The role it plays
A feature store is a system that computes, stores, and serves features so the same definitions feed both training and online prediction. It exists to kill training serving skew, where a feature is computed one way offline and another way online.
Two access patterns
- An offline store holds large historical feature tables for training, optimized for batch reads.
- An online store holds the latest feature values per entity, optimized for low latency lookups during serving.
Why one definition matters
A feature is defined once and materialized into both stores. Training pulls a point in time correct snapshot; serving pulls the freshest value by key. Because both come from the same definition, the model sees consistent inputs in both worlds.
Point in time correctness
For training labels you must join features as they were at the event time, not as they are now, or you leak future information. A good feature store enforces this join.
Key idea
A feature store defines each feature once and materializes it to an offline store for training and an online store for serving, eliminating training serving skew.