Ray Internals: Object Management with the Ownership Model
summary
Ray restructures object metadata by shards the master across workers so the creator of an object owns its metadata, keeping location and lineage local to reduce latency for many short tasks (millisecond-scale) and avoid a central scheduler bottleneck. The runtime uses immutable object refs, lineage-based reexecution for fault recovery, and a low-overhead lease/owner-detection protocol; Ray v1.0 implements this and shows substantially better tail latency. Ongoing work adds ownership transfer and automatic transfer detection to handle upward and lateral sharing by sharing ownership among workers.
AI-written summary · about
chapters (6)
- 0:00Introduction to Ray ownership
- 2:52Understanding object refs
- 6:06Challenges of distributed systems
- 9:28Implementing the ownership model
- 13:24Ownership transfer patterns
- 18:15Design for ownership transfer
related talks
22mRay Internals: A Peek at `ray.get`
12mOwnership: A Distributed Futures System for Fine-Grained Tasks
26mRay: A General Purpose Serverless Substrate?
14mKeynote: The Future of Ray
11mKeynote Q&A
22mDebugging and Observability for Distributed Ray Applications
37mEvolving Ray Core: Scalability, Reliability, and Compiled Graphs
29mAdopting Ray @ LinkedIn