Snowflake on Snowflake: Driving Product Strategy with Data Science

Julian Ferreira: my name is julian ferreira and i'm a data science product marketing manager here at snowflake and i'll be moderating today's events uh i think first of all i want to thank everyone for being here and i appreciate the fact that you're taking time uh to learn more about how our team uses snowflake internally a few house rules uh do know that the webinar is being recorded and so the recording will be available to you tomorrow And as we go along the way, if you have any questions, please do submit those questions through the Q&A panel on the side of the ON24 window. We will be going over these questions since we have some time reserved to go over Q&A at the end. Another quick house rule is that please do note that today we're going to be discussing features that are not yet generally available. And so this is just a friendly reminder to make sure that you don't make any purchasing decisions based on what's currently being shown and only on what's available in Snowflake today. For today's agenda, we're first going to introduce our speakers who are part of our product data science team here at Snowflake.

We'll then talk a little bit about sort of the product landscape and some of the challenges that we're facing today as we try to provide value to our customers. And then we'll go into a deep dive as to what it means to be customer zero for our data science team. After that, we'll sort of talk about some of the success we've seen, what we're looking forward to in the future. And then, as mentioned, we are going to do some Q&A towards the end. With that, it's my pleasure to introduce our speakers today. Chris, why don't you get started?

Chris Pouliot: Yeah, hi. I'm glad everyone can join today's webinar. I'm Chris Pouliot, the head of data science and analytics for the product team. I joined Snowflake in July of 2019 to run the product data science team, which is trying to help both product and engineering be as data driven as possible. Previous to Snowflake, I was at companies such as Google, Netflix, Lyft, and Neo, and really enjoying my time at Snowflake, super exciting place to work. And then here, Margaret is on my team, and she can introduce herself.

Margaret Campbell: Hey, I'm Margaret. I'm one of the data scientists on Chris's team. I came to data science by way of economics and I guess started out doing consulting, then moved actually into the nonprofit field for a long time, but ended up feeling like I wanted to get to work with a company that had the resources to be really cutting edge. And Snowflake has absolutely been that. I joined in January 2019, and it has been an exciting and wonderful ride.

Julian Ferreira: So, yeah, thank you guys for sharing a little bit more about yourselves. I think as we get started, something that we want to talk about is where the product strategy landscape is today. And I think some of these things should be familiar to you, but we're going to give you sort of a little bit of a different perspective as we go through each of these. As we all know, being customer-centric is sort of one of the most important things, and it's been a focus of many product teams, primarily in sort of the B2C space, but it's also now an important piece of the B2B, which is for a company like Snowflake. And I think as we're thinking about who is the customer, one of the things that we need to think about is that our customer base can be very diverse. in a way to be able to represent what they're thinking, their needs, and what they're looking for, we also need to have a diverse product team and a diverse set of backgrounds that make part of that product team in order to drive more customer value through either accessibility, representation, and so many other ways where we can engage more with different customers in different regions and sort of having a more local approach to everything.

A second piece is gonna be sort of moving into an anti-fragile product strategy. And just to sort of explain where this comes from is that in the past, the focus had been around being agile, right? And agile meant being able to adapt quickly, respond quickly in order to react to adversity. I think in the next spectrum of that path is where you come into, how do you become anti-fragile? And really what this means is that you're not only able to adapt, but you're also able to improve as you face new adversity. I think this is where embedding data science into the product teams really adds a lot of value because you're sort of taking and going through the exercise of taking everything you've learned and then move those past learnings into improve and grow into the future. And so really taking that enterprise approach with a data science embedded piece to it. And then of course, culture is changing, right? We've been working from home for a little bit over a year now for most of us or many of us. And so product teams have had to redefine the way they work and collaborate. And so in this sort of remote world, we have to think of ways to be transparent and find new methods of communicating with each other, right?

I think one way to do this and sort of our philosophy at Snowflake is sort of having a central source of truth so that everyone can access and see what's going on within each one of those teams. And so that way, by having transparent and clear communication, it increases the collaboration between both the product managers and everyone that works around the product team, whether it's data science, engineering, and so forth, just to reduce a little bit of that video conferencing overload as much as we can. and so uh just wondering chris maybe if you have uh any additional thoughts of sort of what's your experience been so far around any of these pillars here at snowflake

