- Home
- Feature Store Software
- Databricks Feature Store
Databricks Feature Store
San Francisco, California, United States
What it is
Databricks Feature Store is a registry for organizing features used by AI and machine-learning models. With a Unity Catalog-enabled workspace, teams can register features and models for governance, lineage, point-in-time joins, and sharing and discovery across workspaces. Features may be defined as declarative Feature Views managed through Databricks pipelines or stored in feature tables populated with values in Delta tables. Feature Views and their materialization capability are in Public Preview. Models trained using Feature Store features track feature lineage and retrieve the latest feature values during inference. Databricks says using the same feature computations for training and inference prevents training/serving skew. Feature serving endpoints and online stores support real-time applications; the main overview describes endpoint latency in milliseconds. Supported third-party online stores include Amazon DynamoDB, Amazon Aurora (MySQL-compatible), and Amazon RDS MySQL, with availability depending on Feature Store mode. The paid cloud service has a free trial. Materialization uses serverless compute, endpoints use the Model Serving SKU, and online stores use Lakebase compute priced by capacity units and replicas.
For and against
For
- Tracks feature lineage from model training.
- Supports point-in-time joins and cross-workspace discovery.
- Online stores and endpoints support real-time serving.
- Free trial is available.
Against
- Requires a Unity Catalog-enabled workspace.
- Feature Store APIs cannot run outside Databricks environments.
- Feature Views and materialization are in Public Preview.
- Paid pricing varies by product, cloud provider, and region.
The verdict
Databricks Feature Store brings feature registration, governance, and serving into a Unity Catalog-based workflow. Check the workspace requirement and API limitation against your development setup before choosing it.
Who it suits
It suits teams building AI or ML models that need managed feature governance, lineage, and real-time serving. It requires a workspace enabled for Unity Catalog; its APIs cannot be called from local or other non-Databricks environments, though local IDE development and unit testing are supported.
Plans and prices
Pay as you go
No up-front costs · Pay for products used · Rates vary by product, cloud provider, and region
Read 30 Sep 2026 · databricks.com
On the record
- Client limitation
- Feature Engineering and Feature Store APIs cannot be called from local or other non-Databricks environments, although local IDE development and unit testing are supported.docs.databricks.com · read 30 Sep 2026
- Feature authoring
- Features can be authored as declarative Feature Views managed through Databricks pipelines or as feature tables populated by writing values to Delta tables.docs.databricks.com · read 30 Sep 2026
- Feature Views status
- Feature Views and their feature materialization capability are in Public Preview.docs.databricks.com · read 30 Sep 2026
- Governance and discovery
- Registering features and models in Unity Catalog provides governance, lineage, point-in-time joins, and cross-workspace feature sharing and discovery.docs.databricks.com · read 30 Sep 2026
- Integrations
- Supported third-party online stores include Amazon DynamoDB, Amazon Aurora (MySQL-compatible), and Amazon RDS MySQL, with availability differing by Feature Store mode.docs.databricks.com · read 30 Sep 2026
- Model training and inference
- Models trained with Feature Store features automatically track feature lineage and look up the latest feature values at inference time.docs.databricks.com · read 30 Sep 2026
- Pricing components
- Materialization uses serverless compute, feature serving endpoints use the Model Serving SKU, and online stores use Lakebase compute priced by capacity units and replicas.docs.databricks.com · read 30 Sep 2026
- Pricing model
- Feature Store is billed at cost for underlying serverless compute, online store, and serving infrastructure, with no premium on top.docs.databricks.com · read 30 Sep 2026
- Purpose
- Databricks Feature Store is a central registry for the features used in AI and ML models.docs.databricks.com · read 30 Sep 2026
- Python client
- The Feature Engineering Python client is distributed as databricks-feature-engineering on PyPI and is pre-installed in Databricks Runtime 13.3 LTS ML and above.docs.databricks.com · read 30 Sep 2026
- Real-time serving
- Feature serving endpoints and online stores support serving features for real-time applications, with the main overview describing endpoints as providing millisecond latency.docs.databricks.com · read 30 Sep 2026
- Security and access
- Access to Unity Catalog feature tables is managed through Unity Catalog access controls.docs.databricks.com · read 30 Sep 2026
- Training and serving consistency
- Databricks says using the same feature computations at inference as during training eliminates training/serving skew.docs.databricks.com · read 30 Sep 2026
- Workspace requirement
- The current Databricks Feature Store requires a workspace enabled for Unity Catalog.docs.databricks.com · read 30 Sep 2026
