FEAST

What it is

Feast (Feature Store) is an open-source system for managing and serving machine learning features consistently across training and inference. It solves the training-serving skew problem: the same feature definitions and transformation logic get used to build training datasets (offline) and to serve low-latency features to models in production (online), instead of two teams reimplementing the same logic twice and getting subtly different numbers.

It's a Python-based control plane that sits on top of storage you already have — offline stores like BigQuery, Snowflake, Redshift, or a data lake, and online stores like Redis, DynamoDB, or Postgres. Feast doesn't store your data; it manages metadata, materialization jobs, and a consistent API for reading/writing features.

Who builds it and why

Feast originated at Gojek, was developed further with Google, and is now hosted under the Linux Foundation AI & Data Foundation (not CNCF, despite that category placement in some directories). Tecton (a commercial feature platform vendor) was a major early contributor and effectively drove the project for a few years before scaling back involvement as their commercial product diverged.

381 contributors and continued commits through mid-2026 indicate it's not a single-vendor side project — it's a genuine community effort, though like most infra projects in this space, a small number of maintainers likely do the bulk of the design work. CNCF maturity level: not publicly available (and not applicable, since it's LF AI & Data, not CNCF).

Production readiness signal

• 7,111 GitHub stars, active commits as recently as July 2026 — actively maintained, not abandoned.
• 381 contributors is a healthy number for this category.
• Latest release v0.64.0 — a high version number suggests frequent iterative releases rather than a slow, heavily-gated release cadence. Check the changelog before upgrading; breaking changes between minor versions have happened historically.
• Apache-2.0 license — no vendor lock-in risk from licensing.
• Founded date: not publicly available, but the project has been around since roughly 2019 based on public history, so it has multiple years of production use behind it at companies like Gojek, Salesforce, Twitter/X, and Robinhood (per public case studies).

Who should use this

• Teams running multiple ML models in production that share features across models and need to avoid recomputing the same aggregations in five different pipelines.
• Organizations with an existing data stack (warehouse + a KV store) who want a thin metadata/serving layer instead of a full platform rewrite.
• ML platform teams building internal MLOps tooling who want an open, extensible base rather than a black box.

Who should NOT use this

• Small teams with one or two models and no real feature-reuse problem. Feast adds operational overhead (registry, materialization jobs, online store to run) that isn't worth it if you're just calling model.predict() on a dataframe.
• Teams wanting a fully managed, zero-ops solution — Feast is self-hosted infrastructure; you own the online store, the materialization scheduler, and the registry. If you don't want that operational burden, a managed platform is the better call.
• Deep learning shops doing mostly unstructured data (image, text, audio) with embeddings generated on the fly — Feast's model is built around tabular, batch/streaming structured features, and doesn't map cleanly onto embedding-heavy pipelines.
• Anyone expecting built-in transformation/feature engineering compute at scale — Feast is a store and serving layer, not a transformation engine. You still need Spark/dbt/Flink upstream.

Alternatives

• Tecton — commercial, managed feature platform built by former Feast maintainers; less ops overhead, but paid and less flexible for custom infra.
• Hopsworks Feature Store — open-core alternative with tighter integration to its own Hopsworks platform and stronger built-in support for streaming features.
• AWS SageMaker Feature Store — fully managed, AWS-native option if you're already committed to that ecosystem and want to avoid running Feast's infrastructure yourself.

Pricing

Fully open source, Apache-2.0. No license fees. Your cost is entirely infrastructure (offline/online store compute and storage) and engineering time to operate it. No official managed/hosted version from the Feast project itself — commercial managed offerings come from third parties (e.g., Tecton, though their product has diverged from upstream Feast).