Feature Store

What is a feature store?

In machine learning, "features" are the columns you feed a model: a user's age, orders in the last 30 days, a product's price. A feature store is a warehouse that keeps all these features in one place, managed centrally. Training and online serving both pull from the same pool — whoever needs a feature just comes and takes it.

What core pain does it solve?

Training vs. serving mismatch
You compute features with one script for training, then recompute them with different logic in production. The two don't line up, and the model's live performance falls apart. A feature store keeps features in one place so training and serving both read the same thing.
Reinventing the wheel
The same feature gets written by different teams, slowly and with bugs. Centralize it, compute once, and use it everywhere.
Features are hard to reuse and govern
Features scattered across scripts are hard to find, change or trace. A feature store gives them a home, plus versioning and lineage.

What does one look like?

Online store
Serving needs low latency, so features live in an in-memory store like Redis — fetched in milliseconds.
Offline store
Big historical batches for training go in a data warehouse or object storage.
Feature registry
Records each feature's definition, source and version, so it's easy to discover and reuse.

When should you reach for one?

When you notice your team computing "the same feature" several times over, or models that train fine but break in production, a feature store is often the fix. It's a growing piece of MLOps infrastructure.

Bottom line: a feature store is the central warehouse for features — one place for training and serving, compute once, use everywhere, and no offline/online mismatch.