Skip to content

Latest commit

 

History

History
303 lines (218 loc) · 14.9 KB

TROUBLESHOOTING.md

File metadata and controls

303 lines (218 loc) · 14.9 KB

Troubleshooting

Tools

A list of tools which are helpful for troubleshooting

Installation

Troubeshooting runbooks that relate to ARC installation problems

InternalError when calling webhook: context deadline exceeded

Problem

This issue can come up for various reasons like leftovers from previous installations or not being able to access the K8s service's clusterIP associated with the admission webhook server (of ARC).

Internal error occurred: failed calling webhook "mutate.runnerdeployment.actions.summerwind.dev":
Post "https://actions-runner-controller-webhook.actions-runner-system.svc:443/mutate-actions-summerwind-dev-v1alpha1-runnerdeployment?timeout=10s": context deadline exceeded

Solution

First we will try the common solution of checking webhook leftovers from previous installations:

  1. kubectl get validatingwebhookconfiguration -A
    kubectl get mutatingwebhookconfiguration -A
  2. If you see any webhooks related to actions-runner-controller, delete them:
    kubectl delete mutatingwebhookconfiguration actions-runner-controller-mutating-webhook-configuration
    kubectl delete validatingwebhookconfiguration actions-runner-controller-validating-webhook-configuration

If that didn't work then probably your K8s control-plane is somehow unable to access the K8s service's clusterIP associated with the admission webhook server:

  1. You're running apiserver as a binary and you didn't make service cluster IPs available to the host network.
  2. You're running the apiserver in the pod but your pod network (i.e. CNI plugin installation and config) is not good so your pods(like kube-apiserver) in the K8s control-plane nodes can't access ARC's admission webhook server pod(s) in probably data-plane nodes.

Another reason could be due to GKEs firewall settings you may run into the following errors when trying to deploy runners on a private GKE cluster:

To fix this, you may either:

  1. Configure the webhook to use another port, such as 443 or 10250, each of which allow traffic by default.

    # With helm, you'd set `webhookPort` to the port number of your choice
    # See https://github.com/actions/actions-runner-controller/pull/1410/files for more information
    helm upgrade --install --namespace actions-runner-system --create-namespace \
                 --wait actions-runner-controller actions/actions-runner-controller \
                 --set webhookPort=10250
  2. Set up a firewall rule to allow the master node to connect to the default webhook port. The exact way to do this may vary, but the following script should point you in the right direction:

    # 1) Retrieve the network tag automatically given to the worker nodes
    # NOTE: this only works if you have only one cluster in your GCP project. You will have to manually inspect the result of this command to find the tag for the cluster you want to target
    WORKER_NODES_TAG=$(gcloud compute instances list --format='text(tags.items[0])' --filter='metadata.kubelet-config:*' | grep tags | awk '{print $2}' | sort | uniq)
    
    # 2) Take note of the VPC network in which you deployed your cluster
    # NOTE this only works if you have only one network in which you deploy your clusters
    NETWORK=$(gcloud compute instances list --format='text(networkInterfaces[0].network)' --filter='metadata.kubelet-config:*' | grep networks | awk -F'/' '{print $NF}' | sort | uniq)
    
    # 3) Get the master source ip block
    SOURCE=$(gcloud container clusters describe <cluster-name> --region <region> | grep masterIpv4CidrBlock| cut -d ':' -f 2 | tr -d ' ')
    
    gcloud compute firewall-rules create k8s-cert-manager --source-ranges $SOURCE --target-tags $WORKER_NODES_TAG  --allow TCP:9443 --network $NETWORK

Invalid header field value

Problem

2020-11-12T22:17:30.693Z	ERROR	controller-runtime.controller	Reconciler error	
{
  "controller": "runner",
  "request": "actions-runner-system/runner-deployment-dk7q8-dk5c9",
  "error": "failed to create registration token: Post \"https://api.github.com/orgs/$YOUR_ORG_HERE/actions/runners/registration-token\": net/http: invalid header field value \"Bearer $YOUR_TOKEN_HERE\\n\" for key Authorization"
}

Solution

Your base64'ed PAT token has a new line at the end, it needs to be created without a \n added, either:

  • echo -n $TOKEN | base64
  • Create the secret as described in the docs using the shell and documented flags

Helm chart install failure: certificate signed by unknown authority

Problem

Error: UPGRADE FAILED: failed to create resource: Internal error occurred: failed calling webhook "webhook.cert-manager.io": failed to call webhook: Post "https://cert-manager-webhook.cert-manager.svc:443/mutate?timeout=10s": x509: certificate signed by unknown authority

Apparently, it's failing while helm is creating one of resources defined in the ARC chart and the cause was that cert-manager's webhook is not working correctly, due to the missing or the invalid CA certficate.

