State Management in Distributed Online Learning with Ray
summary
Experimental framework uses Ray actors and Ray Serve with gRPC to manage continuously updating distributed models for suicide prevention, overdose and infectious disease surveillance. It demonstrates a 75-hour training pipeline reduced to 15 minutes, explores hierarchical local→regional→global parameter propagation, and tests conflict resolution strategies including peer-to-peer vs coordinator tradeoffs. Key work targets are scalable state propagation, transactional consistency of parameter commits, and performance scaling on large clusters; unresolved items include collision resolution policies and full distributed scaling experiments.
AI-written summary · about
chapters (6)
- 0:00Introduction and motivation
- 3:51Motivation for real-time analytics
- 13:04Realm architecture overview
- 20:23Ray integration and usage
- 25:20Observations and future work
- 27:46Summary and acknowledgments
related talks
14mBuilding High Availability and Scalability Online Computing Applications on Ray
01h 31mIntroduction to Ray for distributed and machine learning applications in Python
15mRay: A System for Distributed Applications
42mRay: A distributed system for emerging AI applications
39mScaling Python for Machine Learning: Beyond Data Parallelism
24mRay: A System for High-performance, Distributed ML Applications
33mOvercoming Distributed ML Challenges with Ray Train
28mA Deep Dive into Ray’s Scheduling Policy