If you've ever built a machine learning model that worked perfectly in a notebook but behaved differently once it went live, you've run into the problem that feature stores exist to solve. A model is only as good as the data (called "features" in ML jargon — things like "average order size in the last 30 days") it's fed. If the code that computes that number during training is even slightly different from the code that computes it live in production, your model's predictions quietly drift from what it was trained to do. A feature store is the piece of infrastructure that fixes this: one place to define, store and serve those numbers, so training and live predictions always see the same data computed the same way.
Two of the most established open-source options for this are Feast and Hopsworks. They solve the same core problem but take almost opposite approaches to how much you build yourself versus how much comes bundled.
The short version
Feast is a free, unopinionated library: you plug it into whatever databases you already run (Redis, BigQuery, Snowflake, DynamoDB and a dozen others), and it handles the plumbing — but you're on your own for hosting, scaling and support. Hopsworks is a managed platform: it bundles the feature store together with a full data lakehouse and MLOps tooling, with sub-millisecond feature serving, but it's a heavier, more opinionated system and its open-source core is AGPL-3.0 licensed (a stricter copyleft license than Feast's Apache-2.0). Pick Feast if you already have a data platform and just need point-in-time-correct joins on top of it. Pick Hopsworks if you want the feature store, the lakehouse and the serving layer to come as one product you don't have to stitch together yourself.
Feast: bring your own infrastructure
Feast is a fully open-source (Apache-2.0) feature store with no paid tier at all — there's nothing to buy, because there's no product being sold, just a library and a spec. It's designed to sit on top of infrastructure you likely already have: 10+ supported offline stores (for training data) and 15+ online stores (for real-time serving), from Redis and DynamoDB to Snowflake and BigQuery. Its main technical contribution is point-in-time-correct joins — a way of pulling historical feature values as they looked at the exact moment in the past you're training on, which is exactly the mechanism that prevents training/serving skew.
- Best for: ML and data engineering teams who already run a Spark/warehouse/Redis-type stack and want a thin, vendor-neutral layer on top of it.
- Watch out for: there's no managed hosting or commercial support bundled in — running it at scale is your data engineering team's job, not a vendor's.
Hopsworks: the managed all-in-one
Hopsworks is a managed feature store bundled with what it calls an "AI lakehouse" — meaning the feature store, the data lakehouse (with support for Delta, Iceberg and Hudi table formats) and MLOps tooling (with Ray, Spark and Flink compute) all ship as one platform instead of separate pieces you integrate yourself. It's freemium: a genuinely free tier (1 project, no credit card) to start, a pay-as-you-go SaaS tier for unlimited projects, and custom enterprise pricing that includes on-premise deployment for regulated industries. Its headline technical claim is sub-millisecond online feature serving, aimed at latency-sensitive use cases like fraud scoring or real-time recommendations.
- Best for: teams (often in finance, retail or government, per its own customer base) that want one platform covering the feature store, the lakehouse and serving, without stitching together separate open-source projects.
- Watch out for: the open-source core is AGPL-3.0, a stricter copyleft license than Feast's Apache-2.0 — worth a legal read if you plan to build commercial software around it. It's also a heavier, more opinionated system than a single-purpose library.
| Criteria | Feast | Hopsworks |
|---|---|---|
| Pricing model | Free (open source, no paid tier) | Freemium: free tier, pay-as-you-go SaaS, custom enterprise |
| License | Apache-2.0 | AGPL-3.0 (core), commercial tiers on top |
| Scope | Feature store only — bring your own stores/compute | Feature store + lakehouse + MLOps compute, bundled |
| Serving latency | Depends on the online store you plug in | Sub-millisecond, purpose-built serving layer |
| Ops burden | Self-managed — your team runs it | Managed SaaS option available |
| Best for | Teams with an existing data stack, wanting a thin layer | Teams wanting one bundled platform, less integration work |
Both are legitimate, widely-used choices, and notably each project's own comparison page lists the other as a direct alternative — this isn't a case of one being a knockoff of the other. If your team already has strong data engineering and just needs the point-in-time-join logic, Feast costs nothing and won't lock you in. If you'd rather pay for a platform that removes the integration work — and can live with AGPL-3.0 — Hopsworks gets you further faster, especially if low-latency serving matters to your use case.