Chris Pouliot: yeah thanks julian um i think you know uh my background was a lot of consumer facing companies you know like google and netflix and lyft and and at those companies we always you know is like you know think of the user experience optimize the user experience and that helps optimize the business um and this is like my one of my this is my first enterprise company that i'm working at but that that philosophy doesn't change we have to make sure that our customers are happy and happy and and at snowflake means their query their their queries are fast reliable consistent um you know with data governance you know layered on top uh all those things are super important so you know At Snowflake, we think of the customer and the user experience first, as these consumer companies also do. And then I think in the shifting culture too is related. If you have a really great data platform where you can ask a question and get an answer really quickly, that transforms you know the way that the business functions where you rely upon the data you become that that uh you know the data-driven company which people aspire to be and you know in order for that to happen you need to have a great data platform which Snowflake provides

Julian Ferreira: yeah no i totally agree with all those uh thanks for sharing chris i think as we uh as the next part of our for uh webinar we also want to talk about sort of the challenges that comes with when you're trying to take on these pillars and make them real. And so it's a combination of both the technology and the team challenges that happen across those. And so to begin with, to have a central source of truth, sort of data needs to be collected and it needs to be cleaned up for analytical use. And so how do you go from this disparate silo environments to a more centralized point of view and central source of truth. That's a challenge that everyone goes through at some point. The second one is having to work with different systems. I think depending on when the organization started, where the different teams were started, they may be dealing with one or more different platforms. And sometimes these can add complexity depending on performance or scalability barriers that they may bring. And so having to deal with those can slow down the team which is not something that you wanna do if you're running a lean team, right?

That you're trying to make them to efficient as possible, as productive as possible and provide as much value as they can without focusing on sort of what may be the more boring part of data science, which is collecting data, cleaning it up, and making a model ready. I think Margaret, from your perspective, coming from different organizations and joining Snowflake, have you felt these challenges before? What are some of the things you've seen here at Snowflake before we jump into the details of everything you guys are working on?

Margaret Campbell: Yeah. So, I mean, I think everybody has kind of dealt with the disparate data problem. You know, each team kind of has their set of spreadsheets that they live with. And then the data team is often one of the only ones that's using the data warehouse. Data and engineering kind of access that. And then everybody ends up having to ask questions of the data science team. And then there's kind of the process of, okay, you have your data on a spreadsheet. I have all of the company data in this data warehouse. How do I get your spreadsheet data in? How do I get it updated so that it's like regular and I can actually build important information off of this? And that kind of challenge of making sure all of the teams are working together often ends up meaning that the data science team is kind of a bottleneck for getting business questions answered. So rather than having everybody have access to the data they need, it's siloed and you have to kind of go through the data gatekeepers.

Julian Ferreira: Yeah, I think not a data scientist myself, but I know that when I have product questions, it's sort of very easy to go to everything that you guys have put together and bring my own questions and then get answers quickly without having to jump on a simple going into the whole not overloading teams with Zoom. And so with that, now that we've sort of addressed the opportunities and the challenges, I wanted to go into a deep dive about what you guys are working on and what it means to be customer zero at Snowflake. And so with that, I'll pass it down to you, Margaret and Chris.

Chris Pouliot: Yeah, great. One thing about being customer zero is that Margaret and I are data scientists and then Snowflake is made for data scientists, right? So we get to use Snowflake on a daily basis and we get the dog food feature to be the first people to try a new feature before anyone else. And that's really exciting for us because, you know, our ideas and feedback and integrated back into the product, which is super exciting. So we're part of the product development lifecycle. And, you know, our input is heard and, you know, what's good for us is good for our customers or what doesn't work for us doesn't work for our customers. So we're the first line of quality assurance. So that's a really gratifying part of the job. So at Snowflake, this is kind of the process we go through in any data project. You have to ingest the data. You have to model that data within the database. We do a lot of analysis, visualize, and then also we want to be able to predict the future versus just describe what's happening in the past. We're going to go into much detail in this presentation on each of these five steps.

And one thing to note, too, within the data science team is that our data scientists are full stack. So we don't have a separate data engineer, data scientist, and architect or QA. All our data scientists do all these five steps end to end so that they have the best context within a given area that they support.

Margaret Campbell: So throughout this, we're gonna really dive in on a specific use case, which is server forecasting. I think it's really helpful sometimes to have an example. And in this specific example, our problem is that we're now large enough that Amazon actually cares if we scale up very quickly. So we have to be sure that not only for our own purposes, for planning and making provisioning requests, our cloud service providers know where we're going and how quickly we're gonna grow. And obviously, they actually have access to all of the log data. They know what we've been using. And so the question really is, how can we provide value added to them? I mean, they can draw a straight line forecast off of what we've been using. And we want to provide them with something that is more useful so that they have signals that will tell them a little bit more about how we're growing. But before we dive in on all of that, I wanted to set the stage by talking about the data lifecycle. Wow.

