Feature Views centralize feature definitions so data scientists specify data sources, entities, time-series columns, and computation logic once, enabling consistent training and online inference.

Governance and integration: materialized features are managed as Unity Catalog objects for discoverability, access control, and lineage tracking, with MLflow integration that records feature dependencies when models are logged to aid deployment and inference.

Databricks is launching Feature Views, a managed framework to simplify the creation, serving, and governance of ML features for production and real-time applications.

Positioned as part of an end-to-end ML lifecycle platform, Feature Views cover feature definition, experimentation, production pipelines, and governance in a unified approach.

Genie Code enables rapid iteration by using natural language prompts to generate feature definitions, analyze feature importance, and build training sets directly within notebooks.

The unified feature definition helps reduce training-serving skew, ensures historically accurate data for experiments, and streamlines moving features from notebook experiments into production pipelines.

For real-time use cases like fraud detection and personalization, Feature Views support streaming data sources with aims for end-to-end p99 latency around 200 milliseconds from event ingestion to online availability, including backfilling historical data and updating streaming features.