Troubleshoot cluster connections

This page describes how to troubleshoot common errors that you might encounter when registering clusters to a fleet or connecting to clusters outside of Google Cloud using the Google Cloud console, the Google Cloud CLI, or kubectl through the Connect gateway.

On-premises clusters and clusters on other public clouds depend on the Connect Agent to establish and maintain a connection between the cluster and your Google Cloud project, and to handle Kubernetes requests. If you see errors such as "Unreachable Agent" or "Failed to connect to cluster's control plane" this could indicate a problem with the Connect Agent.

Collecting Connect Agent logs

When you register a cluster outside Google Cloud, it uses the Connect Agent to handle communication between your cluster and your fleet host project. The Connect Agent is a Deployment, gke-connect-agent, typically installed in your cluster in the namespace gke-connect. Collecting logs from this Connect Agent can be useful for troubleshooting registration and connection issues.

You can retrieve the Agent's logs by running the following command (adjust the count of lines if necessary):

kubectl logs -n gke-connect -l app=gke-connect-agent --tail=-1

To get information about every Connect Agent running in your project clusters:

kubectl describe deployment --all-namespaces -l app=gke-connect-agent

A successful connection should have entries similar to the below example:

2019/02/16 17:28:43.312056 dialer.go:244: dialer: dial: connected to gkeconnect.googleapis.com:443
2019/02/16 17:28:43.312279 tunnel.go:234: serve: opening egress stream...
2019/02/16 17:28:43.312439 tunnel.go:248: serve: registering endpoint="442223602236", shard="88d2bca5-f40a-11e8-983e-42010a8000b2" {"Params":{"GkeConnect":{"endpoint_class":1,"metadata":{"Metadata":{"Default":{"manifest_version":"234227867"}}}}}} ...
2019/02/16 17:28:43.312656 tunnel.go:259: serve: serving requests...

Collecting GKE Identity Service logs

Inspecting the GKE Identity Service logs may be helpful if you are having issues with Google Groups or third-party support for Connect gateway. This method of generating logs is only applicable to GKE on VMware or Google Distributed Cloud Virtual for Bare Metal clusters.

  1. Increase the verbosity of GKE Identity Service logs by making an edit to the clientconfig custom resource with the following command:

    kubectl edit deployment -n anthos-identity-service
    

    and adding a vmodule flag under the containers field like so:

    spec:
      containers:
      ...
      - command:
        - --vmodule=cloud/identity/hybrid/charon/*=9
    
  2. Restart the GKE Identity Service pod by deleting it with the following command:

    kubectl delete pods -l k8s-app=ais -n anthos-identity-service
    

    A pod should turn up again within a few seconds.

  3. Once the pod has restarted, run the original command which was returning an unexpected response to populate the GKE Identity Service pod logs with more details.

  4. Save the output of these logs to a file using the following command:

    kubectl logs -l k8s-app=ais -n anthos-identity-service --tail=-1 > gke_id_service_logs.txt
    

If expected groups are missing from the GKE Identity Service pod logs, verify the setup for the cluster is correct. If there are other GKE Identity Service related issues, refer to Troubleshoot user access issues or Troubleshooting fleet-level setup issues.

tls: oversized record errors

Symptom

You might encounter an error like this one:

... dialer: dial: connection to gkeconnect.googleapis.com:443 failed after
388.080605ms: serve: egress call failed: rpc error: code = Unauthenticated
desc = transport: oauth2: cannot fetch token: Post
https://oauth2.googleapis.com/token: proxyconnect tcp: tls: oversized record
received with length 20527
Possible causes

This can mean that the Connect Agent is trying to connect via HTTPS to an HTTP-only proxy. The Connect Agent only supports CONNECT-based HTTP proxies.