Julian Ferreira: Everyone looks like we lost Margo for a second. She's going to be trying to refresh in the pretty soon and so she'll be back to us in the meantime. I don't know Chris, maybe if you have a couple of things to share about the data lifecycle instead of how this process fits within your team.

Chris Pouliot: Sure. So what Margaret was about to say is, you know, we have source data, right? In these cases, they're like server logs, usage logs, which is for every query, all the metadata that it produces. And then we also have a lot of business data, right? So from finance support and sales. We ingest that data so that it's in one place. One thing to note here is ELT versus ETL, right? So Snowflake, it's better to load the data and then transform it versus transform and then load in some more traditional databases because of the power of Snowflake that it can handle all this data at scale very efficiently. So then the data is into Snowflake, or Snowhouse, which is our internal deployment of Snowflake. And then we talk to the business. What business questions are we trying to answer? And then we explore the data. and go back and forth, back and forth, and then that helps us determine what the data model should be, how should we lay out the tables and schemas so that we'll be able to easily answer those questions quickly and reliably and consistently. Once we have the data model, then the data scientists might wanna predict the future, like saying like, hey, what's the forecast gonna be in the next six months versus the last six months?

And that's where we're gonna talk about more like, hey, how do you do machine learning? on this data in order to help predict the future.

Margaret Campbell: So... Hey, guys. Sorry.

Chris Pouliot: I could have gotten picked up. All right, so I just went over the slide, Margaret, if you heard me. OK. So that's in general. And now Margaret's going to talk about some features. So some features to help make that happen is one of them is SnowPipe. So SnowPipe loads the data into Snowflake as soon as it's available in micro-batches. So say, for example, once your data lands in S3, Snowflake realizes that, hey, there's some new data. Let me automatically load it into Snowflake. The next one is semi-structured data, right? So we can load and query semi-structured JSON data without any kinds of transformation. So we just handle the, you know, the JSONs natively, which is very useful. The next one is Marketplace. So Marketplace, you know, there's a lot of third party data that's out there that's super useful for both analysis and machine learning. So for example, weather, stock market data, COVID data, there's a lot of other different providers and those providers are growing super rapidly. So this is like a one button push. Hey, I want weather data. And if the weather providers on Snowflake and you're on Snowflake, it's basically a permissions issue that the weather company grants you access to their table with the weather data.

And there's no APIs or maintenance or anything. It's just like very seamless. The next one is BI tools. We both have a very robust ecosystem of BI tool providers, like for example, Tableau and ThoughtSpot that you're able to connect to with a connector to Snowflake. But we also have Snowsight. So Snowsight has our own BI tool that we're gonna go into more detail in a few slides, which allows you some really great visualizations and dashboard capabilities that are super simple and easy to use. Streams and tasks has to do with like, hey, you know, data pipelines and data pipelines using SQL, which is, you know, super useful. Snowpark, so Snowpark is able to run Java and Scala, including data frames, which is super exciting to the data scientists, you know, within the Snowflake compute ecosystem. So you don't have to manage any of the compute outside of Snowflake. which personally I'm very, very excited about. And then the last one is zero copy cloning. So as a data scientist, I wanna do something destructive to a table, but usually you just have to copy that table and then do your destructive stuff to it.

Here, able to take a snapshot of the table without physically copying. So it's basically behind the scenes as just keeping track of the changes that you made to that original table. But you can do all that stuff without making a physical copy, which is nice. So Margaret's going to talk about then, hey, for our example in server forecasting,

Margaret Campbell: Yes. So hopefully I don't get cut off this time. Sorry, guys. So in this specific example, we had server logs. So basically, we have information about every server that we have at every minute across all of our deployments. And then we also have information about kind of what growth we're expecting. And we can imply that from contracts and leads and billing. The interesting thing here is I, as the expert in the server logs, had lots of knowledge about that. I didn't have so much information about how to use contracts and billing, but fortunately, our finance team had already done a lot of work to model that data and make it useful. And then because Snowflake allows world-based access control, I don't have to have access to the entire finance database. I can get access to the specific pieces of data that I need, and I can make a case for why they're useful. And then with those specific signals, we're able to answer the question of how will peak server usage grow over six months? So this was kind of an interesting back and forth over some period of time with obviously our business owners and, you know, capacity planning and finance team to try to come up with a useful data set that we then turned into a forecast about exactly that we then pulled into Python and turned into a forecast about what growth we're expecting to see over the next six months.

Oh, no, Chris, I think we've lost your sound.

