Deploy Ray Cluster in ACK - Container Service for Kubernetes

Prerequisites

Ensure the following:

An ACK managed Pro cluster running Kubernetes v1.24 or later (create | upgrade).
At least one node with 8 vCPUs and 32 GB memory (test). Size to your workload in production. For GPU workloads, use supported Elastic Compute Service (ECS) instance families.
kubectl installed and connected to your cluster.
(Optional) An ApsaraDB for Tair instance for Global Control Store (GCS) fault tolerance.

(Optional) Set up GCS fault tolerance

Create an ApsaraDB for Tair (Redis-compatible) instance:

Same region and Virtual Private Cloud (VPC) as your ACK cluster (create an instance).
A whitelist allowing access from the VPC CIDR block.
A VPC endpoint (recommended). View connection addresses.
The instance password (change or reset).

Install the KubeRay Operator

Log on to the ACK console.
In the left-side navigation pane, click Clusters. Click your cluster name.
Go to Operations > Add-ons > Manage Applications.
Under Kuberay-Operator, click Install.

Deploy the Ray Cluster

Important

This example uses rayproject/ray:2.36.1 from Docker Hub. If the pull fails, use one of these alternatives:

Mirror the image via Container Registry from outside the Chinese mainland (subscribe to images outside China).
Create a Global Accelerator instance to pull images from overseas.

Create a Ray Cluster named myfirst-ray-cluster:

The manifest configures:

Component	Setting	Notes
Head node	`num-cpus: "0"`	Reserves the head for cluster management
Worker group `work1`	1 replica, scalable up to 1,000	Adjust `replicas` to match your workload
Autoscaling	Disabled (`enableInTreeAutoscaling: false`)	Enable after profiling your workload resource patterns

Full manifest

cat <<EOF | kubectl apply -f -
apiVersion: ray.io/v1
kind: RayCluster
metadata:
  name: myfirst-ray-cluster
  namespace: default
spec:
  suspend: false
  autoscalerOptions:
    env: []
    envFrom: []
    idleTimeoutSeconds: 60
    imagePullPolicy: Always
    resources:
      limits:
        cpu: 2000m
        memory: 2024Mi
      requests:
        cpu: 2000m
        memory: 2024Mi
    securityContext: {}
    upscalingMode: Default
  enableInTreeAutoscaling: false
  headGroupSpec:
    rayStartParams:
      dashboard-host: 0.0.0.0
      num-cpus: "0"
    serviceType: ClusterIP
    template:
      spec:
        containers:
        - image: rayproject/ray:2.36.1
          imagePullPolicy: Always
          name: ray-head
          resources:
            limits:
              cpu: "4"
              memory: 4G
            requests:
              cpu: "1"
              memory: 1G
  workerGroupSpecs:
  - groupName: work1
    maxReplicas: 1000
    minReplicas: 0
    numOfHosts: 1
    rayStartParams: {}
    replicas: 1
    template:
      spec:
        containers:
        - image: rayproject/ray:2.36.1
          imagePullPolicy: Always
          name: ray-worker
          resources:
            limits:
              cpu: "4"
              memory: 4G
            requests:
              cpu: "4"
              memory: 4G
EOF

Verify the deployment

Confirm the cluster is running:

Check the Ray Cluster status:

kubectl get raycluster

Expected output:

NAME                  DESIRED WORKERS   AVAILABLE WORKERS   CPUS   MEMORY   GPUS   STATUS   AGE
myfirst-ray-cluster   1                 1                   5      5G       0      ready    4m19s

Check the Pods:

kubectl get pod

Expected output:

NAME                                     READY   STATUS    RESTARTS   AGE
myfirst-ray-cluster-head-5q2hk           1/1     Running   0          4m37s
myfirst-ray-cluster-work1-worker-zkjgq   1/1     Running   0          4m31s

Check the Services:

kubectl get svc

Expected output:

NAME                           TYPE        CLUSTER-IP    EXTERNAL-IP   PORT(S)                                         AGE
kubernetes                     ClusterIP   192.168.0.1   <none>        443/TCP                                         21d
myfirst-ray-cluster-head-svc   ClusterIP   None          <none>        10001/TCP,8265/TCP,8080/TCP,6379/TCP,8000/TCP   6m57s

When the Ray Cluster status is ready and all Pods show 1/1 Running, the deployment is complete.