Enabling Multi-User Clusters with
summary
Runhouse adds a Python-first layer that orchestrates work across remote Ray clusters (multi-region and multi-cloud) via HTTP so a single script can call multiple clusters and reuse services; it supports Ray and PyTorch distributed jobs, exposes a lightweight cluster/service registry plus UI, ships OpenTelemetry metrics, and keeps no local runtime so failures don’t cascade. A customer cut compute costs ~50% compute savings and dropped job failures from ~40% to failure rate <1%, with ~1s developer iteration.
AI-written summary · about
chapters (10)
- 0:00Introduction and motivation
- 1:50Ray's strengths and limitations
- 6:00Solving extra-cluster problems
- 12:44Dispatching Ray with Runhouse
- 18:24Modularizing workflows
- 20:51Reusing clusters and services
- 21:51Case study and impact
- 22:12Runhouse UI and observability
- 25:43Built-in collaboration features
- 26:48Closing and next steps
related talks
14mBuilding High Availability and Scalability Online Computing Applications on Ray
24mRay: A System for High-performance, Distributed ML Applications
21mPowering Open Data Hub with Ray
29mAdopting Ray @ LinkedIn
26mData Processing on Ray
26mRay: A General Purpose Serverless Substrate?
28mScaling Multimodal Data and Reinforcement Learning with Ray
15mRay: A Distributed Compute Engine for AI