KubeRay Operator v1.3RayCluster · RayJob · RayServiceGCS Fault ToleranceAutoscaler v2Zero-Downtime Serve Upgrades

KubeRay Operator Studio

Generate production-ready KubeRay custom resources for Kubernetes: long-lived RayClusters, batch RayJobs with automatic teardown, and RayServices with zero-downtime upgrades. Includes GCS fault tolerance with external Redis, in-tree autoscaler v2, GPU worker groups and Prometheus scraping.

Worker Replicas

1 → 8

autoscaler min → max

Peak GPUs

8 × A100

nvidia.com/gpu at max scale

Head Recovery (RTO)

~30 s

GCS state restored from Redis

Idle GPU Cost Avoided

-87%

vs. static max-size cluster

⚙️ Studio Parameters Live Sync

Custom Resource Kind

Worker Group Profile

Autoscaler Max Replicas

GCS Fault Tolerance

🛡️ Production Guardrails Applied

Head pod runs with num-cpus: "0" so no tasks are scheduled on it
Ray image tag is pinned (no :latest)
Ports, resource requests and limits set on every container
RayJob sets shutdownAfterJobFinishes and a TTL so GPUs are released
RayService uses upgradeStrategy: NewCluster for blue/green rollout

💰

KubeRay FinOps Impact

Scaling GPU workers from 1 to 8 on demand, and tearing RayJob clusters down when the job finishes, avoids paying for idle A100 nodes between training runs.

// Compiling KubeRay manifests...

Interactive SRE Terminal Emulator

Quick Commands: • • •

$ # SRE terminal initialized. Enter command or click above.

📐 Architectural Execution Blueprint

Operator Reconcile Loop

KubeRay Operator Studio Architecture Flow