The feature engineering problem
ML teams re-implement the same feature transformations for training (batch), serving (real-time), and experimentation. The same logic diverges over time causing training-serving skew — the model sees different features at train time vs serve time.
What a feature store solves
- Single definition of a feature used by both training pipelines and the serving path.
- Reuse across teams — the orders team's "customer 30-day spend" feature is available to the fraud team.
- Point-in-time correct training data — retrieves feature values as they existed at training event time, preventing label leakage.
Offline vs online store
- Offline store — historical feature values, used for training. Backed by a data warehouse or S3.
- Online store — latest feature values, used for low-latency inference. Backed by Redis or DynamoDB.
Options
- Open-source: Feast, Hopsworks Community.
- Managed: Databricks Feature Store, AWS SageMaker Feature Store, Tecton.