u003C/li\u003E\u003Cli\u003Ea \u003Cstrong\u003Echallenger candidate\u003C/strong\u003E: the challenger configuration, refit.\u003C/li\u003E\u003C/ul\u003E\n","\u003Cp\u003EBoth candidates are registered with \u003Ccode\u003ECHAMPION_CANDIDATE\u003C/code\u003E / \u003Ccode\u003ECHALLENGER_CANDIDATE\u003C/code\u003E aliases and scored on the same portfolio as the production champion. An advisory \u003Cstrong\u003Evalidation gate\u003C/strong\u003E checks Gini lift ≥ 0 and Brier ≤ 1.02 × champion. A reviewer then gets one email and approves or rejects each candidate separately. The automated gate only \u003Cem\u003Erecommends\u003C/em\u003E. Only a user with the \u003Ccode\u003EMODEL_REVIEWER\u003C/code\u003E role can move the production aliases, which gives you the four-eyes control expected under SR 11-7 / SS1/23.\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode\u003EPD_MODEL_TRAIN_PIPELINE (root, 1st of month 02:00)   SP_REFRESH_FEATURES\n  └─ ..._GENERATE_DATASET                              SP_GENERATE_DATASET\n       └─ ..._TRAIN_CANDIDATES                         SP_TRAIN_CANDIDATES\n            └─ ..._SCORE_CANDIDATES                    SP_SCORE_CANDIDATES\n                 └─ ..._VALIDATION_GATE                SP_VALIDATION_GATE   -> PROMOTION_LOG (PENDING_REVIEW)\n                      └─ ..._NOTIFY_REVIEWER           SP_NOTIFY_REVIEWER   -> review email\n\nReviewer (MODEL_REVIEWER), outside the DAG, once per candidate:\n  CALL SP_REVIEW_MODEL('<version>', 'APPROVE' | 'REJECT', '<comments>')\n    -> SP_APPLY_REVIEW: moves the alias, re-points the monitor, writes the audit trail\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Ch3\u003EPipeline stored procedures\u003C/h3\u003E\n","\u003Cp\u003EThis cell uploads \u003Ccode\u003Epython_pd_model.py\u003C/code\u003E to the \u003Ccode\u003EML_ARTIFACTS\u003C/code\u003E stage and registers each pipeline step as a stored procedure with that file as an import. \u003Ccode\u003ESP_TRAIN_CANDIDATES\u003C/code\u003E trains with the module's \u003Ccode\u003Eprepare_features\u003C/code\u003E, \u003Ccode\u003Esplit_train_test\u003C/code\u003E, \u003Ccode\u003Etrain_pd_model\u003C/code\u003E and hyperparameters, and every step computes metrics with its \u003Ccode\u003Epd_metrics\u003C/code\u003E. The monthly retraining therefore runs the same converted code as the notebook. To change the model, edit \u003Ccode\u003Epython_pd_model.py\u003C/code\u003E and re-run this cell. The pipeline steps are:\u003C/p\u003E\n\u003Ctable\u003E\u003Cthead\u003E\u003Ctr\u003E\u003Cth colspan=\"1\" rowspan=\"1\"\u003EProcedure\u003C/th\u003E\u003Cth colspan=\"1\" rowspan=\"1\"\u003EWhat it does\u003C/th\u003E\u003C/tr\u003E\u003C/thead\u003E\u003Ctbody\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003ESP_REFRESH_FEATURES\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003ERefreshes the behavioral and macro Feature Views\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003ESP_GENERATE_DATASET\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003EWrites a new timestamped version of \u003Ccode\u003EPD_TRAINING_DS\u003C/code\u003E\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003ESP_TRAIN_CANDIDATES\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003ERetrains both candidates, logs runs to Experiment Tracking, registers them with candidate aliases\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003ESP_SCORE_CANDIDATES\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003EScores the production champion and both candidates on the same portfolio\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003ESP_VALIDATION_GATE\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003EComputes Gini/Brier vs the champion, writes one \u003Ccode\u003EPENDING_REVIEW\u003C/code\u003E row per candidate to \u003Ccode\u003EPROMOTION_LOG\u003C/code\u003E\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003ESP_NOTIFY_REVIEWER\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003ESends one HTML email covering both candidates, with approve/reject commands for each\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003ESP_APPLY_REVIEW\u003C/code\u003E\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003EApplies a decision: moves the alias, re-points the monitor, comments the versions, updates the audit trail\u003C/td\u003E\u003C/tr\u003E\u003C/tbody\u003E\u003C/table\u003E\n","\u003Cp\u003EIt also creates \u003Ccode\u003EPROMOTION_LOG\u003C/code\u003E, the model risk review queue and audit trail.\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Pipeline steps as Python stored procedures — each Task in the DAG calls one of these\n# Every retraining cycle produces TWO candidates, each approved (or rejected) separately by the reviewer:\n#   champion candidate   — the current champion's hyperparameters refit on the latest Dataset\n#   challenger candidate — the challenger configuration (CHALLENGER_PARAMS) refit on the latest Dataset\n# Candidates carry CHAMPION_CANDIDATE / CHALLENGER_CANDIDATE aliases until reviewed; only an approval\n# moves the CHAMPION / CHALLENGER alias that weekly scoring uses.\nfrom snowflake.snowpark import Session\nimport python_pd_model as pdm\n\nDB, SCHEMA, WH = "CREDIT_RISK_ML", "PD_MODEL", "CREDIT_RISK_WH"\nMODEL_NAME = "PD_XGB_MODEL"\nSP_PACKAGES = ["snowflake-ml-python", "xgboost", "scikit-learn", "scipy", "pandas"]\nROLES = ["CHAMPION", "CHALLENGER"]\n\nsession.sql(f"CREATE STAGE IF NOT EXISTS {DB}.{SCHEMA}.ML_ARTIFACTS").collect()\n# Ship the model module to the stage; every pipeline procedure imports it from there\nMODEL_MODULE = f"@{DB}.{SCHEMA}.ML_ARTIFACTS/code/python_pd_model.py"\nsession.file.put("python_pd_model.py", f"@{DB}.{SCHEMA}.ML_ARTIFACTS/code",\n                 auto_compress=False, overwrite=True)\n# PROMOTION_LOG is the model-risk review queue and audit trail, one row per candidate:\n#   TARGET_ROLE        = role the candidate would take if approved (CHAMPION / CHALLENGER)\n#   CHALLENGER_VERSION = the candidate version under review (column name kept for compatibility)\n#   CHAMPION_VERSION   = the champion in production when the candidate was evaluated (the benchmark)\n#   DECISION           = the automated gate's recommendation (PASS / REJECT)\n#   STATUS             = PENDING_REVIEW -> APPROVED / REJECTED (by a human) or SUPERSEDED (newer candidate)\nsession.sql(f"""\n    CREATE TABLE IF NOT EXISTS {DB}.{SCHEMA}.PROMOTION_LOG (\n        EVALUATED_AT TIMESTAMP_NTZ, CHAMPION_VERSION STRING, CHALLENGER_VERSION STRING,\n        CHAMPION_GINI FLOAT, CHALLENGER_GINI FLOAT, CHAMPION_BRIER FLOAT, CHALLENGER_BRIER FLOAT,\n        DECISION STRING, PROMOTED BOOLEAN\n    )\n""").collect()\nfor column, dtype in [("STATUS", "STRING"), ("REVIEWED_BY", "STRING"),\n                      ("REVIEWED_AT", "TIMESTAMP_NTZ"), ("REVIEW_COMMENTS", "STRING"),\n                      ("NOTIFIED_AT", "TIMESTAMP_NTZ"), ("TARGET_ROLE", "STRING")]:\n    session.sql(f"ALTER TABLE {DB}.{SCHEMA}.PROMOTION_LOG ADD COLUMN IF NOT EXISTS {column} {dtype}").collect()\n# Rows written before TARGET_ROLE existed were challenger -> champion promotions\nsession.sql(f"UPDATE {DB}.{SCHEMA}.PROMOTION_LOG SET TARGET_ROLE = 'CHAMPION' WHERE TARGET_ROLE IS NULL").collect()\n\n# Reviewer emails and the email integration name are configured in the Setup cell (section 1)\n\n\ndef _version_by_alias(model, alias):\n    import json\n    for _, row in model.show_versions().iterrows():\n        aliases = json.loads(row["aliases"]) if row["aliases"] else []\n        if alias.upper() in [a.upper() for a in aliases]:\n            return model.version(row["name"])\n    return None\n\n\ndef _move_alias(model, alias, version_name):\n    """Point an alias at version_name, removing it from whichever version holds it now (idempotent)."""\n    holder = _version_by_alias(model, alias)\n    if holder is not None and holder.version_name == version_name:\n        return None\n    if holder is not None:\n        holder.unset_alias(alias)\n    model.version(version_name).set_alias(alias)\n    return holder.version_name if holder is not None else None\n\n\ndef _latest_dataset_version(session):\n    versions = session.sql(f"SHOW VERSIONS IN DATASET {DB}.{SCHEMA}.PD_TRAINING_DS").collect()\n    return sorted(versions, key=lambda r: r["created_on"])[-1]["version"]\n\n\ndef _scoring_features(session, fs, limit=100000):\n    spine = session.table(f"{DB}.{SCHEMA}.LOAN_DATA").select(\n        "TRADE_ID", "ARCHIVE_DATE", "DEFAULT_FLAG", "PRODUCT_CATEGORY")\n    if limit is not None:\n        spine = spine.limit(limit)\n    return fs.retrieve_feature_values(\n        spine_df=spine,\n        features=[fs.get_feature_view("LOAN_BEHAVIORAL_FEATURES", "V1"),\n                  fs.get_feature_view("LOAN_MACRO_FEATURES", "V1")],\n        spine_timestamp_col="ARCHIVE_DATE",\n    )\n\n\ndef _set_version_comment(session, version_name, text):\n    """Role-prefixed comment shown next to each version in Snowsight (AI & ML > Models).\n    DDL does not accept bind variables, so the text is escaped as a string literal."""\n    escaped = text.replace("\\\\", "\\\\\\\\").replace("'", "''")\n    session.sql(f"ALTER MODEL {DB}.{SCHEMA}.{MODEL_NAME} MODIFY VERSION {version_name} "\n                f"SET COMMENT = '{escaped}'").collect()\n\n\ndef _upsert_monitor(session, role, version_name):\n    """(Re)point the role's monitor at a version, with a baseline scored on the latest training Dataset."""\n    import json\n    from snowflake.snowpark.functions import col\n    from snowflake.snowpark.types import TimestampType\n    from snowflake.ml.dataset import load_dataset\n    from snowflake.ml.registry import Registry\n    train_df = load_dataset(session, f"{DB}.{SCHEMA}.PD_TRAINING_DS",\n                            _latest_dataset_version(session)).read.to_snowpark_dataframe()\n    baseline_input = train_df.join(\n        session.table(f"{DB}.{SCHEMA}.LOAN_DATA").select("TRADE_ID", "PRODUCT_CATEGORY"), on="TRADE_ID")\n    mv = Registry(session=session, database_name=DB, schema_name=SCHEMA).get_model(MODEL_NAME).version(version_name)\n    mv.run(baseline_input, function_name="predict_proba").with_column(\n        "ARCHIVE_DATE", col("ARCHIVE_DATE").cast(TimestampType())\n    ).write.save_as_table(f"{DB}.{SCHEMA}.PD_BASELINE_{role}", mode="overwrite")\n    # Snowflake allows one monitor per model version: drop any other monitor already on this version\n    for mon in session.sql(f"SHOW MODEL MONITORS IN SCHEMA {DB}.{SCHEMA}").collect():\n        on_version = json.loads(mon["model"]).get("version_name", "").upper() == version_name.upper()\n        if on_version and mon["name"].upper() != f"PD_MODEL_MONITOR_{role}":\n            session.sql(f"DROP MODEL MONITOR {DB}.{SCHEMA}.{mon['name']}").collect()\n    session.sql(f"""\n        CREATE OR REPLACE MODEL MONITOR {DB}.{SCHEMA}.PD_MODEL_MONITOR_{role}\n        WITH\n            MODEL = {DB}.{SCHEMA}.{MODEL_NAME}\n            VERSION = '{version_name}'\n            FUNCTION = 'predict_proba'\n            SOURCE = {DB}.{SCHEMA}.PD_SCORES_{role}\n            BASELINE = {DB}.{SCHEMA}.PD_BASELINE_{role}\n            TIMESTAMP_COLUMN = ARCHIVE_DATE\n            PREDICTION_SCORE_COLUMNS = ('output_feature_1')\n            ACTUAL_CLASS_COLUMNS = (DEFAULT_FLAG)\n            ID_COLUMNS = (TRADE_ID)\n            SEGMENT_COLUMNS = (PRODUCT_CATEGORY)\n            WAREHOUSE = {WH}\n            AGGREGATION_WINDOW = '1 DAY'\n            REFRESH_INTERVAL = '1 DAY'\n    """).collect()\n\n\ndef sp_refresh_features(session: Session) -> str:\n    from snowflake.ml.feature_store import FeatureStore\n    fs = FeatureStore(session, DB, SCHEMA, WH)\n    for name in ["LOAN_BEHAVIORAL_FEATURES", "LOAN_MACRO_FEATURES"]:\n        fs.refresh_feature_view(fs.get_feature_view(name, "V1"))\n    return "Feature views refreshed"\n\n\ndef sp_generate_dataset(session: Session) -> str:\n    import datetime\n    from snowflake.ml.feature_store import FeatureStore\n    fs = FeatureStore(session, DB, SCHEMA, WH)\n    version = datetime.datetime.utcnow().strftime("V_%Y%m%d_%H%M")\n    spine = session.table(f"{DB}.{SCHEMA}.LOAN_DATA").select("TRADE_ID", "ARCHIVE_DATE", "DEFAULT_FLAG")\n    fs.generate_dataset(\n        name="PD_TRAINING_DS", version=version, spine_df=spine,\n        features=[fs.get_feature_view("LOAN_BEHAVIORAL_FEATURES", "V1"),\n                  fs.get_feature_view("LOAN_MACRO_FEATURES", "V1")],\n        spine_timestamp_col="ARCHIVE_DATE", spine_label_cols=["DEFAULT_FLAG"], output_type="dataset",\n    )\n    return version\n\n\ndef sp_train_candidates(session: Session) -> str:\n    """Retrain both models on the latest Dataset and register them as candidates awaiting review."""\n    import datetime\n    from snowflake.ml.dataset import load_dataset\n    from snowflake.ml.experiment import ExperimentTracking\n    from snowflake.ml.registry import Registry\n\n    ds_version = _latest_dataset_version(session)\n    sdf = load_dataset(session, f"{DB}.{SCHEMA}.PD_TRAINING_DS", ds_version).read.to_snowpark_dataframe()\n    pdf = sdf.to_pandas()\n    X, y = pdm.prepare_features(pdf)\n    X_tr, X_te, y_tr, y_te = pdm.split_train_test(X, y)\n\n    reg = Registry(session=session, database_name=DB, schema_name=SCHEMA)\n    m = reg.get_model(MODEL_NAME)\n\n    # Champion candidate keeps the production champion's hyperparameters; challenger uses the challenger config\n    champion = _version_by_alias(m, "champion")\n    champion_params, params_source = pdm.SAS_MAPPED_PARAMS, "SAS-mapped baseline (no champion found)"\n    if champion is not None:\n        champion_params = {k: v for k, v in champion.load(force=True).get_params().items() if v is not None}\n        params_source = f"hyperparameters of champion {champion.version_name}"\n    candidates = [("CHAMPION", champion_params, params_source),\n                  ("CHALLENGER", pdm.CHALLENGER_PARAMS, "challenger configuration")]\n\n    exp = ExperimentTracking(session=session, database_name=DB, schema_name=SCHEMA)\n    exp.set_experiment("PD_MODEL_DEVELOPMENT")\n    stamp = datetime.datetime.utcnow().strftime("%Y%m%d_%H%M")\n    created = []\n    for role, run_params, source in candidates:\n        version_name = f"V_{stamp}_{role[:5]}"          # e.g. V_20261101_0200_CHAMP / _CHALL\n        with exp.start_run(f"retrain_{role.lower()}_{stamp}"):\n            exp.log_params({**run_params, "dataset_version": ds_version, "target_role": role})\n            model = pdm.train_pd_model(X_tr, y_tr, run_params)\n            metrics = pdm.pd_metrics(y_te.values, model.predict_proba(X_te)[:, 1])\n            exp.log_metrics(metrics)\n        reg.log_model(\n            model=model, model_name=MODEL_NAME, version_name=version_name,\n            sample_input_data=sdf.select(pdm.FEATURE_COLS).limit(5),  # Snowpark input -> lineage to the Dataset\n            metrics=metrics, target_platforms=["WAREHOUSE"],\n            comment=f"{role} CANDIDATE — pending review | {source}, dataset {ds_version}",\n        )\n        superseded = _move_alias(reg.get_model(MODEL_NAME), f"{role}_CANDIDATE", version_name)\n        if superseded:\n            _set_version_comment(session, superseded,\n                                 f"SUPERSEDED {role.lower()} candidate — replaced by {version_name} before review")\n        created.append(f"{role.lower()} candidate {version_name}")\n    return "Registered " + " and ".join(created)\n\n\ndef sp_score_candidates(session: Session) -> str:\n    """Score the production champion and both candidates on the identical portfolio."""\n    from snowflake.snowpark.functions import col\n    from snowflake.snowpark.types import TimestampType\n    from snowflake.ml.feature_store import FeatureStore\n    from snowflake.ml.registry import Registry\n    fs = FeatureStore(session, DB, SCHEMA, WH)\n    m = Registry(session=session, database_name=DB, schema_name=SCHEMA).get_model(MODEL_NAME)\n    features = _scoring_features(session, fs)\n    scored = []\n    for alias, table in [("champion", "PD_SCORES_CHAMPION"),\n                         ("champion_candidate", "PD_SCORES_CHAMPION_CANDIDATE"),\n                         ("challenger_candidate", "PD_SCORES_CHALLENGER_CANDIDATE")]:\n        mv = _version_by_alias(m, alias)\n        mv.run(features, function_name="predict_proba").with_column(\n            "ARCHIVE_DATE", col("ARCHIVE_DATE").cast(TimestampType())\n        ).write.save_as_table(f"{DB}.{SCHEMA}.{table}", mode="overwrite")\n        scored.append(f"{alias} {mv.version_name}")\n    return "Scored " + ", ".join(scored)\n\n\ndef sp_validation_gate(session: Session) -> str:\n    """Gate each candidate against the production champion and queue both for human review."""\n    import datetime\n    from snowflake.ml.registry import Registry\n    m = Registry(session=session, database_name=DB, schema_name=SCHEMA).get_model(MODEL_NAME)\n\n    def metrics_for(table):\n        pdf = session.table(f"{DB}.{SCHEMA}.{table}").select("DEFAULT_FLAG", '"output_feature_1"').to_pandas()\n        return pdm.pd_metrics(pdf["DEFAULT_FLAG"].values, pdf["output_feature_1"].values)\n\n    bench = metrics_for("PD_SCORES_CHAMPION")\n    champion_version = _version_by_alias(m, "champion").version_name\n    now = datetime.datetime.utcnow()\n    rows, summary = [], []\n    for role in ROLES:\n        candidate = _version_by_alias(m, f"{role}_candidate").version_name\n        cand = metrics_for(f"PD_SCORES_{role}_CANDIDATE")\n        passed = (cand["gini"] >= bench["gini"] + GATE_MIN_GINI_LIFT and\n                  cand["brier"] <= bench["brier"] * GATE_MAX_BRIER_RATIO)\n        decision = "PASS" if passed else "REJECT"\n        # An older candidate for the same role still waiting for review is superseded by this one\n        session.sql(f"""\n            UPDATE {DB}.{SCHEMA}.PROMOTION_LOG SET STATUS = 'SUPERSEDED'\n            WHERE STATUS = 'PENDING_REVIEW' AND TARGET_ROLE = ? AND CHALLENGER_VERSION != ?\n        """, params=[role, candidate]).collect()\n        rows.append([now, champion_version, candidate, bench["gini"], cand["gini"], bench["brier"],\n                     cand["brier"], decision, False, "PENDING_REVIEW", role])\n        summary.append(f"{role.lower()} candidate {candidate}: {decision}")\n    session.create_dataframe(rows, schema=[\n        "EVALUATED_AT", "CHAMPION_VERSION", "CHALLENGER_VERSION", "CHAMPION_GINI", "CHALLENGER_GINI",\n        "CHAMPION_BRIER", "CHALLENGER_BRIER", "DECISION", "PROMOTED", "STATUS", "TARGET_ROLE",\n    ]).write.mode("append").save_as_table(f"{DB}.{SCHEMA}.PROMOTION_LOG", column_order="name")\n    return "Gate recommendations — " + "; ".join(summary) + " — pending human review"\n\n\ndef sp_notify_reviewer(session: Session) -> str:\n    """One email covering every candidate awaiting review, each with its own approve / reject commands."""\n    from snowflake.ml.registry import Registry\n    pending = session.sql(f"""\n        SELECT * FROM {DB}.{SCHEMA}.PROMOTION_LOG\n        WHERE STATUS = 'PENDING_REVIEW' AND NOTIFIED_AT IS NULL\n        ORDER BY IFF(TARGET_ROLE = 'CHAMPION', 0, 1)\n    """).collect()\n    if not pending:\n        return "Nothing pending review"\n    m = Registry(session=session, database_name=DB, schema_name=SCHEMA).get_model(MODEL_NAME)\n    versions = m.show_versions()\n    trained = {row["name"]: str(row["created_on"])[:16] for _, row in versions.iterrows()}\n    current = {role: (_version_by_alias(m, role.lower()) or None) for role in ROLES}\n    current = {role: (mv.version_name if mv else "none") for role, mv in current.items()}\n\n    cell = "padding:6px 10px;border:1px solid #d1d5db"\n    colours = {"CHAMPION": "#1e3a8a", "CHALLENGER": "#b45309"}\n\n    def metric_row(label, bench_val, cand_val, higher_is_better):\n        delta = cand_val - bench_val\n        better = (delta > 0) == higher_is_better and delta != 0\n        verdict = ("Candidate better" if better else "Production champion better" if delta != 0 else "Equal")\n        colour = "#15803d" if better else "#b91c1c" if delta != 0 else "#6b7280"\n        return (f"<tr><td style='{cell}'><b>{label}</b></td><td style='{cell}'>{bench_val:.4f}</td>"\n                f"<td style='{cell}'>{cand_val:.4f}</td>"\n                f"<td style='{cell};color:{colour}'>{delta:+.4f} ({verdict})</td></tr>")\n\n    sections = []\n    for r in pending:\n        role, v, bench = r["TARGET_ROLE"], r["CHALLENGER_VERSION"], r["CHAMPION_VERSION"]\n        gate_colour = "#15803d" if r["DECISION"] == "PASS" else "#b91c1c"\n        sections.append(f"""\n        <h3 style="color:{colours[role]}">{role} candidate {v}</h3>\n        <p>If approved, <b>{v}</b> becomes the <b>{role}</b> (replacing {current[role]}) and is used by\n           weekly batch scoring as the {role.lower()} from the next run.</p>\n        <table style="border-collapse:collapse">\n          <tr><th style="{cell}"></th>\n              <th style="{cell};background:#1e3a8a;color:#ffffff">PRODUCTION CHAMPION — benchmark</th>\n              <th style="{cell};background:{colours[role]};color:#ffffff">{role} CANDIDATE — awaiting approval</th>\n              <th style="{cell}">Candidate vs production champion</th></tr>\n          <tr><td style="{cell}"><b>Version</b></td><td style="{cell}">{bench}</td><td style="{cell}">{v}</td><td style="{cell}"></td></tr>\n          <tr><td style="{cell}"><b>Trained</b></td><td style="{cell}">{trained.get(bench, 'n/a')}</td><td style="{cell}">{trained.get(v, 'n/a')}</td><td style="{cell}"></td></tr>\n          {metric_row("Gini (higher is better)", r["CHAMPION_GINI"], r["CHALLENGER_GINI"], True)}\n          {metric_row("Brier (lower is better)", r["CHAMPION_BRIER"], r["CHALLENGER_BRIER"], False)}\n        </table>\n        <p>Automated gate recommendation: <b style="color:{gate_colour}">{r['DECISION']}</b> (advisory).</p>\n        <pre>-- Approve {v} as {role} (replaces {current[role]})\nCALL {DB}.{SCHEMA}.SP_REVIEW_MODEL('{v}', 'APPROVE', '<comments>');\n-- Reject {v}; {current[role]} stays {role.lower()}\nCALL {DB}.{SCHEMA}.SP_REVIEW_MODEL('{v}', 'REJECT', '<comments>');</pre>""")\n\n    roles_pending = " + ".join(f"{r['TARGET_ROLE']} {r['CHALLENGER_VERSION']}" for r in pending)\n    subject = f"[PD model review] {len(pending)} candidate(s) awaiting approval — {roles_pending}"\n    body = f"""\n        <p>The monthly retraining run produced new candidates. Each is reviewed <b>separately</b>:\n           you can approve both, one, or neither. Both are benchmarked against the production champion\n           <b>{current['CHAMPION']}</b> on the same portfolio.</p>\n        {''.join(sections)}\n        <p>Gate rule: Gini lift ≥ {GATE_MIN_GINI_LIFT} and Brier ≤ {GATE_MAX_BRIER_RATIO}× production champion.\n           In Snowsight (AI & ML > Models > {MODEL_NAME}) candidates carry the CHAMPION_CANDIDATE /\n           CHALLENGER_CANDIDATE aliases and a comment starting with "CHAMPION CANDIDATE" / "CHALLENGER CANDIDATE".\n           Run the commands as a user with the MODEL_REVIEWER role.</p>\n    """\n    session.sql("CALL SYSTEM$SEND_EMAIL(?, ?, ?, ?, 'text/html')",\n                params=[REVIEW_EMAIL_INTEGRATION, ",".join(REVIEWER_EMAILS), subject, body]).collect()\n    session.sql(f"""\n        UPDATE {DB}.{SCHEMA}.PROMOTION_LOG SET NOTIFIED_AT = CURRENT_TIMESTAMP()::TIMESTAMP_NTZ\n        WHERE STATUS = 'PENDING_REVIEW' AND NOTIFIED_AT IS NULL\n    """).collect()\n    return f"Review request for {roles_pending} sent to {', '.join(REVIEWER_EMAILS)}"\n\n\ndef sp_apply_review(session: Session, challenger_version: str, decision: str,\n                    comments: str, reviewer: str) -> str:\n    """Owner's-rights step: apply a reviewer's decision on one candidate. Only reachable via SP_REVIEW_MODEL.\n    (The first argument is the candidate version; the name is kept so SP_REVIEW_MODEL's signature is unchanged.)"""\n    import datetime\n    from snowflake.ml.registry import Registry\n    candidate = challenger_version\n    decision = decision.strip().upper()\n    if decision not in ("APPROVE", "REJECT"):\n        raise ValueError("decision must be 'APPROVE' or 'REJECT'")\n    if not comments or not comments.strip():\n        raise ValueError("review comments are required for the audit trail")\n    pending = session.sql(f"""\n        SELECT TARGET_ROLE FROM {DB}.{SCHEMA}.PROMOTION_LOG\n        WHERE CHALLENGER_VERSION = ? AND STATUS = 'PENDING_REVIEW'\n    """, params=[candidate]).collect()\n    if not pending:\n        raise ValueError(f"No pending review for version {candidate}")\n    role = pending[0]["TARGET_ROLE"]\n    today = datetime.datetime.utcnow().strftime("%Y-%m-%d")\n    m = Registry(session=session, database_name=DB, schema_name=SCHEMA).get_model(MODEL_NAME)\n\n    promoted = decision == "APPROVE"\n    if promoted:\n        # Each step is idempotent so a retry after a partial failure completes cleanly.\n        # The role's monitor reads PD_SCORES_<ROLE>, so seed it with this candidate's gate scores first.\n        candidate_scores = f"{DB}.{SCHEMA}.PD_SCORES_{role}_CANDIDATE"\n        if session.sql(f"SHOW TABLES LIKE 'PD_SCORES_{role}_CANDIDATE' IN SCHEMA {DB}.{SCHEMA}").collect():\n            session.sql(f"CREATE OR REPLACE TABLE {DB}.{SCHEMA}.PD_SCORES_{role} CLONE {candidate_scores}").collect()\n        _upsert_monitor(session, role, candidate)\n        replaced = _move_alias(m, role, candidate)\n        holder = _version_by_alias(m, f"{role}_candidate")\n        if holder is not None and holder.version_name == candidate:\n            holder.unset_alias(f"{role}_candidate")\n        if role == "CHAMPION":\n            m.default = m.version(candidate)\n        # Role-prefixed comments make each version's status obvious in Snowsight\n        _set_version_comment(session, candidate, f"{role} — approved {today} by {reviewer}: {comments}")\n        if replaced:\n            _set_version_comment(session, replaced, f"RETIRED {role.lower()} — replaced by {candidate} on {today}")\n    else:\n        holder = _version_by_alias(m, f"{role}_candidate")\n        if holder is not None and holder.version_name == candidate:\n            holder.unset_alias(f"{role}_candidate")\n        _set_version_comment(session, candidate,\n                             f"REJECTED {role.lower()} candidate — {today} by {reviewer}: {comments}")\n\n    session.sql(f"""\n        UPDATE {DB}.{SCHEMA}.PROMOTION_LOG\n        SET STATUS = ?, PROMOTED = ?, REVIEWED_BY = ?,\n            REVIEWED_AT = CURRENT_TIMESTAMP()::TIMESTAMP_NTZ, REVIEW_COMMENTS = ?\n        WHERE CHALLENGER_VERSION = ? AND STATUS = 'PENDING_REVIEW'\n    """, params=["APPROVED" if promoted else "REJECTED", promoted, reviewer, comments, candidate]).collect()\n    return (f"{candidate} approved as {role.lower()} by {reviewer}" if promoted\n            else f"{candidate} rejected by {reviewer}; current {role.lower()} retained")\n\n\n# Validation-gate thresholds (candidate must match the champion's discrimination without degrading calibration)\nGATE_MIN_GINI_LIFT = 0.0\nGATE_MAX_BRIER_RATIO = 1.02\n\nfor fn in [sp_refresh_features, sp_generate_dataset, sp_train_candidates,\n           sp_score_candidates, sp_validation_gate, sp_notify_reviewer, sp_apply_review]:\n    session.sproc.register(\n        fn, name=f"{DB}.{SCHEMA}.{fn.__name__.upper()}",\n        packages=SP_PACKAGES, imports=[MODEL_MODULE], is_permanent=True, replace=True,\n        stage_location=f"@{DB}.{SCHEMA}.ML_ARTIFACTS",\n    )\n    print(f"Registered procedure {fn.__name__.upper()}")\n\n# The single-challenger procedures are replaced by the candidate procedures above\nfor old_proc in ["SP_TRAIN_CHALLENGER()", "SP_SCORE_CHAMPION_CHALLENGER()", "SP_PROMOTE_MODEL()"]:\n    session.sql(f"DROP PROCEDURE IF EXISTS {DB}.{SCHEMA}.{old_proc}").collect()\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Ch3\u003EHuman-in-the-loop review\u003C/h3\u003E\n","\u003Cp\u003ECreate the email notification integration and the reviewer entry point. \u003Ccode\u003ESP_REVIEW_MODEL\u003C/code\u003E runs with \u003Cstrong\u003Ecaller's rights\u003C/strong\u003E, so \u003Ccode\u003ECURRENT_USER()\u003C/code\u003E records the actual reviewer. It then hands off to the owner's-rights \u003Ccode\u003ESP_APPLY_REVIEW\u003C/code\u003E, which holds the registry privileges. The cell also creates the \u003Ccode\u003EREVIEW_QUEUE\u003C/code\u003E view and grants \u003Ccode\u003EMODEL_REVIEWER\u003C/code\u003E only what it needs. The pipeline role cannot approve its own models.\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Human-in-the-loop review: notification channel, reviewer role, and the approval entry point\n# REVIEWER_USERS / REVIEWER_EMAILS are configured in the Setup cell (section 1)\n\n# 1. Email notification integration — recipients must be verified emails of users in this account.\n#    ALLOWED_RECIPIENTS is re-synced from the Setup-cell config every run, so editing the lists there is enough.\nallowed = ", ".join(f"'{e}'" for e in dict.fromkeys(REVIEWER_EMAILS + SCORING_NOTIFY_EMAILS))\nsession.sql(f"""\n    CREATE NOTIFICATION INTEGRATION IF NOT EXISTS {REVIEW_EMAIL_INTEGRATION}\n        TYPE = EMAIL\n        ENABLED = TRUE\n        ALLOWED_RECIPIENTS = ({allowed})\n""").collect()\nsession.sql(f"ALTER NOTIFICATION INTEGRATION {REVIEW_EMAIL_INTEGRATION} SET ALLOWED_RECIPIENTS = ({allowed})").collect()\n\n# 2. Reviewer approval entry point — caller's rights so CURRENT_USER() is the actual reviewer,\n#    then hands off to the owner's-rights SP_APPLY_REVIEW, which holds the registry privileges\nsession.sql(f"""\n    CREATE OR REPLACE PROCEDURE {DB}.{SCHEMA}.SP_REVIEW_MODEL(\n        CHALLENGER_VERSION STRING, DECISION STRING, COMMENTS STRING)\n    RETURNS STRING\n    LANGUAGE SQL\n    EXECUTE AS CALLER\n    AS\n    $$\n    DECLARE\n        reviewer STRING DEFAULT CURRENT_USER();\n        result STRING;\n    BEGIN\n        CALL {DB}.