Source: Databricks Blog — 2026-07-10
Summary
Databricks announced public preview of Feature Views, a managed framework that lets teams define an ML feature once and reuse it for both historical training data and production batch or real-time inference, with features governed as Unity Catalog objects.
Key Takeaways
- Eliminates training/serving skew and duplicated feature logic by unifying the definition across experimentation and production.
- Streaming features are served at 200ms end-to-end p99 latency.
- Features become governed Unity Catalog objects with managed materialization pipelines, removing the need for self-managed streaming or online-store infrastructure.
- Aimed at teams currently maintaining separate feature-engineering code paths for offline training and online inference.