Ray + KubeRay on GKE

A persistent Ray cluster on GKE with CPU and GPU worker groups, shared Filestore storage, and GCS object-store spillover. Config Connector manages the GCP infrastructure declaratively; KubeRay manages the Ray cluster lifecycle.

What you’ll build

GCP infra layer (Config Connector → config.yaml)
  GKE cluster (regional, 3-zone, private nodes)
  ├── CPU node pool (n2-standard-4, autoscaling 1–5)
  └── GPU node pool (n1-standard-8 + T4, scales to zero)
  Filestore ENTERPRISE (1Ti, ReadWriteMany)
  GCS spillover bucket (7-day lifecycle)
  Artifact Registry (pre-built Ray images, ray:2.54.0)
  IAM: GSA ray-workload → Workload Identity → K8s SA ray-head
       ray-workload → roles/storage.objectAdmin → spillover bucket

K8s layer (KubeRay → k8s.yaml)
  Namespace ray-system + ResourceQuota
  FilestoreStorageClass → PVC ray-shared (/mnt/ray-data)
  RayCluster composite:
    head pod (2 CPU, 8Gi, shmSize 4Gi)
    cpu workers (2 replicas → max 8, 2 CPU, 4Gi)
    gpu workers (0 → max 4, 4 CPU, 16Gi, T4)
    NetworkPolicy (podSelector + RFC1918 GCS egress)
    PodDisruptionBudget (head minAvailable: 1)
    ClusterRole autoscaler (pod get/list/watch/create/delete/patch)

Source layout:

Layer	Source	Output
GCP infra	src/infra/	config.yaml
K8s workloads	src/k8s/	k8s.yaml
Local k3d smoke test	k3d/src/	k3d/k3d.yaml

Caution

GKE cluster (3× n2-standard-4) + Filestore ENTERPRISE (1Ti) + NAT gateway ≈ $300+/mo while running (us-central1 list price, checked 2026-08 against chant 0.46.0; GCP list prices change and vary by region — re-price before relying on this number). The GPU node pool scales to zero. Tear down after testing.

What you’ll learn

• The two-phase deploy pattern: GCP infra via Config Connector, then K8s manifests via KubeRay
• Why RayCluster uses podSelector-only NetworkPolicy: the GKE pod CIDR mismatch problem
• How GCS object-store spillover prevents head OOM on large models and shuffled datasets
• Why pre-built Artifact Registry images matter at scale vs runtimeEnv pip installs
• Workload Identity wiring: GSA -> K8s ServiceAccount -> GCS access without key files

Key patterns

Two-phase deploy with Config Connector

The example splits GCP infrastructure and K8s workloads into two separate chant build targets. The split exists because the GKE cluster must exist before you can apply K8s manifests to it:

npm run build:gcp   # → config.yaml (Config Connector resources)
kubectl apply -f config.yaml
# wait for GKE cluster ready (~10 min)
npm run build:k8s   # → k8s.yaml (K8s + KubeRay resources)
kubectl apply -f k8s.yaml

Config Connector resources live in the same TypeScript source (src/infra/) as the rest of the config. The single chant build src --lexicon gcp command extracts only the GCP resources; --lexicon k8s extracts only the K8s resources. The shared config.ts passes values (storage class name, namespace, GSA email) between layers without a separate outputs mechanism.

NetworkPolicy: podSelector-only strategy

The RayCluster composite emits a NetworkPolicy that uses podSelector for all intra-cluster ingress and egress rules, with no IP CIDR blocks for Ray traffic. This is intentional.

GKE allocates pod IPs from secondary IP ranges that differ from the declared subnet CIDRs. A rule like ipBlock: 10.128.0.0/9 silently fails when GKE places pods on a secondary range like 10.64.0.0/14. Using podSelector: { ray.io/cluster-name: ray } matches the actual pods regardless of their IP, so the rule works correctly even when secondary ranges shift between node pools or cluster upgrades.