{SCHEMA}.SP_APPLY_REVIEW(:CHALLENGER_VERSION, :DECISION, :COMMENTS, :reviewer) INTO :result;\n        RETURN result;\n    END;\n    $$\n""").collect()\n\n# 3. Review queue for reviewers and MRM reporting\nsession.sql(f"""\n    CREATE OR REPLACE VIEW {DB}.{SCHEMA}.REVIEW_QUEUE AS\n    SELECT TARGET_ROLE, CHALLENGER_VERSION AS CANDIDATE_VERSION, CHAMPION_VERSION AS BENCHMARK_CHAMPION,\n           EVALUATED_AT, NOTIFIED_AT,\n           CHAMPION_GINI, CHALLENGER_GINI, CHAMPION_BRIER, CHALLENGER_BRIER,\n           DECISION AS GATE_RECOMMENDATION, STATUS, REVIEWED_BY, REVIEWED_AT, REVIEW_COMMENTS,\n           DATEDIFF('day', EVALUATED_AT, COALESCE(REVIEWED_AT, CURRENT_TIMESTAMP())) AS DAYS_IN_REVIEW\n    FROM {DB}.{SCHEMA}.PROMOTION_LOG\n    WHERE STATUS IS NOT NULL\n""").collect()\n\n# 4. Segregation of duties: only MODEL_REVIEWER can approve; the pipeline role gets no grant on these\nsession.sql("CREATE ROLE IF NOT EXISTS MODEL_REVIEWER").collect()\nfor grant in [\n    f"GRANT USAGE ON DATABASE {DB} TO ROLE MODEL_REVIEWER",\n    f"GRANT USAGE ON SCHEMA {DB}.{SCHEMA} TO ROLE MODEL_REVIEWER",\n    f"GRANT USAGE ON WAREHOUSE {WH} TO ROLE MODEL_REVIEWER",\n    f"GRANT SELECT ON VIEW {DB}.{SCHEMA}.REVIEW_QUEUE TO ROLE MODEL_REVIEWER",\n    f"GRANT USAGE ON PROCEDURE {DB}.{SCHEMA}.SP_REVIEW_MODEL(STRING, STRING, STRING) TO ROLE MODEL_REVIEWER",\n    f"GRANT USAGE ON PROCEDURE {DB}.{SCHEMA}.SP_APPLY_REVIEW(STRING, STRING, STRING, STRING) TO ROLE MODEL_REVIEWER",\n]:\n    session.sql(grant).collect()\nfor user in REVIEWER_USERS:\n    session.sql(f"GRANT ROLE MODEL_REVIEWER TO USER {user}").collect()\n\nprint(f"Reviewer notifications: {REVIEW_EMAIL_INTEGRATION} -> {', '.join(REVIEWER_EMAILS)}")\nprint(f"MODEL_REVIEWER granted to: {', '.join(REVIEWER_USERS)}")\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Ch3\u003ECreate the training Task DAG\u003C/h3\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Create Task DAG for the monthly retrain of BOTH models (ends at human review of each candidate)\ndag = [\n    # (task name, predecessor, procedure)\n    ("PD_MODEL_TRAIN_PIPELINE", None, "SP_REFRESH_FEATURES"),\n    ("PD_MODEL_TRAIN_PIPELINE_GENERATE_DATASET", "PD_MODEL_TRAIN_PIPELINE", "SP_GENERATE_DATASET"),\n    ("PD_MODEL_TRAIN_PIPELINE_TRAIN_CANDIDATES", "PD_MODEL_TRAIN_PIPELINE_GENERATE_DATASET", "SP_TRAIN_CANDIDATES"),\n    ("PD_MODEL_TRAIN_PIPELINE_SCORE_CANDIDATES", "PD_MODEL_TRAIN_PIPELINE_TRAIN_CANDIDATES", "SP_SCORE_CANDIDATES"),\n    ("PD_MODEL_TRAIN_PIPELINE_VALIDATION_GATE", "PD_MODEL_TRAIN_PIPELINE_SCORE_CANDIDATES", "SP_VALIDATION_GATE"),\n    ("PD_MODEL_TRAIN_PIPELINE_NOTIFY_REVIEWER", "PD_MODEL_TRAIN_PIPELINE_VALIDATION_GATE", "SP_NOTIFY_REVIEWER"),\n]\n\n# The graph can only be changed while the root is suspended\nsession.sql(f"ALTER TASK IF EXISTS {DB}.{SCHEMA}.PD_MODEL_TRAIN_PIPELINE SUSPEND").collect()\n# Remove tasks from earlier designs (single challenger, automated promotion), children first\nfor old_task in ["TASK_PROMOTE_MODEL", "TASK_SCORE_CHAMPION_CHALLENGER", "TASK_TRAIN_CHALLENGER", "TASK_TRAIN_MODEL",\n                 # Pre-rename names of both DAGs\n                 "TASK_NOTIFY_REVIEWER", "TASK_VALIDATION_GATE", "TASK_SCORE_CANDIDATES", "TASK_TRAIN_CANDIDATES",\n                 "TASK_GENERATE_DATASET", "TASK_FEATURE_REFRESH", "TASK_WEEKLY_SCORING_NOTIFY",\n                 "TASK_WEEKLY_SCENARIO_ECL", "TASK_WEEKLY_BATCH_SCORE", "TASK_WEEKLY_FEATURE_REFRESH",\n                 "TRAIN_PIPELINE_NOTIFY_REVIEWER", "TRAIN_PIPELINE_VALIDATION_GATE", "TRAIN_PIPELINE_SCORE_CANDIDATES",\n                 "TRAIN_PIPELINE_TRAIN_CANDIDATES", "TRAIN_PIPELINE_GENERATE_DATASET", "TRAIN_PIPELINE_FEATURE_REFRESH"]:\n    session.sql(f"DROP TASK IF EXISTS {DB}.{SCHEMA}.{old_task}").collect()\n\nfor task, after, proc in dag:\n    trigger = (f"SCHEDULE = 'USING CRON 0 2 1 * * America/New_York'" if after is None\n               else f"AFTER {DB}.{SCHEMA}.{after}")\n    session.sql(f"""\n        CREATE OR REPLACE TASK {DB}.{SCHEMA}.{task}\n            WAREHOUSE = {WH}\n            {trigger}\n        AS\n            CALL {DB}.{SCHEMA}.{proc}()\n    """).collect()\nprint(f"Task DAG created ({len(dag)} tasks, suspended)")\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003EView the graph:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# View the task graph\ntask_graph = session.sql("""\n    SELECT NAME, STATE, SCHEDULE, COMMENT\n    FROM TABLE(INFORMATION_SCHEMA.TASK_DEPENDENTS(\n        TASK_NAME => 'CREDIT_RISK_ML.PD_MODEL.PD_MODEL_TRAIN_PIPELINE',\n        RECURSIVE => TRUE\n    ))\n    ORDER BY NAME\n""")\ntask_graph.show()\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003EActivate the graph. The children are resumed along with the root:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-sql\"\u003ESELECT SYSTEM$TASK_DEPENDENTS_ENABLE('CREDIT_RISK_ML.PD_MODEL.PD_MODEL_TRAIN_PIPELINE');\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003ETo run one cycle on demand instead of waiting for the 1st of the month, resume the child tasks and execute the root:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-sql\"\u003EALTER TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_TRAIN_PIPELINE_NOTIFY_REVIEWER RESUME;\nALTER TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_TRAIN_PIPELINE_VALIDATION_GATE RESUME;\nALTER TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_TRAIN_PIPELINE_SCORE_CANDIDATES RESUME;\nALTER TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_TRAIN_PIPELINE_TRAIN_CANDIDATES RESUME;\nALTER TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_TRAIN_PIPELINE_GENERATE_DATASET RESUME;\nEXECUTE TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_TRAIN_PIPELINE;\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003EYou can follow the run in Snowsight under \u003Cstrong\u003EMonitoring > Task History\u003C/strong\u003E, or open the root task and select the \u003Cstrong\u003EGraph\u003C/strong\u003E tab to see each step's status:\u003C/p\u003E\n","\u003Cp\u003E\u003Cimg src=\"https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/train-pipeline-task-dag-run.png?v=54b7b1dd\" alt=\"Training pipeline Task DAG run\"\u003E\u003C/p\u003E\n","\u003Ch3\u003EReview the candidates\u003C/h3\u003E\n","\u003Cp\u003EWhen the run finishes, the reviewer receives one email. It has a section for each candidate with its metrics against the production champion, the gate's recommendation, and ready-to-run commands:\u003C/p\u003E\n","\u003Cp\u003E\u003Cimg src=\"https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/model-approval-email-notification.png?v=54b7b1dd\" alt=\"Model approval email\"\u003E\u003C/p\u003E\n","\u003Cp\u003EThe reviewer decides on each candidate separately:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-sql\"\u003EUSE ROLE MODEL_REVIEWER;\nSELECT * FROM CREDIT_RISK_ML.PD_MODEL.REVIEW_QUEUE WHERE STATUS = 'PENDING_REVIEW';\nCALL CREDIT_RISK_ML.PD_MODEL.SP_REVIEW_MODEL('<champion_candidate_version>', 'APPROVE', '<comments>');\nCALL CREDIT_RISK_ML.PD_MODEL.SP_REVIEW_MODEL('<challenger_candidate_version>', 'REJECT', '<comments>');\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003EApproving a candidate moves the \u003Ccode\u003Echampion\u003C/code\u003E or \u003Ccode\u003Echallenger\u003C/code\u003E alias, re-points that role's monitor, and records who approved it, when and why. That record is kept in \u003Ccode\u003EPROMOTION_LOG\u003C/code\u003E and in the model version's comment. A candidate left unreviewed is marked \u003Ccode\u003ESUPERSEDED\u003C/code\u003E by the next cycle's candidate.\u003C/p\u003E\n<!-- ------------------------ -->\n","\u003Ch2\u003EAutomate Weekly Scoring\u003C/h2\u003E\n","\u003Cp\u003EA second, independent Task DAG produces the weekly portfolio PDs and scenario-weighted ECL. It never picks a model version itself. It reads whichever versions hold the \u003Ccode\u003Echampion\u003C/code\u003E and \u003Ccode\u003Echallenger\u003C/code\u003E aliases when it runs, so an approval in the training pipeline takes effect from the next weekly run.\u003C/p\u003E\n","\u003Cp\u003E\u003Cimg src=\"https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/scoring-pipeline.png?v=54b7b1dd\" alt=\"Scoring pipeline\"\u003E\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode\u003EPD_MODEL_SCORING_PIPELINE (root, Mondays 03:00)   SP_REFRESH_FEATURES\n  └─ ..._BATCH_SCORE                                SP_WEEKLY_BATCH_SCORE   -> PD_PORTFOLIO_SCORES (append)\n       └─ ..._SCENARIO_ECL                          SP_WEEKLY_SCENARIO_ECL  -> PD_WEIGHTED_HISTORY, PD_ECL_HISTORY (append)\n                                                                            -> PD_CHAMPION_CHALLENGER_TREND (view)\n            └─ ..._NOTIFY                           SP_WEEKLY_SCORING_NOTIFY -> summary email\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003EHow the weekly results are kept:\u003C/p\u003E\n\u003Cul\u003E\u003Cli\u003EEvery row is appended with \u003Ccode\u003ESCORE_TS\u003C/code\u003E, \u003Ccode\u003EMODEL_VERSION\u003C/code\u003E and \u003Ccode\u003EMODEL_ROLE\u003C/code\u003E, so any week's provision traces back to an approved model.\u003C/li\u003E\u003Cli\u003EThe champion's numbers are the official provision. The challenger's build up comparison evidence for the next review.\u003C/li\u003E\u003Cli\u003E\u003Ccode\u003EPD_CHAMPION_CHALLENGER_TREND\u003C/code\u003E summarizes both models week by week: mean PD, observed default rate, PD correlation and weighted ECL.\u003C/li\u003E\u003C/ul\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Weekly scoring steps as stored procedures (reuse helpers and constants from the cells above)\n# Champion AND challenger score the same portfolio every week; MODEL_ROLE tells them apart.\n# The champion's numbers are the official provision; the challenger's are for comparison only.\nsession.sql(f"""\n    CREATE TABLE IF NOT EXISTS {DB}.{SCHEMA}.PD_WEIGHTED_HISTORY (\n        SCORE_TS TIMESTAMP_NTZ, MODEL_VERSION STRING, TRADE_ID NUMBER,\n        WEIGHTED_PD FLOAT, WEIGHTED_ECL FLOAT\n    )\n""").collect()\nsession.sql(f"""\n    CREATE TABLE IF NOT EXISTS {DB}.{SCHEMA}.PD_ECL_HISTORY (\n        SCORE_TS TIMESTAMP_NTZ, MODEL_VERSION STRING, SCENARIO_ID STRING,\n        SCENARIO_WEIGHT FLOAT, LOANS NUMBER, MEAN_PD FLOAT, ECL FLOAT\n    )\n""").collect()\n# MODEL_ROLE (CHAMPION / CHALLENGER) on every history table, added in place if they already exist.\n# Widen it too: a table first created from champion-only scores infers VARCHAR(8), too short for 'CHALLENGER'.\nfor table in ["PD_PORTFOLIO_SCORES", "PD_WEIGHTED_HISTORY", "PD_ECL_HISTORY"]:\n    session.sql(f"ALTER TABLE IF EXISTS {DB}.{SCHEMA}.{table} ADD COLUMN IF NOT EXISTS MODEL_ROLE STRING").collect()\n    session.sql(f"ALTER TABLE IF EXISTS {DB}.{SCHEMA}.{table} ALTER COLUMN MODEL_ROLE SET DATA TYPE STRING").collect()\n\n\ndef _weekly_models(session):\n    """(role, ModelVersion) pairs for this run: always the champion, plus the challenger if one exists."""\n    from snowflake.ml.registry import Registry\n    m = Registry(session=session, database_name=DB, schema_name=SCHEMA).get_model(MODEL_NAME)\n    champion, challenger = _version_by_alias(m, "champion"), _version_by_alias(m, "challenger")\n    models = [("CHAMPION", champion)]\n    if challenger is not None and challenger.version_name != champion.version_name:\n        models.append(("CHALLENGER", challenger))\n    return models\n\n\ndef sp_weekly_batch_score(session: Session) -> str:\n    """Score the full portfolio with champion and challenger; append with run timestamp, version and role."""\n    import datetime\n    from snowflake.snowpark.functions import col, lit\n    from snowflake.snowpark.types import StringType, TimestampType\n    from snowflake.ml.feature_store import FeatureStore\n    fs = FeatureStore(session, DB, SCHEMA, WH)\n    run_ts = datetime.datetime.utcnow().replace(microsecond=0)   # one timestamp shared by both models\n    features = _scoring_features(session, fs, limit=None)\n\n    scored = []\n    for role, mv in _weekly_models(session):\n        preds = (mv.run(features, function_name="predict_proba")\n                 .with_column("ARCHIVE_DATE", col("ARCHIVE_DATE").cast(TimestampType()))\n                 .with_column("SCORE_TS", lit(run_ts).cast(TimestampType()))\n                 .with_column("MODEL_VERSION", lit(mv.version_name).cast(StringType()))\n                 .with_column("MODEL_ROLE", lit(role).cast(StringType())))   # unbounded, not VARCHAR(len(role))\n        # Full feature + prediction columns are kept so a model monitor can track drift on this table\n        preds.write.mode("append").save_as_table(f"{DB}.{SCHEMA}.PD_PORTFOLIO_SCORES", column_order="name")\n        scored.append(f"{role.lower()} {mv.version_name}")\n    n = session.table(f"{DB}.{SCHEMA}.PD_PORTFOLIO_SCORES").filter(\n        (col("SCORE_TS") == lit(run_ts)) & (col("MODEL_ROLE") == lit("CHAMPION"))).count()\n    return f"Scored {n:,} loans with {' and '.join(scored)} at {run_ts:%Y-%m-%d %H:%M} UTC"\n\n\ndef sp_weekly_scenario_ecl(session: Session) -> str:\n    """Score each loan under every macro scenario, per model, and append probability-weighted PD and ECL."""\n    from snowflake.snowpark.functions import max as max_\n    from snowflake.ml.feature_store import FeatureStore\n    # Registered with caller's rights (needed for the temporary scratch table), so the session has no\n    # database/schema of its own; Feature Store metadata lookups need one.\n    session.use_schema(f"{DB}.{SCHEMA}")\n    fs = FeatureStore(session, DB, SCHEMA, WH)\n    # Tie this ECL run to the portfolio scoring run that just completed\n    run_ts = session.table(f"{DB}.{SCHEMA}.PD_PORTFOLIO_SCORES").select(max_("SCORE_TS")).collect()[0][0]\n\n    scenarios = fs.read_feature_view(fs.get_feature_view("MACRO_SCENARIO_FEATURES", "V1"))\n    scenario_input = (_scoring_features(session, fs, limit=None)\n                      .select("TRADE_ID", *pdm.BEHAVIORAL_COLS).cross_join(scenarios)\n                      .select("TRADE_ID", "SCENARIO_ID", "SCENARIO_WEIGHT", *pdm.FEATURE_COLS))\n\n    results = []\n    for role, mv in _weekly_models(session):\n        mv.run(scenario_input, function_name="predict_proba").write.save_as_table(\n            f"{DB}.{SCHEMA}.PD_SCENARIO_SCRATCH", mode="overwrite", table_type="temporary")\n        params = [run_ts, mv.version_name, role]\n        session.sql(f"""\n            INSERT INTO {DB}.{SCHEMA}.PD_WEIGHTED_HISTORY\n                (SCORE_TS, MODEL_VERSION, MODEL_ROLE, TRADE_ID, WEIGHTED_PD, WEIGHTED_ECL)\n            SELECT ?::TIMESTAMP_NTZ, ?, ?, TRADE_ID,\n                   SUM(SCENARIO_WEIGHT * "output_feature_1"),\n                   SUM(SCENARIO_WEIGHT * "output_feature_1" * BALANCE * {LGD_ASSUMPTION})\n            FROM {DB}.{SCHEMA}.PD_SCENARIO_SCRATCH\n            GROUP BY TRADE_ID\n        """, params=params).collect()\n        session.sql(f"""\n            INSERT INTO {DB}.{SCHEMA}.PD_ECL_HISTORY\n                (SCORE_TS, MODEL_VERSION, MODEL_ROLE, SCENARIO_ID, SCENARIO_WEIGHT, LOANS, MEAN_PD, ECL)\n            SELECT ?::TIMESTAMP_NTZ, ?, ?, SCENARIO_ID, ANY_VALUE(SCENARIO_WEIGHT), COUNT(*),\n                   AVG("output_feature_1"), SUM("output_feature_1" * BALANCE * {LGD_ASSUMPTION})\n            FROM {DB}.{SCHEMA}.PD_SCENARIO_SCRATCH\n            GROUP BY SCENARIO_ID\n        """, params=params).collect()\n        weighted = session.sql(f"""\n            SELECT SUM(SCENARIO_WEIGHT * ECL) FROM {DB}.{SCHEMA}.PD_ECL_HISTORY\n            WHERE SCORE_TS = ?::TIMESTAMP_NTZ AND MODEL_ROLE = ?\n        """, params=[run_ts, role]).collect()[0][0]\n        results.append(f"{role.lower()} {mv.version_name}: {weighted:,.0f}")\n    _ensure_trend_view(session)\n    return f"Scenario-weighted ECL — {'; '.join(results)} (run {run_ts})"\n\n\n# Week-by-week champion vs challenger comparison on the identical portfolio.\n# Created by the pipeline (after the first run writes the history tables it reads).\ndef _ensure_trend_view(session):\n    session.sql(f"""\n    CREATE OR REPLACE VIEW {DB}.{SCHEMA}.PD_CHAMPION_CHALLENGER_TREND AS\n    WITH paired AS (\n        SELECT c.SCORE_TS, c.MODEL_VERSION AS CHAMPION_VERSION, h.MODEL_VERSION AS CHALLENGER_VERSION,\n               c.TRADE_ID, c.DEFAULT_FLAG,\n               c."output_feature_1" AS CHAMPION_PD, h."output_feature_1" AS CHALLENGER_PD\n        FROM {DB}.{SCHEMA}.PD_PORTFOLIO_SCORES c\n        JOIN {DB}.{SCHEMA}.PD_PORTFOLIO_SCORES h\n          ON c.SCORE_TS = h.SCORE_TS AND c.TRADE_ID = h.TRADE_ID\n        WHERE c.MODEL_ROLE = 'CHAMPION' AND h.MODEL_ROLE = 'CHALLENGER'\n    ),\n    ecl AS (\n        SELECT SCORE_TS,\n               SUM(IFF(MODEL_ROLE = 'CHAMPION', SCENARIO_WEIGHT * ECL, 0))   AS CHAMPION_WEIGHTED_ECL,\n               SUM(IFF(MODEL_ROLE = 'CHALLENGER', SCENARIO_WEIGHT * ECL, 0)) AS CHALLENGER_WEIGHTED_ECL\n        FROM {DB}.{SCHEMA}.PD_ECL_HISTORY\n        GROUP BY SCORE_TS\n    )\n    SELECT p.SCORE_TS, p.CHAMPION_VERSION, p.CHALLENGER_VERSION,\n           COUNT(*)                                     AS LOANS,\n           AVG(p.CHAMPION_PD)                           AS CHAMPION_MEAN_PD,\n           AVG(p.CHALLENGER_PD)                         AS CHALLENGER_MEAN_PD,\n           AVG(p.DEFAULT_FLAG)                          AS OBSERVED_DEFAULT_RATE,\n           AVG(ABS(p.CHALLENGER_PD - p.CHAMPION_PD))    AS MEAN_ABS_PD_DIFF,\n           CORR(p.CHAMPION_PD, p.CHALLENGER_PD)         AS PD_CORRELATION,\n           AVG(IFF(ABS(p.CHALLENGER_PD - p.CHAMPION_PD) > 0.05, 1, 0)) AS SHARE_LOANS_DIFF_GT_5PP,\n           ANY_VALUE(e.CHAMPION_WEIGHTED_ECL)           AS CHAMPION_WEIGHTED_ECL,\n           ANY_VALUE(e.CHALLENGER_WEIGHTED_ECL)         AS CHALLENGER_WEIGHTED_ECL\n    FROM paired p\n    LEFT JOIN ecl e ON e.SCORE_TS = p.SCORE_TS\n    GROUP BY p.SCORE_TS, p.CHAMPION_VERSION, p.CHALLENGER_VERSION\n    """).collect()\n\n\ndef sp_weekly_scoring_notify(session: Session) -> str:\n    """Email the latest weekly run: champion provision (official) plus challenger comparison."""\n    from snowflake.snowpark.functions import max as max_\n    run_ts = session.table(f"{DB}.{SCHEMA}.PD_ECL_HISTORY").select(max_("SCORE_TS")).collect()[0][0]\n    scen = session.sql(f"""\n        SELECT SCENARIO_ID, SCENARIO_WEIGHT, MEAN_PD, ECL, MODEL_VERSION\n        FROM {DB}.{SCHEMA}.PD_ECL_HISTORY\n        WHERE SCORE_TS = ?::TIMESTAMP_NTZ AND MODEL_ROLE = 'CHAMPION' ORDER BY MEAN_PD\n    """, params=[run_ts]).collect()\n    runs = session.sql(f"""\n        SELECT SCORE_TS, SUM(SCENARIO_WEIGHT * ECL) AS WEIGHTED_ECL\n        FROM {DB}.{SCHEMA}.PD_ECL_HISTORY\n        WHERE SCORE_TS <= ?::TIMESTAMP_NTZ AND MODEL_ROLE = 'CHAMPION'\n        GROUP BY SCORE_TS ORDER BY SCORE_TS DESC LIMIT 2\n    """, params=[run_ts]).collect()\n    portfolio = session.sql(f"""\n        SELECT COUNT(*) AS LOANS, AVG("output_feature_1") AS MEAN_PD\n        FROM {DB}.{SCHEMA}.PD_PORTFOLIO_SCORES\n        WHERE SCORE_TS = ?::TIMESTAMP_NTZ AND MODEL_ROLE = 'CHAMPION'\n    """, params=[run_ts]).collect()[0]\n    trend = session.sql(f"""\n        SELECT * FROM {DB}.{SCHEMA}.PD_CHAMPION_CHALLENGER_TREND WHERE SCORE_TS = ?::TIMESTAMP_NTZ\n    """, params=[run_ts]).collect()\n\n    weighted = runs[0]["WEIGHTED_ECL"]\n    if len(runs) > 1 and runs[1]["WEIGHTED_ECL"]:\n        prev = runs[1]["WEIGHTED_ECL"]\n        change = f"{(weighted - prev) / prev:+.1%} vs previous run ({runs[1]['SCORE_TS']:%Y-%m-%d})"\n    else:\n        change = "first run — no prior week to compare"\n    version = scen[0]["MODEL_VERSION"]\n\n    cell = "padding:6px 10px;border:1px solid #d1d5db"\n    rows = "".join(\n        f"<tr><td style='{cell}'>{r['SCENARIO_ID']}</td><td style='{cell}'>{r['SCENARIO_WEIGHT']:.0%}</td>"\n        f"<td style='{cell}'>{r['MEAN_PD']:.4f}</td><td style='{cell}'>{r['ECL']:,.0f}</td></tr>" for r in scen)\n\n    if trend:\n        t = trend[0]\n        ecl_delta = t["CHALLENGER_WEIGHTED_ECL"] - t["CHAMPION_WEIGHTED_ECL"]\n        comparison = f"""\n        <h3>Champion vs challenger (same portfolio, comparison only)</h3>\n        <table style="border-collapse:collapse">\n          <tr><th style="{cell}"></th>\n              <th style="{cell};background:#1e3a8a;color:#ffffff">CHAMPION {t['CHAMPION_VERSION']}</th>\n              <th style="{cell};background:#b45309;color:#ffffff">CHALLENGER {t['CHALLENGER_VERSION']}</th></tr>\n          <tr><td style="{cell}"><b>Mean PD</b></td><td style="{cell}">{t['CHAMPION_MEAN_PD']:.4f}</td>\n              <td style="{cell}">{t['CHALLENGER_MEAN_PD']:.4f}</td></tr>\n          <tr><td style="{cell}"><b>Weighted ECL</b></td><td style="{cell}">{t['CHAMPION_WEIGHTED_ECL']:,.0f}</td>\n              <td style="{cell}">{t['CHALLENGER_WEIGHTED_ECL']:,.0f} ({ecl_delta:+,.0f})</td></tr>\n        </table>\n        <p>Observed default rate {t['OBSERVED_DEFAULT_RATE']:.4f} · mean |PD difference|\n           {t['MEAN_ABS_PD_DIFF']:.4f} · PD correlation {t['PD_CORRELATION']:.3f} ·\n           {t['SHARE_LOANS_DIFF_GT_5PP']:.1%} of loans differ by more than 5 pp.\n           History: {DB}.{SCHEMA}.PD_CHAMPION_CHALLENGER_TREND.</p>"""\n    else:\n        comparison = "<p>No challenger is currently registered, so only the champion was scored this week.</p>"\n\n    subject = f"[PD model] Weekly scoring complete — {run_ts:%Y-%m-%d}, champion weighted ECL {weighted:,.0f}"\n    body = f"""\n        <p>The weekly PD batch scoring run has completed.</p>\n        <ul>\n          <li>Run: {run_ts:%Y-%m-%d %H:%M} UTC</li>\n          <li>Champion model (official provision): {MODEL_NAME} {version}</li>\n          <li>Loans scored: {portfolio['LOANS']:,} — mean PD {portfolio['MEAN_PD']:.4f}</li>\n          <li>Probability-weighted ECL: <b>{weighted:,.0f}</b> ({change})</li>\n        </ul>\n        <table style="border-collapse:collapse">\n          <tr><th style="{cell}">Scenario</th><th style="{cell}">Weight</th>\n              <th style="{cell}">Mean PD</th><th style="{cell}">ECL</th></tr>{rows}\n        </table>\n        {comparison}\n        <p>ECL uses LGD = {LGD_ASSUMPTION} and BALANCE as EAD (illustrative).\n           Results: {DB}.{SCHEMA}.PD_PORTFOLIO_SCORES, PD_WEIGHTED_HISTORY, PD_ECL_HISTORY (filter on MODEL_ROLE).</p>\n    """\n    session.sql("CALL SYSTEM$SEND_EMAIL(?, ?, ?, ?, 'text/html')",\n                params=[REVIEW_EMAIL_INTEGRATION, ",".join(SCORING_NOTIFY_EMAILS), subject, body]).collect()\n    return f"Scoring summary for {run_ts} sent to {', '.join(SCORING_NOTIFY_EMAILS)}"\n\n\n# SCORING_NOTIFY_EMAILS is configured in the Setup cell (section 1)\n\nfor fn in [sp_weekly_batch_score, sp_weekly_scenario_ecl, sp_weekly_scoring_notify]:\n    session.sproc.register(\n        fn, name=f"{DB}.{SCHEMA}.{fn.__name__.upper()}",\n        packages=SP_PACKAGES, imports=[MODEL_MODULE], is_permanent=True, replace=True,\n        stage_location=f"@{DB}.{SCHEMA}.ML_ARTIFACTS",\n        # Owner's-rights procedures cannot create temporary tables; the scenario step uses one\n        execute_as="caller" if fn is sp_weekly_scenario_ecl else "owner",\n    )\n    print(f"Registered procedure {fn.__name__.upper()}")\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003ECreate the weekly DAG:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Weekly scoring Task DAG — Mondays 03:00 New York, independent of the monthly retraining DAG\nweekly_dag = [\n    # (task name, predecessor, procedure)\n    ("PD_MODEL_SCORING_PIPELINE", None, "SP_REFRESH_FEATURES"),\n    ("PD_MODEL_SCORING_PIPELINE_BATCH_SCORE", "PD_MODEL_SCORING_PIPELINE", "SP_WEEKLY_BATCH_SCORE"),\n    ("PD_MODEL_SCORING_PIPELINE_SCENARIO_ECL", "PD_MODEL_SCORING_PIPELINE_BATCH_SCORE", "SP_WEEKLY_SCENARIO_ECL"),\n    ("PD_MODEL_SCORING_PIPELINE_NOTIFY", "PD_MODEL_SCORING_PIPELINE_SCENARIO_ECL", "SP_WEEKLY_SCORING_NOTIFY"),\n]\n\n# The graph can only be changed while the root is suspended\nsession.sql(f"ALTER TASK IF EXISTS {DB}.{SCHEMA}.PD_MODEL_SCORING_PIPELINE SUSPEND").collect()\n# Remove the pre-rename names of this DAG, children first\nfor old_task in ["SCORING_PIPELINE_NOTIFY", "SCORING_PIPELINE_SCENARIO_ECL", "SCORING_PIPELINE_BATCH_SCORE",\n                 "SCORING_PIPELINE_FEATURE_REFRESH"]:\n    session.sql(f"DROP TASK IF EXISTS {DB}.{SCHEMA}.{old_task}").collect()\n\nfor task, after, proc in weekly_dag:\n    trigger = ("SCHEDULE = 'USING CRON 0 3 * * 1 America/New_York'" if after is None\n               else f"AFTER {DB}.{SCHEMA}.{after}")\n    session.sql(f"""\n        CREATE OR REPLACE TASK {DB}.{SCHEMA}.{task}\n            WAREHOUSE = {WH}\n            {trigger}\n        AS\n            CALL {DB}.{SCHEMA}.{proc}()\n    """).collect()\nprint(f"Weekly scoring DAG created ({len(weekly_dag)} tasks, suspended)")\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003EActivate it:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-sql\"\u003ESELECT SYSTEM$TASK_DEPENDENTS_ENABLE('CREDIT_RISK_ML.PD_MODEL.PD_MODEL_SCORING_PIPELINE');\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003ETo run one cycle on demand:\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-sql\"\u003EALTER TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_SCORING_PIPELINE_NOTIFY RESUME;\nALTER TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_SCORING_PIPELINE_SCENARIO_ECL RESUME;\nALTER TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_SCORING_PIPELINE_BATCH_SCORE RESUME;\nEXECUTE TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_SCORING_PIPELINE;\n\u003C/code\u003E\u003C/pre\u003E\n","\u003Cp\u003EA completed weekly run looks like this in the task graph:\u003C/p\u003E\n","\u003Cp\u003E\u003Cimg src=\"https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/score-pipeline-task-dag-run.png?v=54b7b1dd\" alt=\"Scoring pipeline Task DAG run\"\u003E\u003C/p\u003E\n","\u003Cp\u003EEach run ends with a summary email. It shows the champion's weighted ECL and change since the previous week, the per-scenario breakdown, and the challenger comparison:\u003C/p\u003E\n","\u003Cp\u003E\u003Cimg src=\"https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/scoring-completion-email-notification.png?v=54b7b1dd\" alt=\"Weekly scoring email\"\u003E\u003C/p\u003E\n<!-- ------------------------ -->\n","\u003Ch2\u003EClean Up\u003C/h2\u003E\n","\u003Cp\u003EThe final cell removes everything the notebook created:\u003C/p\u003E\n\u003Cul\u003E\u003Cli\u003Esuspends both Task DAGs;\u003C/li\u003E\u003Cli\u003Edrops the account-level notification integration and the \u003Ccode\u003EMODEL_REVIEWER\u003C/code\u003E role;\u003C/li\u003E\u003Cli\u003Edrops the \u003Ccode\u003ECREDIT_RISK_ML\u003C/code\u003E database and the \u003Ccode\u003ECREDIT_RISK_WH\u003C/code\u003E warehouse.\u003C/li\u003E\u003C/ul\u003E\n","\u003Cp\u003E\u003Cstrong\u003EThis is irreversible.\u003C/strong\u003E It deletes the model versions, Feature Store, Datasets, experiments, score history and the \u003Ccode\u003EPROMOTION_LOG\u003C/code\u003E audit trail. Set \u003Ccode\u003ECONFIRM_CLEANUP = True\u003C/code\u003E only when you mean it. The Marketplace databases are left untouched.\u003C/p\u003E\n\u003Cpre\u003E\u003Ccode class=\"language-python\"\u003E# Tear down all resources created by this notebook — set CONFIRM_CLEANUP = True to run\nCONFIRM_CLEANUP = False\n\nDB, SCHEMA, WH = "CREDIT_RISK_ML", "PD_MODEL", "CREDIT_RISK_WH"\n\nif not CONFIRM_CLEANUP:\n    print("Cleanup skipped — set CONFIRM_CLEANUP = True to permanently delete the resources listed above.")\nelse:\n    def run(sql):\n        try:\n            session.sql(sql).collect()\n            print(f"OK    {sql}")\n        except Exception as e:  # keep going so one missing object doesn't block the rest\n            print(f"SKIP  {sql}  ({str(e).splitlines()[0]})")\n\n    # 1. Suspend both task graphs (root tasks first) so no run starts during cleanup\n    for root in ["PD_MODEL_TRAIN_PIPELINE", "PD_MODEL_SCORING_PIPELINE"]:\n        run(f"ALTER TASK IF EXISTS {DB}.{SCHEMA}.{root} SUSPEND")\n\n    # 2. Account-level objects, which dropping the database does not remove\n    run("DROP NOTIFICATION INTEGRATION IF EXISTS PD_MODEL_REVIEW_EMAIL")\n    run("DROP ROLE IF EXISTS MODEL_REVIEWER")\n\n    # 3. The database: registry models and monitors, Feature Store, Datasets, experiments,\n    #    tasks, procedures, tables, views and stages all live inside it\n    run(f"DROP DATABASE IF EXISTS {DB}")\n\n    # 4. The warehouse (last, since the statements above run on it)\n    run(f"DROP WAREHOUSE IF EXISTS {WH}")\n\n    print("\\nCleanup complete.")\n\u003C/code\u003E\u003C/pre\u003E\n<!-- ------------------------ -->\n","\u003Ch2\u003EConclusion and Resources\u003C/h2\u003E\n","\u003Cp\u003EYou migrated a SAS \u003Ccode\u003EPROC GRADBOOST\u003C/code\u003E PD model to XGBoost on Snowflake ML and ran its full lifecycle in one governed platform. Data never left Snowflake, every model is traceable to the exact data that trained it, and promotion requires an audited human decision.