Feast

✓ Editorially verified

Open-source feature store that serves consistent features to ML training and online inference, with RAG vector search built in.

Free· Free, open source (Apache 2.0); self-hostedRAG

Visit website →

Best for

Pick Feast if you're running production ML or RAG at scale and need one consistent feature definition across offline training and online serving.

Skip if

Skip it if you just want a hosted vector database for a small RAG prototype and don't care about offline/online consistency.

Feast is an open-source feature store (Apache 2.0) that sits between your data warehouse and your ML/LLM serving layer. It defines features once, then materializes them into a low-latency online store for inference and pulls point-in-time-correct historical features for training, so the two never drift. The newer releases also expose vector similarity search, turning Feast into a RAG-friendly feature platform rather than a pure tabular tool.

It's aimed at ML and platform engineers running production systems who are tired of bespoke pipelines duct-taping Snowflake/BigQuery to Redis/DynamoDB. Feast doesn't store data itself; it orchestrates the stores you already have, with adapters for Snowflake, BigQuery, Redshift, Postgres, DuckDB, and Spark on the offline side, and Redis, DynamoDB, Cassandra, MySQL, Milvus, and Qdrant on the online side. The project is free to run yourself, with 290+ contributors and adoption at Robinhood, NVIDIA, Discord, Walmart, Shopify, and Salesforce.

There's a Python SDK, REST APIs, data-quality monitoring, RED metrics, and SOX-style audit logging. Commercial managed Feast offerings exist via third parties (e.g. Tecton, Expedia's contributions), but the upstream project itself is self-hosted only.

Editor's take

Feast is the closest thing the OSS world has to a standard feature store, and the recent pivot to embrace vector search keeps it relevant in the LLM era. It's plumbing, not magic, but if your team is already wiring Snowflake to Redis by hand, adopting Feast is almost always cheaper than building it again.

— The AI Tool Bible editorial team

Pros

✅ Solves train/serve skew with point-in-time-correct historical retrieval
✅ Broad adapter ecosystem across warehouses, KV stores, and vector DBs
✅ Production-proven at Robinhood, NVIDIA, Shopify, Walmart
✅ Vector similarity search makes it usable as a RAG feature layer
✅ Permissive Apache 2.0 license with active community

Cons

⚠️ You operate the underlying stores yourself; Feast is orchestration, not storage
⚠️ Steeper learning curve than a hosted vector DB for simple RAG demos
⚠️ No first-party managed cloud; SaaS is via third parties like Tecton