kuberay

Why we included this project

Teams that have adopted Ray for distributed training and inference often end up writing the same Kubernetes plumbing themselves: controllers to keep clusters alive and job runners that wait for capacity. KubeRay is the operator that packages that work into three custom resource definitions. RayCluster handles cluster creation, deletion, autoscaling, and fault tolerance. RayJob spins up a cluster, submits your workload once it is ready, and can tear the cluster down when the job finishes. RayService pairs a cluster with a Ray Serve deployment graph so upgrades happen without downtime. The job submission flow alone is worth a look if you run shared GPU fleets, since a whole run goes from submission to cleanup without you writing controller logic. A kubectl plugin covers the common workflows for people who would rather not hand-roll the YAML themselves.