You'd try to tail logs from the cert-manager-cainjector and see it's failing with an error like:

$ kubectl -n cert-manager logs cert-manager-cainjector-7cdbb9c945-g6bt4
I0703 03:31:55.159339       1 start.go:91] "starting" version="v1.1.1" revision="3ac7418070e22c87fae4b22603a6b952f797ae96"
I0703 03:31:55.615061       1 leaderelection.go:243] attempting to acquire leader lease  kube-system/cert-manager-cainjector-leader-election...
I0703 03:32:10.738039       1 leaderelection.go:253] successfully acquired lease kube-system/cert-manager-cainjector-leader-election
I0703 03:32:10.739941       1 recorder.go:52] cert-manager/controller-runtime/manager/events "msg"="Normal"  "message"="cert-manager-cainjector-7cdbb9c945-g6bt4_88e4bc70-eded-4343-a6fb-0ddd6434eb55 became leader" "object"={"kind":"ConfigMap","namespace":"kube-system","name":"cert-manager-cainjector-leader-election","uid":"942a021e-364c-461a-978c-f54a95723cdc","apiVersion":"v1","resourceVersion":"1576"} "reason"="LeaderElection"
E0703 03:32:11.192128       1 start.go:119] cert-manager/ca-injector "msg"="manager goroutine exited" "error"=null
I0703 03:32:12.339197       1 request.go:645] Throttling request took 1.047437675s, request: GET:https://10.96.0.1:443/apis/storage.k8s.io/v1beta1?timeout=32s
E0703 03:32:13.143790       1 start.go:151] cert-manager/ca-injector "msg"="Error registering certificate based controllers. Retrying after 5 seconds." "error"="no matches for kind \"MutatingWebhookConfiguration\" in version \"admissionregistration.k8s.io/v1beta1\""
Error: error registering secret controller: no matches for kind "MutatingWebhookConfiguration" in version "admissionregistration.k8s.io/v1beta1"

Solution

Your cluster is based on a new enough Kubernetes of version 1.22 or greater which does not support the legacy admissionregistration.k8s.io/v1beta1 API anymore, and your cert-manager is not up-to-date hence it's still trying to use the leagcy Kubernetes API.

In many cases, it's not an option to downgrade Kubernetes. So, just upgrade cert-manager to a more recent version that does have have the support for the specific Kubernetes version you're using.

See https://cert-manager.io/docs/installation/supported-releases/ for the list of available cert-manager versions.

Operations

Troubeshooting runbooks that relate to ARC operational problems

Stuck runner kind or backing pod

Problem

Sometimes either the runner kind (kubectl get runners) or it's underlying pod can get stuck in a terminating state for various reasons. You can get the kind unstuck by removing its finaliser using something like this:

Solution

Remove the finaliser from the relevent runner kind or pod

# Get all kind runners and remove the finalizer
$ kubectl get runners --no-headers | awk {'print $1'} | xargs kubectl patch runner --type merge -p '{"metadata":{"finalizers":null}}'

# Get all pods that are stuck terminating and remove the finalizer
$ kubectl -n get pods | grep Terminating | awk {'print $1'} | xargs kubectl patch pod -p '{"metadata":{"finalizers":null}}'

Note the code assumes you have already selected the namespace your runners are in and that they are in a namespace not shared with anything else

Delay in jobs being allocated to runners

Problem

ARC isn't involved in jobs actually getting allocated to a runner. ARC is responsible for orchestrating runners and the runner lifecycle. Why some people see large delays in job allocation is not clear however it has been confirmed actions#1387 (comment) that this is caused from the self-update process somehow.

