Starting and stopping Ray clusters on Kubernetes fails

Kun_Hwi_Ko · February 12, 2022, 12:35am

Problem
I found a problem on creating and destroying Ray clusters on Kubernetes, and was hoping to get some advice here.

I typically bring up resources via a REST call to the Kubernetes API server rather than through kubectl create. Most of the times, the call will go through and I see that the resources are created/deleted when I run kubectl get raycluster. I also see that the Pods and services are brought up. However, sometimes I see that the resources do not get created/deleted.

Suspect
I am suspecting this has something to do with the Ray finalizers? One possible workaround I am aware of is to patch the finalizer to be null". This does bring down the resource but it prevents me from bringing up resources afterwards for some reason.

The second workaround that I have done is to restart the Ray operator (e.g. kubectl delete), which seems to make all the REST calls valid again.

Question
Obviously, I don’t think it is a good idea to restart the Ray operator every time. Is it alright to disable the finalizer condition, and if so, what are some problems and how could I do it? Also what is the finalizer waiting for that prevents users from deleting the ray cluster resource?

Any other possible suggestion will also be great!

Kun_Hwi_Ko · February 14, 2022, 8:18pm

This is the YAML output of the raycluster resource:

apiVersion: cluster.ray.io/v1
kind: RayCluster
metadata:
  annotations:
    kopf.zalando.org/last-handled-configuration: |
      {"spec":{"headPodType":"head-node","headServicePorts":[{"name":"client","port":10001,"targetPort":10001},{"name":"dashboard","port":8265,"targetPort":8265},{"name":"ray-serve","port":8000,"targetPort":8000}],"headStartRayCommands":["ray stop","ulimit -n 65536; ray start --head --no-monitor --dashboard-host 0.0.0.0 &> /tmp/raylogs"],"idleTimeoutMinutes":5,"maxWorkers":2,"podTypes":[{"name":"head-node","podConfig":{"apiVersion":"v1","kind":"Pod","metadata":{"generateName":"head-"},"spec":{"containers":[{"args":["trap : TERM INT; touch /tmp/raylogs; tail -f /tmp/raylogs; sleep infinity & wait;"],"command":["/bin/bash","-c","--"],"image":"ci-artifacts.c3.ai/c3:custom-ray","imagePullPolicy":"Always","name":"ray-node","ports":[{"containerPort":6379,"protocol":"TCP"},{"containerPort":10001,"protocol":"TCP"},{"containerPort":8265,"protocol":"TCP"},{"containerPort":8000,"protocol":"TCP"}],"resources":{"limits":{"cpu":"1200m","memory":"4Gi","nvidia.com/gpu":"0"},"requests":{"cpu":"1200m","ephemeral-storage":"1Gi","memory":"4Gi","nvidia.com/gpu":"0"}},"volumeMounts":[{"mountPath":"/dev/shm","name":"dshm"},{"mountPath":"/usr/local/share","name":"environment-shared"}]}],"restartPolicy":"Always","volumes":[{"emptyDir":{"medium":"Memory"},"name":"dshm"},{"name":"environment-shared","persistentVolumeClaim":{"claimName":"test"}}]}},"rayResources":{"CPU":0}},{"maxWorkers":2,"minWorkers":2,"name":"worker-node","podConfig":{"apiVersion":"v1","kind":"Pod","metadata":{"generateName":"worker-"},"spec":{"containers":[{"args":["trap : TERM INT; touch /tmp/raylogs; tail -f /tmp/raylogs; sleep infinity & wait;"],"command":["/bin/bash","-c","--"],"image":"ci-artifacts.c3.ai/c3:custom-ray","imagePullPolicy":"Always","name":"ray-node","resources":{"limits":{"cpu":"1200m","memory":"2Gi","nvidia.com/gpu":"0"},"requests":{"cpu":"1200m","ephemeral-storage":"1Gi","memory":"2Gi","nvidia.com/gpu":"0"}},"volumeMounts":[{"mountPath":"/dev/shm","name":"dshm"},{"mountPath":"/usr/local/share","name":"environment-shared"}]}],"restartPolicy":"Always","volumes":[{"emptyDir":{"medium":"Memory"},"name":"dshm"},{"name":"environment-shared","persistentVolumeClaim":{"claimName":"test"}}]}}}],"upscalingSpeed":1,"workerStartRayCommands":["ray stop","ulimit -n 65536; ray start --address=$RAY_HEAD_IP:6379 &> /tmp/raylogs"]},"status":{"autoscalerRetries":0}}