Chris Pouliot: Oh, yeah. Sorry. So you have the data model. You have the data ingested. You have the data model. You know, you run queries on the data you just ingested, you know, and it's well modeled. So you applied the business transformations and logic to it. Now, as a data scientist, you want to do some simple... you know, exploratory data analysis. So the kind of questions you might want to ask is like who, what, where, why, and when. This is kind of like the basic questions and answers and basic information gathering, right? If you think of data analysis, you maybe can decompose it into three areas. Like I said, the data model, fast, accurate, consistent queries. If it's fast, you can answer more questions, more hypotheses, more answers, back and forth, back and forth. You build upon your learnings from the past question and answer that you just did, and you help find the answer. Let's answer quickly. There's a big part of data analysis is actual statistics, math and statistics. So for example, outlier detection. Outlier detection, you can boil it down to what is expected. And then you have an observation, and then you ask yourself, how does that differ from my expectation?

And then you want to ask yourself, is it statistically significant? Could it just happen by random chance, or is there something that's actually going on? So this is where like a lot of the statistics is needed. And then the third area has to do with visualization, right? Because a lot of times, you know, for humans, you know, things are abstract, you can't see, right? So you see it, you can very quickly, the human mind is really good at seeing trends over, you know, trends over time. Looking at the outliers, it helps you. determine, you know, outliers or patterns that you might not, you know, realize otherwise. So it's visualizing the data is always, you know, necessary and super useful.

Margaret Campbell: So I think Chris had made an excellent point at the end, especially about visualizations. I think one of the key problems is that oftentimes the people who are the most expert in the data, which may be, for example, with operational data, an engineer, don't have access to the ability to very quickly look at their data, right? Just visualize it. And that's one of the amazing things about Snowsight is really you just have to know SQL. This is part of our UI where you can write a very simple basic SQL query and it will automatically pop up a chart and you can modify the chart and that gives you very good, very quick insight into what's happening with your data. In addition, Snowsight provides some really useful insights. For example, here, this is an anonymized dataset and I've done a trend on it, so I have some data going forward that's null. And very quickly, I can click on this value and take a look at it, and it will tell us, sure enough, we have blank values. And more than that, it'll tell us the distribution of data down here. So we can very quickly see what's the frequency of different distributions, and that we also have some outliers out there.

Also, again, to kind of highlight the fact that charting is really simple, you know, if we're looking at the results, we can switch quickly from looking at the tabulated form to looking at the charted form. And with the click of a button, Snowsight actually knows that we have created this column. It recognizes the variable and will chart it for us. And then obviously after you have all of those nice insights, you may want to take a look at them with other pieces of data. So what are the trends in sales compared to customers? This should be highly correlated. And so really quickly here, you can take all of those individual queries that you have written and put them together on a dashboard so that you can look at the data together. And each of these tiles, again, is that query. And then it's very easy to modify the query with some templatized SQL that is native in Snowsight. And it automatically will add filters so that you can update these dashboards and get insights into your data as you go. which is really a helpful thing. I always, when I'm exploring new data sets, I like to do what I call live with my data for a little while to see whether I've extracted all of the useful insights to kind of watch it change over some period of time so that I really understand exactly what's happening before I start trying to build something massive off of it.

Chris Pouliot: Okay, yeah, thanks, Margaret. So the next area has to do with, you know, machine learning, which, again, is supposed to, like, hey, you know, it's not just, like, what's happening, you know, yesterday, the last month, the last six months, but it's really trying to say what's going to happen tomorrow, the next month, the next six months, right? So it's looking forward versus being descriptive and looking in the past. So the machine learning process, a lot of it, maybe I'll dive into an example. Say I'm trying to predict the weather tomorrow, and is it going to rain or not? And I have a lot of historical data, and the main hypothesis is that the past predicts the future. And then in this data set, I'm going to try to say, OK, what are some features or what are some columns that I think are going to be useful in predicting whether or not it's going to rain tomorrow? And I might have some idea of, like, hey, I'm going to look at today's barometric pressure. I remember from elementary school that barometric pressure has to do with whether or not it's going to rain or not. And I'm asking maybe then I'm also going to look at, you know, is it raining today?

