Panel: ML Platform on Ray
summary
Companies replace Spark with Ray for unified training and serving, running XGBoost and GPU deep learning with measurable wins: a 150M-parameter BERT variant trains on 400M records / 10B images, prototypes move from notebook to distributed runs in 2–3 weeks, and some orgs run hundreds of models retrained daily/weekly. Ray runs on Kubernetes, gives notebook decoupling via Ray Client, autoscaling per-team clusters, and drop-in Python replacements; current gaps include Beam/TF-Transform integration and richer production metrics and history services.
AI-written summary · about
chapters (29)
- 0:00Introduction
- 1:00Welcome
- 2:59Panel Introduction
- 3:30Uber
- 5:34Robinhood
- 6:16ML Landscape
- 6:39ML Problems
- 7:20Shopify ML Platform
- 8:42Journey in Organization
- 9:12History of ML Platform
- 11:30Benefits of Ray
- 12:43Infrastructure
- 13:01Uber ML Platform
- 15:01Summary
- 16:21Collaboration
- 17:35Open Source
- 19:31Positive Impacts
- 23:18Ray vs Spark
- 24:24Spark on Ray
- 26:17Additional features
- 28:49Productivity
- 29:41Use Cases
- 31:08Time Consuming Parts
- 32:45ML Engineers
- 33:54ML Platform
- 39:17Ubers use case
- 40:06Advice for other companies
- 41:07Engage with the team
- 42:45Thank you
related talks
28mScaling AI & Machine Learning Workloads With Ray on AWS, Kubernetes, & BERT
24mRay: A System for High-performance, Distributed ML Applications
32mAn Overview of CloudKitchens's Ray-Powered ML Platform
22mRay, a Unified Distributed Framework for the Modern AI Stack
28mScaling Multimodal Data and Reinforcement Learning with Ray
29mScaling ML Pipelines for Autonomous Driving with Ray
31mNext-Generation Data Science Workflows Using Ray
27mPinterest's ML Evolution: Distributed Training with Ray