How Runhouse Orchestrates Multi-Cluster Ray Workloads

Explore a 15-minute conference talk from Ray Summit 2025 where Donny Greenberg from Runhouse demonstrates how Kubetorch introduces a revolutionary Ray-inspired distributed programming paradigm for Kubernetes. Learn how this next-generation framework enables teams to build complex AI workloads "serverlessly" in Python without managing YAML configurations or manually handling Ray clusters. Discover how modern ML infrastructure trends have made Ray a first-class distributed computing primitive, while surfacing new challenges in rapid debugging, programmatic orchestration, fault-tolerant workflows, and the growing demand for ephemeral, per-task Ray clusters. Understand Kubetorch's innovative programming model that extends Ray's familiar Task and Actor abstractions to Kubernetes-native resources, where Actors represent full Kubernetes resources including KubeRay RayCluster instances. Examine how entire Ray programs, services, and pipelines can be composed as higher-order workflows directly in Python, providing dramatically improved developer experience with fast iteration, fault tolerance, and minimal operational overhead. See how workloads scale elastically and portably across Kubernetes environments while supporting incremental adoption by wrapping existing Ray workloads. Gain insights into how Kubetorch delivers serverless Ray capabilities with instant cluster provisioning, ephemeral execution, and scalable workflow composition, all while maintaining the familiar Ray programming model that ML practitioners already know and use.