Troubleshooting

The following sections describe issues you might encounter while using GKE on-prem, and how to resolve them.

Before you begin

Check the following sections before you begin troubleshooting an issue.

Diagnosing cluster issues using gkectl

Use gkectl diagnosecommands to identify cluster issues and share cluster information with Google. See Diagnosing cluster issues.

Default logging behavior

For gkectl and gkeadm it is sufficient to use the default logging settings:

  • By default, log entries are saved as follows:

    • For gkectl, the default log file is /home/ubuntu/.config/gke-on-prem/logs/gkectl-$(date).log, and the file is symlinked with the logs/gkectl-$(date).log file in the local directory where you run gkectl.
    • For gkeadm, the default log file is logs/gkeadm-$(date).log in the local directory where you run gkeadm.
  • All log entries are saved in the log file, even if they are not printed in the terminal (when --alsologtostderr is false).
  • The -v5 verbosity level (default) covers all the log entries needed by the support team.
  • The log file also contains the command executed and the failure message.

We recommend that you send the log file to the support team when you need help.

Specifying a non-default location for the log file

To specify a non-default location for the gkectl log file, use the --log_file flag. The log file that you specify will not be symlinked with the local directory.

To specify a non-default location for the gkeadm log file, use the --log_file flag.

Locating Cluster API logs in the admin cluster

If a VM fails to start after the admin control plane has started, you can try debugging this by inspecting the Cluster API controllers' logs in the admin cluster:

  1. Find the name of the Cluster API controllers Pod in the kube-system namespace, where [ADMIN_CLUSTER_KUBECONFIG] is the path to the admin cluster's kubeconfig file:

    kubectl --kubeconfig [ADMIN_CLUSTER_KUBECONFIG] -n kube-system get pods | grep clusterapi-controllers
  2. Open the Pod's logs, where [POD_NAME] is the name of the Pod. Optionally, use grep or a similar tool to search for errors:

    kubectl --kubeconfig [ADMIN_CLUSTER_KUBECONFIG] -n kube-system logs [POD_NAME] vsphere-controller-manager

Installation

Debugging F5 BIG-IP issues using the admin cluster control plane node's kubeconfig

After an installation, GKE on-prem generates a kubeconfig file in the home directory of your admin workstation named internal-cluster-kubeconfig-debug. This kubeconfig file is identical to your admin cluster's kubeconfig, except that it points directly at the admin cluster's control plane node, where the admin control plane runs. You can use the internal-cluster-kubeconfig-debug file to debug F5 BIG-IP issues.

gkectl check-config validation fails: can't find F5 BIG-IP partitions

Symptoms

Validation fails because F5 BIG-IP partitions can't be found, even though they exist.

Potential causes

An issue with the F5 BIG-IP API can cause validation to fail.

Resolution

Try running gkectl check-config again.

gkectl prepare --validate-attestations fails: could not validate build attestation

Symptoms

Running gkectl prepare with the optional --validate-attestations flag returns the following error:

could not validate build attestation for gcr.io/gke-on-prem-release/.../...: VIOLATES_POLICY
Potential causes

An attestation might not exist for the affected image(s).

Resolution

Try downloading and deploying the admin workstation OVA again, as instructed in Creating an admin workstation. If the issue persists, reach out to Google for assistance.

Debugging using the bootstrap cluster's logs

During installation, GKE on-prem creates a temporary bootstrap cluster. After a successful installation, GKE on-prem deletes the bootstrap cluster, leaving you with your admin cluster and user cluster. Generally, you should have no reason to interact with this cluster.

If something goes wrong during an installation, and you did pass --cleanup-external-cluster=false to gkectl create cluster, you might find it useful to debug using the bootstrap cluster's logs. You can find the Pod, and then get its logs:

kubectl --kubeconfig /home/ubuntu/.kube/kind-config-gkectl get pods -n kube-system
kubectl --kubeconfig /home/ubuntu/.kube/kind-config-gkectl -n kube-system get logs [POD_NAME]

Admin workstation

openssl can't validate admin workstation OVA

Symptoms

Running openssl dgst against the admin workstation OVA file doesn't return Verified OK

Potential causes

An issue is present in the OVA file that prevents successful validation.

Resolution

Try downloading and deploying the admin workstation OVA again, as instructed in Download the admin workstation OVA . If the issue persists, reach out to Google for assistance.

Connect

Unable to register a user cluster

If you encounter issues with registering user clusters, reach out to Google for assistance.

Cluster created during alpha was deregistered

Refer to Registering a user cluster in the Connect documentation.

You might also choose to delete and recreate the cluster.

