A feature store is a system that defines, stores, and serves features, the input variables a machine learning model reads. It usually pairs an offline store for training data with an online store that serves the same features at low latency during inference.
A fraud team, for example, registers a feature such as average transaction value, then reuses the identical definition for both model training and real-time scoring.
Feature stores exist largely to keep training and serving consistent and to enforce point-in-time correctness. Even so, replaying a past prediction requires the exact feature values and data version present at run time. Versioning features as part of a reproducible data state makes a prediction auditable after the fact.