Resolution

You need to reconfigure your proxy environment variables to the following:

http_proxy=http://[PROXY_URL]:[PROXY_PORT]
https_proxy=http://[PROXY_URL]:[PROXY_PORT]

oauth2: cannot fetch token errors

Symptom

You might encounter an error like this one:

...  dialer: dial: connection to gkeconnect.googleapis.com:443 failed
after 388.080605ms: serve: egress call failed: rpc error: code =
Unauthenticated desc = transport: oauth2: cannot fetch token: Post
https://oauth2.googleapis.com/token: read tcp 192.168.1.40:5570->1.1.1.1:80
read: connection reset by peer
Possible causes

This can mean that the upstream HTTP proxy reset the connection, most likely because this particular URL is not allowed by your HTTP proxy. In the example above, the 1.1.1.1:80 is the HTTP proxy address.

Resolution

Check that your HTTP proxy allowlist includes the following URLs/domains:

gkeconnect.googleapis.com
oauth2.googleapis.com/token
www.googleapis.com/oauth2/v1/certs

Connect Agent pod crash and restart errors

Symptom

You might encounter intermittent "Unreachable Agent" errors in the Google Cloud console for your cluster and/or you might find the Pod has restarted multiple times:

$ kubectl get pods -n gke-connect
NAME                                                READY   STATUS    RESTARTS   AGE
gke-connect-agent-20230706-03-00-6b8f75dd58-dzwmt   1/1     Running   5          99m

To troubleshoot this behavior, describe the Pod to see if its last state was terminated due to an out of memory error (OOMKilled).:

  kubectl describe pods/gke-connect-agent-20230706-03-00-6b8f75dd58-dzwmt -n gke-connect
        <some details skipped..>
        Last State:     Terminated
        Reason:       OOMKilled
Possible causes
By default, Connect Agent Pods have a 256MiB RAM limit. If the cluster has installed a lot of workloads, it's possible that some of the requests and responses cannot be handled as expected.
Resolution

Update Connect Agent deployment and grant it a higher memory limit, for example:

containers:
  name: gke-connect-agent-20230706-03-00
  resources:
    limits:
      memory: 512Mi

PermissionDenied errors

Symptom

You might encounter an error like this one:

tunnel.go:250: serve: recv error: rpc error: code = PermissionDenied
desc = The caller does not have permission
dialer.go:210: dialer: dial: connection to gkeconnect.googleapis.com:443
failed after 335.153278ms: serve: receive request failed: rpc error:
code = PermissionDenied desc = The caller does not have permission
dialer.go:150: dialer: connection done: serve: receive request failed:
rpc error: code = PermissionDenied desc = The caller does not have permission
dialer.go:228: dialer: backoff: 1m14.1376766s
Possible causes

This can mean that you have not bound the required Identity and Access Management (IAM) role to the Google Cloud service account you created to authorize the Connect Agent to connect to Google. The Google Cloud service account requires the gkehub.connect IAM role.

This can also happen if you delete and recreate the Google Cloud service account with the same name. You also need to delete and recreate the IAM role binding in that case. Refer to Deleting and recreating service accounts for more information.

Resolution

Bind the gkehub.connect role to your service account (note that the gkehub.admin role does not have the proper permissions for connecting and is not meant to be used by service accounts).

For example, for a project called my-project and a Google Cloud service account called gkeconnect@my-project.iam.gserviceaccount.com, you'd run the following command to bind the role to the service account:

gcloud projects add-iam-policy-binding my-project --member \
serviceAccount:gkeconnect@my-project.iam.gserviceaccount.com \
--role "roles/gkehub.connect"

You can view and verify the service account permissions have been applied to a Google Cloud service account by examining the output of the following command, and you should see the role: roles/gkehub.connect bound to the associated Google Cloud service Account.

gcloud projects get-iam-policy my-project

Error when binding IAM role to the Google Cloud service account

Symptom

You might encounter an error like this one:

ERROR: (gcloud.projects.add-iam-policy-binding) PERMISSION_DENIED:
Service Management API has not been used in project [PROJECT_ID] before or it
is disabled. Enable it by visiting
https://console.developers.google.com/apis/api/servicemanagement.googleapis.com/overview?project=[PROJECT_ID]
then retry. If you enabled this API recently, wait a few minutes for the
action to propagate to our systems and retry.
Possible causes

You might not have the IAM permissions to run the command gcloud projects add-iam-policy-binding.

