Feature Store
A central system for computing, storing, and serving ML features consistently between training and production.
A feature store is infrastructure that computes, stores, and serves the features an ML model uses — with the same definitions for offline training and online serving. It solves a core MLOps problem: keeping training and production features identical, and sharing features across teams.
Worked example: a “user’s 7-day purchase count” feature is defined once; the feature store computes it in batch for training and serves the fresh value at low latency for online prediction — so the model sees the same feature logic in both places. Gotcha: the reason it exists is training-serving skew — when the offline feature (computed in a SQL job) subtly differs from the online one (computed in app code), the model silently degrades in production; a feature store enforces one definition, at the cost of real operational complexity many teams do not need until scale.