GCS egress is the one place that needs an ipBlock, because storage.googleapis.com resolves to Google’s public IP space. The composite allows 0.0.0.0/0 port 443 with RFC1918 ranges excluded. This permits Google APIs while preventing lateral movement to other services in the VPC.

GCS object-store spillover

Ray’s shared object store lives in the head pod’s memory. When large objects (model weights, shuffled datasets) exceed available RAM, Ray falls back to local disk, then OOMs or slows to a crawl on large jobs.

Setting spilloverBucket on the RayCluster composite injects RAY_object_spilling_config into the head container:

{
  "type": "smart_open",
  "params": { "uri": "gs://ray-spill/spill", "num_threads": 16 }
}

Ray spills objects to GCS transparently. The head pod needs GCS write access, which Workload Identity provides by binding the K8s ray-head ServiceAccount to a GCP SA with roles/storage.objectAdmin on the bucket. The 7-day lifecycle rule on the bucket auto-deletes orphaned spill files.

Pre-built images vs runtimeEnv pip installs

Ray’s runtimeEnv mechanism re-runs pip install on every worker restart. On a 4-worker cluster this adds 2–5 minutes to cold start; at 100+ workers it serializes across all nodes on the cluster Redis channel, adding 20+ minutes. The Tencent Weixin team confirmed pre-built images as the production-grade approach.

The example uses an Artifact Registry image (ray:2.40.0) with all dependencies pre-installed. The RAY_IMAGE env var points builds at your registry image. The Artifact Registry repository in the infra layer stores it:

# Build and push your image (example)
docker build -t us-central1-docker.pkg.dev/my-project/ray-images/ray:2.40.0 .
docker push us-central1-docker.pkg.dev/my-project/ray-images/ray:2.40.0

PreStop hooks and graceful drain

The composite injects a preStop lifecycle hook on all worker containers:

lifecycle:
  preStop:
    exec:
      command: ["ray", "stop"]
terminationGracePeriodSeconds: 120

When a worker pod is evicted (spot instance reclaim, node drain, rolling upgrade), Kubernetes sends SIGTERM and waits up to 120 seconds before force-killing. The ray stop preStop hook runs first, draining in-flight tasks from the worker’s local task queue. Without this, workers on preemptible/spot nodes lose their in-flight tasks instantly, and any job with more tasks than the remaining workers must restart from scratch.

Workload Identity for GCS access

The head pod needs GCS credentials for spillover. The composite emits a ServiceAccount named ${name}-head. The infra layer creates a GCP service account (ray-workload) and two IAM bindings:

1. roles/storage.objectAdmin on the spillover bucket -> scoped to the minimum needed
2. roles/iam.workloadIdentityUser binding [ray-system/ray-head] to the GCP SA

The defaults.serviceAccount prop on RayCluster injects the Workload Identity annotation:

defaults: {
  serviceAccount: {
    metadata: {
      annotations: { "iam.gke.io/gcp-service-account": config.rayGsaEmail },
    },
  },
},

The GKE metadata server handles token exchange, so there are no key files or Secrets.

Observability: Prometheus + Grafana via Helm

The same helm install kube-prometheus-stack command installs Prometheus + Grafana in both k3d and GKE:

just install-monitoring   # works in k3d and on GKE
just grafana              # port-forward → http://localhost:3000

The RayCluster composite sets RAY_GRAFANA_HOST on the head pod, which enables the Ray dashboard’s Metrics tab to embed the Grafana iframe. The env var points at the in-cluster kube-prometheus-stack Grafana service. The same DNS name resolves in both environments because both use the same Helm release name.

Ray ships six pre-built Grafana dashboard JSONs at /tmp/ray/session_latest/metrics/grafana/dashboards/ inside the head pod. Copy them into Grafana for per-cluster, per-worker, and per-job panels.

