Accelerating ML at Uber with the Palette Feature Store
summary
Uber describes Palette as a feature store inside the Michelangelo platform that automates pipeline generation, point‑in‑time snapshotting and online/offline consistency, cutting feature onboarding and deployment time. It uses AthenaX for streaming SQL, an internal “dental” snapshot store to shrink joins, Michelangelo AutoML for automatic feature search, shadow testing and capacity planning to validate at millions of queries per second with P99 ≈700ms. Search examples show 58 keyword hits; teams report dozens of internal adopters and end‑to‑end monitoring that detects drift within minutes.
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:49Michelangelo Palette
- 3:20Feature Engineering: User journey
- 6:08Feature Discovery
- 10:15Feature Extraction
- 14:42Model Training
- 18:57Online Model Deployment
- 21:18Model Monitoring
- 24:01Accelerating end-to-end ML
related talks
19mMichelangelo ML Platform at Uber: Past, Present and Future
17mKeynote: Machine Learning & A.I. At Uber
33mEvolution and Scaling of Feature Store at Uber- 33mEvolution and Scaling of Feature Store at Uber
31mUse-Case Driven ML Platforms
26mUber’s Intelligent Insights Assistant
38mFantastic ML apps and how to build them
33mScaling PyTorch Training with Ray Data on Uber's Michelangelo AI Platform