Feast is an open-source feature store for supplying structured data to AI and LLM applications during training and inference. It manages machine-learning features for batch and real-time use, and its point-in-time joins help keep future values out of training data. Feature services support discovery, collaboration, and versioning of feature sets. The Python SDK and CLI manage version-controlled definitions, materialize values, build training datasets, and retrieve online features. A Python feature server exposes an HTTP endpoint with JSON input and output, allowing use by clients that can make HTTP requests. Feast integrates with offline and online stores and data sources, including community and custom integrations. It can be deployed on Kubernetes, where feature servers and scheduled or ad-hoc jobs run as workloads. Authorization support includes OIDC and Kubernetes RBAC, but the default configuration is no_auth. Feast does not provide authentication capabilities, so clients are responsible for managing and passing authentication tokens. Batch transformations require a separate transformation engine.
Who it is for
Feast is designed for data scientists, MLOps engineers, data engineers, and AI engineers who need to manage and serve features. Its open-source and self-hosted deployment options suit teams that can manage their own authentication and deployment.
What is good
- Supports batch and real-time feature serving.
- Point-in-time joins guard against future-value leakage.
- Python SDK and CLI manage feature workflows.
- HTTP feature server accepts JSON input and output.
- Open source and free.
What to know first
- Default authorization configuration is no_auth.
- Clients must manage and pass authentication tokens.
- Batch transformations need a separate transformation engine.
- Spark processing is described as experimental.
Verdict
Feast provides a broad feature-management workflow, from versioned definitions and training datasets to online serving. Teams should account for its authentication responsibilities and the separate engine required for batch transformations.
Feast plans and pricing
All plansCompared on feature store software
- Online store
- Yes
- Offline store
- Yes
- Point-in-time joins
- Yes
- Feature monitoring
- Yes
- Deployment model
- both
- Serving modes
- both