Resolution

You need to have the permission resourcemanager.projects.setIamPolicy If you have the Project IAM Admin, Owner or Editor roles, you should be able to run the command. If an internal security policy prohibits you from running the command, check with your administrator.

Error from invalid service account key

Symptom

You might encounter an error like this one:

2020/05/08 01:22:21.435104 environment.go:214: Got ExternalID 3770f509-b89b-48c4-96e0-860bb70b3a58 from namespace kube-system.
2020/05/08 01:22:21.437976 environment.go:485: Using gcp Service Account key
2020/05/08 01:22:21.438140 gkeconnect_agent.go:50: error creating kubernetes connect agent: failed to get tunnel config: unexpected end of JSON input
Possible causes

These logs indicate that the Connect Agent was provided with an invalid service account key during installation.

Resolution

Create a new JSON file containing service account credentials, and then reinstall the Connect Agent by following the steps to register a cluster.

Error from expired service account key

Symptom

You might encounter an error like this one:

2020/05/08 01:22:21.435104 dialer.go:277: dialer: dial: connection to gkeconnect.googleapis.com:443 failed after 37.901608ms:
serve: egress call failed: rpc error: code = Unauthenticated desc = transport: oauth2: cannot fetch token: 400 Bad Request
Response: {"error":"invalid_grant","error_description":"Invalid JWT Signature."}
Possible causes

These logs indicate that the Connect Agent was dialing Connect with an invalid service account key. The service account key file might contain errors, or the key might have expired.

To check whether the key has expired, use the Google Cloud console to list your service account keys and their expiration dates.

Resolution

Create a new JSON file containing service account credentials, and then reinstall the Connect Agent by following the steps to register a cluster.

Error from skewed system clock

Symptom

You might encounter an error like this one:

acceptCall: failed to parse token in req [rpc_id=1]: Token used before issued [rpc_id=1]
Possible causes

The log message usually indicates that there is a clock skew on the cluster. The cluster-issued token has an out-of-sync timestamp and thus the token is rejected.

Resolution

To see if the clock is not properly synced, you can run the date command on your cluster and compare it with the standard time. Usually a few seconds drift will cause this problem. To resolve this issue, please re-sync your cluster's clock.

Unable to see workloads in Google Cloud console

Symptoms

In the Connect Agent logs, you might observe the following errors:

"https://10.0.10.6:443/api/v1/nodes" YYYY-MM-DDTHH mm:ss.sssZ http.go:86: GET
"https://10.0.10.6:443/api/v1/pods" YYYY-MM-DDTHH mm:ss.sssZ http.go:139:
Response status: "403 Forbidden" YYYY-MM-DDTHH mm:ss.sssZ http.go:139:
Response status: "403 Forbidden"`
Possible causes

These logs indicate that Google Cloud is attempting to access the cluster using the credentials you provided during registration. 403 errors indicate that the credentials don't have the permissions required to access the cluster.

Resolution

Check that the token and the account it's bound to, and be sure that it has the appropriate permissions on the cluster.

Context deadline exceeded

Symptom

You might encounter an error like this one:

2019/03/06 21:08:43.306625 dialer.go:235: dialer: dial: connecting to gkeconnect.googleapis.com:443...
2019/03/06 21:09:13.306893 dialer.go:240: dialer: dial: unable to connect to gkeconnect.googleapis.com:443: context deadline exceeded
2019/03/06 21:09:13.306943 dialer.go:183: dialer: connection done: context deadline exceeded
Possible causes

This error indicates a low-level TCP networking issue where the Connect Agent cannot talk with gkeconnect.googleapis.com.

Resolution

Verify that Pod workloads within this cluster can resolve and have outbound connectivity to gkeconnect.googleapis.com on port 443.

