feature store

Feature stores centralize the storage, management, and serving of ML features for training and inference consistency. **Core problem**: Features computed differently in training vs serving leads to training-serving skew. Feature logic duplicated across teams. **Key capabilities**: **Feature registry**: Catalog of available features with metadata. **Offline store**: Historical features for training (data warehouse, parquet). **Online store**: Low-latency feature retrieval for inference (Redis, DynamoDB). **Feature serving**: APIs to fetch features by entity ID. **Transformation**: Feature engineering pipelines, consistent transformation. **Benefits**: Reuse features across models, ensure consistency, reduce redundant computation, enable discovery. **Architecture**: Transform raw data into features, store in offline/online stores, serve to training and inference. **Popular options**: Feast (open source), Tecton (commercial), Vertex AI Feature Store, Databricks Feature Store, SageMaker Feature Store. **Entity concept**: Features organized by entity (user_id, product_id). Fetch features by entity key. **Time travel**: Retrieve historical feature values as they were at specific times for accurate training. Essential infrastructure for production ML at scale.