Feast
Open-source feature store that serves the same features for model training and online inference, with vector retrieval for RAG and pluggable stores
Overview
Feast is an open-source feature store. You define features once, and Feast serves them consistently to model training (point-in-time historical retrieval from an offline store) and to production inference (low-latency lookups from an online store). It does not lock you into one backend: offline stores include Snowflake, BigQuery and Spark, and online stores include Redis, DynamoDB, Cassandra, ScyllaDB and vector databases such as Milvus, Qdrant and Pinecone. Recent releases lean into AI applications, with document retrieval by vector similarity for RAG; v0.68.0 (November 10, 2026) added a Pinecone online store and deeper Milvus support. Red Hat ships Feast in OpenShift AI. Feast is the main open-source choice left in this space now that Tecton, once a core backer, has been absorbed into Databricks.
The Verdict
Who Should Use Feast?
Best For
- Platform teams that want an open, self-managed feature store
- Teams that need training and serving features to match
- Multi-cloud setups that already run Redis, DynamoDB or a warehouse
- RAG and recommendation apps that combine features with vector search
Not Ideal For
- Teams that want a managed service with support contracts
- Streaming feature computation without Spark or Flink
- Small projects with a single model and no serving path
What's Great
- Free and Apache-2.0, with no vendor lock-in
- Pluggable offline and online stores across clouds
- Vector retrieval for RAG built in
- Steady release cadence (three releases from August to November 2026)
- Backed in production by Red Hat OpenShift AI
Watch Out For
- You run and scale the infrastructure yourself
- Feast does not compute features: pipelines and backfills are yours
- Real-time freshness depends on your own streaming stack
- Its own terminology takes time to learn
Pricing
View all features & details
Key Features
- Point-in-time correct historical features for training
- Low-latency online feature retrieval
- Vector document retrieval for RAG
- Feature services and a web UI
- Pluggable credential resolution and connection pre-warming
- Python SDK and feature server
Use Cases
- Fraud detection and risk scoring
- Recommendations and personalization
- Grounding LLM apps with user and document features
How It Compares
| Feature | Feast | Databricks Feature Store | Hopsworks |
|---|---|---|---|
| License | Open source (Apache 2.0) | Proprietary | See vendor |
| Managed service | No | Yes | See vendor |
| Feature computation | No (bring your own) | Yes | Yes |
| Backend choice | Many pluggable stores | Databricks only | See vendor |
Sources: Feast · Kanerika comparison · Databricks on Tecton (Aug 2025)
Is this your tool? Add a badge