Feature Store
lastmod 2026-04-24created 2025-12-304 edits
A feature store is a central repository for features. A feature store transforms raw data into features, stores them, and serves them for model training and inference.
About
Typical machine learning code for structured data includes a part that loads data and extracts and processes the needed features.
The data usually comes from a data warehouse or the original data source, but instead of connecting directly to the data source, you can build a separate data store for what machine learning needs.
There you can store the needed features preprocessed in advance, and you can also introduce data versioning.
The Feature Store emerged from this concept. Think of a Feature Store as a place that gathers Feature data for machine learning.
Overview
A Feature Store is a centralized repository for efficiently managing, storing, and serving the data (features) used in training and inference of machine learning (ML) models. It bridges data engineering and modeling, and is considered one of the core pieces of infrastructure in MLOps.
Background
It emerged to solve chronic problems in existing ML workflows.
- Duplicate work: several data scientists each create the same features, wasting resources.
- Training-Serving Skew: the data processing logic used in model training differs from the logic in production (real-time), degrading performance.
- Data fragmentation: it is hard to track which model used features from which point in time.
Key components
A feature store generally includes two types of storage plus management functions.
1. Offline Store
- Purpose: dedicated to model training and batch scoring.
- Characteristics: stores large volumes of historical data, usually on a data lake or warehouse such as S3, BigQuery, or HDFS.
2. Online Store
- Purpose: dedicated to serving for real-time inference.
- Characteristics: must guarantee ultra-low-latency data lookups, and usually uses a key-value store such as Redis, DynamoDB, or Cassandra.
3. Entity & Feature View
- Entity: the primary key of the data (e.g. user_id, product_id).
- Feature view: a definition of a set of specific features associated with an entity, plus their transformation logic.
Key features and benefits
- Reusability: once registered, a feature can be shared with other teams or projects in the organization.
- Consistency: the same feature pipeline provides the same data to training and serving, resolving the mismatch.
- Time Travel: lets you reproduce the exact state of features at a specific past point in time for model training.
- Governance: centrally manages each feature's creator, change history, usage, and so on.
Categories
- Feast - Feast is an open-source feature store. It is the fastest way to operationalize analytic data for model training and online inference.
- Hopsworks - Hopsworks is a data-intensive open-source AI platform for developing and operating machine learning models at scale.
- Tecton - Tecton is a fully managed feature platform built to orchestrate the complete lifecycle of features, from transformation to online serving, with enterprise-grade SLAs.
- Rasgo - Rasgo is a web app for exploring, transforming, and visualizing data within a cloud data warehouse.
| Category | Solution |
| Open source | |
| Cloud managed | AWS SageMaker Feature Store, Vertex AI Feature Store (Google Cloud) |
| Enterprise/commercial | Tecton, Databricks Feature Store |
See also
- MLOps
- Korean: 데이터 레이크
- Model serving
Favorite site
- How Toss builds diverse ML models: Feature Store & Trainkit
- What is a Feature Store in Machine Learning? · Eojjeonji Oneureun
- Personalized recommendation system #4. Feature Store - Ohouse blog
- What is a feature store? | Databricks
- The heart of the recommendation system, the Feature Store story (1)—Finding order in chaos: building a Feature Store | by mingrammer | Daangn Tech Blog | Medium