Feast
The open-source feature store for consistent ML training and serving.
About
Feast is the most widely adopted open-source feature store for machine learning. Originally developed at Gojek and Tecton, it is now maintained under the Linux Foundation's LF AI & Data umbrella. Its Python SDK lets teams define features once and serve them consistently across training and inference, eliminating training-serving skew via offline stores (BigQuery, Snowflake, DuckDB) and low-latency online stores (Redis, DynamoDB).
The project is actively maintained with roughly 7,000 GitHub stars and commits through April 2026, and it integrates with the Kubernetes MLOps stack (Kubeflow, KServe, Argo).
As a self-hosted project there is no managed cloud offering — you operate the registry, offline, and online stores yourself. The decoupled architecture is powerful but adds real operational surface compared to managed feature platforms.
What it can do
- Consistent feature definitions across training and serving
- Point-in-time correct historical feature retrieval
- Low-latency online feature serving
- Pluggable offline and online store backends
Setup
pip install feast, then feast init my_feature_repo to scaffold a feature repository.
Prerequisites
Python, plus offline storage (e.g. BigQuery, Snowflake, DuckDB) and an online store (e.g. Redis).
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 72; Trending score 84/100 from researcher assessment.
★ 6,974 GitHub stars
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.