\u003C/p\u003E\n\u003Ctable\u003E\u003Cthead\u003E\u003Ctr\u003E\u003Cth colspan=\"1\" rowspan=\"1\"\u003ESnowflake ML capability\u003C/th\u003E\u003Cth colspan=\"1\" rowspan=\"1\"\u003EWhat it replaces in a SAS stack\u003C/th\u003E\u003C/tr\u003E\u003C/thead\u003E\u003Ctbody\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003EFeature Store (behavioral, macro, scenario Feature Views)\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003ESAS macros generating feature datasets\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003EVersioned Datasets\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003EManual dataset freezes\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003EExperiment Tracking\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003EExcel/Word model development trackers\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003EModel Registry with \u003Ccode\u003Echampion\u003C/code\u003E / \u003Ccode\u003Echallenger\u003C/code\u003E aliases\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003EVersion-named \u003Ccode\u003E.sas7bdat\u003C/code\u003E files and spreadsheet inventories\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003ERegistry SHAP \u003Ccode\u003Eexplain\u003C/code\u003E and ML Lineage\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003EExternal explainability tooling and provenance documents\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003E\u003Ccode\u003Erun()\u003C/code\u003E / \u003Ccode\u003Erun_batch()\u003C/code\u003E batch scoring and scenario-weighted ECL\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003ESAS batch scoring jobs and spreadsheet IFRS 9 overlays\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003EModel Version Monitors\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003EQuarterly manual monitoring reports\u003C/td\u003E\u003C/tr\u003E\u003Ctr\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003ETask DAGs with \u003Ccode\u003EMODEL_REVIEWER\u003C/code\u003E approval\u003C/td\u003E\u003Ctd colspan=\"1\" rowspan=\"1\"\u003ECron + shell scripts and email sign-off chains\u003C/td\u003E\u003C/tr\u003E\u003C/tbody\u003E\u003C/table\u003E\n","\u003Ch3\u003EWhat You Learned\u003C/h3\u003E\n\u003Cul\u003E\u003Cli\u003EMapping SAS model code and hyperparameters to XGBoost with CoCo\u003C/li\u003E\u003Cli\u003EBuilding point-in-time correct features and immutable training Datasets\u003C/li\u003E\u003Cli\u003ETracking experiments and governing models with registry aliases, SHAP and lineage\u003C/li\u003E\u003Cli\u003EScoring offline champion/challenger and computing scenario-weighted ECL\u003C/li\u003E\u003Cli\u003EMonitoring drift and performance per model version\u003C/li\u003E\u003Cli\u003EOrchestrating retraining and scoring with a role-gated, auditable approval step\u003C/li\u003E\u003C/ul\u003E"],"description":"","title":"From SAS to Snowflake: Productionizing Credit Risk Models on a Modern ML Platform",":items":{},":type":"snowflake-site/components/contentfragment",":itemsOrder":[],"elements":{"quickstartArticleBody":{"dataType":"string","value":"\u003C!-- ------------------------ --\u003E\n## Overview\n\nBanks have built Probability of Default (PD), Loss Given Default (LGD) and Exposure at Default (EAD) models in SAS for decades. The models work, but the infrastructure around them (data extracts to SAS Grid, macro-generated datasets, cron-driven batch jobs, spreadsheet model inventories and email sign-offs) makes reproducibility, lineage and monitoring expensive to prove to regulators.\n\nIn this Quickstart you migrate a SAS `PROC GRADBOOST` PD model to **XGBoost on Snowflake ML** and run the whole model lifecycle in one governed platform: features, training data, experiments, registry, explainability, batch scoring, scenario-weighted Expected Credit Loss (ECL) scoring, monitoring, human-in-the-loop promotion and scheduled pipelines.\n\n![Solution architecture](https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/solution-architecture.png?v=54b7b1dd)\n\nThe complete notebook is in the Quickstart assets on GitHub: **[credit_risk_probability_of_default_model.ipynb](https://github.com/Snowflake-Labs/sfquickstarts/blob/master/site/sfguides/src/credit-risk-pd-model-snowflake-ml/assets/credit_risk_probability_of_default_model.ipynb)**. Each step below shows the notebook code with an explanation, so you can follow along cell by cell.\n\n### What You'll Learn\n- How to convert a SAS `PROC GRADBOOST` PD model to Python with CoCo, and map its hyperparameters to `XGBClassifier`\n- How to build Feature Views (backed by Dynamic Tables) for behavioral, macroeconomic and forward-looking scenario features\n- How to generate immutable, point-in-time training Datasets\n- How to log runs with Experiment Tracking and register models on Snowflake Model Registry\n- How to explain predictions with the Model Registry's built-in SHAP `explain` function and trace ML Lineage\n- How to batch score with `run()` and compute probability-weighted ECL across macro scenarios\n- How to attach Model Monitors for drift and performance\n- How to automate monthly retraining and weekly scoring with Task DAGs, including an email-driven, role-gated model approval step\n\n### What You'll Build\n- A Feature Store with three Feature Views: `LOAN_BEHAVIORAL_FEATURES`, `LOAN_MACRO_FEATURES` and `MACRO_SCENARIO_FEATURES`\n- A versioned training Dataset, `PD_TRAINING_DS`\n- A registered model, `PD_XGB_MODEL`, with champion and challenger versions, SHAP explainability and lineage\n- Model Monitors for the champion and challenger models\n- A monthly **training pipeline** Task DAG that retrains both models, scores them against production, gates them, and emails a reviewer for approval\n- A weekly **scoring pipeline** Task DAG that scores the portfolio with the approved models, computes scenario-weighted ECL, keeps history, and emails a summary\n\n### What You'll Need\n- A Snowflake account with access to **Snowflake Notebooks** and a role that can create databases, warehouses, roles and notification integrations (the notebook was built with `ACCOUNTADMIN`)\n- A verified email address on your Snowflake user, for the review and scoring notification emails\n- Two free Snowflake Marketplace listings (installed in the next step):\n  - [Equifax Analytic Dataset (Demo/Sample)](https://app.snowflake.com/marketplace/listing/GZT0ZOSU8UC)\n  - [FRED - Federal Reserve Economic Data](https://app.snowflake.com/marketplace/listing/GZTYZ40XYNE)\n- Familiarity with Python and SQL. Credit risk knowledge helps but is not required.\n\n\u003C!-- ------------------------ --\u003E\n## Get Data and Notebook\n\n### Install the Marketplace data\n1. In Snowsight, open **Data Products \u003E Marketplace**.\n2. Search for **Equifax Analytic Dataset (Demo/Sample)** and select **Get**. Keep the database name `ANALYTIC_DATASET_DEMOSAMPLE`.\n3. Search for **FRED - Federal Reserve Economic Data** (provided by Fruit Stand) and select **Get**. Keep the database name `FRED_FEDERAL_RESERVE_ECONOMIC_DATA_SERIES_OBSERVATIONS`.\n\nThe notebook reads `ANALYTIC_DATASET_DEMOSAMPLE.DATASET.ANALYTICSSAMPLE_LATEST` and `FRED_FEDERAL_RESERVE_ECONOMIC_DATA_SERIES_OBSERVATIONS.FRED.FRED_SERIES_OBSERVATIONS_VW`. If you choose different database names, update those references.\n\n### Import the notebook\n1. Download `credit_risk_probability_of_default_model.ipynb` from the [Quickstart assets](https://github.com/Snowflake-Labs/sfquickstarts/blob/master/site/sfguides/src/credit-risk-pd-model-snowflake-ml/assets/credit_risk_probability_of_default_model.ipynb).\n2. In Snowsight, go to **Projects \u003E Notebooks**, choose **Import .ipynb file**, and select the downloaded notebook.\n3. Make sure these packages are available to the notebook: `snowflake-ml-python`, `xgboost`, `scikit-learn`, `scipy`, `matplotlib`, `altair`.\n4. Upload [`python_pd_model.py`](https://github.com/Snowflake-Labs/sfquickstarts/blob/master/site/sfguides/src/credit-risk-pd-model-snowflake-ml/assets/python_pd_model.py) from the Quickstart assets to the same notebook, next to the `.ipynb` file. The notebook and the training pipeline import the model from it (see the next step).\n\nThe Quickstart assets also contain [`sas_pd_model.sas`](https://github.com/Snowflake-Labs/sfquickstarts/blob/master/site/sfguides/src/credit-risk-pd-model-snowflake-ml/assets/sas_pd_model.sas) (the original SAS model) and [`python_pd_model.py`](https://github.com/Snowflake-Labs/sfquickstarts/blob/master/site/sfguides/src/credit-risk-pd-model-snowflake-ml/assets/python_pd_model.py) (the Python conversion produced with CoCo). The next step shows how that conversion is done.\n\n\u003C!-- ------------------------ --\u003E\n## Convert SAS with CoCo\n\nThe starting point is a SAS PD model, `sas_pd_model.sas`. It follows a pattern found in many banks:\n1. `PROC SQL` joins Equifax loan data to FRED macro indicators, with `%let` macro variables for libraries and the observation window.\n2. A `DATA` step imputes missing values with 0.\n3. `PROC SURVEYSELECT` makes an 80/20 split.\n4. `PROC GRADBOOST` trains a gradient boosting model.\n5. `PROC ASTORE` scores the test set from the saved analytic store.\n6. `PROC LOGISTIC` computes AUC, and the metrics go into an Excel tracker by hand.\n\nThe core of the model is the `PROC GRADBOOST` step. Because it is gradient boosting, like XGBoost, its options map closely onto `XGBClassifier`:\n\n```sas\nproc gradboost data=&OUT_LIB..PD_TRAIN\n    ntrees=200\n    learningrate=0.1\n    maxdepth=6\n    vars_to_try=5\n    minleafsize=20\n    seed=42;\n\n    target DEFAULT_FLAG / level=nominal;\n\n    input VANTAGE_SCORE_3 NUMBER_OF_ACCOUNTS NUMBER_ACCOUNTS_PAST_DUE\n          NUMBER_ACCOUNTS_MAJOR_DEROGATORY HIGH_UTIL_ACCTS\n          AGE_OLDEST_ACCOUNT TOTAL_PAST_DUE_AMOUNT BALANCE\n          / level=interval;\n    input UNEMPLOYMENT_RATE FED_FUNDS_RATE CPI_INDEX / level=interval;\n\n    savestate rstore=&OUT_LIB..PD_GRADBOOST_STORE;\n    ods output VariableImportance = &OUT_LIB..PD_IMPORTANCE;\n    ods output FitStatistics      = &OUT_LIB..PD_FITSTATS;\nrun;\n```\n\n### Prompt CoCo\n**[CoCo](https://www.snowflake.com/en/product/snowflake-coco/)** is Snowflake's AI coding assistant. It's available in Snowsight and as a desktop app. Open the repository folder in CoCo (or attach `sas_pd_model.sas` to the chat) and ask for the conversion in plain language. Name the target frameworks and the Snowflake features you want each SAS construct to become:\n\n```text\nConvert @sas_pd_model.sas into a Python script, python_pd_model.py, that runs on Snowflake.\n\n- Replace PROC GRADBOOST with xgboost.XGBClassifier. Map every GRADBOOST option to its closest\n  XGBoost hyperparameter and explain each mapping in a comment.\n- Replace the %let macros and PROC SQL join with Snowflake Feature Store Feature Views\n  (one for the Equifax behavioral features, one for the FRED macro features), keyed on TRADE_ID\n  with ARCHIVE_DATE as the timestamp.\n- Replace PROC SURVEYSELECT and the manual dataset freeze with a versioned Dataset from\n  fs.generate_dataset, then a stratified 80/20 split with seed 42.\n- Keep the DATA step behavior: impute missing numeric values with 0.\n- Replace PROC LOGISTIC ROC with AUC, Gini and KS computed in Python.\n- Replace the saved analytic store with Snowflake Model Registry registration and a \"champion\" alias.\n- Replace PROC ASTORE with model.run() batch scoring on a warehouse.\nUse the Snowpark session from get_active_session(); data is in CREDIT_RISK_ML.PD_MODEL.\n```\n\nCoCo reads the SAS file and returns [`python_pd_model.py`](https://github.com/Snowflake-Labs/sfquickstarts/blob/master/site/sfguides/src/credit-risk-pd-model-snowflake-ml/assets/python_pd_model.py) (in the Quickstart assets). Ask it to structure the result as a reusable module, not just a one-off script: the feature lists, the SAS-mapped hyperparameters, and a small function for each SAS step. The notebook and the training pipeline then import the same code, so the converted model is the one that is actually trained in production:\n\n```python\n# XGBoost hyperparameters mapped from SAS PROC GRADBOOST:\n#   ntrees=200        →  n_estimators=200\n#   learningrate=0.1  →  learning_rate=0.1\n#   maxdepth=6        →  max_depth=6\n#   vars_to_try=5     →  colsample_bynode = 5/11 (features tried per split)\n#   minleafsize=20    →  min_child_weight (approx)\n#   seed=42           →  random_state=42\nSAS_MAPPED_PARAMS = {\n    \"n_estimators\": 200,\n    \"max_depth\": 6,\n    \"learning_rate\": 0.1,\n    \"colsample_bynode\": 5 / len(FEATURE_COLS),\n    \"min_child_weight\": 20,\n    \"objective\": \"binary:logistic\",\n    \"eval_metric\": \"auc\",\n    \"random_state\": 42,\n}\n\n\ndef prepare_features(pdf):\n    \"\"\"Feature matrix and label from a pandas frame. Replaces the SAS DATA step (missing → 0).\"\"\"\n    return pdf[FEATURE_COLS].fillna(0), pdf[TARGET_COL]\n\n\ndef split_train_test(X, y):\n    \"\"\"Stratified 80/20 split with seed 42. Replaces PROC SURVEYSELECT samprate=0.8 seed=42.\"\"\"\n    return train_test_split(X, y, test_size=0.2, random_state=42, stratify=y)\n\n\ndef train_pd_model(X, y, params=None):\n    \"\"\"Fit the PD model. Replaces PROC GRADBOOST.\"\"\"\n    return XGBClassifier(**(params or SAS_MAPPED_PARAMS)).fit(X, y)\n```\n\nThe module also defines `BEHAVIORAL_COLS`, `MACRO_COLS`, `FEATURE_COLS`, `TARGET_COL`, `CHALLENGER_PARAMS` and `pd_metrics()` (AUC, Gini, KS and Brier, replacing `PROC LOGISTIC`). Running the file directly (`main(session)`) still performs the converted SAS workflow end to end.\n\nTo get this shape, add a line like this to the prompt:\n\n```text\nStructure python_pd_model.py as an importable module: constants for the feature lists and the\nSAS-mapped hyperparameters, one function per SAS step (prepare, split, train, evaluate), and the\nend-to-end workflow in main(session) under if __name__ == \"__main__\".\n```\n\nThe notebook in this guide imports this module and builds the rest of the lifecycle around it: Experiment Tracking, a challenger model, SHAP explainability, monitoring and orchestration. The remaining steps walk through it cell by cell.\n\n\u003C!-- ------------------------ --\u003E\n## Set Up the Environment\n\nEdit the configuration block at the top of the first cell before running anything:\n\n- `REVIEWER_USERS`: Snowflake users who will receive the `MODEL_REVIEWER` role and approve or reject models.\n- `REVIEWER_EMAILS`: verified email addresses of users in your account. They become the allowed recipients of the email notification integration.\n- `SCORING_NOTIFY_EMAILS`: recipients of the weekly scoring summary (must also be in `REVIEWER_EMAILS`).\n- `REVIEW_EMAIL_INTEGRATION`: name of the email notification integration the notebook creates.\n\nThe cell then creates the `CREDIT_RISK_ML` database, the `PD_MODEL` schema and an extra-small `CREDIT_RISK_WH` warehouse.\n\n```python\n# ---------------------------------------------------------------------------\n# Configuration — edit these before running the notebook\n# ---------------------------------------------------------------------------\n# Snowflake users who review and approve models (granted the MODEL_REVIEWER role in section 10)\nREVIEWER_USERS = [\"\u003CYOUR_SNOWFLAKE_USERNAME\u003E\"]\n# Email addresses for model review requests (section 10). Each must be the verified email\n# of a user in this Snowflake account; they become the email integration's ALLOWED_RECIPIENTS.\nREVIEWER_EMAILS = [\"\u003Cyour.email@example.com\u003E\"]\n# Recipients of the weekly scoring summary (section 10b) — must also appear in REVIEWER_EMAILS\nSCORING_NOTIFY_EMAILS = REVIEWER_EMAILS\n# Email notification integration used for both emails (created in section 10)\nREVIEW_EMAIL_INTEGRATION = \"PD_MODEL_REVIEW_EMAIL\"\n# ---------------------------------------------------------------------------\n\nfrom snowflake.snowpark.context import get_active_session\n\nsession = get_active_session()\n\nsession.sql(\"CREATE DATABASE IF NOT EXISTS CREDIT_RISK_ML\").collect()\nsession.sql(\"CREATE SCHEMA IF NOT EXISTS CREDIT_RISK_ML.PD_MODEL\").collect()\nsession.sql(\"\"\"\n    CREATE WAREHOUSE IF NOT EXISTS CREDIT_RISK_WH\n    WAREHOUSE_SIZE = 'XSMALL' AUTO_SUSPEND = 120 AUTO_RESUME = TRUE\n\"\"\").collect()\nsession.sql(\"USE SCHEMA CREDIT_RISK_ML.PD_MODEL\").collect()\nsession.sql(\"USE WAREHOUSE CREDIT_RISK_WH\").collect()\nprint(\"Setup complete: CREDIT_RISK_ML.PD_MODEL\")\n```\n\n\u003C!-- ------------------------ --\u003E\n## Explore the Data\n\nStart by profiling the two Marketplace sources. The Equifax demo sample contains 500K loan-level records across product categories such as auto, credit card, mortgage, student and personal loans.\n\n```python\nexplore_equifax = session.sql(\"\"\"\n    SELECT COUNT(*) AS total_rows,\n           COUNT(DISTINCT CONSUMER_ID) AS unique_consumers,\n           COUNT(DISTINCT TRADE_ID) AS unique_trades,\n           MIN(ARCHIVE_DATE) AS min_date,\n           MAX(ARCHIVE_DATE) AS max_date,\n           COUNT(DISTINCT PRODUCT_CATEGORY) AS product_categories\n    FROM ANALYTIC_DATASET_DEMOSAMPLE.DATASET.ANALYTICSSAMPLE_LATEST\n\"\"\")\nexplore_equifax.show()\n```\n\nThe target is **default**, defined as 90+ days past due (`STATUS_CATEGORY \u003E= 4`). This is the standard PD definition and it mirrors the SAS model. Check the class balance:\n\n```python\n# Target variable distribution\n# STATUS_CATEGORY: 1=Current, 2=30DPD, 3=60DPD, 4=90DPD, 5=120+DPD\n# Default flag: STATUS_CATEGORY \u003E= 4 (90+ days past due)\ntarget_dist = session.sql(\"\"\"\n    SELECT CASE WHEN STATUS_CATEGORY \u003E= 4 THEN 1 ELSE 0 END AS DEFAULT_FLAG,\n           COUNT(*) AS count,\n           ROUND(COUNT(*) * 100.0 / SUM(COUNT(*)) OVER(), 1) AS pct\n    FROM ANALYTIC_DATASET_DEMOSAMPLE.DATASET.ANALYTICSSAMPLE_LATEST\n    GROUP BY DEFAULT_FLAG ORDER BY DEFAULT_FLAG\n\"\"\")\ntarget_dist.show()\n```\n\nFrom FRED, the model uses three monthly macro series: unemployment rate (`UNRATE`), the federal funds rate (`FEDFUNDS`) and CPI (`CPIAUCSL`).\n\n```python\n# FRED macro series available for this model\nexplore_fred = session.sql(\"\"\"\n    SELECT SERIES_ID, SERIES_NAME, FREQUENCY, UNITS,\n           MIN(DATE) AS min_date, MAX(DATE) AS max_date, COUNT(*) AS obs_count\n    FROM FRED_FEDERAL_RESERVE_ECONOMIC_DATA_SERIES_OBSERVATIONS.FRED.FRED_SERIES_OBSERVATIONS_VW\n    WHERE SERIES_ID IN ('UNRATE', 'FEDFUNDS', 'CPIAUCSL')\n    GROUP BY SERIES_ID, SERIES_NAME, FREQUENCY, UNITS\n    ORDER BY SERIES_ID\n\"\"\")\nexplore_fred.show()\n```\n\n\u003C!-- ------------------------ --\u003E\n## Prepare the Data\n\nManaged Feature Views are backed by Dynamic Tables, which can only read a shared Marketplace view if the provider has enabled change tracking on its underlying tables. Consumers can't enable change tracking on shared objects. To avoid depending on the provider, first materialize the data into the project schema with change tracking enabled.\n\n`LOAN_DATA` keeps the loan key, observation date, the derived `DEFAULT_FLAG` and the behavioral attributes. `CHANGE_TRACKING = TRUE` keeps the downstream Dynamic Tables refreshable after the table is re-created.\n\n```python\n# Materialize Equifax loan data\n# CHANGE_TRACKING keeps the Feature View dynamic tables refreshable after this table is re-created\nsession.sql(\"\"\"\n    CREATE OR REPLACE TABLE CREDIT_RISK_ML.PD_MODEL.LOAN_DATA CHANGE_TRACKING = TRUE AS\n    SELECT \n        TRADE_ID, CONSUMER_ID, ARCHIVE_DATE,\n        CASE WHEN STATUS_CATEGORY \u003E= 4 THEN 1 ELSE 0 END AS DEFAULT_FLAG,\n        VANTAGE_SCORE_3, NUMBER_OF_ACCOUNTS, NUMBER_ACCOUNTS_PAST_DUE,\n        NUMBER_ACCOUNTS_MAJOR_DEROGATORY,\n        BANKCARD_ACCTS_OVER_75_PERCENT_UTILIZATION AS HIGH_UTIL_ACCTS,\n        AGE_OLDEST_ACCOUNT, TOTAL_PAST_DUE_AMOUNT, BALANCE,\n        PRODUCT_CATEGORY, STATUS_CATEGORY\n    FROM ANALYTIC_DATASET_DEMOSAMPLE.DATASET.ANALYTICSSAMPLE_LATEST\n\"\"\").collect()\nprint(f\"LOAN_DATA: {session.table('CREDIT_RISK_ML.PD_MODEL.LOAN_DATA').count():,} rows\")\n```\n\n`MACRO_INDICATORS` pivots the three FRED series from long to wide, with one row per month:\n\n```python\n# Pivot FRED monthly series into a wide table\nsession.sql(\"\"\"\n    CREATE OR REPLACE TABLE CREDIT_RISK_ML.PD_MODEL.MACRO_INDICATORS CHANGE_TRACKING = TRUE AS\n    SELECT DATE AS OBSERVATION_MONTH,\n           MAX(CASE WHEN SERIES_ID = 'UNRATE' THEN VALUE END) AS UNEMPLOYMENT_RATE,\n           MAX(CASE WHEN SERIES_ID = 'FEDFUNDS' THEN VALUE END) AS FED_FUNDS_RATE,\n           MAX(CASE WHEN SERIES_ID = 'CPIAUCSL' THEN VALUE END) AS CPI_INDEX\n    FROM FRED_FEDERAL_RESERVE_ECONOMIC_DATA_SERIES_OBSERVATIONS.FRED.FRED_SERIES_OBSERVATIONS_VW\n    WHERE SERIES_ID IN ('UNRATE', 'FEDFUNDS', 'CPIAUCSL') AND DATE \u003E= '2022-01-01'\n    GROUP BY DATE HAVING UNEMPLOYMENT_RATE IS NOT NULL\n    ORDER BY DATE\n\"\"\").collect()\nprint(f\"MACRO_INDICATORS: {session.table('CREDIT_RISK_ML.PD_MODEL.MACRO_INDICATORS').count()} rows\")\nsession.table(\"CREDIT_RISK_ML.PD_MODEL.MACRO_INDICATORS\").show(5)\n```\n\n### Feature exploration\nBefore engineering features, run a few standard credit risk checks. Each aggregation runs in Snowflake, and only small summaries come back to pandas for plotting.\n\nClass balance and default rate by product:\n\n```python\n# Class balance + default rate by product category\nimport matplotlib.pyplot as plt\n\nclass_pdf = session.sql(\"\"\"\n    SELECT DEFAULT_FLAG, COUNT(*) AS N\n    FROM CREDIT_RISK_ML.PD_MODEL.LOAN_DATA GROUP BY DEFAULT_FLAG ORDER BY DEFAULT_FLAG\n\"\"\").to_pandas()\n\nproduct_pdf = session.sql(\"\"\"\n    SELECT PRODUCT_CATEGORY, AVG(DEFAULT_FLAG) AS DEFAULT_RATE, COUNT(*) AS N\n    FROM CREDIT_RISK_ML.PD_MODEL.LOAN_DATA\n    GROUP BY PRODUCT_CATEGORY ORDER BY DEFAULT_RATE DESC\n\"\"\").to_pandas()\n\nfig, axes = plt.subplots(1, 2, figsize=(14, 5))\naxes[0].bar([\"Non-default (0)\", \"Default (1)\"], class_pdf[\"N\"], color=[\"#4C9BE8\", \"#E8684C\"])\naxes[0].set_title(\"Class Balance\")\naxes[0].set_ylabel(\"Loans\")\n\naxes[1].bar(product_pdf[\"PRODUCT_CATEGORY\"], product_pdf[\"DEFAULT_RATE\"], color=\"#7A6FD6\")\naxes[1].set_title(\"Default Rate by Product Category\")\naxes[1].set_ylabel(\"Default rate\")\naxes[1].tick_params(axis=\"x\", rotation=45)\nplt.tight_layout()\nplt.show()\n```\n\n![Class balance and default rate by product category](https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/class-balance-default-rate-by-product.png?v=54b7b1dd)\n\nMacro trends against the portfolio default rate over time:\n\n```python\n# Macro trends vs portfolio default rate over time\ntrend_pdf = session.sql(\"\"\"\n    SELECT m.OBSERVATION_MONTH, m.UNEMPLOYMENT_RATE, m.FED_FUNDS_RATE, m.CPI_INDEX,\n           AVG(l.DEFAULT_FLAG) AS DEFAULT_RATE\n    FROM CREDIT_RISK_ML.PD_MODEL.MACRO_INDICATORS m\n    LEFT JOIN CREDIT_RISK_ML.PD_MODEL.LOAN_DATA l\n        ON DATE_TRUNC('MONTH', l.ARCHIVE_DATE) = m.OBSERVATION_MONTH\n    GROUP BY 1, 2, 3, 4 ORDER BY 1\n\"\"\").to_pandas()\n\nfig, axes = plt.subplots(2, 2, figsize=(14, 8), sharex=True)\nseries = [(\"UNEMPLOYMENT_RATE\", \"Unemployment rate (%)\"),\n          (\"FED_FUNDS_RATE\", \"Fed funds rate (%)\"),\n          (\"CPI_INDEX\", \"CPI index\"),\n          (\"DEFAULT_RATE\", \"Portfolio default rate\")]\nfor ax, (c, title) in zip(axes.flat, series):\n    ax.plot(trend_pdf[\"OBSERVATION_MONTH\"], trend_pdf[c], marker=\".\")\n    ax.set_title(title)\n    ax.tick_params(axis=\"x\", rotation=45)\nplt.suptitle(\"FRED Macro Indicators and Portfolio Default Rate (Monthly)\")\nplt.tight_layout()\nplt.show()\n```\n\n![FRED macro indicators and portfolio default rate](https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/macro-trends-vs-default-rate.png?v=54b7b1dd)\n\nA correlation matrix on a 20K-row sample, taken in Snowflake before collecting:\n\n```python\n# Correlation heatmap on a 20K-row sample (sampled in Snowflake before collecting)\nimport numpy as np\n\ncorr_cols = behavioral_cols + [\"UNEMPLOYMENT_RATE\", \"FED_FUNDS_RATE\", \"CPI_INDEX\", \"DEFAULT_FLAG\"]\nsample_pdf = session.sql(f\"\"\"\n    SELECT * FROM (\n        SELECT {\", \".join(\"l.\" + c for c in behavioral_cols + [\"DEFAULT_FLAG\"])},\n               m.UNEMPLOYMENT_RATE, m.FED_FUNDS_RATE, m.CPI_INDEX\n        FROM CREDIT_RISK_ML.PD_MODEL.LOAN_DATA l\n        JOIN CREDIT_RISK_ML.PD_MODEL.MACRO_INDICATORS m\n            ON DATE_TRUNC('MONTH', l.ARCHIVE_DATE) = m.OBSERVATION_MONTH\n    ) SAMPLE (20000 ROWS)\n\"\"\").to_pandas()\n\ncorr = sample_pdf[corr_cols].corr()\nfig, ax = plt.subplots(figsize=(11, 9))\nim = ax.imshow(corr, cmap=\"RdBu_r\", vmin=-1, vmax=1)\nax.set_xticks(range(len(corr_cols)))\nax.set_xticklabels(corr_cols, rotation=90)\nax.set_yticks(range(len(corr_cols)))\nax.set_yticklabels(corr_cols)\nfor i in range(len(corr_cols)):\n    for j in range(len(corr_cols)):\n        ax.text(j, i, f\"{corr.iloc[i, j]:.2f}\", ha=\"center\", va=\"center\", fontsize=7)\nfig.colorbar(im, ax=ax, shrink=0.8)\nax.set_title(\"Feature Correlation Matrix (incl. DEFAULT_FLAG)\")\nplt.tight_layout()\nplt.show()\n```\n\n![Feature correlation matrix](https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/correlation-heatmap.png?v=54b7b1dd)\n\n\u003C!-- ------------------------ --\u003E\n## Build the Feature Store\n\nIn SAS, features usually live in macro-generated datasets refreshed on custom schedules. Here they become managed **Feature Views**: you declare the transformation once, and Snowflake handles refresh, versioning and point-in-time correct retrieval.\n\nCreate the Feature Store in the project schema and register the `LOAN` entity, keyed on `TRADE_ID`:\n\n```python\nfrom snowflake.ml.feature_store import FeatureStore, Entity, FeatureView, CreationMode\n\nfs = FeatureStore(\n    session=session,\n    database=\"CREDIT_RISK_ML\",\n    name=\"PD_MODEL\",\n    default_warehouse=\"CREDIT_RISK_WH\",\n    creation_mode=CreationMode.CREATE_IF_NOT_EXIST\n)\n\n# Entity: one row per loan trade\nloan_entity = Entity(name=\"LOAN\", join_keys=[\"TRADE_ID\"])\nfs.register_entity(loan_entity)\nprint(f\"Registered entity: {loan_entity.name}\")\n```\n\n**Behavioral Feature View.** Loan-level credit attributes from Equifax, time-stamped by `ARCHIVE_DATE` so later joins are point-in-time correct:\n\n```python\n# Behavioral Feature View — loan-level attributes from Equifax\nbehavioral_source = session.sql(\"\"\"\n    SELECT TRADE_ID, ARCHIVE_DATE,\n           VANTAGE_SCORE_3, NUMBER_OF_ACCOUNTS, NUMBER_ACCOUNTS_PAST_DUE,\n           NUMBER_ACCOUNTS_MAJOR_DEROGATORY, HIGH_UTIL_ACCTS,\n           AGE_OLDEST_ACCOUNT, TOTAL_PAST_DUE_AMOUNT, BALANCE\n    FROM CREDIT_RISK_ML.PD_MODEL.LOAN_DATA\n\"\"\")\n\nbehavioral_fv = FeatureView(\n    name=\"LOAN_BEHAVIORAL_FEATURES\",\n    entities=[loan_entity],\n    feature_df=behavioral_source,\n    timestamp_col=\"ARCHIVE_DATE\",\n    refresh_freq=\"1 day\",\n    desc=\"Loan-level behavioral features from Equifax\"\n)\nbehavioral_fv = fs.register_feature_view(feature_view=behavioral_fv, version=\"V1\")\nprint(f\"Registered: {behavioral_fv.name} v{behavioral_fv.version}\")\n```\n\n**Macro Feature View.** FRED indicators joined to each loan's observation month. These historical values are used for training:\n\n```python\n# Macro Feature View — FRED economic indicators joined to loan population\nmacro_source = session.sql(\"\"\"\n    SELECT l.TRADE_ID, l.ARCHIVE_DATE,\n           m.UNEMPLOYMENT_RATE, m.FED_FUNDS_RATE, m.CPI_INDEX\n    FROM CREDIT_RISK_ML.PD_MODEL.LOAN_DATA l\n    JOIN CREDIT_RISK_ML.PD_MODEL.MACRO_INDICATORS m\n        ON DATE_TRUNC('MONTH', l.ARCHIVE_DATE) = m.OBSERVATION_MONTH\n\"\"\")\n\nmacro_fv = FeatureView(\n    name=\"LOAN_MACRO_FEATURES\",\n    entities=[loan_entity],\n    feature_df=macro_source,\n    timestamp_col=\"ARCHIVE_DATE\",\n    refresh_freq=\"1 day\",\n    desc=\"Macroeconomic indicators from FRED joined to loan population\"\n)\nmacro_fv = fs.register_feature_view(feature_view=macro_fv, version=\"V1\")\nprint(f\"Registered: {macro_fv.name} v{macro_fv.version}\")\n```\n\n**Macro Scenario Feature View.** IFRS 9 and CECL require forward-looking, scenario-weighted loss estimates. This view holds four illustrative scenarios (baseline, upside, downside and severe downside), each a shock applied to the latest FRED observation with a probability weight. In production these would come from your economics team, regulator-prescribed scenarios or a Marketplace provider. The view is keyed by `SCENARIO_ID` rather than by loan, because each loan is scored once per scenario at scoring time.\n\n| Scenario | Weight | Unemployment shift | Fed funds shift | CPI factor |\n|---|---|---|---|---|\n| BASELINE | 0.50 | 0.0 | 0.00 | 1.000 |\n| UPSIDE | 0.15 | -0.5 | +0.25 | 1.010 |\n| DOWNSIDE | 0.30 | +2.0 | -1.00 | 0.990 |\n| SEVERE_DOWNSIDE | 0.05 | +4.5 | -2.00 | 0.970 |\n\n```python\n# Macro Scenario Feature View — forward-looking scenarios for IFRS 9 / CECL scenario-weighted ECL\n# Illustrative shocks applied to the latest FRED observation; in production these come from the\n# economics team, regulator-prescribed scenarios or a Marketplace provider.\nfrom python_pd_model import MACRO_COLS  # shared with the training module\n\nsession.sql(\"\"\"\n    CREATE OR REPLACE TABLE CREDIT_RISK_ML.PD_MODEL.MACRO_SCENARIOS CHANGE_TRACKING = TRUE AS\n    WITH latest AS (\n        SELECT * FROM CREDIT_RISK_ML.PD_MODEL.MACRO_INDICATORS\n        QUALIFY ROW_NUMBER() OVER (ORDER BY OBSERVATION_MONTH DESC) = 1\n    ),\n    shocks (SCENARIO_ID, SCENARIO_WEIGHT, UNRATE_SHIFT, FEDFUNDS_SHIFT, CPI_FACTOR) AS (\n        SELECT * FROM VALUES\n            ('BASELINE',        0.50,  0.0,   0.00, 1.000),\n            ('UPSIDE',          0.15, -0.5,   0.25, 1.010),\n            ('DOWNSIDE',        0.30,  2.0,  -1.00, 0.990),\n            ('SEVERE_DOWNSIDE', 0.05,  4.5,  -2.00, 0.970)\n    )\n    SELECT s.SCENARIO_ID, s.SCENARIO_WEIGHT,\n           l.OBSERVATION_MONTH AS SCENARIO_ANCHOR_MONTH,\n           l.UNEMPLOYMENT_RATE + s.UNRATE_SHIFT               AS UNEMPLOYMENT_RATE,\n           GREATEST(l.FED_FUNDS_RATE + s.FEDFUNDS_SHIFT, 0)   AS FED_FUNDS_RATE,\n           l.CPI_INDEX * s.CPI_FACTOR                         AS CPI_INDEX\n    FROM shocks s CROSS JOIN latest l\n\"\"\").collect()\n\nscenario_entity = Entity(name=\"MACRO_SCENARIO\", join_keys=[\"SCENARIO_ID\"])\nfs.register_entity(scenario_entity)\n\nscenario_fv = FeatureView(\n    name=\"MACRO_SCENARIO_FEATURES\",\n    entities=[scenario_entity],\n    feature_df=session.table(\"CREDIT_RISK_ML.PD_MODEL.MACRO_SCENARIOS\").select(\n        \"SCENARIO_ID\", \"SCENARIO_WEIGHT\", *MACRO_COLS),\n    refresh_freq=\"1 day\",  # in practice quarterly, or ad hoc when economics revises forecasts\n    desc=\"Forward-looking macro scenarios (baseline/upside/downside/severe) with probability weights\",\n)\nscenario_fv = fs.register_feature_view(feature_view=scenario_fv, version=\"V1\")\nprint(f\"Registered: {scenario_fv.name} v{scenario_fv.version}\")\nfs.read_feature_view(scenario_fv).show()\n```\n\nList the registered Feature Views:\n\n```python\n# List all registered feature views\nfv_list = fs.list_feature_views().to_pandas()\nprint(f\"Feature Views registered: {len(fv_list)}\")\nfv_list[[\"NAME\", \"VERSION\", \"DESC\"]]\n```\n\n\u003C!