What a feature store does
A feature store is a central system that computes, stores, and serves the input features your models depend on. Its core job is to make the same feature definition available both for training and for live serving.
The two halves
- The offline store holds large historical tables, often in a data warehouse or object storage. It powers batch training jobs where you need months of rows and latency does not matter.
- The online store holds the latest feature values in a fast key value database. It serves a single prediction request in a few milliseconds.
Why split them
- Training reads millions of rows in bulk, so it wants cheap columnar storage and high throughput.
- Serving reads one entity at a time under a tight latency budget, so it wants a low latency lookup keyed by entity id.
- One storage engine cannot be great at both, so the feature store keeps both and syncs values between them.
The shared definition
The key benefit is a single feature definition that both halves read. That consistency is what prevents subtle mismatches between how a feature looked in training and how it looks in production.
Key idea
A feature store keeps one feature definition behind two backends, a bulk offline store for training and a low latency online store for serving.