Definition
Feature stores prevent training–serving skew by reusing the same feature definitions for batch training and low-latency online lookup.
They version features, manage point-in-time correctness, and often integrate with warehouses and stream processors.
In simple terms
A shared spice pantry with labeled jars — every chef (model) uses the same measured ingredients instead of reinventing recipes inconsistently.
Where you see it
- Fraud models reading fresh user risk features online.
- Recommendation systems sharing embedding and activity features.
How it works
1.Define features
Code or config declares transformations.
2.Materialize
Batch and/or streaming compute writes values.
3.Serve
Training jobs and online APIs read consistent features.
Why it matters
- Feature stores scale ML teams past one-off notebook feature scripts.
Often confused
A feature store is just a database table.
The value is contracts, point-in-time joins, and training/serving consistency — not storage alone.