-- ------------------------ --\u003E\n## Generate a Versioned Training Dataset\n\nRegulators ask: *\"What data trained this model?\"* A Snowflake **Dataset** answers that with a single immutable, versioned reference.\n\nBuild a spine of loan IDs, observation dates and labels:\n\n```python\nfrom snowflake.snowpark.functions import col\n\n# Build spine: loan IDs + observation dates + labels\nspine_df = session.table(\"CREDIT_RISK_ML.PD_MODEL.LOAN_DATA\").select(\n    col(\"TRADE_ID\"), col(\"ARCHIVE_DATE\"), col(\"DEFAULT_FLAG\")\n)\nprint(f\"Spine rows: {spine_df.count():,}\")\n```\n\nThen let the Feature Store join the behavioral and macro features as of each loan's `ARCHIVE_DATE`, so no feature value from the future leaks into training:\n\n```python\n# Generate training dataset with point-in-time joins (no look-ahead bias)\ndataset = fs.generate_dataset(\n    name=\"PD_TRAINING_DS\",\n    spine_df=spine_df,\n    features=[behavioral_fv, macro_fv],\n    spine_timestamp_col=\"ARCHIVE_DATE\",\n    spine_label_cols=[\"DEFAULT_FLAG\"],\n    output_type=\"dataset\"\n)\nprint(f\"Dataset: {dataset.fully_qualified_name}\")\nprint(f\"Version: {dataset.selected_version.name}\")\n```\n\nInspect the result:\n\n```python\n# Inspect the versioned dataset\nds_df = dataset.read.to_snowpark_dataframe()\nprint(f\"Rows: {ds_df.count():,}\")\nprint(f\"Columns: {ds_df.columns}\")\nds_df.show(5)\n```\n\n\u003C!-- ------------------------ --\u003E\n## Train with Experiment Tracking\n\nThe PD model is an `XGBClassifier`, the Python equivalent of the SAS `PROC GRADBOOST` model (`sas_pd_model.sas`). It was converted with **CoCo**. The hyperparameters map directly from the SAS source:\n\n| SAS PROC GRADBOOST | XGBClassifier | Value |\n|---|---|---|\n| `ntrees=200` | `n_estimators` | 200 |\n| `learningrate=0.1` | `learning_rate` | 0.1 |\n| `maxdepth=6` | `max_depth` | 6 |\n| `vars_to_try=5` | `colsample_bynode` | 5/11 (~0.45), per split |\n| `minleafsize=20` | `min_child_weight` | 20 (approximate) |\n| `seed=42` | `random_state` | 42 |\n\nImport the model definition from `python_pd_model.py`, then load the Dataset and create the stratified 80/20 split over the 11 features (8 behavioral, 3 macro) with the module's `prepare_features` and `split_train_test`:\n\n```python\n# The model definition lives in python_pd_model.py (converted from SAS by Cortex Code).\n# The notebook and the training pipeline both import it, so they train exactly the same model.\nimport numpy as np\nfrom python_pd_model import (FEATURE_COLS, TARGET_COL, SAS_MAPPED_PARAMS, CHALLENGER_PARAMS,\n                             prepare_features, split_train_test, train_pd_model, pd_metrics)\n\n# Load dataset to pandas\ntrain_pdf = ds_df.to_pandas()\n\nX, y = prepare_features(train_pdf)                           # SAS DATA step: missing -\u003E 0\nX_train, X_test, y_train, y_test = split_train_test(X, y)    # SAS PROC SURVEYSELECT: 80/20, seed 42\nprint(f\"Train: {X_train.shape[0]:,} rows | Test: {X_test.shape[0]:,} rows\")\nprint(f\"Default rate — Train: {y_train.mean():.3f} | Test: {y_test.mean():.3f}\")\n```\n\nTrain two candidates with the module's `train_pd_model` and log both to the `PD_MODEL_DEVELOPMENT` experiment: the SAS-mapped baseline, `SAS_MAPPED_PARAMS` (the future **champion**), and a tuned configuration, `CHALLENGER_PARAMS` (the **challenger**). Each run records its hyperparameters, the exact Dataset version it consumed, and the standard credit risk metrics:\n- **Gini** (2 × AUC − 1): discrimination\n- **KS**: the maximum separation between the score distributions of defaulters and non-defaulters\n- **Brier score**: calibration\n\nThis replaces the Word or Excel tracker in which modelers record what they tried.\n\n```python\n# Train XGBoost — converted from SAS PROC GRADBOOST (see sas_pd_model.sas) — with Experiment Tracking\n# Every run (params, dataset version, Gini/KS/Brier) is logged to the PD_MODEL_DEVELOPMENT experiment\nfrom snowflake.ml.experiment import ExperimentTracking\n\nexp = ExperimentTracking(session=session, database_name=\"CREDIT_RISK_ML\", schema_name=\"PD_MODEL\")\nexp.set_experiment(\"PD_MODEL_DEVELOPMENT\")\n\n# Champion candidate: hyperparameters mapped from SAS PROC GRADBOOST (see SAS_MAPPED_PARAMS in python_pd_model.py)\nparams = SAS_MAPPED_PARAMS\n# Challenger candidate: CHALLENGER_PARAMS (slower learning rate, more trees, row subsampling)\n\nruns = {}\nfor run_name, run_params in [(\"sas_mapped_baseline\", params), (\"challenger_lr005_d5\", CHALLENGER_PARAMS)]:\n    with exp.start_run(run_name):\n        exp.log_params({**run_params, \"dataset\": dataset.fully_qualified_name,\n                        \"dataset_version\": dataset.selected_version.name})\n        clf = train_pd_model(X_train, y_train, run_params)\n        run_metrics = pd_metrics(y_test.values, clf.predict_proba(X_test)[:, 1])\n        exp.log_metrics(run_metrics)\n    runs[run_name] = (clf, run_metrics)\n    print(f\"{run_name:\u003C22} AUC={run_metrics['auc']:.4f}  Gini={run_metrics['gini']:.4f}  \"\n          f\"KS={run_metrics['ks']:.4f}  Brier={run_metrics['brier']:.4f}\")\n\nmodel, champion_metrics = runs[\"sas_mapped_baseline\"]\nchallenger_model, challenger_metrics = runs[\"challenger_lr005_d5\"]\ny_pred_proba = model.predict_proba(X_test)[:, 1]\nprint(\"\\nCompare runs in Snowsight: AI & ML \u003E Experiments \u003E PD_MODEL_DEVELOPMENT\")\n```\n\nPlot the champion's feature importance:\n\n```python\n# Feature importance\nimport matplotlib.pyplot as plt\n\nimportances = model.feature_importances_\nsorted_idx = np.argsort(importances)\n\nfig, ax = plt.subplots(figsize=(10, 6))\nax.barh(range(len(sorted_idx)), importances[sorted_idx])\nax.set_yticks(range(len(sorted_idx)))\nax.set_yticklabels([FEATURE_COLS[i] for i in sorted_idx])\nax.set_xlabel(\"Feature Importance (Gain)\")\nax.set_title(\"PD Model — XGBoost Feature Importance\")\nplt.tight_layout()\nplt.show()\n```\n\n![XGBoost feature importance](https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/xgboost-feature-importance.png?v=54b7b1dd)\n\nYou can compare the runs in Snowsight under **AI & ML \u003E Experiments \u003E PD_MODEL_DEVELOPMENT**.\n\n### Optional: scale out with ML Jobs and Many Model Training\nThe demo portfolio trains comfortably in the notebook kernel. For larger portfolios, the same code runs as an **ML Job** on a CPU compute pool. **Many Model Training** fits one PD model per segment (here `PRODUCT_CATEGORY`) in a single distributed run. Both cells are gated by `RUN_ON_COMPUTE_POOL = False`, so the notebook runs end to end without a compute pool. Set it to `True` to try them.\n\n```python\n# ML Job: run the challenger training on a compute pool instead of the notebook kernel\nRUN_ON_COMPUTE_POOL = False\nCOMPUTE_POOL = \"SYSTEM_COMPUTE_POOL_CPU\"\n\nsession.sql(\"CREATE STAGE IF NOT EXISTS CREDIT_RISK_ML.PD_MODEL.ML_ARTIFACTS\").collect()\n\nif RUN_ON_COMPUTE_POOL:\n    from snowflake.ml.jobs import remote\n\n    @remote(COMPUTE_POOL, stage_name=\"CREDIT_RISK_ML.PD_MODEL.ML_ARTIFACTS\",\n            pip_requirements=[\"xgboost\", \"scikit-learn\"], session=session)\n    def train_pd_on_pool(dataset_name: str, dataset_version: str, run_params: dict, feature_cols: list):\n        from snowflake.snowpark import Session\n        from snowflake.ml.dataset import load_dataset\n        from xgboost import XGBClassifier\n        from sklearn.metrics import roc_auc_score\n\n        s = Session.builder.getOrCreate()\n        pdf = load_dataset(s, dataset_name, dataset_version).read.to_pandas()\n        X_all, y_all = pdf[feature_cols].fillna(0), pdf[\"DEFAULT_FLAG\"]\n        clf = XGBClassifier(**run_params).fit(X_all, y_all)\n        print(f\"In-sample Gini: {2 * roc_auc_score(y_all, clf.predict_proba(X_all)[:, 1]) - 1:.4f}\")\n        return clf\n\n    job = train_pd_on_pool(dataset.fully_qualified_name, dataset.selected_version.name,\n                           CHALLENGER_PARAMS, FEATURE_COLS)\n    print(f\"ML Job {job.id} submitted to {COMPUTE_POOL}\")\n    job.wait()\n    pool_model = job.result()\n    print(job.get_logs())\nelse:\n    print(\"Skipped ML Job — set RUN_ON_COMPUTE_POOL = True to train on a compute pool\")\n```\n\n```python\n# Many Model Training: one PD model per product segment, trained in parallel on the compute pool\nif RUN_ON_COMPUTE_POOL:\n    from snowflake.ml.modeling.distributors.many_model import ManyModelTraining\n\n    segment_train_df = ds_df.join(\n        session.table(\"CREDIT_RISK_ML.PD_MODEL.LOAN_DATA\").select(\"TRADE_ID\", \"PRODUCT_CATEGORY\"),\n        on=\"TRADE_ID\",\n    )\n\n    def train_segment_model(data_connector, context):\n        from xgboost import XGBClassifier\n        seg = data_connector.to_pandas()\n        print(f\"Training PD model for segment {context.partition_id}: {len(seg):,} loans\")\n        return XGBClassifier(**params).fit(seg[FEATURE_COLS].fillna(0), seg[TARGET_COL])\n\n    mmt = ManyModelTraining(train_segment_model, \"CREDIT_RISK_ML.PD_MODEL.ML_ARTIFACTS\")\n    mmt_run = mmt.run(\n        partition_by=\"PRODUCT_CATEGORY\",\n        snowpark_dataframe=segment_train_df,\n        run_id=\"pd_by_product_category\",\n    )\n    print(f\"Many Model Training finished: {mmt_run.wait()}\")\nelse:\n    print(\"Skipped Many Model Training — set RUN_ON_COMPUTE_POOL = True to train per-segment models\")\n```\n\n\u003E aside negative\n\u003E The per-segment models are a demonstration only. They are not registered or used by the pipelines later in this guide.\n\n\u003C!-- ------------------------ --\u003E\n## Register and Explain Model\n\nRegister both models in the **Model Registry** as versions of `PD_XGB_MODEL`: **V1** gets the `champion` alias (production) and **V2** gets the `challenger` alias (candidate). Logging each model with a Snowpark sample from the versioned Dataset does two things:\n- it records **ML Lineage** from the model back to the Dataset and Feature Views;\n- with `target_platforms=[\"WAREHOUSE\"]`, it enables the Registry's built-in SHAP `explain` function.\n\n```python\nfrom snowflake.ml.registry import Registry\n\nreg = Registry(session=session, database_name=\"CREDIT_RISK_ML\", schema_name=\"PD_MODEL\")\n\n# Drop existing model if re-running so V1/V2 can be re-registered\ntry:\n    reg.delete_model(\"PD_XGB_MODEL\")\n    print(\"Dropped existing PD_XGB_MODEL for re-registration\")\nexcept Exception:\n    pass\n\n# Snowpark sample from the versioned Dataset -\u003E records lineage Dataset -\u003E Model\nsample_input = ds_df.select(FEATURE_COLS).limit(5)\n\nregistered = {}\nfor version_name, alias, mdl, mtr, desc in [\n    (\"V1\", \"champion\", model, champion_metrics, \"SAS-mapped hyperparameters\"),\n    (\"V2\", \"challenger\", challenger_model, challenger_metrics, \"tuned candidate\"),\n]:\n    mv_i = reg.log_model(\n        model=mdl,\n        model_name=\"PD_XGB_MODEL\",\n        version_name=version_name,\n        sample_input_data=sample_input,\n        comment=f\"{alias.upper()} — PD model, XGBoost migrated from SAS PROC GRADBOOST ({desc})\",\n        metrics=mtr,\n        target_platforms=[\"WAREHOUSE\"],  # warehouse target + sample input enables the explain function\n    )\n    mv_i.set_alias(alias)\n    registered[alias] = mv_i\n    print(f\"Registered {version_name} as '{alias}': Gini={mtr['gini']:.4f}, KS={mtr['ks']:.4f}, Brier={mtr['brier']:.4f}\")\n\nmv = registered[\"champion\"]\npd_model = reg.get_model(\"PD_XGB_MODEL\")\npd_model.default = mv\npd_model.show_versions()\n```\n\n### SHAP explainability\nCredit decisions must be explainable, for audit purposes. The `explain` function runs on the warehouse next to the data, with no external SHAP tooling. Start with global importance:\n\n```python\n# SHAP Explainability from the Registry — critical for adverse action notices and SR 11-7 compliance\n# The explain function runs on the warehouse next to the data; no local SHAP tooling required\nimport altair as alt\nfrom snowflake.ml.monitoring.explain_visualize import plot_violin\n\nexplain_input = ds_df.select(\"TRADE_ID\", *FEATURE_COLS).limit(200)\nexplanations = mv.run(explain_input, function_name=\"explain\").to_pandas()\n\nshap_cols = [f\"{c}_explanation\" for c in FEATURE_COLS]\nshap_df = explanations[shap_cols].rename(columns=lambda c: c.replace(\"_explanation\", \"\"))\nfeature_df = explanations[FEATURE_COLS]\n\n# Global importance: mean |SHAP| per feature\nprint(shap_df.abs().mean().sort_values(ascending=False).round(4).to_string())\n\nalt.renderers.enable(\"mimetype\")\nplot_violin(shap_df=shap_df, feature_df=feature_df)\n```\n\n![SHAP values per feature](https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/shap-explainability.png?v=54b7b1dd)\n\nThen explain a single prediction. This is the input to an adverse action notice (\"why was this applicant declined?\"):\n\n```python\n# Per-prediction explanation (adverse action reporting) from the Registry's explain output\nsample_idx = 0\nsample_trade = explanations[\"TRADE_ID\"].iloc[sample_idx].item()  # numpy scalar -\u003E native Python for Snowpark literal\nsample_pd = mv.run(\n    explain_input.filter(col(\"TRADE_ID\") == sample_trade), function_name=\"predict_proba\"\n).to_pandas()[\"output_feature_1\"].iloc[0]\n\nprint(f\"Loan {sample_trade}: PD = {sample_pd:.4f}\\n\")\nprint(\"Top adverse action factors:\")\nfactors = sorted(\n    zip(FEATURE_COLS, shap_df.iloc[sample_idx].values, feature_df.iloc[sample_idx].values),\n    key=lambda x: abs(x[1]), reverse=True,\n)\nfor feat, shap_val, feat_val in factors[:5]:\n    direction = \"increases\" if shap_val \u003E 0 else \"decreases\"\n    print(f\"  {feat} = {feat_val:.1f} ({direction} PD by {abs(shap_val):.4f})\")\n```\n\n### ML Lineage\nAnswer \"what data trained this model?\" with a query instead of a document hunt. Walk upstream from the champion to its Dataset, Feature Views and source tables:\n\n```python\n# ML Lineage — trace the champion back to its Dataset, Feature Views and source tables\n# Answers the validator question \"what data trained this model?\" with a query, not a document hunt\nfor node in mv.lineage(direction=\"upstream\"):\n    print(f\"model {mv.version_name} \u003C- {node}\")\n    if hasattr(node, \"lineage\"):\n        for parent in node.lineage(direction=\"upstream\"):\n            print(f\"    \u003C- {parent}\")\n```\n\n\u003C!-- ------------------------ --\u003E\n## Batch Score and ECL\n\nCredit risk models run in batch, for monthly provisioning and quarterly stress tests. Features come straight from the Feature Store through `retrieve_feature_values`, so scoring uses exactly the same feature definitions as training.\n\nScore the same portfolio with the champion and the challenger. This is **offline champion/challenger**: both models see the identical population, and nothing reaches production decisions until a model is explicitly promoted.\n\n```python\n# Score the same portfolio with champion and challenger — features served by the Feature Store\nfrom snowflake.snowpark.types import TimestampType\n\npd_model = reg.get_model(\"PD_XGB_MODEL\")\n\nscoring_spine = session.table(\"CREDIT_RISK_ML.PD_MODEL.LOAN_DATA\").select(\n    \"TRADE_ID\", \"ARCHIVE_DATE\", \"DEFAULT_FLAG\", \"PRODUCT_CATEGORY\"\n).limit(100000)\n\nscoring_df = fs.retrieve_feature_values(\n    spine_df=scoring_spine,\n    features=[behavioral_fv, macro_fv],\n    spine_timestamp_col=\"ARCHIVE_DATE\",\n)\n\nfor alias in [\"champion\", \"challenger\"]:\n    mv_alias = pd_model.version(alias)\n    preds = mv_alias.run(scoring_df, function_name=\"predict_proba\")\n    # Model monitors require a TIMESTAMP_NTZ timestamp column\n    preds = preds.with_column(\"ARCHIVE_DATE\", col(\"ARCHIVE_DATE\").cast(TimestampType()))\n    preds.write.save_as_table(f\"CREDIT_RISK_ML.PD_MODEL.PD_SCORES_{alias.upper()}\", mode=\"overwrite\")\n    print(f\"{alias:\u003C10} ({mv_alias.version_name}): scored {preds.count():,} loans \"\n          f\"-\u003E PD_SCORES_{alias.upper()}\")\n\nsession.table(\"CREDIT_RISK_ML.PD_MODEL.PD_SCORES_CHAMPION\").show(5)\n```\n\nCompare discrimination and calibration on that population:\n\n```python\n# Champion vs challenger on the identical population: discrimination (Gini, KS) and calibration (Brier)\nfrom sklearn.metrics import brier_score_loss\n\ncomparison = []\nfor alias in [\"CHAMPION\", \"CHALLENGER\"]:\n    s = session.table(f\"CREDIT_RISK_ML.PD_MODEL.PD_SCORES_{alias}\").select(\n        \"DEFAULT_FLAG\", '\"output_feature_1\"').to_pandas()\n    y_s, p_s = s[\"DEFAULT_FLAG\"].values, s[\"output_feature_1\"].values\n    a = roc_auc_score(y_s, p_s)\n    comparison.append({\n        \"model\": alias, \"auc\": a, \"gini\": 2 * a - 1,\n        \"ks\": ks_2samp(p_s[y_s == 0], p_s[y_s == 1]).statistic,\n        \"brier\": brier_score_loss(y_s, p_s),\n        \"mean_pd\": p_s.mean(), \"observed_dr\": y_s.mean(),\n    })\n\nimport pandas as pd\npd.DataFrame(comparison).set_index(\"model\").round(4)\n```\n\nFor very large portfolios, `run_batch()` distributes inference across a compute pool. This step is optional and gated by the same flag:\n\n```python\n# Large portfolios: job-based batch inference with run_batch() on an SPCS compute pool\n# (warehouse run() above is enough for this demo portfolio; RUN_ON_COMPUTE_POOL is set in section 6b)\nif RUN_ON_COMPUTE_POOL:\n    from snowflake.ml.model.batch import JobSpec, OutputSpec, SaveMode\n\n    batch_job = pd_model.version(\"champion\").run_batch(\n        X=scoring_df,\n        compute_pool=COMPUTE_POOL,\n        output_spec=OutputSpec(\n            stage_location=\"@CREDIT_RISK_ML.PD_MODEL.ML_ARTIFACTS/batch_scores/\",\n            mode=SaveMode.OVERWRITE,\n        ),\n        job_spec=JobSpec(function_name=\"predict_proba\"),\n    )\n    batch_job.wait()\n    print(f\"run_batch job {batch_job.id}: {batch_job.status}\")\nelse:\n    print(\"Skipped run_batch() — set RUN_ON_COMPUTE_POOL = True to score on a compute pool\")\n```\n\n### Scenario-weighted Expected Credit Loss\nECL = PD × LGD × EAD, weighted across macro scenarios:\nScenario-conditioned scoring: score each loan once per macro scenario, then probability-weight\nLGD is a flat illustrative assumption and BALANCE is used as the EAD proxy — this guide models PD only\n\n```python\nLGD_ASSUMPTION = 0.45\n\nscenarios_df = fs.read_feature_view(scenario_fv)\nbehavioral_cols = [c for c in FEATURE_COLS if c not in MACRO_COLS]\n\n# Replace each loan's historical macro values with the scenario's forward-looking values\nscenario_scoring_df = scoring_df.select(\"TRADE_ID\", \"ARCHIVE_DATE\", *behavioral_cols).cross_join(scenarios_df)\nscenario_preds = pd_model.version(\"champion\").run(\n    scenario_scoring_df.select(\"TRADE_ID\", \"SCENARIO_ID\", \"SCENARIO_WEIGHT\", *FEATURE_COLS),\n    function_name=\"predict_proba\",\n)\nscenario_preds.write.save_as_table(\"CREDIT_RISK_ML.PD_MODEL.PD_SCORES_SCENARIO\", mode=\"overwrite\")\n\necl_by_scenario = session.sql(f\"\"\"\n    SELECT SCENARIO_ID, ANY_VALUE(SCENARIO_WEIGHT) AS SCENARIO_WEIGHT,\n           AVG(\"output_feature_1\") AS MEAN_PD,\n           SUM(\"output_feature_1\" * BALANCE * {LGD_ASSUMPTION}) AS ECL\n    FROM CREDIT_RISK_ML.PD_MODEL.PD_SCORES_SCENARIO\n    GROUP BY SCENARIO_ID\n    ORDER BY MEAN_PD\n\"\"\").to_pandas()\nweighted_ecl = (ecl_by_scenario[\"SCENARIO_WEIGHT\"] * ecl_by_scenario[\"ECL\"]).sum()\n\n# Loan-level probability-weighted PD for provisioning\nsession.sql(\"\"\"\n    CREATE OR REPLACE TABLE CREDIT_RISK_ML.PD_MODEL.PD_SCORES_WEIGHTED AS\n    SELECT TRADE_ID, SUM(SCENARIO_WEIGHT * \"output_feature_1\") AS WEIGHTED_PD\n    FROM CREDIT_RISK_ML.PD_MODEL.PD_SCORES_SCENARIO\n    GROUP BY TRADE_ID\n\"\"\").collect()\n\nprint(ecl_by_scenario.round(4).to_string(index=False))\nprint(f\"\\nProbability-weighted ECL: {weighted_ecl:,.0f}  (LGD = {LGD_ASSUMPTION}, EAD = BALANCE)\")\n```\n\n\u003C!-- ------------------------ --\u003E\n## Monitor the Models\n\nAttach a **Model Version Monitor** to the champion and to the challenger. Each monitor:\n- reads its model's scoring output (`PD_SCORES_\u003CROLE\u003E`);\n- compares it with a baseline, which is that version's scores on its training Dataset;\n- tracks data drift and prediction drift against that baseline;\n- tracks performance against realized defaults (`DEFAULT_FLAG`);\n- breaks every metric out by `PRODUCT_CATEGORY`.\n\nThis replaces quarterly manual monitoring reports.\n\n```python\n# Model Version Monitors — one per version, each on its own scoring output (champion vs challenger)\n# predict_proba returns lowercase columns: \"output_feature_1\" is P(default), so it is the score column\n# Binary classification: ground truth (0/1 DEFAULT_FLAG) goes in ACTUAL_CLASS_COLUMNS, not ACTUAL_SCORE_COLUMNS\nfor alias, version in [(\"CHAMPION\", \"V1\"), (\"CHALLENGER\", \"V2\")]:\n    # Baseline = this version's scores on its training Dataset; drift (PSI, JS divergence) is measured against it\n    baseline_input = ds_df.join(\n        session.table(\"CREDIT_RISK_ML.PD_MODEL.LOAN_DATA\").select(\"TRADE_ID\", \"PRODUCT_CATEGORY\"), on=\"TRADE_ID\"\n    )\n    pd_model.version(version).run(baseline_input, function_name=\"predict_proba\").with_column(\n        \"ARCHIVE_DATE\", col(\"ARCHIVE_DATE\").cast(TimestampType())\n    ).write.save_as_table(f\"CREDIT_RISK_ML.PD_MODEL.PD_BASELINE_{alias}\", mode=\"overwrite\")\n\n    session.sql(f\"\"\"\n        CREATE OR REPLACE MODEL MONITOR CREDIT_RISK_ML.PD_MODEL.PD_MODEL_MONITOR_{alias}\n        WITH\n            MODEL = CREDIT_RISK_ML.PD_MODEL.PD_XGB_MODEL\n            VERSION = '{version}'\n            FUNCTION = 'predict_proba'\n            SOURCE = CREDIT_RISK_ML.PD_MODEL.PD_SCORES_{alias}\n            BASELINE = CREDIT_RISK_ML.PD_MODEL.PD_BASELINE_{alias}\n            TIMESTAMP_COLUMN = ARCHIVE_DATE\n            PREDICTION_SCORE_COLUMNS = ('output_feature_1')\n            ACTUAL_CLASS_COLUMNS = (DEFAULT_FLAG)\n            ID_COLUMNS = (TRADE_ID)\n            SEGMENT_COLUMNS = (PRODUCT_CATEGORY)\n            WAREHOUSE = CREDIT_RISK_WH\n            AGGREGATION_WINDOW = '1 DAY'\n            REFRESH_INTERVAL = '1 DAY'\n    \"\"\").collect()\n    print(f\"Model Monitor created: PD_MODEL_MONITOR_{alias} (version {version})\")\n\n# Drop the single-version monitor from earlier runs of this notebook\nsession.sql(\"DROP MODEL MONITOR IF EXISTS CREDIT_RISK_ML.PD_MODEL.PD_MODEL_MONITOR\").collect()\nprint(\"Tracking per version: data drift vs training baseline, prediction drift, Gini/KS/calibration vs realized defaults\")\nprint(\"View in Snowsight: AI & ML \u003E Models \u003E PD_XGB_MODEL \u003E Monitors\")\n```\n\nView the monitors in Snowsight under **AI & ML \u003E Models \u003E PD_XGB_MODEL \u003E Monitors**. When the training pipeline promotes a new version, it re-points that role's monitor automatically.\n\n\u003C!-- ------------------------ --\u003E\n## Automate Monthly Retraining\n\nNow automate the lifecycle with **Task DAGs**, replacing the cron and shell scripts common in SAS environments. Each task calls a Python stored procedure.\n\n![Training pipeline](https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/training-pipeline.png?v=54b7b1dd)\n\nEvery month the training pipeline produces **two candidates** on a fresh Dataset:\n- a **champion candidate**: the production champion's hyperparameters, refit;\n- a **challenger candidate**: the challenger configuration, refit.\n\nBoth candidates are registered with `CHAMPION_CANDIDATE` / `CHALLENGER_CANDIDATE` aliases and scored on the same portfolio as the production champion. An advisory **validation gate** checks Gini lift ≥ 0 and Brier ≤ 1.02 × champion. A reviewer then gets one email and approves or rejects each candidate separately. The automated gate only *recommends*. Only a user with the `MODEL_REVIEWER` role can move the production aliases, which gives you the four-eyes control expected under SR 11-7 / SS1/23.\n\n```\nPD_MODEL_TRAIN_PIPELINE (root, 1st of month 02:00)   SP_REFRESH_FEATURES\n  └─ ..._GENERATE_DATASET                              SP_GENERATE_DATASET\n       └─ ..._TRAIN_CANDIDATES                         SP_TRAIN_CANDIDATES\n            └─ ..._SCORE_CANDIDATES                    SP_SCORE_CANDIDATES\n                 └─ ..._VALIDATION_GATE                SP_VALIDATION_GATE   -\u003E PROMOTION_LOG (PENDING_REVIEW)\n                      └─ ..._NOTIFY_REVIEWER           SP_NOTIFY_REVIEWER   -\u003E review email\n\nReviewer (MODEL_REVIEWER), outside the DAG, once per candidate:\n  CALL SP_REVIEW_MODEL('\u003Cversion\u003E', 'APPROVE' | 'REJECT', '\u003Ccomments\u003E')\n    -\u003E SP_APPLY_REVIEW: moves the alias, re-points the monitor, writes the audit trail\n```\n\n### Pipeline stored procedures\nThis cell uploads `python_pd_model.py` to the `ML_ARTIFACTS` stage and registers each pipeline step as a stored procedure with that file as an import. `SP_TRAIN_CANDIDATES` trains with the module's `prepare_features`, `split_train_test`, `train_pd_model` and hyperparameters, and every step computes metrics with its `pd_metrics`. The monthly retraining therefore runs the same converted code as the notebook. To change the model, edit `python_pd_model.py` and re-run this cell. The pipeline steps are:\n\n| Procedure | What it does |\n|---|---|\n| `SP_REFRESH_FEATURES` | Refreshes the behavioral and macro Feature Views |\n| `SP_GENERATE_DATASET` | Writes a new timestamped version of `PD_TRAINING_DS` |\n| `SP_TRAIN_CANDIDATES` | Retrains both candidates, logs runs to Experiment Tracking, registers them with candidate aliases |\n| `SP_SCORE_CANDIDATES` | Scores the production champion and both candidates on the same portfolio |\n| `SP_VALIDATION_GATE` | Computes Gini/Brier vs the champion, writes one `PENDING_REVIEW` row per candidate to `PROMOTION_LOG` |\n| `SP_NOTIFY_REVIEWER` | Sends one HTML email covering both candidates, with approve/reject commands for each |\n| `SP_APPLY_REVIEW` | Applies a decision: moves the alias, re-points the monitor, comments the versions, updates the audit trail |\n\nIt also creates `PROMOTION_LOG`, the model risk review queue and audit trail.\n\n```python\n# Pipeline steps as Python stored procedures — each Task in the DAG calls one of these\n# Every retraining cycle produces TWO candidates, each approved (or rejected) separately by the reviewer:\n#   champion candidate   — the current champion's hyperparameters refit on the latest Dataset\n#   challenger candidate — the challenger configuration (CHALLENGER_PARAMS) refit on the latest Dataset\n# Candidates carry CHAMPION_CANDIDATE / CHALLENGER_CANDIDATE aliases until reviewed; only an approval\n# moves the CHAMPION / CHALLENGER alias that weekly scoring uses.\nfrom snowflake.snowpark import Session\nimport python_pd_model as pdm\n\nDB, SCHEMA, WH = \"CREDIT_RISK_ML\", \"PD_MODEL\", \"CREDIT_RISK_WH\"\nMODEL_NAME = \"PD_XGB_MODEL\"\nSP_PACKAGES = [\"snowflake-ml-python\", \"xgboost\", \"scikit-learn\", \"scipy\", \"pandas\"]\nROLES = [\"CHAMPION\", \"CHALLENGER\"]\n\nsession.sql(f\"CREATE STAGE IF NOT EXISTS {DB}.{SCHEMA}.ML_ARTIFACTS\").collect()\n# Ship the model module to the stage; every pipeline procedure imports it from there\nMODEL_MODULE = f\"@{DB}.{SCHEMA}.ML_ARTIFACTS/code/python_pd_model.py\"\nsession.file.put(\"python_pd_model.py\", f\"@{DB}.{SCHEMA}.ML_ARTIFACTS/code\",\n                 auto_compress=False, overwrite=True)\n# PROMOTION_LOG is the model-risk review queue and audit trail, one row per candidate:\n#   TARGET_ROLE        = role the candidate would take if approved (CHAMPION / CHALLENGER)\n#   CHALLENGER_VERSION = the candidate version under review (column name kept for compatibility)\n#   CHAMPION_VERSION   = the champion in production when the candidate was evaluated (the benchmark)\n#   DECISION           = the automated gate's recommendation (PASS / REJECT)\n#   STATUS             = PENDING_REVIEW -\u003E APPROVED / REJECTED (by a human) or SUPERSEDED (newer candidate)\nsession.sql(f\"\"\"\n    CREATE TABLE IF NOT EXISTS {DB}.{SCHEMA}.PROMOTION_LOG (\n        EVALUATED_AT TIMESTAMP_NTZ, CHAMPION_VERSION STRING, CHALLENGER_VERSION STRING,\n        CHAMPION_GINI FLOAT, CHALLENGER_GINI FLOAT, CHAMPION_BRIER FLOAT, CHALLENGER_BRIER FLOAT,\n        DECISION STRING, PROMOTED BOOLEAN\n    )\n\"\"\").collect()\nfor column, dtype in [(\"STATUS\", \"STRING\"), (\"REVIEWED_BY\", \"STRING\"),\n                      (\"REVIEWED_AT\", \"TIMESTAMP_NTZ\"), (\"REVIEW_COMMENTS\", \"STRING\"),\n                      (\"NOTIFIED_AT\", \"TIMESTAMP_NTZ\"), (\"TARGET_ROLE\", \"STRING\")]:\n    session.sql(f\"ALTER TABLE {DB}.{SCHEMA}.PROMOTION_LOG ADD COLUMN IF NOT EXISTS {column} {dtype}\").collect()\n# Rows written before TARGET_ROLE existed were challenger -\u003E champion promotions\nsession.sql(f\"UPDATE {DB}.{SCHEMA}.PROMOTION_LOG SET TARGET_ROLE = 'CHAMPION' WHERE TARGET_ROLE IS NULL\").collect()\n\n# Reviewer emails and the email integration name are configured in the Setup cell (section 1)\n\n\ndef _version_by_alias(model, alias):\n    import json\n    for _, row in model.show_versions().iterrows():\n        aliases = json.loads(row[\"aliases\"]) if row[\"aliases\"] else []\n        if alias.upper() in [a.upper() for a in aliases]:\n            return model.version(row[\"name\"])\n    return None\n\n\ndef _move_alias(model, alias, version_name):\n    \"\"\"Point an alias at version_name, removing it from whichever version holds it now (idempotent).\"\"\"\n    holder = _version_by_alias(model, alias)\n    if holder is not None and holder.version_name == version_name:\n        return None\n    if holder is not None:\n        holder.unset_alias(alias)\n    model.version(version_name).set_alias(alias)\n    return holder.version_name if holder is not None else None\n\n\ndef _latest_dataset_version(session):\n    versions = session.sql(f\"SHOW VERSIONS IN DATASET {DB}.