Solution

Disable the self-update process in your runner manifests

apiVersion: actions.summerwind.dev/v1alpha1
kind: RunnerDeployment
metadata:
  name: example-runnerdeployment-with-sleep
spec:
  template:
    spec:
      ...
      env:
        - name: DISABLE_RUNNER_UPDATE
          value: "true"

Runner coming up before network available

Problem

If you're running your action runners on a service mesh like Istio, you might have problems with runner configuration accompanied by logs like:

....
runner Starting Runner listener with startup type: service
runner Started listener process
runner An error occurred: Not configured
runner Runner listener exited with error code 2
runner Runner listener exit with retryable error, re-launch runner in 5 seconds.
....

This is because the istio-proxy has not completed configuring itself when the configuration script tries to communicate with the network.

More broadly, there are many other circumstances where the runner pod coming up first can cause issues.

Solution

Added originally to help users with older istio instances. Newer Istio instances can use Istio's holdApplicationUntilProxyStarts attribute (istio/istio#11130) to avoid having to delay starting up the runner. Please read the discussion in #592 for more information.

You can add a delay to the runner's entrypoint script by setting the STARTUP_DELAY_IN_SECONDS environment variable for the runner pod. This will cause the script to sleep X seconds, this works with any runner kind.

apiVersion: actions.summerwind.dev/v1alpha1
kind: RunnerDeployment
metadata:
  name: example-runnerdeployment-with-sleep
spec:
  template:
    spec:
      ...
      env:
        - name: STARTUP_DELAY_IN_SECONDS
          value: "5"

Outgoing network action hangs indefinitely

Problem

Some random outgoing network actions hangs indefinitely. This could be because your cluster does not give Docker the standard MTU of 1500, you can check this out by running ip link in a pod that encounters the problem and reading the outgoing interface's MTU value. If it is smaller than 1500, then try the following.

Solution

Add a dockerMTU key in your runner's spec with the value you read on the outgoing interface. For instance:

apiVersion: actions.summerwind.dev/v1alpha1
kind: RunnerDeployment
metadata:
  name: github-runner
  namespace: github-system
spec:
  replicas: 6
  template:
    spec:
      dockerMTU: 1400
      repository: $username/$repo
      env: []

If the issue still persists, you can set the ARC_DOCKER_MTU_PROPAGATION to propagate the host MTU to networks created by the GitHub Runner. For instance:

apiVersion: actions.summerwind.dev/v1alpha1
kind: RunnerDeployment
metadata:
  name: github-runner
  namespace: github-system
spec:
  replicas: 6
  template:
    spec:
      dockerMTU: 1400
      repository: $username/$repo
      env:
        - name: ARC_DOCKER_MTU_PROPAGATION
          value: "true"

You can read the discussion regarding this issue in (#1406)[actions#1046].

Unable to scale to zero with TotalNumberOfQueuedAndInProgressWorkflowRuns

Problem

HRA doesn't scale the RunnerDeployment to zero, even though you did configure HRA correctly, to have a pull-based scaling metric TotalNumberOfQueuedAndInProgressWorkflowRuns, and set minReplicas: 0.

Solution

You very likely have some dangling workflow jobs stuck in queued or in_progress as seen in #1057.

Manually call the "list workflow runs" API, and remove the dangling workflow job(s).

Slow / failure to boot dind sidecar (default runner)

Problem

If you noticed that it takes several minutes for sidecar dind container to be created or it exits with with error just after being created it might indicate that you are experiencing disk performance issue. You might see message failed to reserve container name when scaling up multiple runners at once. When you ssh on kubernetes node that problematic pods were scheduled on you can use tools like atop, htop or iotop to check IO usage and cpu time percentage used on iowait. If you see that disk usage is high (80-100%) and iowaits are taking a significant chunk of you cpu time (normally it should not be higher than 10%) it means that performance is being bottlenecked by slow disk.

Solution

The solution is to switch to using faster storage, if you are experiencing this issue you are probably using hdd, switch to ssh fixed the problem in my case. Most cloud providers have a list of storage options to use just pick something faster that your current disk, for on prem clusters you will need to invest in some ssds.