Rook Ceph or Looker pod stuck in Init state

Error details

The error code for this issue is 30001.

Description

Occasionally, on node restart, an issue causes the Looker or Rook Ceph pod to get stuck in Init state as the volume required for attaching the PVC to a pod is missing.

Verify if the problem is indeed related to Longhorn by running the following command:

kubectl get events -A -o json | jq -r '.items[] | select(.message != null) | select(.message | contains("cannot get resource \"volumeattachments\" in API group \"storage.k8s.io\""))'

If it is related to Longhorn, this command should return a list of pod names affected by the issue. If the command does not return anything, the cause of the problem is different.

Solution

Run the following script to fix the problematic pods if the previous command returns a non-empty output:

#!/bin/bash

function wait_till_rollout() {
    local namespace=$1
    local object_type=$2
    local deploy=$3

    local try=0
    local maxtry=2
    local status="notready"

    while [[ ${status} == "notready" ]]  && (( try != maxtry )) ; do
        kubectl -n "$namespace" rollout status "$deploy" -w --timeout=600s;
        # shellcheck disable=SC2181
        if [[ "$?" -ne 0 ]];
        then
            status="notready"
            try=$((try+1))
        else
            status="ready"
        fi
    done
    if [[ $status == "notready" ]]; then
        echo "$deploy of type $object_type failed in namespace $namespace. Plz re-run the script once again to verify that it's not a transient issue !!!"
        exit 1
    fi
}

function fix_pv_deployments() {
    for pod_name in $(kubectl get events -A -o json | jq -r '.items[]  | select(.message | contains("cannot get resource \"volumeattachments\" in API group \"storage.k8s.io\"")) | select(.involvedObject.kind == "Pod") | .involvedObject.name + "/" + .involvedObject.namespace' | sort | uniq)
    do
        POD_NAME=$(echo "${pod_name}" | cut -d '/' -f1)
        NS=$(echo "${pod_name}" | cut -d '/' -f2)
        controller_data=$(kubectl -n "${NS}" get po "${POD_NAME}" -o json | jq -r '[.metadata.ownerReferences[] | select(.controller==true)][0] | .kind + "=" + .name')
        [[ $controller_data == "" ]] && error "Error: Could not determine owner for pod: ${POD_NAME}" && exit 1
        CONTROLLER_KIND=$(echo "${controller_data}" | cut -d'=' -f1)
        CONTROLLER_NAME=$(echo "${controller_data}" | cut -d'=' -f2)
        if [[ $CONTROLLER_KIND == "ReplicaSet" ]]
        then
            controller_data=$(kubectl  -n "${NS}" get "${CONTROLLER_KIND}" "${CONTROLLER_NAME}" -o json | jq -r '[.metadata.ownerReferences[] | select(.controller==true)][0] | .kind + "=" + .name')
            CONTROLLER_KIND=$(echo "${controller_data}" | cut -d'=' -f1)
            CONTROLLER_NAME=$(echo "${controller_data}" | cut -d'=' -f2)

            replicas=$(kubectl -n "${NS}" get "$CONTROLLER_KIND" "$CONTROLLER_NAME" -o json | jq -r '.status.replicas')
            unavailable_replicas=$(kubectl -n "${NS}" get "$CONTROLLER_KIND" "$CONTROLLER_NAME" -o json | jq -r '.status.unavailableReplicas')

            if [ -n "$unavailable_replicas" ]; then
                available_replicas=$((replicas - unavailable_replicas))
                if [ $available_replicas -eq 0 ]; then
                    kubectl -n "$NS" scale "$CONTROLLER_KIND" "$CONTROLLER_NAME" --replicas=0
                    sleep 15
                    kubectl -n "$NS" scale "$CONTROLLER_KIND" "$CONTROLLER_NAME" --replicas="$replicas"
                    deployment_name="$CONTROLLER_KIND/$CONTROLLER_NAME"
                    wait_till_rollout "$NS" "deploy" "$deployment_name"
                fi
            fi
        fi
    done
}

fix_pv_deployments