{SCHEMA}.PD_TRAINING_DS\").collect()\n    return sorted(versions, key=lambda r: r[\"created_on\"])[-1][\"version\"]\n\n\ndef _scoring_features(session, fs, limit=100000):\n    spine = session.table(f\"{DB}.{SCHEMA}.LOAN_DATA\").select(\n        \"TRADE_ID\", \"ARCHIVE_DATE\", \"DEFAULT_FLAG\", \"PRODUCT_CATEGORY\")\n    if limit is not None:\n        spine = spine.limit(limit)\n    return fs.retrieve_feature_values(\n        spine_df=spine,\n        features=[fs.get_feature_view(\"LOAN_BEHAVIORAL_FEATURES\", \"V1\"),\n                  fs.get_feature_view(\"LOAN_MACRO_FEATURES\", \"V1\")],\n        spine_timestamp_col=\"ARCHIVE_DATE\",\n    )\n\n\ndef _set_version_comment(session, version_name, text):\n    \"\"\"Role-prefixed comment shown next to each version in Snowsight (AI & ML \u003E Models).\n    DDL does not accept bind variables, so the text is escaped as a string literal.\"\"\"\n    escaped = text.replace(\"\\\\\", \"\\\\\\\\\").replace(\"'\", \"''\")\n    session.sql(f\"ALTER MODEL {DB}.{SCHEMA}.{MODEL_NAME} MODIFY VERSION {version_name} \"\n                f\"SET COMMENT = '{escaped}'\").collect()\n\n\ndef _upsert_monitor(session, role, version_name):\n    \"\"\"(Re)point the role's monitor at a version, with a baseline scored on the latest training Dataset.\"\"\"\n    import json\n    from snowflake.snowpark.functions import col\n    from snowflake.snowpark.types import TimestampType\n    from snowflake.ml.dataset import load_dataset\n    from snowflake.ml.registry import Registry\n    train_df = load_dataset(session, f\"{DB}.{SCHEMA}.PD_TRAINING_DS\",\n                            _latest_dataset_version(session)).read.to_snowpark_dataframe()\n    baseline_input = train_df.join(\n        session.table(f\"{DB}.{SCHEMA}.LOAN_DATA\").select(\"TRADE_ID\", \"PRODUCT_CATEGORY\"), on=\"TRADE_ID\")\n    mv = Registry(session=session, database_name=DB, schema_name=SCHEMA).get_model(MODEL_NAME).version(version_name)\n    mv.run(baseline_input, function_name=\"predict_proba\").with_column(\n        \"ARCHIVE_DATE\", col(\"ARCHIVE_DATE\").cast(TimestampType())\n    ).write.save_as_table(f\"{DB}.{SCHEMA}.PD_BASELINE_{role}\", mode=\"overwrite\")\n    # Snowflake allows one monitor per model version: drop any other monitor already on this version\n    for mon in session.sql(f\"SHOW MODEL MONITORS IN SCHEMA {DB}.{SCHEMA}\").collect():\n        on_version = json.loads(mon[\"model\"]).get(\"version_name\", \"\").upper() == version_name.upper()\n        if on_version and mon[\"name\"].upper() != f\"PD_MODEL_MONITOR_{role}\":\n            session.sql(f\"DROP MODEL MONITOR {DB}.{SCHEMA}.{mon['name']}\").collect()\n    session.sql(f\"\"\"\n        CREATE OR REPLACE MODEL MONITOR {DB}.{SCHEMA}.PD_MODEL_MONITOR_{role}\n        WITH\n            MODEL = {DB}.{SCHEMA}.{MODEL_NAME}\n            VERSION = '{version_name}'\n            FUNCTION = 'predict_proba'\n            SOURCE = {DB}.{SCHEMA}.PD_SCORES_{role}\n            BASELINE = {DB}.{SCHEMA}.PD_BASELINE_{role}\n            TIMESTAMP_COLUMN = ARCHIVE_DATE\n            PREDICTION_SCORE_COLUMNS = ('output_feature_1')\n            ACTUAL_CLASS_COLUMNS = (DEFAULT_FLAG)\n            ID_COLUMNS = (TRADE_ID)\n            SEGMENT_COLUMNS = (PRODUCT_CATEGORY)\n            WAREHOUSE = {WH}\n            AGGREGATION_WINDOW = '1 DAY'\n            REFRESH_INTERVAL = '1 DAY'\n    \"\"\").collect()\n\n\ndef sp_refresh_features(session: Session) -\u003E str:\n    from snowflake.ml.feature_store import FeatureStore\n    fs = FeatureStore(session, DB, SCHEMA, WH)\n    for name in [\"LOAN_BEHAVIORAL_FEATURES\", \"LOAN_MACRO_FEATURES\"]:\n        fs.refresh_feature_view(fs.get_feature_view(name, \"V1\"))\n    return \"Feature views refreshed\"\n\n\ndef sp_generate_dataset(session: Session) -\u003E str:\n    import datetime\n    from snowflake.ml.feature_store import FeatureStore\n    fs = FeatureStore(session, DB, SCHEMA, WH)\n    version = datetime.datetime.utcnow().strftime(\"V_%Y%m%d_%H%M\")\n    spine = session.table(f\"{DB}.{SCHEMA}.LOAN_DATA\").select(\"TRADE_ID\", \"ARCHIVE_DATE\", \"DEFAULT_FLAG\")\n    fs.generate_dataset(\n        name=\"PD_TRAINING_DS\", version=version, spine_df=spine,\n        features=[fs.get_feature_view(\"LOAN_BEHAVIORAL_FEATURES\", \"V1\"),\n                  fs.get_feature_view(\"LOAN_MACRO_FEATURES\", \"V1\")],\n        spine_timestamp_col=\"ARCHIVE_DATE\", spine_label_cols=[\"DEFAULT_FLAG\"], output_type=\"dataset\",\n    )\n    return version\n\n\ndef sp_train_candidates(session: Session) -\u003E str:\n    \"\"\"Retrain both models on the latest Dataset and register them as candidates awaiting review.\"\"\"\n    import datetime\n    from snowflake.ml.dataset import load_dataset\n    from snowflake.ml.experiment import ExperimentTracking\n    from snowflake.ml.registry import Registry\n\n    ds_version = _latest_dataset_version(session)\n    sdf = load_dataset(session, f\"{DB}.{SCHEMA}.PD_TRAINING_DS\", ds_version).read.to_snowpark_dataframe()\n    pdf = sdf.to_pandas()\n    X, y = pdm.prepare_features(pdf)\n    X_tr, X_te, y_tr, y_te = pdm.split_train_test(X, y)\n\n    reg = Registry(session=session, database_name=DB, schema_name=SCHEMA)\n    m = reg.get_model(MODEL_NAME)\n\n    # Champion candidate keeps the production champion's hyperparameters; challenger uses the challenger config\n    champion = _version_by_alias(m, \"champion\")\n    champion_params, params_source = pdm.SAS_MAPPED_PARAMS, \"SAS-mapped baseline (no champion found)\"\n    if champion is not None:\n        champion_params = {k: v for k, v in champion.load(force=True).get_params().items() if v is not None}\n        params_source = f\"hyperparameters of champion {champion.version_name}\"\n    candidates = [(\"CHAMPION\", champion_params, params_source),\n                  (\"CHALLENGER\", pdm.CHALLENGER_PARAMS, \"challenger configuration\")]\n\n    exp = ExperimentTracking(session=session, database_name=DB, schema_name=SCHEMA)\n    exp.set_experiment(\"PD_MODEL_DEVELOPMENT\")\n    stamp = datetime.datetime.utcnow().strftime(\"%Y%m%d_%H%M\")\n    created = []\n    for role, run_params, source in candidates:\n        version_name = f\"V_{stamp}_{role[:5]}\"          # e.g. V_20261101_0200_CHAMP / _CHALL\n        with exp.start_run(f\"retrain_{role.lower()}_{stamp}\"):\n            exp.log_params({**run_params, \"dataset_version\": ds_version, \"target_role\": role})\n            model = pdm.train_pd_model(X_tr, y_tr, run_params)\n            metrics = pdm.pd_metrics(y_te.values, model.predict_proba(X_te)[:, 1])\n            exp.log_metrics(metrics)\n        reg.log_model(\n            model=model, model_name=MODEL_NAME, version_name=version_name,\n            sample_input_data=sdf.select(pdm.FEATURE_COLS).limit(5),  # Snowpark input -\u003E lineage to the Dataset\n            metrics=metrics, target_platforms=[\"WAREHOUSE\"],\n            comment=f\"{role} CANDIDATE — pending review | {source}, dataset {ds_version}\",\n        )\n        superseded = _move_alias(reg.get_model(MODEL_NAME), f\"{role}_CANDIDATE\", version_name)\n        if superseded:\n            _set_version_comment(session, superseded,\n                                 f\"SUPERSEDED {role.lower()} candidate — replaced by {version_name} before review\")\n        created.append(f\"{role.lower()} candidate {version_name}\")\n    return \"Registered \" + \" and \".join(created)\n\n\ndef sp_score_candidates(session: Session) -\u003E str:\n    \"\"\"Score the production champion and both candidates on the identical portfolio.\"\"\"\n    from snowflake.snowpark.functions import col\n    from snowflake.snowpark.types import TimestampType\n    from snowflake.ml.feature_store import FeatureStore\n    from snowflake.ml.registry import Registry\n    fs = FeatureStore(session, DB, SCHEMA, WH)\n    m = Registry(session=session, database_name=DB, schema_name=SCHEMA).get_model(MODEL_NAME)\n    features = _scoring_features(session, fs)\n    scored = []\n    for alias, table in [(\"champion\", \"PD_SCORES_CHAMPION\"),\n                         (\"champion_candidate\", \"PD_SCORES_CHAMPION_CANDIDATE\"),\n                         (\"challenger_candidate\", \"PD_SCORES_CHALLENGER_CANDIDATE\")]:\n        mv = _version_by_alias(m, alias)\n        mv.run(features, function_name=\"predict_proba\").with_column(\n            \"ARCHIVE_DATE\", col(\"ARCHIVE_DATE\").cast(TimestampType())\n        ).write.save_as_table(f\"{DB}.{SCHEMA}.{table}\", mode=\"overwrite\")\n        scored.append(f\"{alias} {mv.version_name}\")\n    return \"Scored \" + \", \".join(scored)\n\n\ndef sp_validation_gate(session: Session) -\u003E str:\n    \"\"\"Gate each candidate against the production champion and queue both for human review.\"\"\"\n    import datetime\n    from snowflake.ml.registry import Registry\n    m = Registry(session=session, database_name=DB, schema_name=SCHEMA).get_model(MODEL_NAME)\n\n    def metrics_for(table):\n        pdf = session.table(f\"{DB}.{SCHEMA}.{table}\").select(\"DEFAULT_FLAG\", '\"output_feature_1\"').to_pandas()\n        return pdm.pd_metrics(pdf[\"DEFAULT_FLAG\"].values, pdf[\"output_feature_1\"].values)\n\n    bench = metrics_for(\"PD_SCORES_CHAMPION\")\n    champion_version = _version_by_alias(m, \"champion\").version_name\n    now = datetime.datetime.utcnow()\n    rows, summary = [], []\n    for role in ROLES:\n        candidate = _version_by_alias(m, f\"{role}_candidate\").version_name\n        cand = metrics_for(f\"PD_SCORES_{role}_CANDIDATE\")\n        passed = (cand[\"gini\"] \u003E= bench[\"gini\"] + GATE_MIN_GINI_LIFT and\n                  cand[\"brier\"] \u003C= bench[\"brier\"] * GATE_MAX_BRIER_RATIO)\n        decision = \"PASS\" if passed else \"REJECT\"\n        # An older candidate for the same role still waiting for review is superseded by this one\n        session.sql(f\"\"\"\n            UPDATE {DB}.{SCHEMA}.PROMOTION_LOG SET STATUS = 'SUPERSEDED'\n            WHERE STATUS = 'PENDING_REVIEW' AND TARGET_ROLE = ? AND CHALLENGER_VERSION != ?\n        \"\"\", params=[role, candidate]).collect()\n        rows.append([now, champion_version, candidate, bench[\"gini\"], cand[\"gini\"], bench[\"brier\"],\n                     cand[\"brier\"], decision, False, \"PENDING_REVIEW\", role])\n        summary.append(f\"{role.lower()} candidate {candidate}: {decision}\")\n    session.create_dataframe(rows, schema=[\n        \"EVALUATED_AT\", \"CHAMPION_VERSION\", \"CHALLENGER_VERSION\", \"CHAMPION_GINI\", \"CHALLENGER_GINI\",\n        \"CHAMPION_BRIER\", \"CHALLENGER_BRIER\", \"DECISION\", \"PROMOTED\", \"STATUS\", \"TARGET_ROLE\",\n    ]).write.mode(\"append\").save_as_table(f\"{DB}.{SCHEMA}.PROMOTION_LOG\", column_order=\"name\")\n    return \"Gate recommendations — \" + \"; \".join(summary) + \" — pending human review\"\n\n\ndef sp_notify_reviewer(session: Session) -\u003E str:\n    \"\"\"One email covering every candidate awaiting review, each with its own approve / reject commands.\"\"\"\n    from snowflake.ml.registry import Registry\n    pending = session.sql(f\"\"\"\n        SELECT * FROM {DB}.{SCHEMA}.PROMOTION_LOG\n        WHERE STATUS = 'PENDING_REVIEW' AND NOTIFIED_AT IS NULL\n        ORDER BY IFF(TARGET_ROLE = 'CHAMPION', 0, 1)\n    \"\"\").collect()\n    if not pending:\n        return \"Nothing pending review\"\n    m = Registry(session=session, database_name=DB, schema_name=SCHEMA).get_model(MODEL_NAME)\n    versions = m.show_versions()\n    trained = {row[\"name\"]: str(row[\"created_on\"])[:16] for _, row in versions.iterrows()}\n    current = {role: (_version_by_alias(m, role.lower()) or None) for role in ROLES}\n    current = {role: (mv.version_name if mv else \"none\") for role, mv in current.items()}\n\n    cell = \"padding:6px 10px;border:1px solid #d1d5db\"\n    colours = {\"CHAMPION\": \"#1e3a8a\", \"CHALLENGER\": \"#b45309\"}\n\n    def metric_row(label, bench_val, cand_val, higher_is_better):\n        delta = cand_val - bench_val\n        better = (delta \u003E 0) == higher_is_better and delta != 0\n        verdict = (\"Candidate better\" if better else \"Production champion better\" if delta != 0 else \"Equal\")\n        colour = \"#15803d\" if better else \"#b91c1c\" if delta != 0 else \"#6b7280\"\n        return (f\"\u003Ctr\u003E\u003Ctd style='{cell}'\u003E\u003Cb\u003E{label}\u003C/b\u003E\u003C/td\u003E\u003Ctd style='{cell}'\u003E{bench_val:.4f}\u003C/td\u003E\"\n                f\"\u003Ctd style='{cell}'\u003E{cand_val:.4f}\u003C/td\u003E\"\n                f\"\u003Ctd style='{cell};color:{colour}'\u003E{delta:+.4f} ({verdict})\u003C/td\u003E\u003C/tr\u003E\")\n\n    sections = []\n    for r in pending:\n        role, v, bench = r[\"TARGET_ROLE\"], r[\"CHALLENGER_VERSION\"], r[\"CHAMPION_VERSION\"]\n        gate_colour = \"#15803d\" if r[\"DECISION\"] == \"PASS\" else \"#b91c1c\"\n        sections.append(f\"\"\"\n        \u003Ch3 style=\"color:{colours[role]}\"\u003E{role} candidate {v}\u003C/h3\u003E\n        \u003Cp\u003EIf approved, \u003Cb\u003E{v}\u003C/b\u003E becomes the \u003Cb\u003E{role}\u003C/b\u003E (replacing {current[role]}) and is used by\n           weekly batch scoring as the {role.lower()} from the next run.\u003C/p\u003E\n        \u003Ctable style=\"border-collapse:collapse\"\u003E\n          \u003Ctr\u003E\u003Cth style=\"{cell}\"\u003E\u003C/th\u003E\n              \u003Cth style=\"{cell};background:#1e3a8a;color:#ffffff\"\u003EPRODUCTION CHAMPION — benchmark\u003C/th\u003E\n              \u003Cth style=\"{cell};background:{colours[role]};color:#ffffff\"\u003E{role} CANDIDATE — awaiting approval\u003C/th\u003E\n              \u003Cth style=\"{cell}\"\u003ECandidate vs production champion\u003C/th\u003E\u003C/tr\u003E\n          \u003Ctr\u003E\u003Ctd style=\"{cell}\"\u003E\u003Cb\u003EVersion\u003C/b\u003E\u003C/td\u003E\u003Ctd style=\"{cell}\"\u003E{bench}\u003C/td\u003E\u003Ctd style=\"{cell}\"\u003E{v}\u003C/td\u003E\u003Ctd style=\"{cell}\"\u003E\u003C/td\u003E\u003C/tr\u003E\n          \u003Ctr\u003E\u003Ctd style=\"{cell}\"\u003E\u003Cb\u003ETrained\u003C/b\u003E\u003C/td\u003E\u003Ctd style=\"{cell}\"\u003E{trained.get(bench, 'n/a')}\u003C/td\u003E\u003Ctd style=\"{cell}\"\u003E{trained.get(v, 'n/a')}\u003C/td\u003E\u003Ctd style=\"{cell}\"\u003E\u003C/td\u003E\u003C/tr\u003E\n          {metric_row(\"Gini (higher is better)\", r[\"CHAMPION_GINI\"], r[\"CHALLENGER_GINI\"], True)}\n          {metric_row(\"Brier (lower is better)\", r[\"CHAMPION_BRIER\"], r[\"CHALLENGER_BRIER\"], False)}\n        \u003C/table\u003E\n        \u003Cp\u003EAutomated gate recommendation: \u003Cb style=\"color:{gate_colour}\"\u003E{r['DECISION']}\u003C/b\u003E (advisory).\u003C/p\u003E\n        \u003Cpre\u003E-- Approve {v} as {role} (replaces {current[role]})\nCALL {DB}.{SCHEMA}.SP_REVIEW_MODEL('{v}', 'APPROVE', '<comments>');\n-- Reject {v}; {current[role]} stays {role.lower()}\nCALL {DB}.{SCHEMA}.SP_REVIEW_MODEL('{v}', 'REJECT', '<comments>');\u003C/pre\u003E\"\"\")\n\n    roles_pending = \" + \".join(f\"{r['TARGET_ROLE']} {r['CHALLENGER_VERSION']}\" for r in pending)\n    subject = f\"[PD model review] {len(pending)} candidate(s) awaiting approval — {roles_pending}\"\n    body = f\"\"\"\n        \u003Cp\u003EThe monthly retraining run produced new candidates. Each is reviewed \u003Cb\u003Eseparately\u003C/b\u003E:\n           you can approve both, one, or neither. Both are benchmarked against the production champion\n           \u003Cb\u003E{current['CHAMPION']}\u003C/b\u003E on the same portfolio.\u003C/p\u003E\n        {''.join(sections)}\n        \u003Cp\u003EGate rule: Gini lift ≥ {GATE_MIN_GINI_LIFT} and Brier ≤ {GATE_MAX_BRIER_RATIO}× production champion.\n           In Snowsight (AI & ML > Models > {MODEL_NAME}) candidates carry the CHAMPION_CANDIDATE /\n           CHALLENGER_CANDIDATE aliases and a comment starting with \"CHAMPION CANDIDATE\" / \"CHALLENGER CANDIDATE\".\n           Run the commands as a user with the MODEL_REVIEWER role.\u003C/p\u003E\n    \"\"\"\n    session.sql(\"CALL SYSTEM$SEND_EMAIL(?, ?, ?, ?, 'text/html')\",\n                params=[REVIEW_EMAIL_INTEGRATION, \",\".join(REVIEWER_EMAILS), subject, body]).collect()\n    session.sql(f\"\"\"\n        UPDATE {DB}.{SCHEMA}.PROMOTION_LOG SET NOTIFIED_AT = CURRENT_TIMESTAMP()::TIMESTAMP_NTZ\n        WHERE STATUS = 'PENDING_REVIEW' AND NOTIFIED_AT IS NULL\n    \"\"\").collect()\n    return f\"Review request for {roles_pending} sent to {', '.join(REVIEWER_EMAILS)}\"\n\n\ndef sp_apply_review(session: Session, challenger_version: str, decision: str,\n                    comments: str, reviewer: str) -\u003E str:\n    \"\"\"Owner's-rights step: apply a reviewer's decision on one candidate. Only reachable via SP_REVIEW_MODEL.\n    (The first argument is the candidate version; the name is kept so SP_REVIEW_MODEL's signature is unchanged.)\"\"\"\n    import datetime\n    from snowflake.ml.registry import Registry\n    candidate = challenger_version\n    decision = decision.strip().upper()\n    if decision not in (\"APPROVE\", \"REJECT\"):\n        raise ValueError(\"decision must be 'APPROVE' or 'REJECT'\")\n    if not comments or not comments.strip():\n        raise ValueError(\"review comments are required for the audit trail\")\n    pending = session.sql(f\"\"\"\n        SELECT TARGET_ROLE FROM {DB}.{SCHEMA}.PROMOTION_LOG\n        WHERE CHALLENGER_VERSION = ? AND STATUS = 'PENDING_REVIEW'\n    \"\"\", params=[candidate]).collect()\n    if not pending:\n        raise ValueError(f\"No pending review for version {candidate}\")\n    role = pending[0][\"TARGET_ROLE\"]\n    today = datetime.datetime.utcnow().strftime(\"%Y-%m-%d\")\n    m = Registry(session=session, database_name=DB, schema_name=SCHEMA).get_model(MODEL_NAME)\n\n    promoted = decision == \"APPROVE\"\n    if promoted:\n        # Each step is idempotent so a retry after a partial failure completes cleanly.\n        # The role's monitor reads PD_SCORES_\u003CROLE\u003E, so seed it with this candidate's gate scores first.\n        candidate_scores = f\"{DB}.{SCHEMA}.PD_SCORES_{role}_CANDIDATE\"\n        if session.sql(f\"SHOW TABLES LIKE 'PD_SCORES_{role}_CANDIDATE' IN SCHEMA {DB}.{SCHEMA}\").collect():\n            session.sql(f\"CREATE OR REPLACE TABLE {DB}.{SCHEMA}.PD_SCORES_{role} CLONE {candidate_scores}\").collect()\n        _upsert_monitor(session, role, candidate)\n        replaced = _move_alias(m, role, candidate)\n        holder = _version_by_alias(m, f\"{role}_candidate\")\n        if holder is not None and holder.version_name == candidate:\n            holder.unset_alias(f\"{role}_candidate\")\n        if role == \"CHAMPION\":\n            m.default = m.version(candidate)\n        # Role-prefixed comments make each version's status obvious in Snowsight\n        _set_version_comment(session, candidate, f\"{role} — approved {today} by {reviewer}: {comments}\")\n        if replaced:\n            _set_version_comment(session, replaced, f\"RETIRED {role.lower()} — replaced by {candidate} on {today}\")\n    else:\n        holder = _version_by_alias(m, f\"{role}_candidate\")\n        if holder is not None and holder.version_name == candidate:\n            holder.unset_alias(f\"{role}_candidate\")\n        _set_version_comment(session, candidate,\n                             f\"REJECTED {role.lower()} candidate — {today} by {reviewer}: {comments}\")\n\n    session.sql(f\"\"\"\n        UPDATE {DB}.{SCHEMA}.PROMOTION_LOG\n        SET STATUS = ?, PROMOTED = ?, REVIEWED_BY = ?,\n            REVIEWED_AT = CURRENT_TIMESTAMP()::TIMESTAMP_NTZ, REVIEW_COMMENTS = ?\n        WHERE CHALLENGER_VERSION = ? AND STATUS = 'PENDING_REVIEW'\n    \"\"\", params=[\"APPROVED\" if promoted else \"REJECTED\", promoted, reviewer, comments, candidate]).collect()\n    return (f\"{candidate} approved as {role.lower()} by {reviewer}\" if promoted\n            else f\"{candidate} rejected by {reviewer}; current {role.lower()} retained\")\n\n\n# Validation-gate thresholds (candidate must match the champion's discrimination without degrading calibration)\nGATE_MIN_GINI_LIFT = 0.0\nGATE_MAX_BRIER_RATIO = 1.02\n\nfor fn in [sp_refresh_features, sp_generate_dataset, sp_train_candidates,\n           sp_score_candidates, sp_validation_gate, sp_notify_reviewer, sp_apply_review]:\n    session.sproc.register(\n        fn, name=f\"{DB}.{SCHEMA}.{fn.__name__.upper()}\",\n        packages=SP_PACKAGES, imports=[MODEL_MODULE], is_permanent=True, replace=True,\n        stage_location=f\"@{DB}.{SCHEMA}.ML_ARTIFACTS\",\n    )\n    print(f\"Registered procedure {fn.__name__.upper()}\")\n\n# The single-challenger procedures are replaced by the candidate procedures above\nfor old_proc in [\"SP_TRAIN_CHALLENGER()\", \"SP_SCORE_CHAMPION_CHALLENGER()\", \"SP_PROMOTE_MODEL()\"]:\n    session.sql(f\"DROP PROCEDURE IF EXISTS {DB}.{SCHEMA}.{old_proc}\").collect()\n```\n\n### Human-in-the-loop review\nCreate the email notification integration and the reviewer entry point. `SP_REVIEW_MODEL` runs with **caller's rights**, so `CURRENT_USER()` records the actual reviewer. It then hands off to the owner's-rights `SP_APPLY_REVIEW`, which holds the registry privileges. The cell also creates the `REVIEW_QUEUE` view and grants `MODEL_REVIEWER` only what it needs. The pipeline role cannot approve its own models.\n\n```python\n# Human-in-the-loop review: notification channel, reviewer role, and the approval entry point\n# REVIEWER_USERS / REVIEWER_EMAILS are configured in the Setup cell (section 1)\n\n# 1. Email notification integration — recipients must be verified emails of users in this account.\n#    ALLOWED_RECIPIENTS is re-synced from the Setup-cell config every run, so editing the lists there is enough.\nallowed = \", \".join(f\"'{e}'\" for e in dict.fromkeys(REVIEWER_EMAILS + SCORING_NOTIFY_EMAILS))\nsession.sql(f\"\"\"\n    CREATE NOTIFICATION INTEGRATION IF NOT EXISTS {REVIEW_EMAIL_INTEGRATION}\n        TYPE = EMAIL\n        ENABLED = TRUE\n        ALLOWED_RECIPIENTS = ({allowed})\n\"\"\").collect()\nsession.sql(f\"ALTER NOTIFICATION INTEGRATION {REVIEW_EMAIL_INTEGRATION} SET ALLOWED_RECIPIENTS = ({allowed})\").collect()\n\n# 2. Reviewer approval entry point — caller's rights so CURRENT_USER() is the actual reviewer,\n#    then hands off to the owner's-rights SP_APPLY_REVIEW, which holds the registry privileges\nsession.sql(f\"\"\"\n    CREATE OR REPLACE PROCEDURE {DB}.{SCHEMA}.SP_REVIEW_MODEL(\n        CHALLENGER_VERSION STRING, DECISION STRING, COMMENTS STRING)\n    RETURNS STRING\n    LANGUAGE SQL\n    EXECUTE AS CALLER\n    AS\n    $$\n    DECLARE\n        reviewer STRING DEFAULT CURRENT_USER();\n        result STRING;\n    BEGIN\n        CALL {DB}.{SCHEMA}.SP_APPLY_REVIEW(:CHALLENGER_VERSION, :DECISION, :COMMENTS, :reviewer) INTO :result;\n        RETURN result;\n    END;\n    $$\n\"\"\").collect()\n\n# 3. Review queue for reviewers and MRM reporting\nsession.sql(f\"\"\"\n    CREATE OR REPLACE VIEW {DB}.{SCHEMA}.REVIEW_QUEUE AS\n    SELECT TARGET_ROLE, CHALLENGER_VERSION AS CANDIDATE_VERSION, CHAMPION_VERSION AS BENCHMARK_CHAMPION,\n           EVALUATED_AT, NOTIFIED_AT,\n           CHAMPION_GINI, CHALLENGER_GINI, CHAMPION_BRIER, CHALLENGER_BRIER,\n           DECISION AS GATE_RECOMMENDATION, STATUS, REVIEWED_BY, REVIEWED_AT, REVIEW_COMMENTS,\n           DATEDIFF('day', EVALUATED_AT, COALESCE(REVIEWED_AT, CURRENT_TIMESTAMP())) AS DAYS_IN_REVIEW\n    FROM {DB}.{SCHEMA}.PROMOTION_LOG\n    WHERE STATUS IS NOT NULL\n\"\"\").collect()\n\n# 4. Segregation of duties: only MODEL_REVIEWER can approve; the pipeline role gets no grant on these\nsession.sql(\"CREATE ROLE IF NOT EXISTS MODEL_REVIEWER\").collect()\nfor grant in [\n    f\"GRANT USAGE ON DATABASE {DB} TO ROLE MODEL_REVIEWER\",\n    f\"GRANT USAGE ON SCHEMA {DB}.{SCHEMA} TO ROLE MODEL_REVIEWER\",\n    f\"GRANT USAGE ON WAREHOUSE {WH} TO ROLE MODEL_REVIEWER\",\n    f\"GRANT SELECT ON VIEW {DB}.{SCHEMA}.REVIEW_QUEUE TO ROLE MODEL_REVIEWER\",\n    f\"GRANT USAGE ON PROCEDURE {DB}.{SCHEMA}.SP_REVIEW_MODEL(STRING, STRING, STRING) TO ROLE MODEL_REVIEWER\",\n    f\"GRANT USAGE ON PROCEDURE {DB}.{SCHEMA}.SP_APPLY_REVIEW(STRING, STRING, STRING, STRING) TO ROLE MODEL_REVIEWER\",\n]:\n    session.sql(grant).collect()\nfor user in REVIEWER_USERS:\n    session.sql(f\"GRANT ROLE MODEL_REVIEWER TO USER {user}\").collect()\n\nprint(f\"Reviewer notifications: {REVIEW_EMAIL_INTEGRATION} -\u003E {', '.join(REVIEWER_EMAILS)}\")\nprint(f\"MODEL_REVIEWER granted to: {', '.join(REVIEWER_USERS)}\")\n```\n\n### Create the training Task DAG\n```python\n# Create Task DAG for the monthly retrain of BOTH models (ends at human review of each candidate)\ndag = [\n    # (task name, predecessor, procedure)\n    (\"PD_MODEL_TRAIN_PIPELINE\", None, \"SP_REFRESH_FEATURES\"),\n    (\"PD_MODEL_TRAIN_PIPELINE_GENERATE_DATASET\", \"PD_MODEL_TRAIN_PIPELINE\", \"SP_GENERATE_DATASET\"),\n    (\"PD_MODEL_TRAIN_PIPELINE_TRAIN_CANDIDATES\", \"PD_MODEL_TRAIN_PIPELINE_GENERATE_DATASET\", \"SP_TRAIN_CANDIDATES\"),\n    (\"PD_MODEL_TRAIN_PIPELINE_SCORE_CANDIDATES\", \"PD_MODEL_TRAIN_PIPELINE_TRAIN_CANDIDATES\", \"SP_SCORE_CANDIDATES\"),\n    (\"PD_MODEL_TRAIN_PIPELINE_VALIDATION_GATE\", \"PD_MODEL_TRAIN_PIPELINE_SCORE_CANDIDATES\", \"SP_VALIDATION_GATE\"),\n    (\"PD_MODEL_TRAIN_PIPELINE_NOTIFY_REVIEWER\", \"PD_MODEL_TRAIN_PIPELINE_VALIDATION_GATE\", \"SP_NOTIFY_REVIEWER\"),\n]\n\n# The graph can only be changed while the root is suspended\nsession.sql(f\"ALTER TASK IF EXISTS {DB}.{SCHEMA}.PD_MODEL_TRAIN_PIPELINE SUSPEND\").collect()\n# Remove tasks from earlier designs (single challenger, automated promotion), children first\nfor old_task in [\"TASK_PROMOTE_MODEL\", \"TASK_SCORE_CHAMPION_CHALLENGER\", \"TASK_TRAIN_CHALLENGER\", \"TASK_TRAIN_MODEL\",\n                 # Pre-rename names of both DAGs\n                 \"TASK_NOTIFY_REVIEWER\", \"TASK_VALIDATION_GATE\", \"TASK_SCORE_CANDIDATES\", \"TASK_TRAIN_CANDIDATES\",\n                 \"TASK_GENERATE_DATASET\", \"TASK_FEATURE_REFRESH\", \"TASK_WEEKLY_SCORING_NOTIFY\",\n                 \"TASK_WEEKLY_SCENARIO_ECL\", \"TASK_WEEKLY_BATCH_SCORE\", \"TASK_WEEKLY_FEATURE_REFRESH\",\n                 \"TRAIN_PIPELINE_NOTIFY_REVIEWER\", \"TRAIN_PIPELINE_VALIDATION_GATE\", \"TRAIN_PIPELINE_SCORE_CANDIDATES\",\n                 \"TRAIN_PIPELINE_TRAIN_CANDIDATES\", \"TRAIN_PIPELINE_GENERATE_DATASET\", \"TRAIN_PIPELINE_FEATURE_REFRESH\"]:\n    session.sql(f\"DROP TASK IF EXISTS {DB}.{SCHEMA}.{old_task}\").collect()\n\nfor task, after, proc in dag:\n    trigger = (f\"SCHEDULE = 'USING CRON 0 2 1 * * America/New_York'\" if after is None\n               else f\"AFTER {DB}.{SCHEMA}.{after}\")\n    session.sql(f\"\"\"\n        CREATE OR REPLACE TASK {DB}.{SCHEMA}.{task}\n            WAREHOUSE = {WH}\n            {trigger}\n        AS\n            CALL {DB}.{SCHEMA}.{proc}()\n    \"\"\").collect()\nprint(f\"Task DAG created ({len(dag)} tasks, suspended)\")\n```\n\nView the graph:\n\n```python\n# View the task graph\ntask_graph = session.sql(\"\"\"\n    SELECT NAME, STATE, SCHEDULE, COMMENT\n    FROM TABLE(INFORMATION_SCHEMA.TASK_DEPENDENTS(\n        TASK_NAME =\u003E 'CREDIT_RISK_ML.PD_MODEL.PD_MODEL_TRAIN_PIPELINE',\n        RECURSIVE =\u003E TRUE\n    ))\n    ORDER BY NAME\n\"\"\")\ntask_graph.show()\n```\n\nActivate the graph. The children are resumed along with the root:\n\n```sql\nSELECT SYSTEM$TASK_DEPENDENTS_ENABLE('CREDIT_RISK_ML.PD_MODEL.PD_MODEL_TRAIN_PIPELINE');\n```\n\nTo run one cycle on demand instead of waiting for the 1st of the month, resume the child tasks and execute the root:\n\n```sql\nALTER TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_TRAIN_PIPELINE_NOTIFY_REVIEWER RESUME;\nALTER TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_TRAIN_PIPELINE_VALIDATION_GATE RESUME;\nALTER TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_TRAIN_PIPELINE_SCORE_CANDIDATES RESUME;\nALTER TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_TRAIN_PIPELINE_TRAIN_CANDIDATES RESUME;\nALTER TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_TRAIN_PIPELINE_GENERATE_DATASET RESUME;\nEXECUTE TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_TRAIN_PIPELINE;\n```\n\nYou can follow the run in Snowsight under **Monitoring \u003E Task History**, or open the root task and select the **Graph** tab to see each step's status:\n\n![Training pipeline Task DAG run](https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/train-pipeline-task-dag-run.png?v=54b7b1dd)\n\n### Review the candidates\nWhen the run finishes, the reviewer receives one email. It has a section for each candidate with its metrics against the production champion, the gate's recommendation, and ready-to-run commands:\n\n![Model approval email](https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/model-approval-email-notification.png?v=54b7b1dd)\n\nThe reviewer decides on each candidate separately:\n\n```sql\nUSE ROLE MODEL_REVIEWER;\nSELECT * FROM CREDIT_RISK_ML.PD_MODEL.REVIEW_QUEUE WHERE STATUS = 'PENDING_REVIEW';\nCALL CREDIT_RISK_ML.PD_MODEL.SP_REVIEW_MODEL('\u003Cchampion_candidate_version\u003E', 'APPROVE', '\u003Ccomments\u003E');\nCALL CREDIT_RISK_ML.PD_MODEL.SP_REVIEW_MODEL('\u003Cchallenger_candidate_version\u003E', 'REJECT', '\u003Ccomments\u003E');\n```\n\nApproving a candidate moves the `champion` or `challenger` alias, re-points that role's monitor, and records who approved it, when and why. That record is kept in `PROMOTION_LOG` and in the model version's comment. A candidate left unreviewed is marked `SUPERSEDED` by the next cycle's candidate.\n\n\u003C!-- ------------------------ --\u003E\n## Automate Weekly Scoring\n\nA second, independent Task DAG produces the weekly portfolio PDs and scenario-weighted ECL. It never picks a model version itself. It reads whichever versions hold the `champion` and `challenger` aliases when it runs, so an approval in the training pipeline takes effect from the next weekly run.