Databricks Online Feature Stores
Databricks Online Feature Stores is a Databricks ai / ml capability, introduced September 2025.
Low-latency serving of feature data to online applications and real-time models, powered by Lakebase and kept consistent with the offline feature tables it publishes from.
- The feature store is a Postgres database wearing a machine-learning hat: it is powered by Databricks Lakebase, and every online store you create provisions a Lakebase project behind the scenes.
- New ones quietly changed shape in 2026 - they are now created as Lakebase Autoscaling projects rather than the provisioned instances the original stores used.
- Publish the same feature table twice and endpoints will not let you choose between the copies - they always resolve to the oldest online table by creation timestamp.
Limitations: Requires Databricks Runtime 16.4 LTS ML or above (or serverless) and only Unity Catalog feature tables; you cannot target a specific online table, a store supports at most 3 read replicas, Lakebase scale-to-zero is unsupported, publishing ignores the filter_condition, checkpoint_location, mode, trigger, and features parameters, Autoscaling instances created through the projects API or UI cannot be updated or given read replicas, and new endpoints that look up features from multiple online stores cannot be created on Lakebase Autoscaling instances.
Open in REbricked →- Category
- AI / ML
- Also known as
- Online Feature Store, online store
- Verified
- 2026-08-20