Feast

Feast is an open source feature store for machine learning. It holds one definition of a feature and serves it to both training and inference: historical retrieval assembles training sets from an offline store, while online retrieval serves the same features at low latency from an online store. That removes the mismatch that appears when training and serving pipelines compute the same feature separately. It is used by teams running production models and, increasingly, retrieval for language model applications. Cases named by the project include recommendations driven by recent user interactions, fraud detection that combines historical patterns with current behavior, risk scoring in financial services, and customer segmentation on shared definitions. The Python API exposes a feature store object with calls to fetch historical features as a dataframe, fetch online features for a set of entities, and retrieve documents by vector similarity for retrieval augmented generation. Offline and online stores are pluggable, with integrations spanning common warehouses and key value stores. Feast is free and open source, and the project publishes no price list or commercial tier. Teams run it on their own infrastructure and pay only for the stores and compute they choose. Support comes through the documentation and a community Slack rather than a paid plan. Two recent additions are worth noting. A closer integration with ScyllaDB adds vector search and improved performance to the online store, and a Chronon integration lets Feast discover and retrieve features that Chronon computes and serves, so the two systems split ownership rather than duplicate work.