Storage

Volume fails to attach

Symptoms

The output of gkectl diagnose cluster looks like the following:

Checking cluster object...PASS
Checking machine objects...PASS
Checking control plane pods...PASS
Checking gke-connect pods...PASS
Checking kube-system pods...PASS
Checking gke-system pods...PASS
Checking storage...FAIL
    PersistentVolume pvc-776459c3-d350-11e9-9db8-e297f465bc84: virtual disk "[datastore_nfs] kubevols/kubernetes-dynamic-pvc-776459c3-d350-11e9-9db8-e297f465bc84.vmdk" IS attached to machine "gsl-test-user-9b46dbf9b-9wdj7" but IS NOT listed in the Node.Status
1 storage errors

One or more Pods is stuck in ContainerCreating state with a warning like the following:

Events:
  Type     Reason              Age               From                     Message
  ----     ------              ----              ----                     -------
  Warning  FailedAttachVolume  6s (x6 over 31s)  attachdetach-controller  AttachVolume.Attach failed for volume "pvc-776459c3-d350-11e9-9db8-e297f465bc84" : Failed to add disk 'scsi0:6'.

Potential causes

If a virtual disk is attached to the wrong virtual machine, it may be due to issue #32727 in Kubernetes 1.12.

Resolution

If a virtual disk is attached to the wrong virtual machine, you might need to manually detach it:

  1. Drain the node. See Safely draining a node. You might want to include the --ignore-daemonsets and --delete-local-data flags in your kubectl drain command.
  2. Power off the VM.
  3. Edit the VM's hardware config in vCenter to remove the volume.
  4. Power on the VM
  5. Uncordon the node.

Volume is lost

Symptoms

The output of gkectl diagnose cluster looks like the following:

Checking cluster object...PASS
Checking machine objects...PASS
Checking control plane pods...PASS
Checking gke-connect pods...PASS
Checking kube-system pods...PASS
Checking gke-system pods...PASS
Checking storage...FAIL
    PersistentVolume pvc-52161704-d350-11e9-9db8-e297f465bc84: virtual disk "[datastore_nfs] kubevols/kubernetes-dynamic-pvc-52161704-d350-11e9-9db8-e297f465bc84.vmdk" IS NOT found
1 storage errors

One or more Pods is stuck in ContainerCreating state with a warning like the following:

Events:
  Type     Reason              Age                   From                                    Message
  ----     ------              ----                  ----                                    -------
  Warning  FailedAttachVolume  71s (x28 over 42m)    attachdetach-controller                 AttachVolume.Attach failed for volume "pvc-52161704-d350-11e9-9db8-e297f465bc84" : File []/vmfs/volumes/43416d29-03095e58/kubevols/
  kubernetes-dynamic-pvc-52161704-d350-11e9-9db8-e297f465bc84.vmdk was not found

Potential causes

If you see a "not found" error related to your VMDK file, it is likely that the virtual disk was permanently deleted. This can happen if an operator manually deletes a virtual disk or the virtual machine it is attached to. To prevent this, manage your virtual machines as described in Resizing a user cluster and Upgrading clusters

Resolution

If a virtual disk was permanently deleted, you might need to manually clean up related Kubernetes resources:

  1. Delete the PVC that referenced the PV by running kubectl delete pvc [PVC_NAME].
  2. Delete the Pod that referenced the PVC by running kubectl delete pod [POD_NAME].
  3. Repeat step 2. (Yes, really. See Kubernetes issue 74374.)

vSphere CSI Volume fails to detach

Symptoms

If you find pods stuck in the ContainerCreating phase with FailedAttachVolume warnings, it could be due to a failed detach on a different node.

Run the following command to find CSI detach errors:

kubectl get volumeattachments -o=custom-columns=NAME:metadata.name,DETACH_ERROR:status.detachError.message

The output should look like the following:

NAME                                                                   DETACH_ERROR
csi-0e80d9be14dc09a49e1997cc17fc69dd8ce58254bd48d0d8e26a554d930a91e5   rpc error: code = Internal desc = QueryVolume failed for volumeID: "57549b5d-0ad3-48a9-aeca-42e64a773469". ServerFaultCode: NoPermission
csi-164d56e3286e954befdf0f5a82d59031dbfd50709c927a0e6ccf21d1fa60192d   
csi-8d9c3d0439f413fa9e176c63f5cc92bd67a33a1b76919d42c20347d52c57435c   
csi-e40d65005bc64c45735e91d7f7e54b2481a2bd41f5df7cc219a2c03608e8e7a8   

Potential causes

The CNS > Searchable privilege has not been granted to the vSphere user.

Resolution

Add the CNS > Searchable privilege to your vCenter user account. The detach operation automatically retries until it succeeds.

Upgrades

About downtime during upgrades

Resource Description
Admin cluster

When an admin cluster is down, user cluster control planes and workloads on user clusters continue to run, unless they were affected by a failure that caused the downtime

User cluster control plane

Typically, you should expect no noticeable downtime to user cluster control planes. However, long-running connections to the Kubernetes API server might break and would need to be re-established. In those cases, the API caller should retry until it establishes a connection. In the worst case, there can be up to one minute of downtime during an upgrade.

User cluster nodes

If an upgrade requires a change to user cluster nodes, GKE on-prem recreates the nodes in a rolling fashion, and reschedules Pods running on these nodes. You can prevent impact to your workloads by configuring appropriate PodDisruptionBudgets and anti-affinity rules.