Job portability

Ray’s HTTP Jobs API is the same whether you’re talking to a local KubeRay cluster or this GKE deployment:

kubectl port-forward -n ray-system svc/ray-head-svc 8265:8265
ray job submit --address http://localhost:8265 -- python your_job.py

Jobs written with standard @ray.remote decorators, Ray Tune, or Ray Data require no changes to run on GKE. The one pattern that diverges: runtime_env YAML for pip dependencies works fine locally but serializes across all workers on cluster startup at scale: 100 workers each running pip install adds 20+ minutes to cold start. Pre-built images (the approach this example uses) remove that overhead.

KubeRay scale ceiling

KubeRay handles clusters up to ~1,000 nodes reliably. Beyond that, the single-controller architecture becomes a bottleneck; the Tencent Weixin team hit this and built a custom distributed scheduler (Starlink) and federated multi-cluster topology. For most production use cases, 1,000 nodes is more than sufficient. Near that scale, choose between federating multiple KubeRay clusters behind a single submission endpoint and adopting a custom scheduler.

Local smoke test (no GCP required)

Before spending $300+/mo on GKE, validate the KubeRay lifecycle locally with k3d. The smoke test runs in ~3 minutes with no cloud credentials.

Paste this to Claude Code from the repo root:

Run the ray-kuberay-gke local smoke test.
The example is in examples/ray-kuberay-gke.

Run npm install from the repo root first, then follow the instructions
in examples/ray-kuberay-gke/README.md.
Prerequisites: k3d and kubectl must be installed.

The agent creates a k3d cluster, install the KubeRay operator, build k3d/k3d.yaml from the local chant sources, apply it, and verify that ray.cluster_resources() returns ≥ 2 CPUs.

What the k3d layer validates:

• KubeRay operator deploys and becomes Available
• RayCluster CR is accepted and reaches state=ready
• Head + 1 CPU worker join the cluster

What it does not validate (all covered by the production GKE deploy): NetworkPolicy enforcement, GPU scheduling, ReadWriteMany shared storage, GCS spillover, Workload Identity.

The k3d config substitutes GCP-specific dependencies with local equivalents: rayproject/ray:2.54.0-py311 from Docker Hub instead of Artifact Registry (aarch64 variant selected automatically on Apple Silicon), k3s local-path StorageClass instead of Filestore, and removes the Workload Identity annotation and GCS spillover config entirely.

Deploy

See examples/ray-kuberay-gke/README.md for the full phase-by-phase deploy guide and paste-ready agent prompt.

Bootstrap Config Connector (one-time)

Config Connector is a GKE addon that manages GCP resources via kubectl apply. You need a management cluster with it running before you can apply config.yaml. Every chant GKE example ships a bootstrap script for this:

export GCP_PROJECT_ID=my-project
cd examples/ray-kuberay-gke
just bootstrap   # ~5 minutes

This creates ray-mgmt, a single-node GKE cluster in us-central1 that runs Config Connector. It enables the required GCP APIs (container, compute, file, storage, artifactregistry), creates a config-connector-sa service account with roles/editor + roles/iam.securityAdmin, and wires Workload Identity so the K8s SA can impersonate it.

If you already have a Config Connector management cluster from another chant GKE example, skip this step.

Quick reference

export GCP_PROJECT_ID=my-project GCP_REGION=us-central1
cd examples/ray-kuberay-gke
npm install         # from repo root first
just bootstrap      # one-time: create Config Connector management cluster
npm run build:gcp && kubectl apply -f config.yaml
# wait for cluster (~10 min)
just get-credentials
just install-operator
npm run build:k8s && kubectl apply -f k8s.yaml
just wait && just test-job
just install-monitoring && just grafana

Tear down

just teardown
# deletes k8s.yaml resources first, then config.yaml (GCP infra)

GCP resources (GKE cluster, Filestore, GCS bucket) are deleted when Config Connector reconciles the deletion. The Filestore ENTERPRISE instance takes ~5 minutes to delete.