Feast is the leading open-source feature store for AI and machine learning. It solves one of ML’s most persistent production problems — keeping the features used in training identical to the ones used at serving time — by giving teams one system to define, store and serve features for both. The project has earned about 7,300 stars on GitHub and is actively maintained in 2026.
Feast sits between your data sources and your models. You define features once in Python, from batch sources like data warehouses and from streaming sources, and Feast materializes them into an online store (such as Redis or DynamoDB) for low-latency serving while also generating point-in-time-correct training datasets from the offline store — the historical joins that prevent label leakage. A central feature registry gives teams a shared catalog of vetted features, so a fraud signal built by one team can be discovered and reused by another instead of being re-engineered. It integrates with the infrastructure teams already run and stays framework-agnostic: Feast serves features to any model, in any serving system. As ML systems have grown from single models to fleets of models and LLM applications, Feast has become standard plumbing for organizations that need feature consistency they can audit.
Feast is free and open source under the Apache 2.0 license. You run it on your own infrastructure and pay only for that.
Feast is for ML platform teams and data scientists who need consistent, reusable features across training and production — and an end to training-serving skew bugs.