Ray libraries and Kubernetes operations

Ray libraries and Kubernetes operations An architecture diagram generated by Archify. Application code · Train / Tune / Serve / Core · Architecture component Application code Train / Tune / Serve / Core Ray runtime · tasks / actors / ObjectRefs · Architecture component Ray runtime tasks / actors / ObjectRefs Ray Pods on Kubernetes · head / workers · Architecture component Ray Pods on Kubernetes head / workers KubeRay operator · RayCluster / RayJob / RayService · Architecture component KubeRay operator RayCluster / RayJob / RayService library / Core APIs runs inside reconcile resources Legend Backend Cloud External

Application

  • • Application selects libraries, work, and checkpoint policies
  • • Framework-specific communication/controllers still exist above Core

Infrastructure

  • • KubeRay does not automatically dispatch between ML libraries
  • • Ray scheduling, Pod placement, and EC2 provisioning are separate