\n\n![Scoring pipeline](https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/scoring-pipeline.png?v=54b7b1dd)\n\n```\nPD_MODEL_SCORING_PIPELINE (root, Mondays 03:00)   SP_REFRESH_FEATURES\n  └─ ..._BATCH_SCORE                                SP_WEEKLY_BATCH_SCORE   -\u003E PD_PORTFOLIO_SCORES (append)\n       └─ ..._SCENARIO_ECL                          SP_WEEKLY_SCENARIO_ECL  -\u003E PD_WEIGHTED_HISTORY, PD_ECL_HISTORY (append)\n                                                                            -\u003E PD_CHAMPION_CHALLENGER_TREND (view)\n            └─ ..._NOTIFY                           SP_WEEKLY_SCORING_NOTIFY -\u003E summary email\n```\n\nHow the weekly results are kept:\n- Every row is appended with `SCORE_TS`, `MODEL_VERSION` and `MODEL_ROLE`, so any week's provision traces back to an approved model.\n- The champion's numbers are the official provision. The challenger's build up comparison evidence for the next review.\n- `PD_CHAMPION_CHALLENGER_TREND` summarizes both models week by week: mean PD, observed default rate, PD correlation and weighted ECL.\n\n```python\n# Weekly scoring steps as stored procedures (reuse helpers and constants from the cells above)\n# Champion AND challenger score the same portfolio every week; MODEL_ROLE tells them apart.\n# The champion's numbers are the official provision; the challenger's are for comparison only.\nsession.sql(f\"\"\"\n    CREATE TABLE IF NOT EXISTS {DB}.{SCHEMA}.PD_WEIGHTED_HISTORY (\n        SCORE_TS TIMESTAMP_NTZ, MODEL_VERSION STRING, TRADE_ID NUMBER,\n        WEIGHTED_PD FLOAT, WEIGHTED_ECL FLOAT\n    )\n\"\"\").collect()\nsession.sql(f\"\"\"\n    CREATE TABLE IF NOT EXISTS {DB}.{SCHEMA}.PD_ECL_HISTORY (\n        SCORE_TS TIMESTAMP_NTZ, MODEL_VERSION STRING, SCENARIO_ID STRING,\n        SCENARIO_WEIGHT FLOAT, LOANS NUMBER, MEAN_PD FLOAT, ECL FLOAT\n    )\n\"\"\").collect()\n# MODEL_ROLE (CHAMPION / CHALLENGER) on every history table, added in place if they already exist.\n# Widen it too: a table first created from champion-only scores infers VARCHAR(8), too short for 'CHALLENGER'.\nfor table in [\"PD_PORTFOLIO_SCORES\", \"PD_WEIGHTED_HISTORY\", \"PD_ECL_HISTORY\"]:\n    session.sql(f\"ALTER TABLE IF EXISTS {DB}.{SCHEMA}.{table} ADD COLUMN IF NOT EXISTS MODEL_ROLE STRING\").collect()\n    session.sql(f\"ALTER TABLE IF EXISTS {DB}.{SCHEMA}.{table} ALTER COLUMN MODEL_ROLE SET DATA TYPE STRING\").collect()\n\n\ndef _weekly_models(session):\n    \"\"\"(role, ModelVersion) pairs for this run: always the champion, plus the challenger if one exists.\"\"\"\n    from snowflake.ml.registry import Registry\n    m = Registry(session=session, database_name=DB, schema_name=SCHEMA).get_model(MODEL_NAME)\n    champion, challenger = _version_by_alias(m, \"champion\"), _version_by_alias(m, \"challenger\")\n    models = [(\"CHAMPION\", champion)]\n    if challenger is not None and challenger.version_name != champion.version_name:\n        models.append((\"CHALLENGER\", challenger))\n    return models\n\n\ndef sp_weekly_batch_score(session: Session) -\u003E str:\n    \"\"\"Score the full portfolio with champion and challenger; append with run timestamp, version and role.\"\"\"\n    import datetime\n    from snowflake.snowpark.functions import col, lit\n    from snowflake.snowpark.types import StringType, TimestampType\n    from snowflake.ml.feature_store import FeatureStore\n    fs = FeatureStore(session, DB, SCHEMA, WH)\n    run_ts = datetime.datetime.utcnow().replace(microsecond=0)   # one timestamp shared by both models\n    features = _scoring_features(session, fs, limit=None)\n\n    scored = []\n    for role, mv in _weekly_models(session):\n        preds = (mv.run(features, function_name=\"predict_proba\")\n                 .with_column(\"ARCHIVE_DATE\", col(\"ARCHIVE_DATE\").cast(TimestampType()))\n                 .with_column(\"SCORE_TS\", lit(run_ts).cast(TimestampType()))\n                 .with_column(\"MODEL_VERSION\", lit(mv.version_name).cast(StringType()))\n                 .with_column(\"MODEL_ROLE\", lit(role).cast(StringType())))   # unbounded, not VARCHAR(len(role))\n        # Full feature + prediction columns are kept so a model monitor can track drift on this table\n        preds.write.mode(\"append\").save_as_table(f\"{DB}.{SCHEMA}.PD_PORTFOLIO_SCORES\", column_order=\"name\")\n        scored.append(f\"{role.lower()} {mv.version_name}\")\n    n = session.table(f\"{DB}.{SCHEMA}.PD_PORTFOLIO_SCORES\").filter(\n        (col(\"SCORE_TS\") == lit(run_ts)) & (col(\"MODEL_ROLE\") == lit(\"CHAMPION\"))).count()\n    return f\"Scored {n:,} loans with {' and '.join(scored)} at {run_ts:%Y-%m-%d %H:%M} UTC\"\n\n\ndef sp_weekly_scenario_ecl(session: Session) -\u003E str:\n    \"\"\"Score each loan under every macro scenario, per model, and append probability-weighted PD and ECL.\"\"\"\n    from snowflake.snowpark.functions import max as max_\n    from snowflake.ml.feature_store import FeatureStore\n    # Registered with caller's rights (needed for the temporary scratch table), so the session has no\n    # database/schema of its own; Feature Store metadata lookups need one.\n    session.use_schema(f\"{DB}.{SCHEMA}\")\n    fs = FeatureStore(session, DB, SCHEMA, WH)\n    # Tie this ECL run to the portfolio scoring run that just completed\n    run_ts = session.table(f\"{DB}.{SCHEMA}.PD_PORTFOLIO_SCORES\").select(max_(\"SCORE_TS\")).collect()[0][0]\n\n    scenarios = fs.read_feature_view(fs.get_feature_view(\"MACRO_SCENARIO_FEATURES\", \"V1\"))\n    scenario_input = (_scoring_features(session, fs, limit=None)\n                      .select(\"TRADE_ID\", *pdm.BEHAVIORAL_COLS).cross_join(scenarios)\n                      .select(\"TRADE_ID\", \"SCENARIO_ID\", \"SCENARIO_WEIGHT\", *pdm.FEATURE_COLS))\n\n    results = []\n    for role, mv in _weekly_models(session):\n        mv.run(scenario_input, function_name=\"predict_proba\").write.save_as_table(\n            f\"{DB}.{SCHEMA}.PD_SCENARIO_SCRATCH\", mode=\"overwrite\", table_type=\"temporary\")\n        params = [run_ts, mv.version_name, role]\n        session.sql(f\"\"\"\n            INSERT INTO {DB}.{SCHEMA}.PD_WEIGHTED_HISTORY\n                (SCORE_TS, MODEL_VERSION, MODEL_ROLE, TRADE_ID, WEIGHTED_PD, WEIGHTED_ECL)\n            SELECT ?::TIMESTAMP_NTZ, ?, ?, TRADE_ID,\n                   SUM(SCENARIO_WEIGHT * \"output_feature_1\"),\n                   SUM(SCENARIO_WEIGHT * \"output_feature_1\" * BALANCE * {LGD_ASSUMPTION})\n            FROM {DB}.{SCHEMA}.PD_SCENARIO_SCRATCH\n            GROUP BY TRADE_ID\n        \"\"\", params=params).collect()\n        session.sql(f\"\"\"\n            INSERT INTO {DB}.{SCHEMA}.PD_ECL_HISTORY\n                (SCORE_TS, MODEL_VERSION, MODEL_ROLE, SCENARIO_ID, SCENARIO_WEIGHT, LOANS, MEAN_PD, ECL)\n            SELECT ?::TIMESTAMP_NTZ, ?, ?, SCENARIO_ID, ANY_VALUE(SCENARIO_WEIGHT), COUNT(*),\n                   AVG(\"output_feature_1\"), SUM(\"output_feature_1\" * BALANCE * {LGD_ASSUMPTION})\n            FROM {DB}.{SCHEMA}.PD_SCENARIO_SCRATCH\n            GROUP BY SCENARIO_ID\n        \"\"\", params=params).collect()\n        weighted = session.sql(f\"\"\"\n            SELECT SUM(SCENARIO_WEIGHT * ECL) FROM {DB}.{SCHEMA}.PD_ECL_HISTORY\n            WHERE SCORE_TS = ?::TIMESTAMP_NTZ AND MODEL_ROLE = ?\n        \"\"\", params=[run_ts, role]).collect()[0][0]\n        results.append(f\"{role.lower()} {mv.version_name}: {weighted:,.0f}\")\n    _ensure_trend_view(session)\n    return f\"Scenario-weighted ECL — {'; '.join(results)} (run {run_ts})\"\n\n\n# Week-by-week champion vs challenger comparison on the identical portfolio.\n# Created by the pipeline (after the first run writes the history tables it reads).\ndef _ensure_trend_view(session):\n    session.sql(f\"\"\"\n    CREATE OR REPLACE VIEW {DB}.{SCHEMA}.PD_CHAMPION_CHALLENGER_TREND AS\n    WITH paired AS (\n        SELECT c.SCORE_TS, c.MODEL_VERSION AS CHAMPION_VERSION, h.MODEL_VERSION AS CHALLENGER_VERSION,\n               c.TRADE_ID, c.DEFAULT_FLAG,\n               c.\"output_feature_1\" AS CHAMPION_PD, h.\"output_feature_1\" AS CHALLENGER_PD\n        FROM {DB}.{SCHEMA}.PD_PORTFOLIO_SCORES c\n        JOIN {DB}.{SCHEMA}.PD_PORTFOLIO_SCORES h\n          ON c.SCORE_TS = h.SCORE_TS AND c.TRADE_ID = h.TRADE_ID\n        WHERE c.MODEL_ROLE = 'CHAMPION' AND h.MODEL_ROLE = 'CHALLENGER'\n    ),\n    ecl AS (\n        SELECT SCORE_TS,\n               SUM(IFF(MODEL_ROLE = 'CHAMPION', SCENARIO_WEIGHT * ECL, 0))   AS CHAMPION_WEIGHTED_ECL,\n               SUM(IFF(MODEL_ROLE = 'CHALLENGER', SCENARIO_WEIGHT * ECL, 0)) AS CHALLENGER_WEIGHTED_ECL\n        FROM {DB}.{SCHEMA}.PD_ECL_HISTORY\n        GROUP BY SCORE_TS\n    )\n    SELECT p.SCORE_TS, p.CHAMPION_VERSION, p.CHALLENGER_VERSION,\n           COUNT(*)                                     AS LOANS,\n           AVG(p.CHAMPION_PD)                           AS CHAMPION_MEAN_PD,\n           AVG(p.CHALLENGER_PD)                         AS CHALLENGER_MEAN_PD,\n           AVG(p.DEFAULT_FLAG)                          AS OBSERVED_DEFAULT_RATE,\n           AVG(ABS(p.CHALLENGER_PD - p.CHAMPION_PD))    AS MEAN_ABS_PD_DIFF,\n           CORR(p.CHAMPION_PD, p.CHALLENGER_PD)         AS PD_CORRELATION,\n           AVG(IFF(ABS(p.CHALLENGER_PD - p.CHAMPION_PD) \u003E 0.05, 1, 0)) AS SHARE_LOANS_DIFF_GT_5PP,\n           ANY_VALUE(e.CHAMPION_WEIGHTED_ECL)           AS CHAMPION_WEIGHTED_ECL,\n           ANY_VALUE(e.CHALLENGER_WEIGHTED_ECL)         AS CHALLENGER_WEIGHTED_ECL\n    FROM paired p\n    LEFT JOIN ecl e ON e.SCORE_TS = p.SCORE_TS\n    GROUP BY p.SCORE_TS, p.CHAMPION_VERSION, p.CHALLENGER_VERSION\n    \"\"\").collect()\n\n\ndef sp_weekly_scoring_notify(session: Session) -\u003E str:\n    \"\"\"Email the latest weekly run: champion provision (official) plus challenger comparison.\"\"\"\n    from snowflake.snowpark.functions import max as max_\n    run_ts = session.table(f\"{DB}.{SCHEMA}.PD_ECL_HISTORY\").select(max_(\"SCORE_TS\")).collect()[0][0]\n    scen = session.sql(f\"\"\"\n        SELECT SCENARIO_ID, SCENARIO_WEIGHT, MEAN_PD, ECL, MODEL_VERSION\n        FROM {DB}.{SCHEMA}.PD_ECL_HISTORY\n        WHERE SCORE_TS = ?::TIMESTAMP_NTZ AND MODEL_ROLE = 'CHAMPION' ORDER BY MEAN_PD\n    \"\"\", params=[run_ts]).collect()\n    runs = session.sql(f\"\"\"\n        SELECT SCORE_TS, SUM(SCENARIO_WEIGHT * ECL) AS WEIGHTED_ECL\n        FROM {DB}.{SCHEMA}.PD_ECL_HISTORY\n        WHERE SCORE_TS \u003C= ?::TIMESTAMP_NTZ AND MODEL_ROLE = 'CHAMPION'\n        GROUP BY SCORE_TS ORDER BY SCORE_TS DESC LIMIT 2\n    \"\"\", params=[run_ts]).collect()\n    portfolio = session.sql(f\"\"\"\n        SELECT COUNT(*) AS LOANS, AVG(\"output_feature_1\") AS MEAN_PD\n        FROM {DB}.{SCHEMA}.PD_PORTFOLIO_SCORES\n        WHERE SCORE_TS = ?::TIMESTAMP_NTZ AND MODEL_ROLE = 'CHAMPION'\n    \"\"\", params=[run_ts]).collect()[0]\n    trend = session.sql(f\"\"\"\n        SELECT * FROM {DB}.{SCHEMA}.PD_CHAMPION_CHALLENGER_TREND WHERE SCORE_TS = ?::TIMESTAMP_NTZ\n    \"\"\", params=[run_ts]).collect()\n\n    weighted = runs[0][\"WEIGHTED_ECL\"]\n    if len(runs) \u003E 1 and runs[1][\"WEIGHTED_ECL\"]:\n        prev = runs[1][\"WEIGHTED_ECL\"]\n        change = f\"{(weighted - prev) / prev:+.1%} vs previous run ({runs[1]['SCORE_TS']:%Y-%m-%d})\"\n    else:\n        change = \"first run — no prior week to compare\"\n    version = scen[0][\"MODEL_VERSION\"]\n\n    cell = \"padding:6px 10px;border:1px solid #d1d5db\"\n    rows = \"\".join(\n        f\"\u003Ctr\u003E\u003Ctd style='{cell}'\u003E{r['SCENARIO_ID']}\u003C/td\u003E\u003Ctd style='{cell}'\u003E{r['SCENARIO_WEIGHT']:.0%}\u003C/td\u003E\"\n        f\"\u003Ctd style='{cell}'\u003E{r['MEAN_PD']:.4f}\u003C/td\u003E\u003Ctd style='{cell}'\u003E{r['ECL']:,.0f}\u003C/td\u003E\u003C/tr\u003E\" for r in scen)\n\n    if trend:\n        t = trend[0]\n        ecl_delta = t[\"CHALLENGER_WEIGHTED_ECL\"] - t[\"CHAMPION_WEIGHTED_ECL\"]\n        comparison = f\"\"\"\n        \u003Ch3\u003EChampion vs challenger (same portfolio, comparison only)\u003C/h3\u003E\n        \u003Ctable style=\"border-collapse:collapse\"\u003E\n          \u003Ctr\u003E\u003Cth style=\"{cell}\"\u003E\u003C/th\u003E\n              \u003Cth style=\"{cell};background:#1e3a8a;color:#ffffff\"\u003ECHAMPION {t['CHAMPION_VERSION']}\u003C/th\u003E\n              \u003Cth style=\"{cell};background:#b45309;color:#ffffff\"\u003ECHALLENGER {t['CHALLENGER_VERSION']}\u003C/th\u003E\u003C/tr\u003E\n          \u003Ctr\u003E\u003Ctd style=\"{cell}\"\u003E\u003Cb\u003EMean PD\u003C/b\u003E\u003C/td\u003E\u003Ctd style=\"{cell}\"\u003E{t['CHAMPION_MEAN_PD']:.4f}\u003C/td\u003E\n              \u003Ctd style=\"{cell}\"\u003E{t['CHALLENGER_MEAN_PD']:.4f}\u003C/td\u003E\u003C/tr\u003E\n          \u003Ctr\u003E\u003Ctd style=\"{cell}\"\u003E\u003Cb\u003EWeighted ECL\u003C/b\u003E\u003C/td\u003E\u003Ctd style=\"{cell}\"\u003E{t['CHAMPION_WEIGHTED_ECL']:,.0f}\u003C/td\u003E\n              \u003Ctd style=\"{cell}\"\u003E{t['CHALLENGER_WEIGHTED_ECL']:,.0f} ({ecl_delta:+,.0f})\u003C/td\u003E\u003C/tr\u003E\n        \u003C/table\u003E\n        \u003Cp\u003EObserved default rate {t['OBSERVED_DEFAULT_RATE']:.4f} · mean |PD difference|\n           {t['MEAN_ABS_PD_DIFF']:.4f} · PD correlation {t['PD_CORRELATION']:.3f} ·\n           {t['SHARE_LOANS_DIFF_GT_5PP']:.1%} of loans differ by more than 5 pp.\n           History: {DB}.{SCHEMA}.PD_CHAMPION_CHALLENGER_TREND.\u003C/p\u003E\"\"\"\n    else:\n        comparison = \"\u003Cp\u003ENo challenger is currently registered, so only the champion was scored this week.\u003C/p\u003E\"\n\n    subject = f\"[PD model] Weekly scoring complete — {run_ts:%Y-%m-%d}, champion weighted ECL {weighted:,.0f}\"\n    body = f\"\"\"\n        \u003Cp\u003EThe weekly PD batch scoring run has completed.\u003C/p\u003E\n        \u003Cul\u003E\n          \u003Cli\u003ERun: {run_ts:%Y-%m-%d %H:%M} UTC\u003C/li\u003E\n          \u003Cli\u003EChampion model (official provision): {MODEL_NAME} {version}\u003C/li\u003E\n          \u003Cli\u003ELoans scored: {portfolio['LOANS']:,} — mean PD {portfolio['MEAN_PD']:.4f}\u003C/li\u003E\n          \u003Cli\u003EProbability-weighted ECL: \u003Cb\u003E{weighted:,.0f}\u003C/b\u003E ({change})\u003C/li\u003E\n        \u003C/ul\u003E\n        \u003Ctable style=\"border-collapse:collapse\"\u003E\n          \u003Ctr\u003E\u003Cth style=\"{cell}\"\u003EScenario\u003C/th\u003E\u003Cth style=\"{cell}\"\u003EWeight\u003C/th\u003E\n              \u003Cth style=\"{cell}\"\u003EMean PD\u003C/th\u003E\u003Cth style=\"{cell}\"\u003EECL\u003C/th\u003E\u003C/tr\u003E{rows}\n        \u003C/table\u003E\n        {comparison}\n        \u003Cp\u003EECL uses LGD = {LGD_ASSUMPTION} and BALANCE as EAD (illustrative).\n           Results: {DB}.{SCHEMA}.PD_PORTFOLIO_SCORES, PD_WEIGHTED_HISTORY, PD_ECL_HISTORY (filter on MODEL_ROLE).\u003C/p\u003E\n    \"\"\"\n    session.sql(\"CALL SYSTEM$SEND_EMAIL(?, ?, ?, ?, 'text/html')\",\n                params=[REVIEW_EMAIL_INTEGRATION, \",\".join(SCORING_NOTIFY_EMAILS), subject, body]).collect()\n    return f\"Scoring summary for {run_ts} sent to {', '.join(SCORING_NOTIFY_EMAILS)}\"\n\n\n# SCORING_NOTIFY_EMAILS is configured in the Setup cell (section 1)\n\nfor fn in [sp_weekly_batch_score, sp_weekly_scenario_ecl, sp_weekly_scoring_notify]:\n    session.sproc.register(\n        fn, name=f\"{DB}.{SCHEMA}.{fn.__name__.upper()}\",\n        packages=SP_PACKAGES, imports=[MODEL_MODULE], is_permanent=True, replace=True,\n        stage_location=f\"@{DB}.{SCHEMA}.ML_ARTIFACTS\",\n        # Owner's-rights procedures cannot create temporary tables; the scenario step uses one\n        execute_as=\"caller\" if fn is sp_weekly_scenario_ecl else \"owner\",\n    )\n    print(f\"Registered procedure {fn.__name__.upper()}\")\n```\n\nCreate the weekly DAG:\n\n```python\n# Weekly scoring Task DAG — Mondays 03:00 New York, independent of the monthly retraining DAG\nweekly_dag = [\n    # (task name, predecessor, procedure)\n    (\"PD_MODEL_SCORING_PIPELINE\", None, \"SP_REFRESH_FEATURES\"),\n    (\"PD_MODEL_SCORING_PIPELINE_BATCH_SCORE\", \"PD_MODEL_SCORING_PIPELINE\", \"SP_WEEKLY_BATCH_SCORE\"),\n    (\"PD_MODEL_SCORING_PIPELINE_SCENARIO_ECL\", \"PD_MODEL_SCORING_PIPELINE_BATCH_SCORE\", \"SP_WEEKLY_SCENARIO_ECL\"),\n    (\"PD_MODEL_SCORING_PIPELINE_NOTIFY\", \"PD_MODEL_SCORING_PIPELINE_SCENARIO_ECL\", \"SP_WEEKLY_SCORING_NOTIFY\"),\n]\n\n# The graph can only be changed while the root is suspended\nsession.sql(f\"ALTER TASK IF EXISTS {DB}.{SCHEMA}.PD_MODEL_SCORING_PIPELINE SUSPEND\").collect()\n# Remove the pre-rename names of this DAG, children first\nfor old_task in [\"SCORING_PIPELINE_NOTIFY\", \"SCORING_PIPELINE_SCENARIO_ECL\", \"SCORING_PIPELINE_BATCH_SCORE\",\n                 \"SCORING_PIPELINE_FEATURE_REFRESH\"]:\n    session.sql(f\"DROP TASK IF EXISTS {DB}.{SCHEMA}.{old_task}\").collect()\n\nfor task, after, proc in weekly_dag:\n    trigger = (\"SCHEDULE = 'USING CRON 0 3 * * 1 America/New_York'\" if after is None\n               else f\"AFTER {DB}.{SCHEMA}.{after}\")\n    session.sql(f\"\"\"\n        CREATE OR REPLACE TASK {DB}.{SCHEMA}.{task}\n            WAREHOUSE = {WH}\n            {trigger}\n        AS\n            CALL {DB}.{SCHEMA}.{proc}()\n    \"\"\").collect()\nprint(f\"Weekly scoring DAG created ({len(weekly_dag)} tasks, suspended)\")\n```\n\nActivate it:\n\n```sql\nSELECT SYSTEM$TASK_DEPENDENTS_ENABLE('CREDIT_RISK_ML.PD_MODEL.PD_MODEL_SCORING_PIPELINE');\n```\n\nTo run one cycle on demand:\n\n```sql\nALTER TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_SCORING_PIPELINE_NOTIFY RESUME;\nALTER TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_SCORING_PIPELINE_SCENARIO_ECL RESUME;\nALTER TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_SCORING_PIPELINE_BATCH_SCORE RESUME;\nEXECUTE TASK CREDIT_RISK_ML.PD_MODEL.PD_MODEL_SCORING_PIPELINE;\n```\n\nA completed weekly run looks like this in the task graph:\n\n![Scoring pipeline Task DAG run](https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/score-pipeline-task-dag-run.png?v=54b7b1dd)\n\nEach run ends with a summary email. It shows the champion's weighted ECL and change since the previous week, the per-scenario breakdown, and the challenger comparison:\n\n![Weekly scoring email](https://www.snowflake.com/content/dam/snowflake-site/developers/guides/credit-risk-pd-model-snowflake-ml/scoring-completion-email-notification.png?v=54b7b1dd)\n\n\u003C!-- ------------------------ --\u003E\n## Clean Up\n\nThe final cell removes everything the notebook created:\n- suspends both Task DAGs;\n- drops the account-level notification integration and the `MODEL_REVIEWER` role;\n- drops the `CREDIT_RISK_ML` database and the `CREDIT_RISK_WH` warehouse.\n\n**This is irreversible.** It deletes the model versions, Feature Store, Datasets, experiments, score history and the `PROMOTION_LOG` audit trail. Set `CONFIRM_CLEANUP = True` only when you mean it. The Marketplace databases are left untouched.\n\n```python\n# Tear down all resources created by this notebook — set CONFIRM_CLEANUP = True to run\nCONFIRM_CLEANUP = False\n\nDB, SCHEMA, WH = \"CREDIT_RISK_ML\", \"PD_MODEL\", \"CREDIT_RISK_WH\"\n\nif not CONFIRM_CLEANUP:\n    print(\"Cleanup skipped — set CONFIRM_CLEANUP = True to permanently delete the resources listed above.\")\nelse:\n    def run(sql):\n        try:\n            session.sql(sql).collect()\n            print(f\"OK    {sql}\")\n        except Exception as e:  # keep going so one missing object doesn't block the rest\n            print(f\"SKIP  {sql}  ({str(e).splitlines()[0]})\")\n\n    # 1. Suspend both task graphs (root tasks first) so no run starts during cleanup\n    for root in [\"PD_MODEL_TRAIN_PIPELINE\", \"PD_MODEL_SCORING_PIPELINE\"]:\n        run(f\"ALTER TASK IF EXISTS {DB}.{SCHEMA}.{root} SUSPEND\")\n\n    # 2. Account-level objects, which dropping the database does not remove\n    run(\"DROP NOTIFICATION INTEGRATION IF EXISTS PD_MODEL_REVIEW_EMAIL\")\n    run(\"DROP ROLE IF EXISTS MODEL_REVIEWER\")\n\n    # 3. The database: registry models and monitors, Feature Store, Datasets, experiments,\n    #    tasks, procedures, tables, views and stages all live inside it\n    run(f\"DROP DATABASE IF EXISTS {DB}\")\n\n    # 4. The warehouse (last, since the statements above run on it)\n    run(f\"DROP WAREHOUSE IF EXISTS {WH}\")\n\n    print(\"\\nCleanup complete.\")\n```\n\n\u003C!-- ------------------------ --\u003E\n## Conclusion and Resources\n\nYou migrated a SAS `PROC GRADBOOST` PD model to XGBoost on Snowflake ML and ran its full lifecycle in one governed platform. Data never left Snowflake, every model is traceable to the exact data that trained it, and promotion requires an audited human decision.\n\n| Snowflake ML capability | What it replaces in a SAS stack |\n|---|---|\n| Feature Store (behavioral, macro, scenario Feature Views) | SAS macros generating feature datasets |\n| Versioned Datasets | Manual dataset freezes |\n| Experiment Tracking | Excel/Word model development trackers |\n| Model Registry with `champion` / `challenger` aliases | Version-named `.sas7bdat` files and spreadsheet inventories |\n| Registry SHAP `explain` and ML Lineage | External explainability tooling and provenance documents |\n| `run()` / `run_batch()` batch scoring and scenario-weighted ECL | SAS batch scoring jobs and spreadsheet IFRS 9 overlays |\n| Model Version Monitors | Quarterly manual monitoring reports |\n| Task DAGs with `MODEL_REVIEWER` approval | Cron + shell scripts and email sign-off chains |\n\n### What You Learned\n- Mapping SAS model code and hyperparameters to XGBoost with CoCo\n- Building point-in-time correct features and immutable training Datasets\n- Tracking experiments and governing models with registry aliases, SHAP and lineage\n- Scoring offline champion/challenger and computing scenario-weighted ECL\n- Monitoring drift and performance per model version\n- Orchestrating retraining and scoring with a role-gated, auditable approval step\n","title":"Quickstart Article Body","multiValue":false,":type":"text/x-markdown"},"quickstartArticleLogoImage":{"dataType":"string","title":"Quickstart Article Logo Image","multiValue":false,":type":"text/plain"}},"elementsOrder":["quickstartArticleBody","quickstartArticleLogoImage"],"isDeveloperGuidesPage":false,"model":"snowflake-site/models/quickstart-article"},"flexible_column_cont":{"id":"flexible-column-container-7c3b6cead1","type":"2-column-75-25","alignColumns":"top","containerMaxWidth":"extra-large","topPadding":"none","bottomPadding":"none","spaceBetween":"none","reverseOnMobile":false,"carouselOnMobile":false,"backgroundImageOption":"none","flexible_column_content_container_1":{"layout":"SIMPLE","id":"container-2f156aaf80",":items":{"quickstart_last_modi":{"id":"quickstart-last-modified-29b41a8870","icon":{"id":"icon","icon":"calendar",":type":"snowflake-site/components/icon","appliedCssClassNames":"snowflake-icon-blue"},"lastModifiedDatePrefix":"Updated","lastModifiedDate":"2026-11-12",":type":"snowflake-site/components/quickstart/quickstart-last-modified","appliedCssClassNames":"snowflake-responsive-component-top-padding-small"},"text":{"id":"text-1177045a58","additionalClasses":"qs-disclaimer-text","text":"\u003Cp\u003E\u003Cspan style=\"color: #666;\"\u003EThis content is provided as is, and is not maintained on an ongoing basis. It may be out of date with current Snowflake instances\u003C/span\u003E\u003C/p\u003E\r\n","richText":true,":type":"snowflake-site/components/text","appliedCssClassNames":"snowflake-responsive-component-top-padding-small"}},":type":"snowflake-site/components/flexible-column-container/flexible-column-content-container",":itemsOrder":["quickstart_last_modi","text"]},"flexible_column_content_container_2":{"layout":"SIMPLE","id":"container-9cad73a472",":items":{},":type":"snowflake-site/components/flexible-column-container/flexible-column-content-container",":itemsOrder":[]},":type":"snowflake-site/components/flexible-column-container","isBlogPage":false,"isActiveTOC":false}},":type":"snowflake-site/components/flexible-column-container/flexible-column-content-container",":itemsOrder":["contentfragment","flexible_column_cont"]},"flexible_column_content_container_2":{"layout":"SIMPLE","id":"container-613a6d598a",":items":{"quickstart_table_of_":{"layout":"SIMPLE","id":"container-1f74570e66","isDeveloperGuidesPage":false,":items":{"quickstart_table_of_":{"id":"quickstart-table-of-content-f8963b7f1a","headings":["\u003Ch2\u003EOverview\u003C/h2\u003E","\u003Ch2\u003EGet Data and Notebook\u003C/h2\u003E","\u003Ch2\u003EConvert SAS with CoCo\u003C/h2\u003E","\u003Ch2\u003ESet Up the Environment\u003C/h2\u003E","\u003Ch2\u003EExplore the Data\u003C/h2\u003E","\u003Ch2\u003EPrepare the Data\u003C/h2\u003E","\u003Ch2\u003EBuild the Feature Store\u003C/h2\u003E","\u003Ch2\u003EGenerate a Versioned Training Dataset\u003C/h2\u003E","\u003Ch2\u003ETrain with Experiment Tracking\u003C/h2\u003E","\u003Ch2\u003ERegister and Explain Model\u003C/h2\u003E","\u003Ch2\u003EBatch Score and ECL\u003C/h2\u003E","\u003Ch2\u003EMonitor the Models\u003C/h2\u003E","\u003Ch2\u003EAutomate Monthly Retraining\u003C/h2\u003E","\u003Ch2\u003EAutomate Weekly Scoring\u003C/h2\u003E","\u003Ch2\u003EClean Up\u003C/h2\u003E","\u003Ch2\u003EConclusion and Resources\u003C/h2\u003E"],":type":"snowflake-site/components/quickstart/quickstart-table-of-content","fragmentPath":"/content/dam/snowflake-site/en/content-fragments/quickstarts/credit-risk-pd-model-snowflake-ml"},"quickstart_button":{"id":"quickstart-button-f88f71da6f",":type":"snowflake-site/components/quickstart/quickstart-button","fragmentPath":"/content/dam/snowflake-site/en/content-fragments/quickstarts/credit-risk-pd-model-snowflake-ml","appliedCssClassNames":"snowflake-responsive-component-top-padding-none"}},":type":"snowflake-site/components/quickstart/quickstart-table-of-content/quickstart-table-of-content-container",":itemsOrder":["quickstart_table_of_","quickstart_button"]}},":type":"snowflake-site/components/flexible-column-container/flexible-column-content-container",":itemsOrder":["quickstart_table_of_"]},":type":"snowflake-site/components/flexible-column-container","isBlogPage":false,"isActiveTOC":false},"markup_editor":{"id":"markup-editor-f1be79a0fa","title":"Page CSS","cssUrl":"/content/snowflake-site/global/en/developers/guides/credit-risk-pd-model-snowflake-ml/_jcr_content/root/responsivegrid/markup_editor.eb24a609da50a43b.css",":type":"snowflake-site/components/markup-editor","isGSAPEnabled":false}},":itemsOrder":["quickstart_hero","flexible_column_cont","markup_editor"],":type":"wcm/foundation/components/responsivegrid"},"modal_container":{"layout":"SIMPLE","id":"container-3f29add385",":items":{},":type":"snowflake-site/components/modal/modal-container",":itemsOrder":[]},"experiencefragment-footer":{"id":"experiencefragment-0f0a6394e8","localizedFragmentVariationPath":"/content/experience-fragments/snowflake-site/language-masters/en/site/footer/master/jcr:content","configured":true,":items":{"root":{"additionalClasses":"sf-footer","layout":"SIMPLE","id":"container-97cabec366",":items":{"container_copy":{"additionalClasses":"sf-footer__inner","layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"flexible_column_cont":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-cc93e50b04",":items":{"flexible_column_cont":{"id":"flexible-column-container-7cd53fe9b4","type":"1-column","alignColumns":"top","containerMaxWidth":"extra-large","topPadding":"medium","bottomPadding":"extra-small","spaceBetween":"small","reverseOnMobile":false,"carouselOnMobile":false,"propertiesCSSClasses":"sf-footer-grid","backgroundImageOption":"none","flexible_column_content_container_1":{"layout":"SIMPLE","id":"container-a28f4d73c6",":items":{"container":{"additionalClasses":"sf-footer-grid__inner","layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"container":"aem-GridColumn aem-GridColumn--default--12","container_1622723482":"aem-GridColumn aem-GridColumn--default--12","container_copy_copy_":"aem-GridColumn aem-GridColumn--default--12","container_copy_copy":"aem-GridColumn aem-GridColumn--default--12","container_copy":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-e89935230c",":items":{"container_1622723482":{"additionalClasses":"sf-footer__column","layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"container":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-ea1c4a40f7",":items":{"container":{"additionalClasses":"sf-footer__newsletter-group","layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"text":"aem-GridColumn aem-GridColumn--default--12","marketo_v2":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-c378dd0da0",":items":{"text":{"id":"text-2422b74d79","additionalClasses":"sf-footer__newsletter-title","text":"\u003Cp\u003E\u003Cb\u003ESubscribe to our monthly newsletter\u003C/b\u003E\u003C/p\u003E\r\n\u003Cp\u003EStay up to date on Snowflake’s latest products, expert insights and resources—right in your inbox!