Our team is growing, and we’re finding that different people are rewriting the same SQL queries to calculate "average user spend" for different models. I’m looking into implementing a Feature Store. Is Feast too complex for a 5-person team, or is it better to just have a shared GitHub repo with standardized Python functions for feature engineering?
3 answers
For a 5-person team, Feast might be overkill in terms of infrastructure maintenance, but the problem it solves is real. If you're constantly recalculating features, you’re wasting money and risking "online-offline skew" where your training features don't match your production features. A middle ground is a "Feature Library"—a central repository of versioned SQL and Python scripts. However, if you plan on scaling your real-time serving soon, start with Feast now. It forces you to define your features as code, making the transition to production much smoother than a bunch of fragmented notebooks or a shared repo of loose functions.
What’s the actual "Latency" requirement for your production features? If you need sub-10ms lookups for user profiles, a shared GitHub repo won't help you. You'll need something that can sync to a low-latency store like Redis. Does Feast manage that sync automatically?
We tried Feast and found it too heavy. We ended up building a simple "Registry" in MLflow to track which datasets were used for which model, which was enough for our size.
That’s a fair alternative, Lisa. However, as the models get more complex, the "Feature Drift" monitoring that comes with a real Feature Store becomes indispensable for maintaining model health.
Mark, yes, Feast manages the "Materialization" from your offline store (like BigQuery or Snowflake) to your online store (like Redis or DynamoDB). Linda, for a small team, the biggest hurdle isn't the code—it’s the "Point-in-Time" joins. Feast handles the logic of making sure you don't leak future data into your training sets by joining features based on the exact timestamp of the event. Doing that manually in SQL is where most errors happen, so Feast might save you more time than it costs in setup.