/
New feature Public Preview

Declarative Feature Engineering

Declarative Feature Engineering is a Databricks ai / ml capability, introduced March 2026.

APIs to define features declaratively from Delta tables, Kafka streams, and request-time data, with time-windowed aggregations that Databricks computes, materializes, and serves for training and online inference.

  • You declare a time-windowed aggregation once and it computes the feature for you - materializing to Delta or Lakebase and wiring into online model serving - the feature-engineering echo of the declarative-pipelines idea.
  • It launched in Beta in March 2026, initially only in the AWS us-east-1 and us-west-2 regions.
  • Graduating to Public Preview left the Beta-era definitions behind, so the client ships an is_beta_feature_view() check and a Feature.clone() to drag them across.

Limitations: Needs serverless or Databricks Runtime 17.0 ML and above plus databricks-feature-engineering 0.16.0 or later; entity columns cannot be DATE or TIMESTAMP, only a limited set of aggregation functions is supported, RequestSource takes scalar types only and no aggregations or time windows, and entity, timeseries, and request feature column names must be globally unique across all sources in a training set or serving endpoint.

Open in REbricked →
Category
AI / ML
Introduced
March 2026
Also known as
Declarative Feature Engineering APIs, Feature Views
Verified
2026-08-20

Sources

Related in AI / ML