Session

A Simple, Managed Platform - Modern Feature Engineering on Databricks

Overview

ExperienceIn Person
TrackArtificial Intelligence & Agents
IndustryEnergy & Utilities, Enterprise Technology, Financial Services
TechnologiesDatabricks Agents
Skill LevelIntermediate

The hardest part of real-time ML is the feature plumbing. Teams spend time creating and maintaining fragmented pipelines rather than focusing on their ML capabilities. Declarative Features breaks that cycle. Define what a feature is (source, entity, aggregation, time window) and let the platform compute it correctly in every context. This session covers Declarative Features, our managed feature pipeline offering, and explains how they can take your ML platform into the next phase of growth.

Session Speakers

Speaker placeholderIMAGE COMING SOON

Ian Ackerman

/Software Engineer
Databricks

Speaker placeholderIMAGE COMING SOON

Nick Joung

/Sr. Product Manager
Databricks

Full Summary

Why Feature Views Replace the Feature Store Frankenstack

Generative AI gets the spotlight, but high-scale, low-latency decisions in production still run on classic machine learning. Fraud detection, recommendations, and pricing depend on fresh, contextual features. The speakers introduced Feature Views, a unified abstraction that collapses duplicate batch and streaming pipelines into one definition on the governed Databricks Platform, where data science and AI workloads run on a unified lakehouse foundation.

FAQ


No. The speakers stated there is no additional premium for Feature Views. Billing reflects the same compute and storage used by Lakebase, Spark Streaming, and feature-serving endpoints.

With Spark Structured Streaming in real-time mode writing directly into Lakebase, end-to-end freshness is around 200 milliseconds. Teams can choose micro-batch mode for lower cost with higher latency in the tens of seconds, using the same Feature View.

Yes. A single declarative definition drives batch materialization for training, streaming computation for online freshness, and serving through existing endpoints. The platform computes the same logical value correctly in each context.

The platform backfills historical portions from the offline store, then hands off to the streaming job for continuous updates. Long windows, such as 30 days, do not require waiting for the stream to fill naturally.

Feature Views are first-class objects in Unity Catalog, with links to online and offline materializations. Lineage integrates with MLflow from features to models to serving endpoints, so access control, discovery, and traceability follow the same patterns as the Databricks Platform.