Ray Data: What's New - GPU Shuffle, Unity Catalog, DataSourceV2
summary
Ray Data 2.53–2.58 adds an actor-only model that replaces mixed tasks to enforce per-actor buffer limits, reducing cross-node spilling and improving pipeline stability; sizing now looks at input and output buffers holistically. Shuffles move to a disk-backed shuffle with per-node files and file-server actors to avoid driver quadratic refs and improve fault tolerance. A GPU path uses UCX/NVLink for an on-device shuffle, showing about 4x faster and 2–3x cheaper runs on large pretraining jobs. Unity Catalog and DataSourceV2 arrive for enterprise sources.
AI-written summary · about
chapters (8)
- 0:00Introduction and agenda
- 4:51Core Ray Data concepts
- 7:17Architectural improvements
- 15:29Per-actor buffer limits
- 22:19Data shuffle enhancements
- 26:31GPU-based shuffles
- 29:22New data sources support
- 30:36Performance results and Q&A
related talks
32mAnyscale's Ray Data: Revolutionizing Batch Inference
22mLarge Scale Data Loading and Data Preprocessing with Ray
22mRay, a Unified Distributed Framework for the Modern AI Stack
28mScaling Multimodal Data and Reinforcement Learning with Ray
26mData Processing on Ray
29mScaling ML Pipelines for Autonomous Driving with Ray
30mHow to Scale Unstructured Data Processing with Ray Data
32mRay @ eBay: Pioneering a Next-Gen AI Platform