Resizing user clusters

Resizing a user cluster fails

Symptoms

A resize operation on a user cluster fails.

Potential causes

Several factors could cause resize operations to fail.

Resolution

If a resize fails, follow these steps:

  1. Check the cluster's MachineDeployment status to see if there are any events or error messages:

    kubectl describe machinedeployments [MACHINE_DEPLOYMENT_NAME]
  2. Check if there are errors on the newly-created Machines:

    kubectl describe machine [MACHINE_NAME]

Error: "no addresses can be allocated"

Symptoms

After resizing a user cluster, kubectl describe machine [MACHINE_NAME] displays the following error:

Events:
   Type     Reason  Age                From                    Message
   ----     ------  ----               ----                    -------
   Warning  Failed  9s (x13 over 56s)  machineipam-controller  ipam: no addresses can be allocated
   
Potential causes

There aren't enough IP addresses available for the user cluster.

Resolution

Allocate more IP addresses for the cluster. Then, delete the affected Machine:

kubectl delete machine [MACHINE_NAME]

If the cluster is configured correctly, a replacement Machine is created with an IP address.

Sufficient number of IP addresses allocated, but Machine fails to register with cluster

Symptoms

Network has enough addresses allocated but the Machine still fails to register with the user cluster.

Possible causes

There might be an IP conflict. The IP might be taken by another Machine or by your load balancer.

Resolution

Check that the affected Machine's IP address is not taken. If there is a conflict, you need to resolve the conflict in your environment.

vSphere

Debugging with govc

If you encounter issues specific to vSphere, you can use govc to troubleshoot. For example, you can easily confirm permissions and access for your vCenter user accounts and collect vSphere logs.

Changing vCenter Certificate

If you are running a vCenter server in evaluation or default setup mode, and it has a generated TLS certificate, this certificate might change over time. If the certificate has changed, you need to let your running cluster(s) know about the new certificate:

  1. Retrieve the new vCenter cert and save to a file:

    true | openssl s_client -connect [VCENTER_IP_ADDRESS]:443 -showcerts 2>/dev/null | sed -ne '/-BEGIN/,/-END/p' > vcenter.pem
  2. Now, for each cluster, delete the ConfigMap containing the vSphere and vCenter certificate for each cluster, and create a new ConfigMap with the new cert. For example:

    kubectl --kubeconfig kubeconfig delete configmap vsphere-ca-certificate -n kube-system
    kubectl --kubeconfig kubeconfig delete configmap vsphere-ca-certificate -n user-cluster1
    kubectl --kubeconfig kubeconfig create configmap -n user-cluster1 --dry-run vsphere-ca-certificate --from-file=ca.crt=vcenter.pem  -o yaml  | kubectl --kubeconfig kubeconfig apply -f -
    kubectl --kubeconfig kubeconfig create configmap -n kube-system --dry-run vsphere-ca-certificate --from-file=ca.crt=vcenter.pem  -o yaml  | kubectl --kubeconfig kubeconfig apply -f -
  3. Delete the clusterapi-controllers Pod for each cluster. When the Pod restarts, it begins using the new certificate. For example:

    kubectl --kubeconfig kubeconfig -n kube-system get pods
    kubectl --kubeconfig kubeconfig -n kube-system delete pod clusterapi-controllers-...

Miscellaneous

Terraform vSphere provider session limit

GKE on-prem uses Terraform's vSphere provider to bring up VMs in your vSphere environment. The provider's session limit is 1000 sessions. The current implementation doesn't close active sessions after use. You might encounter 503 errors if you have too many sessions running.

Sessions are automatically closed after 300 seconds.

Symptoms

If you have too many sessions running, you might encounter the following error:

Error connecting to CIS REST endpoint: Login failed: body:
  {"type":"com.vmware.vapi.std.errors.service_unavailable","value":
  {"messages":[{"args":["1000","1000"],"default_message":"Sessions count is
  limited to 1000. Existing sessions are 1000.",
  "id":"com.vmware.vapi.endpoint.failedToLoginMaxSessionCountReached"}]}},
  status: 503 Service Unavailable
Potential causes

There are too many Terraform provider sessions running in your environment.

Resolution

Currently, this is working as intended. Sessions are automatically closed after 300 seconds. For more information, refer to to GitHub issue #618.

Using a proxy for Docker: oauth2: cannot fetch token

Symptoms

While using a proxy, you encounter the following error:

oauth2: cannot fetch token: Post https://oauth2.googleapis.com/token: proxyconnect tcp: tls: oversized record received with length 20527
Potential causes

You might have provided a HTTPS proxy instead of HTTP.

Resolution

In your Docker configuration, change the proxy address to http:// instead of https://.

Verifying that licenses are valid

Remember to verify that your licenses is valid, especially if you are using trial licenses. You might encounter unexpected failures if your F5, ESXi host, or vCenter licenses have expired.