feast-dev/feast

What it solves

Feast provides a centralized way to manage machine learning features, bridging the gap between offline training data and online serving data. It solves the problem of data leakage during training by ensuring point-in-time correctness and removes the need for ML teams to rebuild complex data retrieval logic for every new model.

How it works

Feast acts as a data access layer that decouples ML models from the underlying data infrastructure. It manages three primary components:

• Offline Store: Processes historical data for batch scoring or model training.
• Online Store: A low-latency storage system for real-time predictions.
• Feature Server: A battle-tested server that serves pre-computed features to models online.

Users define their features in a repository and use the Feast SDK to retrieve historical features for training datasets or fetch online features for real-time inference.

Who it’s for

ML platform teams and data scientists who need to productionize analytic data for model training and online inference across various cloud providers (AWS, GCP, Snowflake) and data sources.

Highlights

• Point-in-time Correctness: Prevents data leakage by generating correct historical feature sets.
• Broad Integration: Supports a vast array of offline and online stores, including Redis, DynamoDB, BigQuery, Snowflake, and vector stores like Qdrant and Milvus.
• Cros-platform Serving: Offers feature servers in Python, Go, and Java.
• Feature Engineering: Supports on-demand and streaming transformations.
• Data Quality: Includes built-in metrics, drift detection, and serving log monitoring.