You know, I think my hypothesis is if it's raining today, it's going to rain tomorrow, right? There's probably some kind of correlation, you know, between the two. So I have to ingest that data, I have to find that data, I have to ingest it. Some of it may be from the data exchange marketplace, you know, to get third-party data. And I ingest the data, I model the data, I might need to do some transformations, and then I have that in Snowflake. And at that point, then I want to train a model, which is applying the math, saying, like, and trying to find what's the connection between those features and whether or not it's going to rain, the probability that it's going to rain tomorrow or not. I iterate, I fit a model, I look at the error and I ask myself, is that acceptable error or not? If it's not acceptable, then I go back, right? I say, okay, what other features can I add? So maybe I'm gonna look at, know time of year you know geographical area like in california in the bay area like it's in july i can tell you for sure it's not going to rain tomorrow because that's just our weather pattern you know uh here um so there's a lot of the so there's other data that i add to it then i refit the model right so my point is there's a lot of back and forth and you know 80 of this whole machine learning you know modeling process is you know getting the data visualizing and understanding it future engineering all the stuff that snowflake does you know really well um and then it comes to the 20 we're actually doing the math and optimization you know fitting the model which we we're going to talk about in a second where we're going to you know we have a very strong partner ecosystem and that supports that has really great snowflake connectors uh you know to to get the data um and then you know then there's productionizing the model and evaluating it So going here, so some of the overall benefits of Snowflake, you have the data, it's in one place, there's no data silos, which we mentioned before.

We have RBAC, which is role-based access control. And in order to make sure that the right people Only the right people have the right access to the data on kind of like a need to know basis. So it's not open to everybody. It's just open to people that need to see that data. There's near zero maintenance. We automate a lot of those different things that you before you need to have like a DBA in order to maintain it. And unlimited performance and scale. If you want more performance, you just create a new warehouse or a larger warehouse in order to meet those compute needs. So going here, talking about the Snowflake features, a lot of these I already talked about. We talked about Snowpipe and JSON in the marketplace. You know, we talked about, you know, Margaret talked about Snowsight and visualizing the data. Now, feature engineering, you know, we support ANSI SQL streams and tasks, which is like helps, you know, the Snowflake feature and that helps, you know, create the data pipeline. Snowpark, which allows, you know, Java and Scala to run natively within Snowflake. which is, you know, super useful.

And then Python is gonna be coming in the future. So as data scientists, you know, we prefer Python versus Java and Scala, but Java and Scala was, you know, prioritized first, because it was easier to implement. um and then i talked about zero copy cloning so these are the 80 you know you're you know you're in here using snowflake features you know um to create the data create your training data set create more features etc and then i mentioned now it comes to like oh i need to actually do the math to do the optimization to come up with the with the model um you know we partner we have strong partnerships so we you know our partners could be like data robot h2o dataiku and all these uh companies have uh you know, integrations, connectors to Snowflake, right? The depth of the integration varies by partner. Most can read and write back to Snowflake and some can use external functions. External functions are functions that are hosted on, say, for example, AWS Lambda, which you have an API that you can interact with. And these external functions can even trigger training, right? Also, I wanna talk about the Snowflake Python.

So the Snowflake Python connector comes with native Panda support and is used very heavily by our data science teams. And then also, we have Java UDFs. So if whatever machine learning platform you're using exports a JAR file for the trained model, we could use it in the Java UDF in order to do bulk scoring and inference in Snowflake. So I think that's also super exciting. Moving on, you know, I said, you know, you get the data in a model, you go back and forth sometimes, you know, five times, 10 times, 50 times until you get to the error that you find acceptable in your model. And then you want to productionize it. and you push it to production, and you want to evaluate it. So a lot of our partners supports model management. Model management might consist of looking for a model drift over time. So you already know your error starts to creep worse and worse over time. And then the action taken would be to retrain that model with newer data. And then they'll have these. ML tools will help facilitate that. And also you might want to A-B test, you know, a new model versus an existing model.

This would also facilitate that. But a lot of times, you know, like these ML partners can write data back to Snowflake. So you can even like create a dashboard of like model performance over time and do some of this tracking within Snowflake itself, which is useful. That's the story on machine learning. I think it's a great story. As a data scientist, I made machine learning models in the past, and a lot of the pain that I felt was actually getting the data, manipulating the data, versus actually training the data. This is where Snowflake Forte comes into play.

Margaret Campbell: So coming back to this example, I wanted to highlight here, you know, this is, again, that anonymized server forecasting data or server data. And Snowflake has internally the ability to do some nice simple regressions. We can build a trend line. But there might be something more interesting that you want to actually build some good machine learning around. So, for example, here we see very strong weekly seasonality. where people spin up a lot of servers during the week, and then they spin them down during the weekend because they take vacation. Hooray, good job, people. We also have annual seasonality. This dip corresponds to Christmas and New Year's, and we see it every year, and it means people are taking time off, which is great, but it's something that, depending on the granularity that we need this data at, we might want to be able to take advantage of, or take advantage of, be aware of as we're planning. So we pull it into Python. And the beautiful thing about this is with the Snowflake Python connector, again, as Chris mentioned, they have built-in Panda support. So pulling in data sets is extremely easy.

