特徴量ストア（Feature Store）

Feature Store is built for data science and ML engineering teams that need to share and reuse features across multiple models instead of recomputing logic in individual notebooks. It treats features as governed, reusable assets tracked with lineage from source data through the features and models that depend on them, so teams can find and reuse trusted features and reproduce how a production model was built. Skyscanner's Senior Director of Product Management, Data, Jules Marshall, credits Feature Views with cutting feature code so its data scientists can focus on what drives traveler value rather than on how to compute features.

Feature Store replaces feature logic buried in individual notebooks and ad hoc pipelines with a single, governed layer for building, sharing, and serving features. Feature Views ensure the same point-in-time accurate values used for training are also served online, which minimizes training-serving skew, a common source of errors when features are computed separately for each stage. Once a Feature View is ready, one API call materializes it into managed pipelines that keep both online and offline stores in sync, removing the need to hand-build and maintain separate batch and serving code paths.

Online Feature Store serves features to production applications and model endpoints with low latency at high scale, keeping data current by syncing from Materialized Feature Views or Feature Tables. It is backed by Lakebase Autoscaling, Databricks' managed Postgres, so it is serverless and scales with traffic without requiring a third-party key-value store. Model Serving endpoints integrate directly with Lakebase to fetch features or serve predictions without a separate feature-fetch step. For streaming use cases, Feature Views also orchestrate Spark Real-Time Mode (RTM) pipelines that write model-ready aggregates to Lakebase with under 200ms freshness.

The Feature Store product page does not publish specific pricing, tiers, or billing details for Feature Store or the Online Feature Store. To get pricing or sizing information for your workload, you can try Databricks for free or schedule a demo to talk with the Databricks team directly.

You can get started by using the Feature Engineering SDK for notebook experimentation and point-in-time accurate training data computation, then defining Feature Views for the features your models need. From there, a single API call materializes managed, production-ready pipelines that write to both online and offline stores, and Unity Catalog handles governance and discovery as you scale usage across teams. To explore hands-on, you can try Databricks for free or review the Feature Store documentation.