Ray: A General Purpose Serverless Substrate?
summary
Ray 1.0 positions Ray as a general-purpose serverless substrate that supports stateful actors, tasks, and a shared-memory object store so ML training and data pipelines run without gluing services. New features include auto scaler node caching (keeps stopped nodes to cut scale‑up time), accelerator type resources with multi-node auto-scaling (request a Tesla V100 and get the right node), and placement groups for gang scheduling. Open-source integrations (Horovod, Hugging Face, Dask) already run on this substrate.
AI-written summary · about
chapters (7)
- 0:00Introduction to Ray
- 2:31What serverless means
- 7:29Limitations of FaaS for stateful apps
- 10:13Requirements for general-purpose serverless
- 16:10Ray vs FaaS APIs
- 21:06Ray 1.0 features for serverless
- 25:15Conclusion and resources
related talks
24mRay: A System for High-performance, Distributed ML Applications
15mRay: A Distributed Compute Engine for AI
29mRay:A Distributed Computing Platform for Reinforcement Learning
11mKeynote Q&A
03mRay: A Distributed Execution Framework for Emerging AI Applications
14mBuilding High Availability and Scalability Online Computing Applications on Ray
26mData Processing on Ray
28mA Deep Dive into Ray’s Scheduling Policy