And I personally use my own little library of useful tools. Um, and then the nice thing is used to be places to work. Um, all of the data is right there for you. You can build as complex or as simple a model as you need. Um, but the, the connection between your data warehouse, where you've been able to munch your data and build something very meaningful to put your analytics on top of, um, is very easy and straightforward. And then as Chris mentioned, um, you know, you, you can take the output of this model. And you can stick it back into Snowhouse and model and watch your model over time, see how it's performing, whether you need to change anything, and also just to be able to report out and have all of those pieces of data, again, back in the same place, which is really helpful.

Julian Ferreira: Yeah, thank you both for sharing all the details around how you guys have been exploring. Some features are already available and some that aren't. And so to close what you guys have already seen is that the reason they're able to leverage all this data in an efficient manner is the fact that there's unified security and governance on all that data. And so as a data scientist, It's easy to ask for data, but then you only get the access you need depending on the exercise that you wanna take, right? And so sort of to the example of, yeah, I needed some data from finance, but I don't need to have all of it, just some specific pieces of it. And so using some of our features such as role-based access control in order to manage that. And then if so, for some reason we had, for example, PI data, or we were the pharmaceutical company, we had some health information or patients, then we could do things like dynamic data masking. And so that way you can still use the data for analytical purposes, but only those with the right we'll be able to see the data exposed in the format. I think the other big important piece, of course, is sort of that access to the data.

And it doesn't matter if it's structure on structure and easily be able to pull that using our query system, which takes this variant format for semi-structured data. It makes it really easy to query files that maybe are in sort of adjacent format or any of those other familiar semi-structured data format. uh and then last of course i think there's a really really ton of value to the snowflake data marketplace in that it could be not only to get access to third-party data sets uh which some of our teams do right so some of the guys building predictive models for cells are bringing in different account information through the data marketplace but it could also be to sort of share data with partners and so what we see this is for example and the industrial or supply chain sort of industries where you have a lot of data that sits in like your customer, your supplier, and then your data, then how do you make that easily accessible with, while still not breaking any sort of data privacy or governance rules, using the same features that your internal organization may be using to manage the security of that data.

And so I think that's becoming a really efficient way for organizations to just increase the amount of relevant data for their models that's beyond their four walls. And then last and not least important is of course the fact that it's going to be really scalable and you're going to really get the performance that you need. And so regardless of whether your data volumes are growing or the number of users that are using that data is growing, Snowflake's gonna be able to support all of that, all with almost zero main instance so that in some systems you have to sort of think about scaling, right? And build their own predictive models to how users are gonna use the platform so that they can manually scale and think about how much my environment needs to go up and down depending on the amount of users or the amount of workloads that I'm planning to have. And so it's completely having to disregard that it's really valuable for data scientists. And sort of close on some of the things that Chris has already highlighted. I think some of these are some of the features that our customers who are also starting their journey and evolving their journey around data science with Snowflake that have been found really valuable.

And so I just want to touch base on a couple of these. I think, of course, Chris already mentioned that the zero copy cloning, which makes it really easy to pull data that's in production, sort of have your own. view of it without actually making a copy. And so the reason it's called zero copy is because you're not actually making a copy, so there's no duplicate storage, there's no duplicate cost for that table view that you get to use for your experimentation for any data science modeling. I think, of course, data marketplace. We just talked about that one right before. The other P is about using SQL to do a lot of the transformations in your data in a very efficient manner before you do something like using our Python connector to pull data into your development interface. And so this way, it sort of frees up a lot of that memory usage that would be taken up otherwise if you had to do a lot of those complex transformations, cleaning up the data and integrating the data that could all be done using NTSQL and Snowflake. Of course, trends and tasks as you're moving from prototype to production is, how do I keep feeding real-time data in these micro-batches to my model?

Using that in order to continuously feed the data and continuously score and put those results back in Snowflake. So that becomes really useful. Of course, our Python connector, as many of you may be working in IDEs, such as Jupyter notebooks. And so being able to grab that data from Snowflake and use your libraries and frameworks of choice, it's going to be really beneficial. And so we spent a lot of time in building that native Python connector. And so we really hope that our customers are enjoying it. And then of course, right now we have external functions, which Chris mentioned, this is a easy way to sort of trigger job sharing. So you may either connect it to a Lambda function and request a training job or even a inference job. And so that's one way in which we can use that. And just wanted to sort of clear up that some of the features that we did talk about are sort of coming. And so we're very excited as to how these will be so beneficial to the data science team, but just know they are coming and so not readily available if you try to use the product today. This includes our support for unstructured data, as well as the ability to use Java and Python within Snowflake, which is what we call our Snowpark.