\u003C/p\u003E\r\n","richText":true,":type":"snowflake-site/components/text","appliedCssClassNames":"text-size-regular text-color-text-04"},"marketo_v2":{"id":"marketo-v2-301a333f52","marketoForm":{"hidden":null,"formId":"45871","successUrl":null,"edit":false,"script":null,"values":null},"munchkinId":"252-RFO-227","serverInstance":"252-RFO-227.mktoweb.com","partnerCampaignPage":false,"marketoConfigured":true,"formConfigured":true,":type":"snowflake-site/components/form/marketo-v2"}},":type":"snowflake-site/components/container",":itemsOrder":["text","marketo_v2"],"appliedCssClassNames":"snowflake-responsive-container-inner-padding-small"}},":type":"snowflake-site/components/container",":itemsOrder":["container"],"appliedCssClassNames":"snowflake-responsive-container-inner-padding-small"},"container":{"layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"text_copy":"aem-GridColumn aem-GridColumn--default--12","text":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-ee9dd62da1",":items":{"text":{"id":"text-1e929f3eb6","additionalClasses":"sf-footer__link-group","text":"\u003Cp class=\"sf-footer__column-title\"\u003EProduct\u003C/p\u003E\r\n\u003Cul\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/product/platform/\"\u003EPlatform\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/product/snowflake-cowork/\"\u003ESnowflake CoWork\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/product/data-engineering/\"\u003EData Engineering\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/product/analytics/\"\u003EAnalytics\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/product/ai/\"\u003EAI\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/product/applications-and-collaboration/\"\u003EApplications & Collaboration\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/pricing-options/\"\u003EPricing\u003C/a\u003E\u003C/li\u003E\r\n\u003C/ul\u003E\r\n","richText":true,":type":"snowflake-site/components/text","appliedCssClassNames":"text-size-small text-color-text-04"},"text_copy":{"id":"text-732187f0fa","additionalClasses":"sf-footer__link-group","text":"\u003Cp class=\"sf-footer__column-title\"\u003ESupport\u003C/p\u003E\r\n\u003Cul\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/support/\"\u003ESupport\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/legal/addenda/priority-support-services-description/\"\u003EPriority Support\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"https://status.snowflake.com/\" target=\"_blank\" rel=\"noopener noreferrer\"\u003EStatus\u003C/a\u003E\u003C/li\u003E\r\n\u003C/ul\u003E\r\n","richText":true,":type":"snowflake-site/components/text","appliedCssClassNames":"text-size-small text-color-text-04"}},":type":"snowflake-site/components/container",":itemsOrder":["text","text_copy"],"appliedCssClassNames":"snowflake-responsive-container-inner-padding-medium"},"container_copy_copy":{"layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"text":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-c9f94f8a82",":items":{"text":{"id":"text-6faf2dee57","additionalClasses":"sf-footer__link-group","text":"\u003Cp class=\"sf-footer__column-title\"\u003E\u003Ca href=\"/en/solutions/industries/\"\u003EIndustries\u003C/a\u003E\u003C/p\u003E\r\n\u003Cul\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/solutions/industries/advertising-media-entertainment/\"\u003EAdvertising, Media & Entertainment\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/solutions/industries/financial-services/\"\u003EFinancial Services\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/solutions/industries/healthcare-and-life-sciences/\"\u003EHealthcare & Life Sciences\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/solutions/industries/manufacturing/\"\u003EManufacturing\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/solutions/industries/public-sector/\"\u003EPublic Sector\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/solutions/industries/retail-consumer-goods/\"\u003ERetail & Consumer Goods\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/solutions/industries/telecom/\"\u003ETelecom\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/solutions/industries/technology/\"\u003ETechnology\u003C/a\u003E\u003C/li\u003E\r\n\u003C/ul\u003E\r\n","richText":true,":type":"snowflake-site/components/text","appliedCssClassNames":"text-size-small text-color-text-04"}},":type":"snowflake-site/components/container",":itemsOrder":["text"],"appliedCssClassNames":"snowflake-responsive-container-inner-padding-small"},"container_copy":{"layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"text":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-91c85eddef",":items":{"text":{"id":"text-2784e2f1f9","additionalClasses":"sf-footer__link-group","text":"\u003Cp class=\"sf-footer__column-title\"\u003ECompany\u003C/p\u003E\r\n\u003Cul\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/company/overview/about-snowflake/\"\u003EAbout Snowflake\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/company/overview/leadership-and-board/\"\u003ELeadership & Board\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca target=\"_blank\" href=\"https://careers.snowflake.com/us/en\" rel=\"noopener noreferrer\"\u003ECareers\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca target=\"_blank\" href=\"https://investors.snowflake.com/overview/default.aspx\" rel=\"noopener noreferrer\"\u003EInvestor Relations\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca target=\"_blank\" href=\"https://trust.snowflake.com/\" rel=\"noopener noreferrer\"\u003ETrust Center\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca target=\"_blank\" href=\"https://www.snowflake.com/brand-guidelines/\" rel=\"noopener noreferrer\"\u003EBrand Guidelines\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/contact/\"\u003EContact\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/news/\"\u003ENewsroom\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/company/overview/esg/\"\u003EEnvironmental, Social & Governance\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/solutions/frontier-engineering/\"\u003EFrontier Engineering\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/company/overview/snowflake-ventures/\"\u003ESnowflake Ventures\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/company/overview/end-data-disparity/\"\u003EEnd Data Disparity\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/summit/\"\u003ESnowflake Summit 27\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E \u003C/li\u003E\r\n\u003C/ul\u003E\r\n","richText":true,":type":"snowflake-site/components/text","appliedCssClassNames":"text-size-small text-color-text-04"}},":type":"snowflake-site/components/container",":itemsOrder":["text"],"appliedCssClassNames":"snowflake-responsive-container-inner-padding-small"},"container_copy_copy_":{"layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"text":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-2f4d3f787d",":items":{"text":{"id":"text-891bf4dad2","additionalClasses":"sf-footer__link-group","text":"\u003Cp class=\"sf-footer__column-title\"\u003ELearn\u003C/p\u003E\r\n\u003Cul\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/resources/\"\u003EResource Library\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/webinars/demo/\"\u003ELive Demos\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/fundamentals/\"\u003EFundamentals\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/resources/learn/training/\"\u003ETraining\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/resources/learn/certifications/\"\u003ECertifications\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca target=\"_blank\" href=\"https://learn.snowflake.com/en/\" rel=\"noopener noreferrer\"\u003ESnowflake University\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/developers/guides/\"\u003EDeveloper Guides\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca target=\"_blank\" href=\"https://docs.snowflake.com/\" rel=\"noopener noreferrer\"\u003EDocumentation\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"/en/data-governance/\"\u003EData Governance\u003C/a\u003E\u003C/li\u003E\r\n\u003C/ul\u003E\r\n","richText":true,":type":"snowflake-site/components/text","appliedCssClassNames":"text-size-small text-color-text-04"}},":type":"snowflake-site/components/container",":itemsOrder":["text"],"appliedCssClassNames":"snowflake-responsive-container-inner-padding-small"}},":type":"snowflake-site/components/container",":itemsOrder":["container_1622723482","container","container_copy_copy","container_copy","container_copy_copy_"],"appliedCssClassNames":"snowflake-responsive-container-inner-padding-small"}},":type":"snowflake-site/components/flexible-column-container/flexible-column-content-container",":itemsOrder":["container"]},":type":"snowflake-site/components/flexible-column-container","isBlogPage":false,"isActiveTOC":false}},":type":"snowflake-site/components/container",":itemsOrder":["flexible_column_cont"],"appliedCssClassNames":"snowflake-container snowflake-responsive-container-inner-padding-small"},"container_573483281_":{"additionalClasses":"sf-footer__bottom","layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"container_112062425":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-960169caaf",":items":{"container_112062425":{"layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"flexible_column_cont":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-b3e0cc85b2",":items":{"flexible_column_cont":{"id":"flexible-column-container-c1e0e04521","type":"1-column","alignColumns":"top","containerMaxWidth":"extra-large","topPadding":"none","bottomPadding":"none","spaceBetween":"small","reverseOnMobile":false,"carouselOnMobile":false,"backgroundImageOption":"none","flexible_column_content_container_1":{"layout":"SIMPLE","id":"container-79dedcdc35",":items":{"container":{"additionalClasses":"sf-footer__legal-container","layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"container":"aem-GridColumn aem-GridColumn--default--12","text_copy_copy_16360":"aem-GridColumn aem-GridColumn--default--12","markup_editor":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-e6c642dc60",":items":{"container":{"layout":"RESPONSIVE_GRID","columnCount":12,"columnClassNames":{"image":"aem-GridColumn aem-GridColumn--default--12"},"gridClassNames":"aem-Grid aem-Grid--12 aem-Grid--default--12","id":"container-04dd31f15b",":items":{"image":{"id":"image-f977909ad8","additionalClasses":"sf-footer__logo","height":"64","alt":"Snowflake logo","imageLink":{"valid":true,"url":"/en/"},"src":"/content/experience-fragments/snowflake-site/language-masters/en/site/footer/master/_jcr_content/root/container_573483281_/container_112062425/flexible_column_cont/flexible_column_content_container_1/container/container/image.coreimg.svg/1747882370694/nav-icon-snowflake-bug.svg","lazyEnabled":true,"isLcpImage":false,"width":"64",":type":"snowflake-site/components/image"}},":type":"snowflake-site/components/container",":itemsOrder":["image"],"appliedCssClassNames":"snowflake-responsive-container-inner-padding-extra-small"},"text_copy_copy_16360":{"id":"text-67093964cf","additionalClasses":"sf-footer__legal-links","text":"\u003Cul\u003E\r\n\u003Cli\u003E© 2026 Snowflake Inc. All Rights Reserved\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"https://www.snowflake.com/en/legal/privacy/privacy-policy/\"\u003EPrivacy Policy\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"https://snowflake.com/en/legal/snowflake-site-terms/\"\u003ESite Terms\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"https://info.snowflake.com/Preference-center.html\"\u003ECommunication Preferences\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Cbutton id=\"ot-sdk-btn\" class=\"ot-sdk-show-settings\"\u003ECookie Settings\u003C/button\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"https://www.snowflake.com/en/legal/privacy/privacy-policy/#12\"\u003EDo Not Share My Personal Information\u003C/a\u003E\u003C/li\u003E\r\n\u003Cli\u003E\u003Ca href=\"https://www.snowflake.com/en/legal/\"\u003ELegal\u003C/a\u003E\u003C/li\u003E\r\n\u003C/ul\u003E\r\n","richText":true,":type":"snowflake-site/components/text","appliedCssClassNames":"text-size-small text-color-text-04"},"markup_editor":{"id":"markup-editor-d12104dc65","title":" ","htmlContent":"\u003Cdiv class=\"sf-footer__social\"\u003E\r\n\u003Cdiv data-testid=\"snowflake-footer-twitter\" class=\"snowflake-button-icon snowflake-button-white snowflake-footer-social-item\"\u003E\u003Cdiv class=\"snowflake-button-icon \"\u003E\u003Ca href=\"https://x.com/Snowflake\" data-testid=\"button-external\" aria-label=\"X (Twitter)\" role=\"button\" class=\"snowflake-button-container\" title=\"X (Twitter)\" tabindex=\"0\" target=\"_blank\" rel=\"noreferrer\"\u003E\u003Cdiv data-testid=\"button-icon-wrapper\"\u003E\u003Csvg xmlns=\"http://www.w3.org/2000/svg\" fill=\"none\" viewBox=\"0 0 59 53\" class=\"button-icon\"\u003E\u003Cpath fill=\"currentColor\" d=\"M46.614 0h9.044L35.8 22.49 59 53H40.795L26.54 34.46 10.223 53H1.18l21.036-24.055L0 0h18.657l12.878 16.937zM43.45 47.72h5.013L16.023 5.085h-5.387z\"\u003E\u003C/path\u003E\u003C/svg\u003E\u003C/div\u003E\u003C/a\u003E\u003Cdiv\u003E\u003C/div\u003E\u003C/div\u003E\u003C/div\u003E\u003Cdiv data-testid=\"snowflake-footer-linkedin\" class=\"snowflake-button-icon snowflake-button-white snowflake-footer-social-item\"\u003E\u003Cdiv class=\"snowflake-button-icon \"\u003E\u003Ca href=\"https://www.linkedin.com/company/3653845\" data-testid=\"button-external\" aria-label=\"LinkedIn\" role=\"button\" class=\"snowflake-button-container\" title=\"LinkedIn\" tabindex=\"0\" target=\"_blank\" rel=\"noreferrer\"\u003E\u003Cdiv data-testid=\"button-icon-wrapper\"\u003E\u003Csvg xmlns=\"http://www.w3.org/2000/svg\" fill=\"currentColor\" viewBox=\"0 0 24 24\" class=\"button-icon\"\u003E\u003Cpath d=\"M22.223 0H1.772C.792 0 0 .773 0 1.73v20.536C0 23.222.792 24 1.772 24h20.451c.98 0 1.777-.778 1.777-1.73V1.73C24 .773 23.203 0 22.223 0ZM7.12 20.452H3.558V8.995H7.12v11.457ZM5.34 7.434a2.064 2.064 0 1 1 0-4.125 2.063 2.063 0 0 1 0 4.125Zm15.112 13.018h-3.558v-5.57c0-1.326-.024-3.037-1.852-3.037-1.851 0-2.133 1.449-2.133 2.944v5.663H9.356V8.995h3.413v1.566h.047c.473-.9 1.636-1.852 3.365-1.852 3.605 0 4.27 2.372 4.27 5.457v6.286Z\"\u003E\u003C/path\u003E\u003C/svg\u003E\u003C/div\u003E\u003C/a\u003E\u003Cdiv\u003E\u003C/div\u003E\u003C/div\u003E\u003C/div\u003E\u003Cdiv data-testid=\"snowflake-footer-facebook\" class=\"snowflake-button-icon snowflake-button-white snowflake-footer-social-item\"\u003E\u003Cdiv class=\"snowflake-button-icon \"\u003E\u003Ca href=\"https://www.facebook.com/snowflakedb/\" data-testid=\"button-external\" aria-label=\"Facebook\" role=\"button\" class=\"snowflake-button-container\" title=\"Facebook\" tabindex=\"0\" target=\"_blank\" rel=\"noreferrer\"\u003E\u003Cdiv data-testid=\"button-icon-wrapper\"\u003E\u003Csvg xmlns=\"http://www.w3.org/2000/svg\" fill=\"currentColor\" viewBox=\"0 0 24 24\" class=\"button-icon\"\u003E\u003Cpath d=\"M24 12c0-6.627-5.373-12-12-12S0 5.373 0 12c0 5.99 4.388 10.954 10.125 11.854V15.47H7.078V12h3.047V9.356c0-3.007 1.792-4.668 4.533-4.668 1.312 0 2.686.234 2.686.234v2.953H15.83c-1.491 0-1.956.925-1.956 1.875V12h3.328l-.532 3.469h-2.796v8.385C19.612 22.954 24 17.99 24 12Z\"\u003E\u003C/path\u003E\u003C/svg\u003E\u003C/div\u003E\u003C/a\u003E\u003Cdiv\u003E\u003C/div\u003E\u003C/div\u003E\u003C/div\u003E\u003Cdiv data-testid=\"snowflake-footer-youtube\" class=\"snowflake-button-icon snowflake-button-white snowflake-footer-social-item\"\u003E\u003Cdiv class=\"snowflake-button-icon \"\u003E\u003Ca href=\"https://www.youtube.com/user/snowflakecomputing\" data-testid=\"button-external\" aria-label=\"YouTube\" role=\"button\" class=\"snowflake-button-container\" title=\"YouTube\" tabindex=\"0\" target=\"_blank\" rel=\"noreferrer\"\u003E\u003Cdiv data-testid=\"button-icon-wrapper\"\u003E\u003Csvg xmlns=\"http://www.w3.org/2000/svg\" fill=\"currentColor\" viewBox=\"0 0 24 24\" class=\"button-icon\"\u003E\u003Cpath d=\"M23.76 7.2s-.233-1.655-.955-2.381c-.914-.956-1.936-.961-2.405-1.017-3.356-.244-8.395-.244-8.395-.244h-.01s-5.039 0-8.395.244c-.469.056-1.49.06-2.405 1.017C.473 5.545.244 7.2.244 7.2S0 9.145 0 11.086v1.819c0 1.94.24 3.886.24 3.886s.233 1.654.95 2.38c.915.957 2.115.924 2.65 1.027 1.92.183 8.16.24 8.16.24s5.044-.01 8.4-.249c.469-.056 1.49-.06 2.405-1.017.722-.727.956-2.381.956-2.381S24 14.85 24 12.905v-1.819c0-1.94-.24-3.886-.24-3.886ZM9.52 15.113V8.367l6.483 3.385-6.483 3.36Z\"\u003E\u003C/path\u003E\u003C/svg\u003E\u003C/div\u003E\u003C/a\u003E\u003Cdiv\u003E\u003C/div\u003E\u003C/div\u003E\u003C/div\u003E\r\n\u003C/div\u003E",":type":"snowflake-site/components/markup-editor","isGSAPEnabled":false}},":type":"snowflake-site/components/container",":itemsOrder":["container","text_copy_copy_16360","markup_editor"],"appliedCssClassNames":"snowflake-responsive-container-inner-padding-none"}},":type":"snowflake-site/components/flexible-column-container/flexible-column-content-container",":itemsOrder":["container"]},":type":"snowflake-site/components/flexible-column-container","isBlogPage":false,"isActiveTOC":false}},":type":"snowflake-site/components/container",":itemsOrder":["flexible_column_cont"],"appliedCssClassNames":"snowflake-container snowflake-responsive-container-inner-padding-small"}},":type":"snowflake-site/components/container",":itemsOrder":["container_112062425"],"appliedCssClassNames":"snowflake-responsive-container-inner-padding-none"},"markup_editor_copy":{"id":"markup-editor-b7c6837009","title":"New css","cssContent":".snowflake-flexible-column-container-items:has(.snowflake-experience-fragment:empty){display:none !important}.snowflake-image-container img{background-color:transparent}div.snowflake-person-chip-avatar{width:80px !important}#snowflake-blog-template-main-container .snowflake-quote-item-card{margin-top:40px}#snowflake-blog-template-main-container .aem-GridColumn:has(.vertical-video){background-color:#000;border-radius:16px;overflow:hidden}#snowflake-blog-template-main-container .is-vertical img{max-width:400px;margin-left:auto;margin-right:auto}#snowflake-blog-template-main-container .vertical-video{max-width:240px;margin-left:auto;margin-right:auto}@media screen and (min-width:1367px){.dynamic .heading-1-v2 .snowflake-title-v2-line{font-size:72px !important;line-height:60px !important}}.snowflake-flexible-column-container-items-alignment-match-height .download-card,.snowflake-flexible-column-container-items-alignment-match-height .download-card\u003E.container{height:100%}.download-card div.code-toolbar\u003E.toolbar .copy-to-clipboard-button{background-color:white;border:1px solid #a9e1f6;margin-right:4px;top:6px;border-radius:16px;height:26px;width:40px}.download-card .snowflake-code-snippet\u003Ediv.code-toolbar\u003E.toolbar\u003E.toolbar-item\u003Ebutton:before{content:'';background-image:url(\"data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' width='16' height='16' viewBox='0 0 24 24' fill='none' stroke-width='2' stroke-linecap='round' stroke-linejoin='round'%3E%3Crect x='9' y='9' width='13' height='13' rx='2' ry='2' style='stroke:%23249EDC;'%3E%3C/rect%3E%3Cpath d='M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1' style='stroke:%23249EDC;'%3E%3C/path%3E%3C/svg%3E\");background-size:auto 65%;background-position:center;background-repeat:no-repeat;top:0;left:0;width:100%;height:100%}.download-card .snowflake-code-snippet\u003Ediv.code-toolbar\u003E.toolbar\u003E.toolbar-item\u003Ebutton:hover:before{background-image:url(\"data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' width='16' height='16' viewBox='0 0 24 24' fill='none' stroke-width='2' stroke-linecap='round' stroke-linejoin='round'%3E%3Crect x='9' y='9' width='13' height='13' rx='2' ry='2' style='stroke:%23fff;'%3E%3C/rect%3E%3Cpath d='M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1' style='stroke:%23fff;'%3E%3C/path%3E%3C/svg%3E\")}.download-card\u003Ediv{background-color:#fff;border:1px solid #ccc;border-radius:8px;padding:24px}.download-chip__headline{border-bottom:1px solid #ccc;padding-bottom:16px;margin-bottom:16px}.download-chip{padding:8px 12px !important;border-radius:4px;transition:300ms ease background-color}.download-chip .black-blue-text-color .snowflake-title-v2-line{color:#000 !important;padding-right:24px;font-family:'Lato',sans-serif;font-size:14px !important;font-weight:500 !important}.download-chip .black-blue-text-color .snowflake-title-v2-line:not(:first-child){opacity:.6;font-style:italic !important}.download-chip .snowflake-content-chip-button{display:none}.download-chip.is-external-link{background-size:16px 16px;background-image:url(\"data:image/svg+xml,%3Csvg width='15' height='15' viewBox='0 0 15 15' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M1.06055 13.0607L11.8605 2.26067M13.0605 10.6607V1.06067H3.46055' stroke='%23249EDC' stroke-width='2.12132' stroke-linecap='round' stroke-linejoin='round'/%3E%3C/svg%3E%0A\")}.download-chip{background-image:url(\"data:image/svg+xml,%3Csvg width='18' height='18' viewBox='0 0 18 18' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cg clip-path='url(%23clip0_883_7979)'%3E%3Cpath d='M3.375 16.875H14.625' stroke='%23249EDC' stroke-width='1.40625' stroke-linecap='round' stroke-linejoin='round'/%3E%3Cpath d='M9 1.125V11.25' stroke='%23249EDC' stroke-width='1.40625' stroke-linecap='round' stroke-linejoin='round'/%3E%3Cpath d='M4.5 7.875L9 12.375L13.5 7.875' stroke='%23249EDC' stroke-width='1.40625' stroke-linecap='round' stroke-linejoin='round'/%3E%3C/g%3E%3Cdefs%3E%3CclipPath id='clip0_883_7979'%3E%3Crect width='18' height='18' fill='white'/%3E%3C/clipPath%3E%3C/defs%3E%3C/svg%3E%0A\");background-size:24px auto;background-repeat:no-repeat;background-position:calc(100% - 12px) center}.download-chip__headline{display:flex;gap:16px;flex-direction:row !important;flex-wrap:nowrap}.download-chip__headline::before{content:'';display:inline-block;width:24px;height:24px;background-position:center;background-image:url(\"data:image/svg+xml,%3Csvg width='21' height='21' viewBox='0 0 21 21' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M7.50005 9.89999C8.13657 9.89999 8.74702 9.64713 9.19711 9.19704C9.64719 8.74696 9.90005 8.13651 9.90005 7.49999V2.69999C9.90005 2.06347 9.64719 1.45302 9.19711 1.00293C8.74702 .552844 8.13657 .299988 7.50005 .299988H2.70005C2.06353 .299988 1.45308 .552844 1.00299 1.00293C.552905 1.45302 .300049 2.06347 .300049 2.69999V7.49999C.300049 8.13651 .552905 8.74696 1.00299 9.19704C1.45308 9.64713 2.06353 9.89999 2.70005 9.89999H7.50005ZM7.50005 7.49999H2.70005V2.69999H7.50005V7.49999Z' fill='%23249EDC' stroke='white' stroke-width='.6'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M7.50005 20.3C8.13657 20.3 8.74702 20.0472 9.19711 19.5971C9.64719 19.147 9.90005 18.5365 9.90005 17.9V13.1C9.90005 12.4635 9.64719 11.853 9.19711 11.403C8.74702 10.9529 8.13657 10.7 7.50005 10.7H2.70005C2.06353 10.7 1.45308 10.9529 1.00299 11.403C.552905 11.853 .300049 12.4635 .300049 13.1V17.9C.300049 18.5365 .552905 19.147 1.00299 19.5971C1.45308 20.0472 2.06353 20.3 2.70005 20.3H7.50005ZM7.50005 17.9H2.70005V13.1H7.50005V17.9Z' fill='%23249EDC' stroke='white' stroke-width='.6'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M17.9001 9.89999C18.5366 9.89999 19.147 9.64713 19.5971 9.19704C20.0472 8.74696 20.3001 8.13651 20.3001 7.49999V2.69999C20.3001 2.06347 20.0472 1.45302 19.5971 1.00293C19.147 .552844 18.5366 .299988 17.9001 .299988H13.1001C12.4636 .299988 11.8531 .552844 11.403 1.00293C10.9529 1.45302 10.7001 2.06347 10.7001 2.69999V7.49999C10.7001 8.13651 10.9529 8.74696 11.403 9.19704C11.8531 9.64713 12.4636 9.89999 13.1001 9.89999H17.9001ZM17.9001 7.49999H13.1001V2.69999H17.9001V7.49999Z' fill='%23249EDC' stroke='white' stroke-width='.6'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M17.9001 20.3C18.5366 20.3 19.147 20.0472 19.5971 19.5971C20.0472 19.147 20.3001 18.5365 20.3001 17.9V13.1C20.3001 12.4635 20.0472 11.853 19.5971 11.403C19.147 10.9529 18.5366 10.7 17.9001 10.7H13.1001C12.4636 10.7 11.8531 10.9529 11.403 11.403C10.9529 11.853 10.7001 12.4635 10.7001 13.1V17.9C10.7001 18.5365 10.9529 19.147 11.403 19.5971C11.8531 20.0472 12.4636 20.3 13.1001 20.3H17.9001ZM17.9001 17.9H13.1001V13.1H17.9001V17.9Z' fill='%23249EDC' stroke='white' stroke-width='.6'/%3E%3C/svg%3E%0A\");background-size:contain;background-repeat:no-repeat}.download-chip__headline.is-cli::before{background-image:url(\"data:image/svg+xml,%3Csvg width='24' height='24' viewBox='0 0 24 24' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M4 17L10 11L4 5' stroke='%23000' stroke-width='2' stroke-linecap='round' stroke-linejoin='round'/%3E%3Cpath d='M12 19H20' stroke='%23000' stroke-width='2' stroke-linecap='round' stroke-linejoin='round'/%3E%3C/svg%3E%0A\")}.download-card pre[class*=language-]{padding:8px 12px;background-color:var(--ui-background-05);overflow:hidden}.download-chip__headline.is-windows,.download-chip__headline.is-mac{gap:12px}.download-chip__headline.is-windows::before{width:16px;height:20px;background-image:url(\"data:image/svg+xml,%3Csvg width='4875' height='4875' viewBox='0 0 4875 4875' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cg clip-path='url(%23clip0_122_201)'%3E%3Cpath d='M0 0H2311V2310H0V0ZM2564 0H4875V2310H2564V0ZM0 2564H2311V4875H0V2564ZM2564 2564H4875V4875H2564' fill='%23000'/%3E%3C/g%3E%3C/svg%3E\")}.download-chip__headline.is-mac::before{width:16px;height:20px;background-image:url(\"data:image/svg+xml,%3Csvg version='1.1' id='Layer_1' xmlns:x='ns_extend;' xmlns:i='ns_ai;' xmlns:graph='ns_graphs;' xmlns='http://www.w3.org/2000/svg' xmlns:xlink='http://www.w3.org/1999/xlink' x='0' y='0' viewBox='0 0 41.5 51' style='enable-background:new 0 0 41.5 51;' xml:space='preserve'%3E%3Cmetadata%3E%3Csfw xmlns='ns_sfw;'%3E%3Cslices%3E%3C/slices%3E%3CsliceSourceBounds bottomLeftOrigin='true' height='51' width='41.5' x='166.1' y='-208.1'%3E%3C/sliceSourceBounds%3E%3C/sfw%3E%3C/metadata%3E%3Cg%3E%3Cpath d='M40.2,17.4c-3.4,2.1-5.5,5.7-5.5,9.7c0,4.5,2.7,8.6,6.8,10.3c-.8,2.6-2,5-3.5,7.2c-2.2,3.1-4.5,6.3-7.9,6.3s-4.4-2-8.4-2 c-3.9,0-5.3,2.1-8.5,2.1s-5.4-2.9-7.9-6.5C2,39.5,.1,33.7,0,27.6c0-9.9,6.4-15.2,12.8-15.2c3.4,0,6.2,2.2,8.3,2.2 c2,0,5.2-2.3,9-2.3C34.1,12.2,37.9,14.1,40.2,17.4z M28.3,8.1C30,6.1,30.9,3.6,31,1c0-.3,0-.7-.1-1c-2.9,.3-5.6,1.7-7.5,3.9 c-1.7,1.9-2.7,4.3-2.8,6.9c0,.3,0,.6,.1,.9c.2,0,.5,.1,.7,.1C24.1,11.6,26.6,10.2,28.3,8.1z'%3E%3C/path%3E%3C/g%3E%3C/svg%3E\")}.download-chip__headline.is-desktop::before{background-image:url(\"data:image/svg+xml,%3Csvg width='24' height='24' viewBox='0 0 24 24' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cg opacity='.8'%3E%3Cpath d='M1.5 21H22.5V18H1.5V21Z' fill='%23000' stroke='white' stroke-width='.75'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M19.5 15C20.2956 15 21.0587 14.6839 21.6213 14.1213C22.1839 13.5587 22.5 12.7956 22.5 12V6C22.5 5.20435 22.1839 4.44129 21.6213 3.87868C21.0587 3.31607 20.2956 3 19.5 3H4.5C3.70435 3 2.94129 3.31607 2.37868 3.87868C1.81607 4.44129 1.5 5.20435 1.5 6V12C1.5 12.7956 1.81607 13.5587 2.37868 14.1213C2.94129 14.6839 3.70435 15 4.5 15H19.5ZM19.5 12H4.5V6H19.5V12Z' fill='%23000' stroke='white' stroke-width='.75'/%3E%3C/g%3E%3C/svg%3E%0A\")}.download-card .snowflake-code-snippet,.download-card .snowflake-code-snippet code,.download-card .snowflake-code-snippet pre{font-size:14px;color:#000;text-shadow:none !important}.download-chip:hover{background-color:var(--ui-background-05) !important;transition:300ms ease background-color}body:has(.snowflake-skip-to-content[style]) #subNav,.pushdown-banner-dismissed #subNav{top:var(--scroll-padding-top) !important;transition:300ms ease top}body:has(.snowflake-skip-to-content[style*=\"58\"]) #subNav{top:34px !important}body:has(.snowflake-skip-to-content[style*=\"82\"]) #subNav{top:58px !important}body:has(.snowflake-skip-to-content[style*=\"130\"]) #subNav{top:106px !important}body:has(.snowflake-skip-to-content[style*=\"138\"]) #subNav{top:114px !important}body:has(.snowflake-skip-to-content[style*=\"146\"]) #subNav{top:122px !important}.is-hidden .snowflake-person-chip-avatar{display:none}.is-small .snowflake-person-chip-avatar{width:56px;height:56px}.ai-summary ul{margin:16px 0 0 0 !important;padding:0 !important;list-style-type:none}.ai-summary li{margin:0;padding:0 0 0 32px;position:relative}.ai-summary li::before{content:\"\";display:block;border-radius:100%;background:#29b5e8;width:18px;height:18px;position:absolute;top:4px;left:0;border:5px solid #e5f2f7;box-sizing:border-box}.ai-summary li:not(:last-child){margin-bottom:1rem}.snowflake-content-chip-image__image{aspect-ratio:5 / 3 !important}.content-chip-new .snowflake-content-chip-image__image{height:100% !important;aspect-ratio:unset !important}.snapshot-card .snowflake-text p:not(:first-child){margin-top:var(--spacing-01)}.snapshot-card\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:nth-child(2) p:has(b){font-family:'Texta',sans-serif;margin-top:24px}.snapshot-card\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:nth-child(2) p b{font-weight:700 !important}.snapshot-card\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:nth-child(2){border-bottom:1px solid #ccc;padding-bottom:24px}.snapshot-card\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:nth-child(3) p:first-child:has(b){font-family:'Texta',sans-serif;font-size:20px !important;margin-bottom:1rem !important}.snapshot-card\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:nth-child(3) li{display:inline-block}.snapshot-card\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:nth-child(3) li a{display:inline-block;text-decoration:none;padding:4px 16px !important;border:1px solid #ccc;border-radius:24px;color:#666 !important}.snapshot-card\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:nth-child(3) ul{list-style-type:none;display:flex;padding:0 !important;margin:0 !important;gap:12px}.snapshot-card\u003E.container\u003E.cmp-container\u003E.aem-container img{width:90%;max-width:240px;margin:0 auto}.snapshot-card\u003E.container\u003E.cmp-container\u003E.aem-container{padding:40px;max-width:450px;margin:0 0 0 auto;background-color:#fff;box-shadow:0 2px 6px 0 rgba(152,162,179,.25),0 10px 20px 0 rgba(152,162,179,.10);border-radius:8px;border-top:4px solid var(--ui-01)}.ai-summary{background-color:#f3fbfe;border-left:2px solid var(--ui-01);padding:40px}.ai-summary\u003Espan p:last-child:has(i){color:#666;font-size:14px !important}.ai-summary\u003Espan p:last-child:has(i) a{color:#666 !important;text-decoration:underline !important}.ai-summary\u003Espan p:last-child:has(i) a:hover{color:var(--ui-01) !Important}.ai-summary\u003Espan p:first-child:has(b)::after{content:'';display:inline-block;width:20px;height:20px;background-image:url(\"data:image/svg+xml,%3Csvg width='24' height='24' viewBox='0 0 24 24' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M9.3158 3.15226C8.6475 6.2258 6.22698 8.64545 3.15232 9.31587C2.94923 9.36072 