Feature Stores

What it is

A feature store centralizes feature computation and storage so the exact same feature values (and computation logic) are used for both model training and live serving — directly targeting a common, hard-to-debug class of ML bugs.

Key points

• Train/serve skew: the classic failure this solves — if training computes a feature one way (e.g. batch, over a week's data) and serving computes it differently (e.g. real-time, slightly different logic), the model behaves worse in production than offline metrics predicted.
• Offline vs. online stores: a feature store typically has both — an offline store (for training, historical/batch data) and an online store (low-latency lookups at serving time) — backed by the same feature definitions.
• Feature reuse across teams: once a feature is defined and computed once, other models/teams can reuse it rather than each team recomputing the same logic slightly differently.
• Feature stores are most valuable as the number of models and teams grows — for a single model with a small team, the coordination overhead may not be worth it yet.