VideoAmp Unifies ML and Empowers Its Engineers on Snowflake

Founded in 2014, VideoAmp has been re-envisioning how media is valued, bought and sold for a decade. Unlike traditional media measurement companies that may rely on extrapolating small sample sizes for analysis, VideoAmp provides a robust software platform that uses huge volumes of data from 39 million households and 63 million devices that span everything from set-top boxes to smart TVs. Using its proprietary commingling methodology, VideoAmp offers the industry its trusted, high-quality TV data set. 

All that data requires a scalable system to process it so the team can develop a meaningful base for its media currency. In the past, VideoAmp used a traditional data center, which produced viewership tables that were sent to Snowflake for downstream applications. Seeking a robust, easy-to-use data platform with flexible infrastructure, strong data governance and secure data sharing that doesn’t require data duplication, VideoAmp consolidated its data warehouse on Snowflake.

The VideoAmp team had already seen what kind of ROI Snowflake could provide for analytics and insights. But given its exceptional performance during the proof-of-concept stage against their previous environment, the team decided to move fully onto Snowflake.

Working with truly ephemeral workloads in Snowflake also means wasting less time waiting on clusters to spin up. What used to take 10 minutes to spin up a cluster now only takes seconds with a Snowpark warehouse. Shorter latency means VideoAmp can spin up clusters in minutes rather than hours. VideoAmp also saved money on egress charges when moving data between separate environments and Snowflake to run queries, effectively eliminating the hidden costs of “double taxation” for data movement.

With Snowflake’s native application framework, the data quality is more reliable and consistent compared to VideoAmp’s previous environment, giving VideoAmp’s customers more power to use their data effectively. 

Now with Snowflake, VideoAmp developers can also create richer data sets that allow them to unlock new capabilities in their platform without a full re-architecture. And since large data sets are cost-effective to use, they can ask very complicated questions of the data, enrich it, surface it in multiple ways such as Dynamic Tables, and optimize delivery by use case — all without creating new, complex data pipelines.

Thanks to Snowflake and its partner ecosystem, business teams across VideoAmp — from revenue to customer support — now have quicker, easier access to this rich data. After moving its BI tooling provider to Snowflake partner Sigma, VideoAmp delivers more interactive dashboards that both customers and employees use — while also reducing complexity and costs for both the cloud and their BI environment.

Snowflake's robust Horizon Catalog governance features like data masking and row access policies allow VideoAmp to maintain the security they want without having to replicate data for customers. All these improvements have translated to happy teams across the board. Business teams interact more seamlessly with their data versus needing to request bespoke reports or export data in many different ways to surface insights they can use. And developers write less custom code and can prioritize product enhancements. "If we can get insights faster, we can do our job more efficiently,” Adams says. “It’s lower risk and lower latency. Who wouldn’t be happy with that?”

For a media measurement platform like VideoAmp, processing speed is key — along with confidence in data quality. Especially for TV events like live sports, for which organizations invest millions of dollars in ad spend. VideoAmp must be able to turn around billions of rows of data in the span of an hour and still be able to perform quality engineering checks, ensuring that the data can be verified and validated. “Snowflake has been incredibly helpful in that, and we can surface our content ratings and be confident in them,” says Adams. 

Data processing speed also matters when VideoAmp has to do foundational tasks like updating methodology for how they measure impressions. When undergoing those processes, the team has entire years’ worth of data to reprocess, which used to take five whole days. With Snowflake, VideoAmp can reprocess a year’s worth of data in mere hours, allowing them to quickly see and verify their outcomes. The speed ultimately frees up two full-time DevOps engineers, who can work on more impactful product enhancements for VideoAmp customers.

As VideoAmp's data teams settled into a fully consolidated Snowflake environment, a new challenge came into focus: ML still depended on a separate platform. At the center of this was “Personification,” a model that takes device-level viewing events and attributes each event to a person in a household, answering the question of “who is watching what?”

In practice, this meant pulling data out of Snowflake for every model training run and, as Andrew Harding, Principal Engineer at VideoAmp, describes it, “paying a third-party cloud provider to orchestrate SQL queries that ultimately end up running on Snowflake.” Resulting ML model artifacts only landed in Snowflake at the end of the model lifecycle, while internal teams were left to construct all the ML inputs themselves. The solution was the same as it had been for the broader data ecosystem: Bring the process to where the data already lives. With compute unified on Snowflake, VideoAmp rearchitected how the team builds and deploys ML.

For the MLOps team, consolidating on Snowflake meant replacing Spark-based pipelines with pure SQL and Snowpark Python, delivering one environment with a single set of dependencies and zero framework switching. Snowflake ML wrappers now handle distributed training across XGBoost, LightGBM and other algorithms automatically, without requiring any manual compute configuration. For the heaviest jobs, the team scales up to Snowpark-optimized warehouses and compute pools in Snowpark Container Services. The Snowflake Model Registry logs every model version with its runtime fully specified and ready for inference, and ML Datasets connect those artifacts to training data and inference results. The team controls complete lineage through the full model lifecycle for the first time.

For the teams building VideoAmp's products on top of those models, the change is equally significant. Where teams once had to navigate constructing complex ML inputs, they can now interact with a modeling service through a simple, stable interface. Changes to the underlying model, including swapping algorithms and updating features, require zero refactoring by the teams consuming the modeling service, who may not even know a model is involved at all. The modeling team can now look at the orchestration and results of the system on their own terms. As Harding puts it, building that kind of encapsulation “really isn't simple, but working with Snowflake ML made it a lot less painful.”

With the Model Registry and distributed training now running in production, the VideoAmp team is turning its attention to the Snowflake Feature Store and continuing to explore the full breadth of Snowflake capabilities. The next chapter will connect source data to training datasets, model artifacts and inference results in a single, observable lineage. With Snowflake ML now embedded in the team's workflow, VideoAmp is steadily expanding which engineering teams can build and own their own ML results. The goal remains the same: providing the infrastructure for media buyers and sellers to find their most valuable audiences, optimize to what’s working, and measure the real-world impact of advertising.