Agent connection fails intermittently

Symptoms

In the Connect Agent logs, you might observe the following errors:

2020/10/06 18:02:34.409749 dialer.go:277: dialer: dial: connection to gkeconnect.googleapis.com:443 failed after 8m0.790286282s: serve: receive request failed: rpc error: code = Unavailable desc = transport is closing
2020/10/06 18:02:34.416618 dialer.go:207: dialer: connection done: serve: receive request failed: rpc error: code = Unavailable desc = transport is closing
2020/10/06 18:02:34.416722 dialer.go:295: dialer: backoff: 978.11948ms
2020/10/06 18:02:34.410097 tunnel.go:651: sendResponse: EOF [rpc_id=52]
2020/10/06 18:02:34.420077 tunnel.go:651: sendResponse: EOF [rpc_id=52]
2020/10/06 18:02:34.420204 tunnel.go:670: sendHalfClose: EOF [rpc_id=52]
2020/10/06 18:02:34.401412 tunnel.go:670: sendHalfClose: EOF [rpc_id=53]
Possible causes

The connection to Connect closes when Connect Agent does not have enough resources, for example on smaller AWS EC2 instances such as t3.medium.

Resolution

If you use AWS and the T3 instance type, enable T3 unlimited or use an instance type with more resources for your node pools.

Fleet cannot access the project

Symptoms

During some Fleet operations (usually cluster registration) you might observe an error similar to the following:

ERROR: (gcloud.container.hub.memberships.register) failed to initialize Feature
"authorizer", the fleet service account (service-PROJECT_NUMBER@gcp-sa-gkehub.iam.gserviceaccount.com) may not have access to your project
Possible causes

Fleet's default service account, gcp-sa-gkehub, can accidentally become unbound from a project. The Fleet Service Agent is an IAM role that grants the service account the permissions to manage cluster resources. If you remove this role binding from the service account, the default service account becomes unbound from the project, which can prevent you from registering clusters and other cluster operations.

You can check to see if the service account has been removed from your project using gcloud CLI or the Google Cloud console. If the command or the dashboard do not display gcp-sa-gkehub among your service accounts, the service account has become unbound.

gcloud

Run the following command:

gcloud projects get-iam-policy PROJECT_NAME

where PROJECT_NAME is the name of the project where you are trying register the cluster.

Console

Visit the IAM & admin page in Google Cloud console.

Resolution

If you removed the Fleet Service Agent role binding, run the following commands to restore the role binding:

PROJECT_NUMBER=$(gcloud projects describe PROJECT_NAME --format "value(projectNumber)")
gcloud projects add-iam-policy-binding PROJECT_NAME \
  --member "serviceAccount:service-${PROJECT_NUMBER}@gcp-sa-gkehub.iam.gserviceaccount.com" \
  --role roles/gkehub.serviceAgent

To confirm that the role binding was granted:

gcloud projects get-iam-policy PROJECT_NAME

If you see the service account name along with the gkehub.serviceAgent role, the role binding has been granted. For example:

- members:
  - serviceAccount:service-1234567890@gcp-sa-gkehub.iam.gserviceaccount.com
  role: roles/gkehub.serviceAgent

Error in registering a GKE cluster from a different project than Fleet

Symptoms

While registering a GKE cluster from a project different than the Fleet project, you might observe an error similar to the following in gcloud CLI:

...
message: 'DeployPatch failed'>
detail: 'DeployPatch failed'
...

It can be verified in logging by applying the following filters:

resource.type="gke_cluster"
resource.labels.cluster_name="my-cluster"
protoPayload.methodName="google.container.v1beta1.ClusterManager.UpdateCluster"
protoPayload.status.code="13"
protoPayload.status.message="Internal error."
severity=ERROR

Possible causes

The Fleet default Service Account does not have the required permissions in the GKE cluster's project.

Resolution

Grant the Fleet default Service Account the required permissions before registering the cluster.

Error in registering/unregistering a GKE cluster or updating fleet membership details for a registered GKE cluster during credentials rotation

Symptoms

While rotating your cluster credentials(https://cloud.google.com/kubernetes-engine/docs/how-to/credential-rotation), you may find errors if registering/unregistering a GKE cluster or updating the membership for a registered GKE cluster.

