How Feature Stores Enable Operational ML
summary
Data teams face duplicated pipelines and stalled models; feature stores automate feature transforms, store both online and offline feature values for training and real-time inference, and provide a central catalog and lifecycle (deploy, monitor, govern). Claim: platforms cut feature-to-production time from months to hours and helped Uber scale to thousands of production models; Atlassian drops model deployment from months to days. Implements batch and streaming ETL, integrates with Feast or Tecton, supports training-serving skew detection and plug‑in monitoring like Great Expectations.
AI-written summary · about
chapters (16)
- 0:00Introduction to operational ML
- 3:16Challenges in operational ML
- 5:47What is a feature store
- 7:35How feature stores work
- 9:19Data scientist workflow with feature stores
- 12:09Benefits of feature stores
- 13:31Data silos and model accuracy
- 15:58Deploying features to production
- 17:49Avoiding feature duplication
- 19:56Case study: Atlassian
- 20:59Getting started with feature stores
- 22:01Q&A on feature monitoring
- 25:55Scaling feature stores
- 27:20Centralized vs team feature stores
- 28:42Feature aggregation and context
- 29:24Closing and resources
related talks
35mPanel 03- How to make the Feature Store become a hidden part of the productionalization of ML models
08mFeature Store Usage Patterns - From a Single Data Scientist to an Enterprise
32mA feature Store for Machine Learning
40mFrom Legacy to Building Our Own Feature Store: Active Steps
30mDataEngBytes 2023 - MEL-T2-01 - Mikiko Bazeley
15mFeature Stores: Powering Machine Learning Infrastructure with Apache Flink
40mCreating and operating ML models from event based data using feature stores
25mBuilding A Feature Store For Hyper Growth