Yieldmo Produces Advertising Predictions 20x Faster at 25% the Cost with Snowpark Container Services

Evolving consumer data privacy requirements are transforming how organizations advertise online. As industry discussions continue about third-party cookie deprecation, advertisers seek innovative solutions to a complex problem. Yieldmo’s programmatic advertising exchange offers a “privacy first” approach to optimizing campaign performance in near real time.

Yieldmo analyzes dozens of signals — such as smartphone scrolls, tilts and swipes — from billions of daily impressions to predict audience, inventory and creative performance. Making these split-second advertising decisions with predictive insights requires AI expertise, low-latency data processing and a finely tuned data architecture that can quickly adapt to meet surging demand and evolving requirements. To achieve these goals, Yieldmo relies on Snowflake to democratize data, control costs, avoid data movement and accelerate AI workloads.

For years, Yieldmo has relied on Snowflake for data storage and processing. The platform’s elastic compute, interoperable storage, near-zero maintenance, native SQL and Python support, and cloud services allow Yieldmo to power a growing number of use cases. 

Centralizing data in Snowflake made it easier to launch Yieldmo’s BI team and enhance business insights. “Data in Snowflake drives the majority of our decisions,” says Yieldmo CTO Duc Chau. “Snowflake has democratized access to intelligence across the organization in a low-stress way that helps us maximize the value of our data.” For example, Yieldmo’s operations team depends on data in Snowflake to optimize advertising campaigns, while partnership teams make data-driven decisions that lead to healthier relationships with partners. 

Yieldmo’s development and engineering teams run simple SQL queries in Snowflake to understand application performance without manually searching log files. Data scientists easily connect ML tools to Snowflake with less development and IT operations or DevOps work. “The turnaround time from ideation to materialization has been much quicker because of Snowflake,” Chau says.

Yieldmo processes most of its data science workloads with Snowflake and uses H2O Driverless AI to develop machine learning models and produce predictions. Initially, moving large amounts of data from Snowflake to H2O Driverless AI posed operational challenges — especially as data volumes grew.

Yieldmo reimagined its ML inference layer by running H2O eScorer as a service with Snowpark Container Services, bypassing data transfers and bolstering compute capacity for predictive workloads. “What I really appreciate about Snowpark Container Services is being able to do the work by myself using SQL,” says Sergei Izrailev, Head of Analytics and Data Science at Yieldmo. “It’s literally just a few SQL statements to set things up and get started.”

Initial testing shows predictions are 20 times faster with 75% lower costs and less infrastructure management. Yieldmo believes this solution will expedite A/B testing and time to market, elevate the customer experience and, ultimately, drive business growth. “Snowflake is helping us achieve our goals of increasing return on ad spend and deploying budgets more efficiently,” says Eric Shiffman, VP of Product Marketing at Yieldmo. “Providing clients with more clicks for the same amount of money encourages them to spend more with us.”

Snowflake Secure Data Sharing streamlines data collaboration between Yieldmo and its vendors. “Data sharing capabilities have been huge for organizations that are on Snowflake,” Chau says.

For example, Yieldmo securely shares data with Snowflake partner and AI software company Kumo — without having to develop and maintain ETL pipelines. Yieldmo also uses Snowflake Secure Data Sharing to provide log file data to another Snowflake partner, HUMAN Security, a cybersecurity organization that verifies digital interactions. According to Chau, “We just grant them secure access through a data share to their Snowflake account, and they can access it without moving data around.”

Yieldmo’s enhanced inference layer could lessen the company’s reliance on sampled data, enriching insights for downstream analysis. “We’ve always sampled the data because the data set is so big,” Izrailev says. “Now, we may be able to make predictions for every impression, for every ad served.” Snowflake Notebooks will also allow Yieldmo’s data scientists to explore data with less complexity.

Developing applications that request data from Snowflake — instead of traditional production databases — could further simplify Yieldmo’s data architecture. “Since Snowflake can operate both as a data warehouse and a low-latent data store, our team is experimenting with pointing the data access layer for apps to Snowflake,” Chau says. These optimizations will play a valuable role as Yieldmo continues to evolve to meet the changing needs of its clients and online advertising as a whole.