  creationTimestamp: "2022-02-11T23:37:45Z"
  finalizers:
  - kopf.zalando.org/KopfFinalizerMarker
  generation: 1
  name: blue-k8sray-cloud
  namespace: khk
  resourceVersion: "79671714"
  uid: e6e5226f-1ab5-4de6-ab5b-233f1dd37568
spec:
  headPodType: head-node
  headServicePorts:
  - name: client
    port: 10001
    targetPort: 10001
  - name: dashboard
    port: 8265
    targetPort: 8265
  - name: ray-serve
    port: 8000
    targetPort: 8000
  headStartRayCommands:
  - ray stop
  - ulimit -n 65536; ray start --head --no-monitor --dashboard-host 0.0.0.0 &> /tmp/raylogs
  idleTimeoutMinutes: 5
  maxWorkers: 2
  podTypes:
  - name: head-node
    podConfig:
      apiVersion: v1
      kind: Pod
      metadata:
        generateName: head-
      spec:
        containers:
        - args:
          - 'trap : TERM INT; touch /tmp/raylogs; tail -f /tmp/raylogs; sleep infinity
            & wait;'
          command:
          - /bin/bash
          - -c
          - --
          image: ci-artifacts.c3.ai/c3:custom-ray
          imagePullPolicy: Always
          name: ray-node
          ports:
          - containerPort: 6379
            protocol: TCP
          - containerPort: 10001
            protocol: TCP
          - containerPort: 8265
            protocol: TCP
          - containerPort: 8000
            protocol: TCP
          resources:
            limits:
              cpu: 1200m
              memory: 4Gi
              nvidia.com/gpu: "0"
            requests:
              cpu: 1200m
              ephemeral-storage: 1Gi
              memory: 4Gi
              nvidia.com/gpu: "0"
          volumeMounts:
          - mountPath: /dev/shm
            name: dshm
          - mountPath: /usr/local/share
            name: environment-shared
        restartPolicy: Always
        volumes:
        - emptyDir:
            medium: Memory
          name: dshm
        - name: environment-shared
          persistentVolumeClaim:
            claimName: ctptest
    rayResources:
      CPU: 0
  - maxWorkers: 2
    minWorkers: 2
    name: worker-node
    podConfig:
      apiVersion: v1
      kind: Pod
      metadata:
        generateName: worker-
      spec:
        containers:
        - args:
          - 'trap : TERM INT; touch /tmp/raylogs; tail -f /tmp/raylogs; sleep infinity
            & wait;'
          command:
          - /bin/bash
          - -c
          - --
          image: ci-artifacts.c3.ai/c3:custom-ray
          imagePullPolicy: Always
          name: ray-node
          resources:
            limits:
              cpu: 1200m
              memory: 2Gi
              nvidia.com/gpu: "0"
            requests:
              cpu: 1200m
              ephemeral-storage: 1Gi
              memory: 2Gi
              nvidia.com/gpu: "0"
          volumeMounts:
          - mountPath: /dev/shm
            name: dshm
          - mountPath: /usr/local/share/c3/server
            name: environment-shared
        restartPolicy: Always
        volumes:
        - emptyDir:
            medium: Memory
          name: dshm
        - name: environment-shared
          persistentVolumeClaim:
            claimName: test
  upscalingSpeed: 1
  workerStartRayCommands:
  - ray stop
  - ulimit -n 65536; ray start --address=$RAY_HEAD_IP:6379 &> /tmp/raylogs
status:
  autoscalerRetries: 1
  kopf:
    progress: {}
  phase: AutoscalingExceptionRecovery