ERROR: (gcloud.container.hub.memberships.unregister) "code": 13,
"message": "an internal error has occurred"
Possible causes

Cluster credentials are in an intermediate state where the Fleet service is unable to access them.

Resolution

Complete the rotation before registering/umpregistering the cluster or updating the membership for a registered GKE cluster.

Error when disabling the Fleet API

Symptoms

When trying to disable the Fleet API (gkehub.googleapis.com), you may encounter an error similar to the following:

Not ready to deactivate the service on this project; ensure there are no more resources managed by this service.
Possible causes

There are still clusters registered to Google Cloud (memberships) or fleet-level features enabled in this project. All memberships or features must be unregistered or disabled before the API can be disabled.

  • To view your current registered clusters, follow the instructions in View fleet members

  • To see all active fleet-level features for your project:

gcloud and cURL

$ curl -H "Authorization: Bearer $(gcloud auth print-access-token)" \
    https://gkehub.googleapis.com/v1alpha1/projects/PROJECT_NAME/locations/global/features

where PROJECT_NAME is the name of the project where you are trying to disable the Fleet API.

Console

If you have GKE Enterprise enabled in your project, visit the features overview page page in Google Cloud console. Features listed as ENABLED are active fleet-level features.

Resolution

First, unregister any clusters still registered to your project fleet. All clusters must be unregistered before some features can be disabled.

Once you have done this, disable all fleet-level features. Currently, this is only possible with the Fleet REST API.

  1. Disable fleet-level features that you have enabled for your project

    $ gcloud alpha container hub FEATURE_COMMAND disable
    
  2. Disable feature authorizer and metering, which are enabled by default.

    $ curl -H "Authorization: Bearer $(gcloud auth print-access-token)" \
        -X "DELETE" \
        https://gkehub.googleapis.com/v1alpha1/projects/PROJECT_NAME/locations/global/features/FEATURE
    

    where FEATURE is the name of the feature to disable (such as authorizer or metering).

Missing cluster permissions when registering a cluster

Symptom:

While trying to register a cluster with a user account or Google Cloud service account, you may get an error similar to the following:

ERROR: (gcloud.container.hub.memberships.register) ResponseError: code=403, message=Required "container.clusters.get" permission(s) for "projects/my-project/zones/zone-a/clusters/my-cluster"
Possible cause:

The account that is trying to register the cluster does not have the required cluster-admin role-based access control (RBAC) role in the cluster.

Resolution:

Grant the cluster-admin RBAC role to the account before registering the cluster.

Error Failed to check if the user is a cluster-admin: Unable to connect to the server when registering a cluster

Symptom:

While trying to register a cluster, you may get an error similar to the following:

ERROR: (gcloud.container.hub.memberships.register) Failed to check if the user is a cluster-admin: Unable to connect to the server: net/http: request canceled while waiting for connection (Client.Timeout exceeded while awaiting headers)

Or

ERROR: (gcloud.container.hub.memberships.register) Failed to check if the user is a cluster-admin: Unable to connect to the server: dial tcp MASTER_ENDPOINT_IP:443: i/o timeout
Possible cause:

The machine on which you are running the registration gcloud command can't connect to the cluster's external endpoint. This usually happens if you have a private cluster with external access/IP disabled, but your machine's external IP address is not allowlisted. Note that registering a GKE cluster doesn't have this requirement after gcloud 407.0.0.

Resolution:

Make sure the machine on which you want to run the gcloud registration command can access the cluster's API server. If your cluster doesn't have external access enabled, file a case with Google Cloud Support.

Getting additional help

You can file a ticket with Google Cloud support for GKE Enterprise, by performing the following steps:

  1. File a case with Google Cloud Support.
  2. Follow the instructions in Collecting Connect Agent logs to save the Connect logs.
  3. If troubleshooting a GKE on VMware or Google Distributed Cloud Virtual for Bare Metal cluster using Google Groups or third-party support, follow the instructions in Collecting GKE Identity Service logs to save the GKE Identity Service logs. Make sure to sanitize the pod logs in the saved file if necessary.
  4. Attach the relevant logs to your case.