Feature Store

A feature store is a central system for computing, storing, and serving the input variables (features) that machine-learning models use — things like “customer's average purchase over the last 30 days.” It lets a feature be defined once and then reused across many models and teams, both for training and for live predictions.

Its main job is to solve a subtle but damaging bug called training–serving skew: when a feature is calculated one way in the training notebook and a slightly different way in the live system, the model silently performs worse than expected. A feature store guarantees the same definition is used in both places, and also avoids data leakage by serving the value as it was known at the right point in time. It is advanced infrastructure — small projects do fine without one.