Also the YAML output for the head pod:

apiVersion: v1
kind: Pod
metadata:
  creationTimestamp: "2022-02-11T23:37:45Z"
  generateName: blue-k8sray-cloud-head-
  labels:
    cluster.ray.io/component: blue-k8sray-cloud-ray-head
    ray-cluster-name: blue-k8sray-cloud
    ray-file-mounts-contents: da39a3ee5e6b4b0d3255bfef95601890afd80709
    ray-launch-config: 4f78c0890df358fb3fed7aa1a7c10e51a68c9dd1
    ray-node-name: ray-blue-k8sray-cloud-head
    ray-node-status: up-to-date
    ray-node-type: head
    ray-node-uuid: 93d8c86a-3382-45ca-9f88-98533af4f9d5
    ray-runtime-config: 6828672a6e8cc7a7ae88f76f458d2a8124f6784f
    ray-user-node-type: head-node
  name: blue-k8sray-cloud-head-kwqb9
  namespace: khk
  ownerReferences:
  - apiVersion: cluster.ray.io/v1
    blockOwnerDeletion: true
    controller: true
    kind: RayCluster
    name: blue-k8sray-cloud
    uid: e6e5226f-1ab5-4de6-ab5b-233f1dd37568
  resourceVersion: "78839290"
  uid: 737e8d15-5e4b-4f74-8eb5-fb1ddbcfeebb
spec:
  containers:
  - args:
    - 'trap : TERM INT; touch /tmp/raylogs; tail -f /tmp/raylogs; sleep infinity &
      wait;'
    command:
    - /bin/bash
    - -c
    - --
    image: ci-artifacts.c3.ai/c3:custom-ray
    imagePullPolicy: Always
    name: ray-node
    ports:
    - containerPort: 6379
      protocol: TCP
    - containerPort: 10001
      protocol: TCP
    - containerPort: 8265
      protocol: TCP
    - containerPort: 8000
      protocol: TCP
    resources:
      limits:
        cpu: 1200m
        memory: 4Gi
        nvidia.com/gpu: "0"
      requests:
        cpu: 1200m
        ephemeral-storage: 1Gi
        memory: 4Gi
        nvidia.com/gpu: "0"
    terminationMessagePath: /dev/termination-log
    terminationMessagePolicy: File
    volumeMounts:
    - mountPath: /dev/shm
      name: dshm
    - mountPath: /usr/local/share/c3/server
      name: environment-shared
    - mountPath: /var/run/secrets/kubernetes.io/serviceaccount
      name: kube-api-access-ltflp
      readOnly: true
  dnsPolicy: ClusterFirst
  enableServiceLinks: true
  nodeName: aks-agentpool-67207258-vmss000008
  preemptionPolicy: PreemptLowerPriority
  priority: 0
  restartPolicy: Always
  schedulerName: default-scheduler
  securityContext: {}
  serviceAccount: default
  serviceAccountName: default
  terminationGracePeriodSeconds: 30
  tolerations:
  - effect: NoExecute
    key: node.kubernetes.io/not-ready
    operator: Exists
    tolerationSeconds: 300
  - effect: NoExecute
    key: node.kubernetes.io/unreachable
    operator: Exists
    tolerationSeconds: 300
  - effect: NoSchedule
    key: node.kubernetes.io/memory-pressure
    operator: Exists
  - effect: NoSchedule
    key: nvidia.com/gpu
    operator: Exists
  volumes:
  - emptyDir:
      medium: Memory
    name: dshm
  - name: environment-shared
    persistentVolumeClaim:
      claimName: test
  - name: kube-api-access-ltflp
    projected:
      defaultMode: 420
      sources:
      - serviceAccountToken:
          expirationSeconds: 3607
          path: token
      - configMap:
          items:
          - key: ca.crt
            path: ca.crt
          name: kube-root-ca.crt
      - downwardAPI:
          items:
          - fieldRef:
              apiVersion: v1
              fieldPath: metadata.namespace
            path: namespace
status:
  conditions:
  - lastProbeTime: null
    lastTransitionTime: "2022-02-11T23:37:45Z"
    status: "True"
    type: Initialized
  - lastProbeTime: null
    lastTransitionTime: "2022-02-11T23:37:47Z"
    status: "True"
    type: Ready
  - lastProbeTime: null
    lastTransitionTime: "2022-02-11T23:37:47Z"
    status: "True"
    type: ContainersReady
  - lastProbeTime: null
    lastTransitionTime: "2022-02-11T23:37:45Z"
    status: "True"
    type: PodScheduled
  containerStatuses:
  - containerID: containerd://e468980660c376910e92169894857baca2a520a963ccd6fb04a8b71176293fc7
    image: ci-artifacts.c3.ai/c3:custom-ray
    imageID: ci-artifacts.c3.ai/c3@sha256:...
    lastState: {}
    name: ray-node
    ready: true
    restartCount: 0
    started: true
    state:
      running:
        startedAt: "2022-02-11T23:37:47Z"
  hostIP: <hidden>
  phase: Running
  podIP: <hidden>
  podIPs:
  - ip: <hidden>
  qosClass: Guaranteed
  startTime: "2022-02-11T23:37:45Z"

