// the find
feast-dev/feast
The Open Source Feature Store for AI/ML
Feast is a battle-tested open source feature store that solves the training-serving skew problem by providing a unified layer for both offline (historical, point-in-time correct) and online (low-latency) feature retrieval. It's aimed at ML platform teams who are tired of data scientists writing ad-hoc join logic that doesn't match what runs in production. At 7k stars with active commits through yesterday, it's not vaporware.
Point-in-time correct historical feature retrieval is the core value proposition and it's genuinely well-implemented — the entity_df join logic prevents future data leakage without you having to think about it. The plugin architecture is real: 15+ offline stores, 20+ online stores, all swappable via config without touching application code. Materialization pipeline (offline → online) is straightforward to operate and the incremental mode means you're not re-processing your entire dataset on every push. The Kubernetes operator and Helm charts show this has been run in actual production environments, not just demoed at conferences.
The Web UI has been 'Beta' since at least 2022 and the feature lineage explorer is still a checklist item with no checkbox — governance tooling is the weakest part of the system. Many of the most useful integrations (Postgres, Spark, Clickhouse, MongoDB) are 'contrib plugins' maintained by the community, which means varying quality and no guarantee they track the core API changes. Streaming transformations are Alpha and batch transformations are still in progress, so if you need real-time feature computation rather than just ingestion you're mostly on your own. The registry is a local file by default, which works fine for one developer and breaks immediately when two people run `feast apply` in the same environment.