And then last, something that we've started to talk about that I think will be really easy and change the way in which we bring models closer to the data is going to be this Java user-defined function in which we export a JVM and make that model easily accessible so that you would score or do the inference of the model as close to the data as possible so there's no additional back and forth of production data that's scoring real-time or close to real-time models. And so with that, I think we want to talk a little bit about some of the outcomes that we've seen from our teams who've been using Snowflake as customer zero and how that's helped us achieve Snowflake goals around growth, cost reduction, as well as risk reduction. And so I think as an organization, we want to be able to manage all those three categories of business outcomes. And with Snowflake, we're able to do all of those in some way. One, by focusing on going from moving into predictive analytics, we can see more ways in which we can innovate our features. We can also help how manage the experience that our customers get through things like doing, making sure that everything works as expected through some of these models that we talked about, such as server forecasting.

And of course, we as users of our data sharing program also influence the way in which that amount of vendors come. As our product teams are requesting third-party data sets and we want to make sure we partner with the companies that can provide those data sets that power our models as well as could power the models of other organizations. In terms of cost, I think, of course, because of the efficiency and the autoscaling that comes with Snowflake, you can save a lot of cost on compute. And because you are driving faster time to value, it's also going to help us reduce some of the cost as well as our employees, right? The more efficient they are and the less time they spent waiting for queries to run or for models to have available and accessible data, then they can focus on those high value activities, which overall is a win-win both for our teams as well as for the company. And then, of course, last is being able to reduce risk. So we talk about some of those features such as data masking that help us protect data. We do have financial information about customers. And so only the people that need to have access to that are going to be able to have access to that.

And so Snowflake provides a great amount of features that help us control that. Same thing. We're providing a standardized data management so that only good data goes into the models because if you have bad data, then you're going to have bad models and you're going to make bad decisions. And so being able to have standardized management procedures throughout the company. And then, of course, roles-based access controls and sort of the policies that need to be implemented so that only the right people have access to the data they need, whether it's a system or a person. uh this is how at snowflake we've been able to drive value uh using our our own uh product and so we're really excited to see how our customers are going to continue to evolve and also start applying data science workloads with the data that they are having in snowflake and so uh I want to do sort of Chris ask you if there's anything else that you wanted to add that's particular to some of the benefits you've seen in your team before jumping into sort of the future state and where you're seeing things going for your team.

Chris Pouliot: Yeah, thanks, Julian. Yeah, I was thinking that, you know, one new feature that's come out recently is column and row masking policies. So, for example, for column masking, you know, say there's a sensitive column. Maybe it's email, right? I mean, you know, so it's, you know. PII data, and only a certain number of people need to see that, see that column. So you can create a masking policy saying, you know, only this role has access to actually see the column. And if you don't, and if you're not in that role, then it's just all kind of like, it's kind of like a blank is blanked out or crossed out. or X dot or whatever you want to call it, right? So I think that's super useful. So instead of having to, you know, it's a super easy way to create a policy to mask that column for the people that don't have the need to know. So that's a great benefit. And the same goes with row masking. So row masking, say, for example, you have a sales organization and you only want the salesperson to see the rows that correspond to their opportunities, then you can create a masking policy that says, hey, salesperson XYZ can only see their opportunities and they can't see others, right?

So I think those two features really enhance the capabilities of Snowflake when it comes to mitigating and minimizing risk in an easy, straightforward way.

Julian Ferreira: Yeah, I think that's one of the great features that are going to build a stronger story around our data governance and data security for global privacy and compliance requirements.

Unidentified speaker: So I can't attest to that.

Julian Ferreira: So I think before we jump into Q&A, we'd love to hear from you guys sort of what you're thinking about doing next without exposing any secret information, and so.

Chris Pouliot: Right, yeah. No, I think, you know, yeah, there are a lot of exciting future features coming, you know, down the pipeline that I can't talk about, but I would say stay tuned for future events, webinars, or we have the Snowflake Summit, coming up in the summer, which is always a good forum for our leadership to announce new features. And then I'm really super excited about data exchange. We talked about data sharing, marketplace for third-party data. It's a marketplace. It means you have both providers and consumers. So you have to grow both at the same time. The more providers you have, the more consumers you have, which will cause more providers. So it's kind of like this viral effect. And I said, I think from my past experience, the power of third-party data to augment your first-party data is very powerful. Having a diverse set of data that you can get very easily at the press of a button in order to get access to the data is very beneficial to the company. Um, yeah, so those, that's kind of like, and then, you know, I'm looking forward to, uh, trying out these features, giving that feedback and, um, you know, helping shape the product, you know, being part of the product development, um, you know, life cycle.

