Feature Stores (2026): 10 Options for Training and Live Serving

Features are the inputs a machine learning model uses, first during training and then again at inference. Amazon's own page gives a simple example: in an app that recommends a music playlist, features could include song ratings, listening duration and listener demographics. Trouble starts when the same feature gets built twice, once for training and once for the live app, and the two quietly drift apart.

This list covers only that serving layer. The ten options below are listed alphabetically and in no particular order, and every description comes from each vendor's own website or documentation, read as of Oct 2026. If you need the whole model lifecycle, from experiments to deployment, my guide to MLOps tools covers the wider stack.

Feature stores side by side: platform and stand-alone

I checked the free plan column on each vendor's own website, as of Oct 2026. Yes means the page names a free plan, a free version or an open source edition. Trial only means the page names a time-limited free trial. Not stated means the page I read named neither, so check the vendor's own site before you sign up.

Tool	Type	Good for	Free plan
Amazon SageMaker Feature Store	Managed AWS feature repository	Storing, sharing and managing ML features on AWS	Not stated
Chalk	Real-time feature platform	Features and live context computed at inference time	Not stated
Databricks Feature Store	Platform feature store	Governed features with online serving on Databricks	Not stated
Feast	Open source feature store	Serving structured data to AI in training and inference	Yes
Featureform	Feature store (now Redis Feature Store)	Feature serving inside the Redis platform	Not stated
Hopsworks	Feature store and AI lakehouse	Sub-millisecond feature retrieval for real-time ML	Not stated
Iguazio	MLOps platform with a feature store	AI pipelines, model monitoring and GPU management	Not stated
Snowflake Feature Store	Platform feature store	Deriving and managing features inside Snowflake	Not stated
Tecton	Feature platform (now on Databricks)	Buyers heading towards the Databricks platform	Not stated
Vertex AI Feature Store	Cloud feature store	Feature management for Google Cloud ML	Not stated

Ten feature stores for training and inference

Amazon SageMaker Feature Store

Best for: Teams on AWS who want a fully managed, purpose-built repository to store, share and manage features across their machine learning models.

Chalk

Best for: Real-time AI apps that assemble live context at inference time, with features defined once and versioned over windows such as 30 and 90 days.

Databricks Feature Store

Best for: Databricks users who want features registered in Unity Catalog, with lineage, point-in-time joins and cross-workspace sharing built in.

Feast

Best for: Teams that want an open source feature store for real-time recommendations, fraud detection and other AI apps running at high scale.

Featureform

Best for: Redis users, since this name's old address now opens the Redis Feature Store page inside the wider Redis data platform.

Hopsworks

Best for: Real-time machine learning that needs features reused across many models and retrieved in under a millisecond, plus training and serving.

Iguazio

Best for: Regulated sectors such as financial services, telecoms and healthcare that want AI pipeline orchestration, model monitoring and a data mesh for MLOps.

Snowflake Feature Store

Best for: Snowflake ML users deriving features from existing columns, such as a day-of-week feature from a timestamp, inside the same data cloud.

Tecton

Best for: Buyers comfortable with Databricks, since this name's old address now lands on the Databricks data and AI platform home page.

Vertex AI Feature Store

Best for: Google Cloud teams who want a feature store documented as part of the Gemini Enterprise Agent Platform rather than a separate product.

Two stand-alone names now live inside bigger platforms

Here's what I'd want any buyer to know before a demo. When I opened the vendor pages, the Tecton address landed on a Databricks home page, and the Featureform address on a page titled Redis Feature Store. Two of the best-known independent feature stores now sit inside larger platforms.

Meanwhile, the big cloud options are bundled with their parent: Databricks with Unity Catalog, Amazon with SageMaker, Google with its agent platform and Snowflake with Snowflake ML. So the real question isn't only features, it's who owns the roadmap. Ask every vendor what happens to your feature definitions if the product gets folded into something else.

Platform store or stand-alone: how to decide

• Platform store. If your data and models already live on one data platform, its built-in store means one less contract and governance you already run. Databricks, for example, gives registered features lineage and point-in-time joins through Unity Catalog.
• Stand-alone or open source. If you run across clouds or want to avoid lock-in, an open source store you host yourself keeps your feature definitions in your own hands.
• Real-time first. If predictions happen in milliseconds, such as fraud checks or recommendations, look hard at online retrieval speed. One vendor here claims sub-millisecond retrieval.

Whichever route you take, reuse is the big win. Hopsworks quotes Meta reporting that its top 100 features are used in over 100 models.

What to check in a feature store demo

1. Point-in-time correctness. Can it rebuild training data exactly as each feature looked at every past moment, so the model never sees the future?
2. Online and offline parity. Ask to see one feature definition feeding both the training set and the live service.
3. Reuse. How do teams find and share existing features across models?
4. Freshness. How quickly does a new event show up in the online store?
5. Ownership. Who'll maintain the product in two years, and how hard is it to export your definitions if you move?

If your features also feed AI apps that search by meaning, my roundup of vector databases covers the storage side of that.

Do you make a feature store? Tool makers can ask to be considered for a featured place on this list.

My plain advice on choosing a feature store

Start where your data already lives. If you're all-in on one data platform, try its built-in feature store first, because governance and lineage come with it. If you run across clouds or want your definitions to stay portable, test an open source store on one real model. Whichever you pick, ask the ownership question before you sign anything long.

And before you build features at all, make sure people can find the data behind them. My guide to data catalog tools covers that layer.

Want your tool on this list? Featured spots and reviews here are paid partnerships. Here's how to partner with me.

Frequently asked questions

What is a feature store?

It's a central place to define, store and serve the features a machine learning model uses, so the same values reach model training and live predictions.

Do I need a feature store?

Not always. It earns its place when several models reuse the same features, or when live predictions need fresh features fast. A single model scored in batches often manages without one.

Is there an open source feature store?

Yes. One option on this list is an open source feature store, shown in the free plan column. Several cloud and data platforms also include a feature store within the wider product.

What's the difference between a feature store and a vector database?

A feature store serves structured model inputs for training and inference. A vector database stores embeddings for similarity search. AI apps often use both.

Who owns the stand-alone feature stores now?

Two well-known names now open pages from bigger platforms, one from Databricks and one from Redis, so check each vendor's own site for current ownership before you commit.