ckw017 · February 15, 2022, 5:55pm

@Kun_Hwi_Ko can you check through the events in the namespace you’re running the operator in when resources creation/deletion fail (kubectl get events -n <namespace>)

cc @Dmitri any guess on what the finalizer might be hanging on

Kun_Hwi_Ko · February 15, 2022, 7:05pm

@ckw017 Thanks for the advice. I can’t find anything obvious through kubectl get events but I can check again. Below is the output that I have.

LAST SEEN   TYPE     REASON                 OBJECT                                   MESSAGE
56m         Normal   EnsuringLoadBalancer   service/c3-c3-nginx-ingress-controller   Ensuring load balancer
56m         Normal   EnsuredLoadBalancer    service/c3-c3-nginx-ingress-controller   Ensured load balancer
56m         Normal   EnsuringLoadBalancer   service/leader                           Ensuring load balancer
56m         Normal   EnsuredLoadBalancer    service/leader                           Ensured load balancer

When I send the REST call, I see a deletionTimestamp. So my guess is something is stopping the deletion until a certain condition is met.

  creationTimestamp: "2022-02-15T01:51:54Z"
  deletionGracePeriodSeconds: 0
  deletionTimestamp: "2022-02-15T18:57:08Z"
  finalizers:
  - kopf.zalando.org/KopfFinalizerMarker

bensonyuan · February 16, 2022, 1:28am

Additional information from kopf documenting this phenomenon.
https://kopf.readthedocs.io/en/latest/troubleshooting/

https://kopf.readthedocs.io/en/stable/handlers/#state-changing-handlers

Kun_Hwi_Ko · February 16, 2022, 11:33pm

Also want to edit this happens when I

Pull the Ray Helm Chart
Do a Helm installation
Leave it idle and go watch a movie
Come back and do kubectl delete raycluster
I see the deletion time stamp has been created, but the finalizer will prevent me from deleting the resource.

When I see the logs of the Ray operator pod, there should be a Send termination request to ... and a SIGTERM if the termination process starts. However, in some cases like above, I don’t see either happening.

Topic		Replies	Views
Best practise for taking down Ray cluster deployments on Kubernetes Kubernetes	1	981	May 12, 2022
Ray controller restart worker pod after head pod restart Kubernetes	0	192	November 19, 2023
Kubernetes cluster only creates head node Ray Clusters	11	587	June 7, 2022
Head pod does not restart after deleting/draining Kubernetes	7	580	August 9, 2022
Unable to delete errored ray head node in minikube Kubernetes	3	526	April 15, 2021

Starting and stopping Ray clusters on Kubernetes fails

Related Topics