Julian Ferreira: Yeah, we appreciate it. We definitely appreciate that from you guys. I think before we jump into Q&A, I just want to say thank you to everyone that stayed with us during this webinar. And while we did cover a lot today, we can always continue the conversation. So in addition to be able to connecting with our Snowflake team, you can also join me in an upcoming webinar that's going to be around how to build a foundation for data science using Snowflake's Data Cloud. And so I'll be joined by our fuel CTO for data science to speak more about that. And also, if you want to learn more about upcoming features as well as customer stories, then you should definitely register for our upcoming June summit. And the links are going to be available in the resource box, so you can learn there and register more by following those links. And with that, let's jump into a couple of Q&A for the remainder of the time. And let me just take a quick look. What are some of the questions that have been coming in? So I think one of the, maybe it wasn't a question, but more of a comment is the idea of sort of a, full stack skill set within the team.

And so I would love maybe for Margaret to answer how Snowflake has enabled you to do as much as possible to maintain that full stack experience.

Margaret Campbell: Yeah, I think really the fact that you can play around with the data immediately once it gets into the warehouse is really huge. Typically, you have to work with a data engineer to make sure that everything's hooked up and that it's coming in in the right format. But the fact that, you know, I actually saw another question that's about, you know, how does the data science team work? We're actually embedded with engineering teams. So I'm embedded with the team whose questions I'm trying to answer. And that means that I get to talk with the people who are actually building the features that I'm trying to evaluate. And we're working together to do that. And that means that I teach the data, you know, from its very raw state. all the way through to test it, to explore it, to talk with the people who are stakeholders, and then to finally build a data model that will allow us to answer those questions very, very quickly. And obviously, as Chris has mentioned, a large part of data science is that data munching. But if you are exploring data or playing with data and you don't know where it comes from or how it was built or what assumptions were made in, for example, defining a certain field, it can be very easy to make mistakes that are very costly and lead to very bad conclusions.

And so I think the thing that I really like about the idea of a full stack data scientist is you become an expert in the questions you're trying to answer. And you really do, you know, go with the data for its entire lifecycle and you become an expert in that area.

Julian Ferreira: Yeah, and I think that totally aligns with what happens a lot of times, right? Different organizations are going to have data scientists sitting in different teams, and the more they sit outside an organization, the longer it's going to take to catch up and sort of understand the What's the business question right to the beginning of your presentation is that you do have to have a business question in mind. And then sort of being a business expert helps a lot with figuring out the relevant data that you're going to be able to pull. And then once you realize I've been talking to the right teams that are going to give you access to that clean data that's already available in Snowflake, whether it's within your team or outside your team. So I think that totally makes a lot of sense. Another question we had was, and also a little bit of a comment, is that in the experience, and maybe, Chris, you can talk about this, once you've seen data silos sort of break, organizations may have this mentality of that the data needs to be available to everyone to facilitate the collaboration. So sort of as your experience working in different companies and different industries, um what's your point of view on sort of the the access to data and the governance of that data once it's all centralized right

Chris Pouliot: yeah so i think you know great observation um i think it depends It's on your company and your governance, you know, policies, right? I mean, obviously, you can say, well, now we have HR data, like, in our data warehouse, and, you know, that includes everyone's security number and salary. We don't want to make that available to everyone. So you also have to think of what's the balance between productivity and then, you know, governance and security. And then, obviously, like, there needs to be a need to know – You know, you have to have a need to know in order to have access to that, to certain data sets that are sensitive, right? So given that you kind of like have the role-based access control or the common road masking, you know, in order to handle that, then, you know, then you should, you know, most companies, in my opinion, you should have access to most of the data as long as you have a great need.

Julian Ferreira: Yeah, I think the important thing there is there's a lot of Snowflake features that sort of help you manage the right level of information that every user should have. And so we'll go down to each individual's company as to what is available and, of course, making sure they're meeting the regulatory and privacy compliance that is relevant to their industry. And so all the tools are available in Snowflake, but, of course, each customer is going to have to implement them in a different fashion to meet their specific industry use case. So we're sort of very close to the top of the hour. And so I want to thank everyone that attended. If you want to get a follow up for a question, definitely connect with the Snowflake team. And of course, remember that the recording is going to be available tomorrow. And don't forget to check out the resources to see more information about upcoming webinars as well as our June summit. where there will be more product announcements and customer stories to learn from. Thank you, Chris. Thank you, Margaret. It's been great to hear the story of how you guys are sort of being a user and an influencer of our data science strategy.

And we hope that you join us again on another Snowflake story soon. Yeah, thanks, everyone. Have a good day, everyone. Bye.