feast-dev/feast

What it solves

Feast provides a centralized way to manage machine learning features, bridging the gap between historical data used for training and real-time data used for inference. It prevents data leakage by ensuring point-in-time correctness and decouples ML models from the underlying data infrastructure, making models more portable across different environments.

How it works

Feast acts as a data access layer that abstracts storage from retrieval. It manages two primary types of stores:

• Offline Store: Processes historical data for batch scoring or model training.
• Online Store: A low-latency store that powers real-time predictions.

It includes a feature server to serve pre-computed features and a registry to track feature definitions. Users can materialize data from the offline store to the online store for real-time use.

Who it’s for

ML platform teams and data scientists who need to productionize analytic data for model training and online inference at scale.

Highlights

• Broad Infrastructure Support: Integrates with a wide range of data sources, offline stores (e.g., Snowflake, BigQuery, DuckDB), and online stores (e.g., Redis, DynamoDB, Qdrant).
• Point-in-Time Correctness: Generates feature sets that avoid data leakage by ensuring future values do not leak into training sets.
• On-Demand Transformations: Supports transformations on read or write to process features dynamically.
• Feature Quality Monitoring: Includes built-in metrics, drift detection, and serving log monitoring.
• Vector Search Support: Includes alpha release support for vector stores like Milvus and Faiss for NLP tasks.