2.94923 9.63928 3.15232 9.68413C6.22698 10.3522 8.6475 12.7742 9.3158 15.8477C9.36067 16.0508 9.63933 16.0508 9.6842 15.8477C10.3525 12.7742 12.773 10.3545 15.8477 9.68413C16.0508 9.63928 16.0508 9.36072 15.8477 9.31587C12.773 8.64781 10.3525 6.2258 9.6842 3.15226C9.63933 2.94925 9.36067 2.94925 9.3158 3.15226Z' fill='%23249EDC'/%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M17.3725 11.5461C16.9098 13.6739 15.2341 15.3491 13.1054 15.8132C12.9649 15.8443 12.9649 16.0371 13.1054 16.0681C15.2341 16.5307 16.9098 18.2074 17.3725 20.3353C17.4035 20.4758 17.5965 20.4758 17.6275 20.3353C18.0902 18.2074 19.7659 16.5323 21.8946 16.0681C22.0352 16.0371 22.0352 15.8443 21.8946 15.8132C19.7659 15.3507 18.0902 13.6739 17.6275 11.5461C17.5965 11.4055 17.4035 11.4055 17.3725 11.5461Z' fill='%23249EDC'/%3E%3C/svg%3E%0A\");background-repeat:no-repeat;background-size:contain;background-position:center;vertical-align:middle;margin-left:8px}.ai-summary\u003Espan p:first-child:has(b){color:var(--ui-01) !important;text-transform:uppercase}.border-top{border-top:1px solid rgba(0,0,0,.2)}.border-top\u003Espan{display:block;padding-top:32px}body .snowflake-card-v2-advanced-image__image{aspect-ratio:16 / 9 !important}.content-chip-new .snowflake-content-chip-image__image{border-radius:0;object-fit:cover;height:100%}.sf-footer #ot-sdk-btn.ot-sdk-show-settings,.sf-footer #ot-sdk-btn.optanon-show-settings{color:rgba(255,255,255,.7) !important;text-underline-offset:4px;border-top:none;border-left:none;border-right:none;border-bottom:1px dotted transparent;background-color:transparent !important;background-image:none !important;transition:300ms ease text-decoration-color;padding:0 !important;font-size:12px;font-family:'Lato',sans-serif}.sf-footer #ot-sdk-btn.ot-sdk-show-settings:hover,.sf-footer #ot-sdk-btn.optanon-show-settings:hover{color:rgba(255,255,255,1) !important;border-bottom:1px dotted var(--ui-01);transition:300ms ease text-decoration-color}.sf-footer__legal-container\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:last-child{flex-shrink:0}.sf-footer__disclaimers{background-color:#042130}.sf-footer__disclaimers .snowflake-simple-stat-disclaimer p a{color:inherit;text-decoration:none !important}.sf-footer__disclaimers .snowflake-simple-stat-disclaimer p sup{margin-right:2px}.sf-footer__disclaimers .snowflake-simple-stat-disclaimer p{text-indent:-5px;padding-left:5px}.sf-footer__disclaimers-inner{border-top:1px solid rgba(255,255,255,.25);padding:40px 0}.sf-footer__disclaimers .snowflake-simple-stat{align-items:flex-start;text-align:left;color:rgba(255,255,255,.7);margin-bottom:10px}.sf-footer__social{display:flex;justify-content:center;gap:12px}.sf-footer .snowflake-footer-social-item{margin:0 !important}.sf-footer .snowflake-footer-social-item a{line-height:0;background-color:rgba(3,24,35,.8);display:inline-block;width:48px !important;height:48px;border-radius:8px;display:inline-flex;justify-content:center;align-items:center;transition:300ms ease background-color}.sf-footer .snowflake-footer-social-item a:hover{background-color:var(--ui-01) !important;transition:300ms ease background-color}.sf-footer__bottom{padding-bottom:40px}.sf-footer .snowflake-marketo-form .mktoFormRow .mktoFieldWrap .mktoError .mktoErrorMsg{max-width:100%;color:#fff}.sf-footer .mktoForm .mktoError .mktoErrorMsg .mktoErrorDetail{display:inline-block}.sf-footer .mktoFormRow:has(.mktoHtmlText:empty){display:none}.sf-footer .mktoFormRow .mktoHtmlText span{color:#fff !important}.sf-footer{background-color:#042130}.sf-footer .optanon-toggle-display:hover{text-decoration-color:var(--ui-01) !important;cursor:pointer !important;text-underline-offset:4px;text-decoration-style:dotted !important;text-decoration-color:var(--ui-01);color:#fff !important;transition:300ms ease text-decoration-color;text-decoration:underline;opacity:1}.sf-footer__logo{width:40px}.sf-footer-grid__inner\u003E.container\u003E.cmp-container\u003E.aem-container{row-gap:32px}.sf-footer__legal-container\u003E.container\u003E.cmp-container\u003E.aem-container{display:flex;justify-content:space-between;align-items:center;text-align:center;row-gap:16px}.sf-footer__legal-container\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:nth-child(2){text-align:center;flex-grow:1}.sf-footer__legal-links li button,.sf-footer__legal-links li a,.sf-footer__legal-links li{margin:0;color:rgba(255,255,255,.7) !important;font-weight:500}.sf-footer__legal-links li a:hover{color:rgba(255,255,255,1) !important}.sf-footer div.sf-footer__copyright p,.sf-footer div.sf-footer__legal-links li,.sf-footer div.sf-footer__legal-links a,.sf-footer div.sf-footer__legal-links p{font-size:12px !important}.sf-footer__legal-links ul{list-style-type:none;margin:0;padding:0;display:flex;gap:20px;row-gap:4px;justify-content:center;flex-wrap:wrap;text-align:center}.sf-footer__legal-links li:last-child{width:100%}.sf-footer .mktoFormRow:has(.mktoPlaceholder),.sf-footer .mktoFormRow:has(input[type=\"hidden\"]){display:none !important}.sf-footer .mktoFormCol{margin-bottom:0 !important}.sf-footer label[for=\"adhoc1\"]{width:auto !important;flex-grow:1;margin-left:16px}.sf-footer .mktoFieldWrap:has(label[for=\"adhoc1\"]){display:flex;flex-direction:row-reverse;margin-top:22px}.sf-footer .snowflake-marketo-form .mktoFormRow .mktoFieldWrap .mktoCheckboxList input[type=checkbox]{background-color:transparent !important;border:1px solid rgba(255,255,255,.4) !important;border-radius:4px !important}.sf-footer .snowflake-marketo-form .mktoFormRow .mktoFieldWrap .mktoEmailField,.sf-footer .snowflake-marketo-form .mktoFormRow .mktoFieldWrap .mktoTelField,.sf-footer .snowflake-marketo-form .mktoFormRow .mktoFieldWrap .mktoTextField,.sf-footer .snowflake-marketo-form .mktoFormRow .mktoFieldWrap select{background-color:transparent !important;color:#fff !important;height:auto !important;border:1px solid rgba(255,255,255,.4) !important;border-radius:4px !important;padding:12px 18px !important}.sf-footer .snowflake-marketo-form .mktoFormRow .mktoFieldWrap .mktoEmailField:focus,.sf-footer .snowflake-marketo-form .mktoFormRow .mktoFieldWrap .mktoTelField:focus,.sf-footer .snowflake-marketo-form .mktoFormRow .mktoFieldWrap .mktoTextField:focus,.sf-footer .snowflake-marketo-form .mktoFormRow .mktoFieldWrap select:focus{border-color:var(--ui-01) !important}.sf-footer .mktoForm *{padding:0 !important}.sf-footer .mktoForm,.sf-footer .snowflake-marketo-form-container{padding:0 !important;background:transparent;margin-bottom:0;box-shadow:none}.sf-footer .mktoHtmlText.mktoHasWidth{width:100% !important;margin:24px 0}.sf-footer .mktoFormRow{flex-direction:column}.sf-footer .mktoForm .mktoButtonWrap{margin:0 !important}.sf-footer select{background-image:url(\"data:image/svg+xml,%3Csvg width='14' height='8' viewBox='0 0 14 8' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M.981445 1.43496L6.90897 7.32496L12.9314 1.33496' stroke='white' stroke-width='1.33333' stroke-miterlimit='10' stroke-linecap='round' stroke-linejoin='round'/%3E%3C/svg%3E%0A\") !important}.sf-footer .snowflake-marketo-form .mktoButtonWrap.mktoNative{justify-content:flex-start}.sf-footer *::placeholder{color:#fff !important;opacity:.8}.sf-footer .mktoForm .mktoButtonWrap.mktoSimple .mktoButton{background-color:var(--ui-01) !important;color:#fff !important;width:100% !important;padding:12px 16px !important;border:1px solid var(--ui-01) !important;background-image:none !important;border-radius:48px;text-transform:uppercase;font-weight:800 !important;font-family:'Texta',sans-serif !important;font-size:16px !important;line-height:1.2}.sf-footer .snowflake-marketo-form .mktoFormRow .mktoFieldWrap .mktoHtmlText\u003Espan,.sf-footer .snowflake-marketo-form .mktoFormRow .mktoFieldWrap .mktoLabel\u003Espan,.sf-footer .snowflake-marketo-form .mktoFormRow .mktoFieldWrap label.mktoLabel{color:#fff !important}.sf-footer__newsletter-title p:not(:first-child){margin-top:8px !important}.sf-footer__newsletter-title p b{font-weight:800 !important;font-family:'Texta',sans-serif !important;font-size:22px !important;line-height:1.2}.sf-footer__newsletter-title p:last-child{font-size:14px !important;opacity:.8}.sf-footer__link-group li a[target=\"_blank\"]::after{content:'';display:inline-block;width:10px;height:10px;margin-left:5px;background-image:url(\"data:image/svg+xml,%3Csvg width='11' height='11' viewBox='0 0 11 11' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath d='M6.72222 1.22222C6.38471 1.22222 6.11111 .948616 6.11111 .611111C6.11111 .273607 6.38471 0 6.72222 0H10.3889C10.551 0 10.7064 .0643867 10.821 .178988C10.9356 .293596 11 .449032 11 .611111V4.27778C11 4.61529 10.7264 4.88889 10.3889 4.88889C10.0514 4.88889 9.77778 4.61529 9.77778 4.27778V2.08647L4.09879 7.76545C3.86013 8.00409 3.4732 8.00409 3.23454 7.76545C2.99589 7.52681 2.99589 7.13986 3.23454 6.90122L8.91355 1.22222H6.72222ZM0 2.44444C0 1.76943 .547207 1.22222 1.22222 1.22222H4.27778C4.61529 1.22222 4.88889 1.49583 4.88889 1.83333C4.88889 2.17084 4.61529 2.44444 4.27778 2.44444H1.22222V9.77778H8.55556V6.72222C8.55556 6.38471 8.82915 6.11111 9.16667 6.11111C9.50418 6.11111 9.77778 6.38471 9.77778 6.72222V9.77778C9.77778 10.4528 9.23059 11 8.55556 11H1.22222C.547207 11 0 10.4528 0 9.77778V2.44444Z' fill='white'/%3E%3C/svg%3E%0A\");background-size:contain;background-repeat:no-repeat;background-position:center}.sf-footer__link-group ul,.sf-footer__link-group li{margin:0;padding:0;list-style-type:none}.sf-footer__link-group ul{margin-top:20px !important}.sf-footer__link-group li{margin-top:15px}.sf-footer div.sf-footer__link-group\u003Espan\u003Ep\u003Ea,.sf-footer div.sf-footer__link-group\u003Espan\u003Ep{color:var(--ui-01) !important;font-weight:800 !important;font-family:'Texta',sans-serif !important;font-size:20px !important;line-height:1.2}.sf-footer__link-group li a{opacity:.9;color:#fff !important;font-weight:500 !important;font-size:15px !important;line-height:1.3}.sf-footer__link-group li a:hover{opacity:1}.sf-footer-grid__inner\u003E.container\u003E.cmp-container\u003E.aem-container::before,.sf-footer-grid__inner\u003E.container\u003E.cmp-container\u003E.aem-container::after{display:none}.sf-footer__column{flex-grow:1}.sf-footer-grid__inner\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:not(:first-child){width:50%}@media (min-width:800px){.sf-footer__legal-links ul{justify-content:flex-start;text-align:left}.sf-footer__social{justify-content:flex-end}.sf-footer__legal-links ul{padding-left:24px}.sf-footer__legal-container\u003E.container\u003E.cmp-container\u003E.aem-container{text-align:right;flex-wrap:nowrap}.sf-footer__legal-links.align-left ul{justify-content:flex-start}.sf-footer-grid__inner\u003E.container\u003E.cmp-container\u003E.aem-container{display:flex;justify-content:space-between;flex-direction:row}.sf-footer-grid__inner\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv{width:auto !important;max-width:200px}.sf-footer-grid__inner:not(.hidden-newsletter)\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:first-child{flex-grow:1;order:2;width:100% !important;max-width:none}.sf-footer__legal-container\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv{width:auto}}@media screen and (min-width:1380px){.sf-footer-grid__inner\u003E.container\u003E.cmp-container\u003E.aem-container{flex-wrap:nowrap}.sf-footer-grid__inner:not(.hidden-newsletter)\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:first-child{padding-right:48px;max-width:380px;background-color:rgba(3,24,35,.4);padding:32px;margin-left:48px;border-radius:16px}.sf-footer__link-group li,.sf-footer__link-group li a{font-size:14px !important;line-height:1.3}}@media screen and (max-width:991px){.sf-footer-grid__inner\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv:first-child{order:2;margin-top:24px !important}}@media screen and (max-width:420px){.is-reduced-mobile .heading-1-v2,.is-reduced-mobile .heading-1-v2-sm{font-size:32px;line-height:28px}}.quote-content-chip{background-color:var(--ui-background-05);padding:24px;border-radius:12px;position:relative}.quote-content-chip .black-blue-text-color .snowflake-title-v2-line\u003Espan{color:rgba(0,0,0,.8) !important;font-size:15px !important;line-height:1.5 !important;font-family:'Lato',sans-serif;font-weight:400 !important}.quote-content-chip .black-blue-text-color .snowflake-title-v2-line\u003Espan:not(:first-child){max-width:calc(100% - 200px)}.quote-content-chip .black-blue-text-color .snowflake-title-v2-line\u003Espan:nth-child(2){font-family:'Texta',sans-serif;color:#000 !important;font-size:20px !important;font-weight:800 !important;margin-top:24px}.quote-content-chip .snowflake-content-chip-image{width:140px !important}@media screen and (min-width:992px){.quote-content-chip .snowflake-content-chip-image{position:absolute !important;bottom:24px;right:16px}}@media screen and (max-width:991px){.quote-content-chip .snowflake-content-chip-image{margin-bottom:40px}.quote-content-chip{flex-direction:column}}#spa-root{background-color:#fff}.lowercase .snowflake-title-v2-line{text-transform:none !important}.centered .snowflake-logo-content-container-inner{justify-content:center}div.snowflake-linklist-dropdown-menu{max-height:380px}.first-line-blue .snowflake-typographyv2 .snowflake-title-v2-line:first-child{color:var(--ui-01) !important}.is-front{position:relative;z-index:2}.use-case-body .snowflake-text h1,.use-case-body .snowflake-text h2,.use-case-body .snowflake-text h3,.use-case-body .snowflake-text h4,.use-case-body .snowflake-text h5,.use-case-body .snowflake-text h6{font-family:'Texta',sans-serif;color:#000;margin:.25rem 0 0 0}.pc-hero .button-group\u003E.container\u003E.cmp-container\u003E.aem-container{justify-content:flex-start}.sf-footer .mktoFormRow .mktoHtmlText span{font-family:'Lato',sans-serif !important}.snowflake-button-primary.snowflake-button-blue .snowflake-button-container{justify-content:center}.related-chip-25{background-color:#fff;border:1px solid rgba(204,204,204,.5);border-radius:8px;padding:20px;position:relative}.related-chip-25:hover{box-shadow:rgba(152,162,179,.1) 0 10px 20px 0}.related-chip-25:hover::after{right:24px;transition:300ms ease right}.related-chip-25::after{content:'';display:block;transition:300ms ease right;background-image:url(\"data:image/svg+xml,%3Csvg width='8' height='14' viewBox='0 0 8 14' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Cpath fill-rule='evenodd' clip-rule='evenodd' d='M7.66699 7C7.66699 6.6571 7.53559 6.32825 7.30169 6.08578L2.34446 .947072C1.84529 .429617 1.0164 .429617 .517219 .947072C.0427878 1.43887 .042788 2.21798 .517219 2.70978L4.65591 7L.51722 11.2902C.0427889 11.782 .0427887 12.5611 .51722 13.0529C1.0164 13.5704 1.84529 13.5704 2.34447 13.0529L7.30169 7.91421C7.53559 7.67175 7.66699 7.34289 7.66699 7Z' fill='%2329B5E8'/%3E%3C/svg%3E%0A\");width:8px;height:14px;display:block;position:absolute;right:30px;top:50%;transform:translateY(-50%);background-size:contain;background-position:center;background-repeat:no-repeat}.related-chip-25 .heading-5-v2{font-size:22px;line-height:1.1}.related-chip-25 .snowflake-content-chip-image{width:48px;flex-shrink:0}.related-chip-25 .snowflake-content-chip-image__image{aspect-ratio:1;height:auto;object-fit:contain}.related-chip-25 .snowflake-content-chip-button{display:none}.related-chip-25 .snowflake-content-chip-content-without-tag{flex-grow:1;padding-right:24px}.case-study-25.small-logo .snowflake-case-study-card-logo img{width:60px !important}.swiper-slide .case-study-25{width:95%;margin-left:auto;margin-right:auto}.case-study-25 .snowflake-case-study-card-logo img{width:140px !important;height:auto !important;transform:none !important;margin:24px 0 8px 0}.case-study-25 .snowflake-case-study-card-image__image{object-position:left center}.case-study-25 .snowflake-case-study-card-information-container{padding-right:24px}.case-study-25 ul{list-style-type:none;padding:0;margin:8px 0 0 0}.case-study-25 li{font-size:15px !important;line-height:1.3 !important;display:flex;flex-direction:column;border-left:4px solid var(--ui-01);padding-left:24px;margin-top:24px;color:#535862;gap:4px}.case-study-25 li b{display:block;font-family:'Texta',sans-serif;font-weight:900 !important;font-size:48px !important;line-height:.9 !important;color:var(--ui-01)}.case-study-25 .snowflake-case-study-card-description p{color:#535862}.case-study-25 .snowflake-case-study-card-description p:nth-child(2):not(:has(a)){color:#000;font-family:Texta;font-size:30px !important;line-height:1 !important;font-style:normal;font-weight:700;text-indent:-8px}.case-study-25.is-story .snowflake-case-study-card-description p:nth-child(2):not(:has(a)){text-indent:0}.case-study-25 .snowflake-case-study-card-key-card{background-color:transparent}.case-study-25 .snowflake-case-study-card-button{display:none}.case-study-25{border-radius:24px;overflow:hidden}@media screen and (min-width:1024px){.case-study-25 .snowflake-case-study-card-left-container{position:static;width:60%;min-height:0}.case-study-25 .snowflake-case-study-card-right-container::after{content:'';display:block;width:60%;max-width:340px;padding-bottom:50%;background-image:url(\"data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' fill='none' viewBox='0 0 22 16' class='snowflake-pushdown-banner-placeholder-arrow'%3E%3Cpath fill='%2329B5E8' fill-rule='evenodd' d='M17.865 8.756c.088-.274.124-.555.118-.834a2.551 2.551 0 0 0-1.3-2.142L7.887.76C6.645.055 5.063.475 4.35 1.7a2.535 2.535 0 0 0 .947 3.494l4.916 2.809-4.916 2.801a2.543 2.543 0 0 0-.947 3.502c.713 1.222 2.295 1.64 3.537.934l8.796-5.024a2.541 2.541 0 0 0 1.182-1.46Z' clip-rule='evenodd'%3E%3C/path%3E%3C/svg%3E\");background-size:contain;background-repeat:no-repeat;position:absolute;top:-10%;left:-20%}.case-study-25 .snowflake-case-study-card-right-container{max-width:none;width:40%;position:absolute;top:-5%;right:-5%;z-index:0;height:110%}}@media screen and (min-width:768px){.case-study-25 li{max-width:50%}.case-study-25 ul{display:flex;gap:48px}}.snowflake-text.section-eyebrow p{margin-left:auto;margin-right:auto;margin-bottom:16px !important}.snowflake-text.section-eyebrow p,.snowflake-text.eyebrow-text p{text-transform:uppercase;font-family:'Texta',sans-serif !important;font-weight:800 !important;letter-spacing:.025em;margin-bottom:12px;line-height:1.1 !important}.snowflake-title-v2.dynamic .heading-2-v2 span.snowflake-title-v2-line{font-size:clamp(2.5rem,4.5vw,4rem) !important;line-height:.82 !important}.checklist ul{padding:0;margin:0}.checklist ul li{list-style-type:none;padding-left:32px;position:relative}.checklist ul li:not(:last-child){margin-bottom:1em}.checklist ul li::before{content:'';display:inline-block;width:20px;height:20px;background-image:url(\"data:image/svg+xml,%3Csvg width='24' height='25' viewBox='0 0 24 25' fill='none' xmlns='http://www.w3.org/2000/svg'%3E%3Crect y='.985352' width='24' height='24' rx='12' fill='%23D4F0FA'/%3E%3Cpath d='M7.28613 13.2967L10.7147 16.7253L17.5718 9.86816' stroke='%2329B5E8' stroke-width='2' stroke-linecap='round' stroke-linejoin='round'/%3E%3C/svg%3E%0A\");background-size:contain;background-repeat:no-repeat;position:absolute;top:3px;left:0}.last-line-blue .snowflake-typographyv2 .snowflake-title-v2-line:last-child{color:var(--ui-01)}.snowflake-text p sup{line-height:0}.snowflake-title-v2.lowercase .heading-3-v2{font-size:28px;line-height:1;text-transform:none;font-weight:700}.snowflake-title-v2.lowercase .heading-2-v2{font-size:32px;line-height:1;text-transform:none;font-weight:700}.content-chip-new{border:1px solid rgba(204,204,204,.5);border-radius:16px;overflow:hidden}.content-chip-new .snowflake-image-container{border-radius:0;display:none}.content-chip-new .snowflake-content-chip-image{margin-right:0;max-width:180px;flex-shrink:0}.content-chip-new .snowflake-content-chip-content{padding:24px}.content-chip-new .black-blue-text-color .snowflake-title-v2-line:first-child{font-size:24px;line-height:1.1}.content-chip-new .black-blue-text-color .snowflake-title-v2-line:not(:first-child){font-family:'Lato',sans-serif;font-size:17px;color:#535862 !important;font-weight:500;line-height:1.45;margin-top:8px;display:none}div.snowflake-text a{font-weight:normal;color:var(--ui-01);text-decoration:underline;text-underline-offset:4px;text-decoration-style:dotted !important;text-decoration-color:transparent;transition:300ms ease text-decoration-color}div.snowflake-text a:hover{text-decoration-color:var(--ui-01);transition:300ms ease text-decoration-color}.footer-nav__link-group .snowflake-button-container,.subnav__item--button,.snowflake-card-v2-advanced-button .snowflake-button-container{justify-content:flex-start}.button-container\u003E.container\u003E.cmp-container\u003E.aem-container{align-items:center}.button-container\u003E.container\u003E.cmp-container\u003E.aem-container .snowflake-button-primary+.snowflake-button-link{margin-left:12px !important}.snowflake-button-regular.snowflake-button-link .snowflake-button-container{font-size:18px !important;text-align:left;justify-content:flex-start;line-height:1.4 !important}body .snowflake-card-v2-advanced{border:1px solid rgba(204,204,204,.5);border-radius:var(--spacing-02);transition:300ms ease all}body .snowflake-card-v2-advanced:hover{box-shadow:rgba(152,162,179,.1) 0 10px 20px 0;transition:300ms ease all}body .snowflake-card-v2-advanced-inner{border-bottom:none}body .snowflake-card-v2-advanced-image{line-height:0}body .snowflake-card-v2-advanced-image__image{aspect-ratio:16 / 9}body .snowflake-card-v2-advanced-content{position:relative}body .snowflake-card-v2-advanced-content::after{content:'';display:block;position:absolute;bottom:0;left:0;transition:300ms ease all;width:20%;height:4px;background-color:var(--ui-01);opacity:0}body .snowflake-card-v2-advanced:hover .snowflake-card-v2-advanced-content::after{width:100%;opacity:1;transition:300ms ease all}body .snowflake-card-v2-advanced .snowflake-button-link.snowflake-button-blue .snowflake-button-container\u003E.link-icon{transition:300ms ease transform}body .snowflake-card-v2-advanced:hover .snowflake-button-link.snowflake-button-blue .snowflake-button-container\u003E.link-icon{transform:translateX(4px);transition:300ms ease transform}.six-columns\u003E.container\u003E.cmp-container\u003E.aem-container,.three-columns\u003E.container\u003E.cmp-container\u003E.aem-container,.four-columns\u003E.container\u003E.cmp-container\u003E.aem-container,.five-columns\u003E.container\u003E.cmp-container\u003E.aem-container{display:flex;flex-wrap:wrap;gap:24px}.six-columns.align-center\u003E.container\u003E.cmp-container\u003E.aem-container,.three-columns.align-center\u003E.container\u003E.cmp-container\u003E.aem-container,.four-columns.align-center\u003E.container\u003E.cmp-container\u003E.aem-container,.five-columns.align-center\u003E.container\u003E.cmp-container\u003E.aem-container{justify-content:center}.three-columns\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv{width:100%;margin:0 !important}.six-columns\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv,.four-columns\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv,.five-columns\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv{width:calc(50% - 12px);margin:0 !important}@media screen and (min-width:768px){.three-columns\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv{width:calc(50% - 12px)}.six-columns\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv,.four-columns\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv,.five-columns\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv{width:calc(33.333% - 16px)}}@media screen and (min-width:1024px){.snowflake-title-v2.lowercase .heading-3-v2{font-size:34px}.snowflake-title-v2.lowercase.larger .heading-2-v2{font-size:44px;line-height:.95}.three-columns\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv{width:calc(33.333% - 16px)}.four-columns\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv{width:calc(25% - 18px)}.five-columns\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv{width:calc(20% - 19.2px)}.six-columns\u003E.container\u003E.cmp-container\u003E.aem-container\u003Ediv{width:calc(16.6666% - 20px)}.snowflake-title-v2.lowercase .heading-3-v2{font-size:28px !important}}@media screen and (min-width:1200px){.snowflake-title-v2.lowercase .heading-2-v2{font-size:40px}.content-chip-new .snowflake-content-chip-content{padding:32px}.content-chip-new .snowflake-image-container,.content-chip-new .black-blue-text-color .snowflake-title-v2-line:not(:first-child){display:block}}.promo-banner-25{border-radius:16px;overflow:hidden}.promo-banner-25 .snowflake-premium-content-banner-image-container{position:relative;max-width:380px}.promo-banner-25 .snowflake-text{color:#535862}.promo-banner-25 .snowflake-premium-content-banner-image__image{transform:translateY(8px);transition:300ms ease transform;border-radius:0;width:85%;margin:0 auto;display:block;position:relative;z-index:1}.promo-banner-25 .snowflake-premium-content-banner-image__link:hover .snowflake-premium-content-banner-image__image{transform:translateY(0);transition:300ms ease transform}.promo-banner-25 .snowflake-premium-content-banner-image__inner{height:auto;padding-top:24px}.promo-banner-25 .snowflake-premium-content-banner-image__link{position:relative;z-index:1;height:auto}.promo-banner-25 .snowflake-premium-content-banner-image__link::after{content:'';display:block;position:absolute;clip-path:polygon(0 0,66% 0,100% 100%,0 100%);bottom:0;left:0;width:100%;height:100%;background:var(--ui-01);transition:300ms ease width}.promo-banner-25 .snowflake-premium-content-banner-image__link:hover::after{width:110%;transition:300ms ease width}.sf-footer .snowflake-marketo-form .mktoFormRow .mktoFieldWrap select{background-position:95% 50%}.sf-footer__disclaimers .text-size-small .snowflake-text p{color:#fff !important;font-size:10px !important;opacity:.8}@media screen and (min-width:768px){.sf-footer__disclaimers .text-size-small .snowflake-text p{font-size:12px !important}}@media screen and (max-width:1023px){.mobile-top-padding{padding-top:64px}}@media (max-width:799px){.sf-footer .snowflake-marketo-form .mktoButtonWrap.mktoNative .mktoButton{width:100% !important}.sf-footer__logo{text-align:center;display:block;margin:0 auto}}.customer-card .snowflake-card-v2-advanced-image{aspect-ratio:4.35 / 1}.customer-card .snowflake-card-v2-advanced-image__image{width:100%;height:100%;padding-left:8px;object-fit:contain;object-position:left center;margin:0 !important;aspect-ratio:initial}.customer-card .snowflake-card-v2-advanced-image__inner{height:110px}.customer-card .snowflake-card-v2-advanced-tag-indicator{display:none}.pc-hero .snowflake-container-arrow-small-gray-image{top:-34% !important;width:18% !important}.pc-hero .snowflake-container-arrow-small-gray-image path{fill:var(--ui-01);opacity:1}@media screen and (max-width:767px){.mobile-padding-top{padding-top:64px}.hide-mobile{display:none !important}.pc-hero{padding-top:52px}.pc-hero .snowflake-text p,.pc-hero .left-alignment .snowflake-title-v2-line,.pc-hero h1 span{text-align:center !important}}div.snowflake-pushdown-banner-button{margin-top:0}.button-group.align-center\u003E.container\u003E.cmp-container\u003E.aem-container{align-items:center;justify-content:center !important}.text-center .snowflake-breadcrumb-swiper .swiper-wrapper{justify-content:center}div.snowflake-breadcrumb a.snowflake-breadcrumb-item,.snowflake-breadcrumb div.snowflake-breadcrumb-item{text-transform:none;font-weight:500}.snowflake-breadcrumb svg{display:none !important}.snowflake-breadcrumb a:has(svg)::after{content:'/';margin:0 12px;color:#666}.hide-filters .snowflake-filterable-and-searchable-grid-top-part{display:none !important}.page-section{padding-left:24px;padding-right:24px}@media screen and (min-width:768px){.page-section{padding-left:48px;padding-right:48px}}.download-card pre[class*=language-]{overflow-x:scroll !important}",":type":"snowflake-site/components/markup-editor","isGSAPEnabled":false}},":type":"snowflake-site/components/container",":itemsOrder":["container_copy","container_573483281_","markup_editor_copy"]}},":type":"snowflake-site/components/experiencefragment","classNames":"aem-xf",":itemsOrder":["root"]},"markup_editor":{"id":"markup-editor-2e052effa1","title":"Quickstarts Overrides","cssUrl":"/conf/snowflake-site/settings/wcm/templates/quickstart-page-template/structure/_jcr_content/root/markup_editor.45e3b816710d8d5a.css",":type":"snowflake-site/components/markup-editor","isGSAPEnabled":false}},":itemsOrder":["experiencefragment-banner","experiencefragment-header","markup_editor_1950346551","responsivegrid","modal_container","experiencefragment-footer","markup_editor"],":type":"wcm/foundation/components/responsivegrid"}},":mappedPath":"/en/developers/guides/credit-risk-pd-model-snowflake-ml/",":type":"snowflake-site/components/structure/page",":itemsOrder":["root"],":hierarchyType":"page",":path":"/content/snowflake-site/global/en/developers/guides/credit-risk-pd-model-snowflake-ml","isPasswordProtected":false,"analyticsContentTags":["snowflake-site:taxonomy/industry/financial-services","snowflake-site:taxonomy/product/ai","snowflake-site:taxonomy/snowflake-feature/model-development"],"analyticsEnabled":true,"coveoConfig":{"apiKey":"xx335921a6-2a0a-40f2-a167-e390b4766c3d","pipeline":"snowflake.com","searchHub":"snowflake.com","organizationId":"snowflakecomputingproduction8neljofn"},"analyticsDebugMode":false,"analyticsData":{"excludeFromAnalytics":false,"subCategory":"guides","pageType":"developer","templateName":"quickstart-page-template","siteName":"snowflake","pageUrl":"/en/developers/guides/credit-risk-pd-model-snowflake-ml/","language":"en","category":"developers","pageName":"From SAS to Snowflake: Productionizing Credit Risk Models on a Modern ML Platform","contentTags":["snowflake-site:taxonomy/industry/financial-services","snowflake-site:taxonomy/product/ai","snowflake-site:taxonomy/snowflake-feature/model-development"]},"publicationThemeClass":"","locale":"en"}