GKE on Bare Metal known issues

This page lists all known issues for GKE on Bare Metal. To filter the known issues by a product version or category, select your filters from the following drop-down menus.

Select your GKE on Bare Metal version:

Select your problem category:

Or, search for your issue:

Category Identified version(s) Issue and workaround
Logging and monitoring 1.15, 1.16, 1.28

kube-state-metrics OOM in large cluster

You might notice that kube-state-metrics or the gke-metrics-agent Pod that exists on the same node as kube-state-metrics is out of memory (OOM).

This can happen in clusters with more that 50 nodes or with many Kubernetes objects.

Workaround

To resolve this issue, update the stackdriver custom resource definition to use the ksmNodePodMetricsOnly feature gate. This feature gate makes sure that only a small number of critical metrics are exposed.

To use this workaround, complete the following steps:

  1. Check the stackdriver custom resource definition for available feature gates:
    kubectl -n kube-system get crd stackdrivers.addons.gke.io -o yaml | grep ksmNodePodMetricsOnly
            
  2. Update the stackdriver custom resource definition to enable ksmNodePodMetricsOnly:
    kind:stackdriver
    spec:
      featureGates:
         ksmNodePodMetricsOnly: true
            
Installation 1.28.0-1.28.200

Preflight check fails on RHEL 9.2 due to missing iptables

When installing a cluster on the Red Hat Enterprise Linux (RHEL) 9.2 operating system, you might experience a failure due to the missing iptables package. The failure occurs during preflight checks and triggers an error message similar to the following:

'check_package_availability_pass': "The following packages are not available: ['iptables']"
        

RHEL 9.2 is in Preview for GKE on Bare Metal version 1.28.

Workaround

Bypass the preflight check error by setting spec.bypassPreflightCheck to true on your Cluster resource.

Operation 1.10, 1.11, 1.12, 1.13, 1.14, 1.15, 1.16

Slow MetalLB failover at high scale

When MetalLB handles a high number of services (over 10,000), failover can take over an hour. This happens because MetalLB uses a rate limited queue that, when under high scale, can take a while to get to the service that needs to fail over.

Workaround

Upgrade your cluster to version 1.28 or later. If you're unable to upgrade, manually editing the service (for example, adding an annotation) causes the service to failover more quickly.

Operation 1.16.0-1.16.6, 1.28.0-1.28.200

Environment variables have to be set on the admin workstation if proxy is enabled

bmctl check cluster can fail due to proxy failures if you don't have the environment variables HTTPS_PROXY and NO_PROXY defined on the admin workstation. The bmctl command reports an error message about failing to call some google services, like the following example:

[2024-01-29 23:49:03+0000] error validating cluster config: 2 errors occurred:
        * GKERegister check failed: 2 errors occurred:
        * Get "https://gkehub.googleapis.com/v1beta1/projects/baremetal-runqi/locations/global/memberships/ci-ec1a14a903eb1fc": oauth2: cannot fetch token: Post "https://oauth2.googleapis.com/token": dial tcp 108.177.98.95:443: i/o timeout
        * Post "https://cloudresourcemanager.googleapis.com/v1/projects/baremetal-runqi:testIamPermissions?alt=json&prettyPrint=false": oauth2: cannot fetch token: Post "https://oauth2.googleapis.com/token": dial tcp 74.125.199.95:443: i/o timeout

Workaround

Manually set the HTTPS_PROXY and NO_PROXY on the admin workstation.

Upgrades and updates 1.28.0-gke.435

Upgrades to version 1.28.0-gke.435 might fail if audit.log has incorrect ownership

In some cases, the /var/log/apiserver/audit.log file on control plane nodes has both group and user ownership set to root. This file ownership setting causes upgrade failures for the control plane nodes when upgrading a cluster from version 1.16.x to version 1.28.0-gke.435. This issue only applies to clusters that were created prior to version 1.11 and that had Cloud Audit Logs disabled. Cloud Audit Logs is enabled by default for clusters at version 1.9 and higher.

Workaround

If you're unable to upgrade your cluster to version 1.28.100-gke.146, use the following steps as a workaround to complete your cluster upgrade to version 1.28.0-gke.435:

  • If Cloud Audit Logs is enabled, remove the /var/log/apiserver/audit.log file.
  • If Cloud Audit Logs is disabled, change /var/log/apiserver/audit.log ownership to the same as the parent directory, /var/log/apiserver.
Networking, Upgrades and updates 1.28.0-gke.435

MetalLB doesn't assign IP addresses to VIP Services

GKE on Bare Metal uses MetalLB for bundled load balancing. In GKE on Bare Metal release 1.28.0-gke.435, the bundled MetalLB is upgraded to version 0.13, which introduces CRD support for IPAddressPools. However, because ConfigMaps allow any name for an IPAddressPool, the pool names had to be converted to a Kubernetes-compliant name by appending a hash to the end of the name of the IPAddressPool. For example, an IPAddressPool with a name default is converted to a name like default-qpvpd when you upgrade your cluster to version 1.28.0-gke.435.

Since MetalLB requires a specific name of an IPPool for selection, the name conversion prevents MetalLB from making a pool selection and assigning IP addresses. Therefore, Services that use metallb.universe.tf/address-pool as an annotation to select the address pool for an IP address no longer receive an IP address from the MetalLB controller.

This issue is fixed in GKE on Bare Metal version 1.28.100-gke.146.

Workaround

If you can't upgrade your cluster to version 1.28.100-gke.146, use the following steps as a workaround:

  1. Get the converted name of the IPAddressPool:
    kubectl get IPAddressPools -n kube-system
    
  2. Update the affected Service to set the metallb.universe.tf/address-pool annotation to the converted name with the hash.

    For example, if the IPAddressPool name was converted from default to a name like default-qpvpd, change the annotation metallb.universe.tf/address-pool: default in the Service to metallb.universe.tf/address-pool: default-qpvpd.

The hash used in the name conversion is deterministic, so the workaround is persistent.

Upgrades and updates 1.14

Orphan pods after upgrading to version 1.14.x

When you upgrade GKE on Bare Metal clusters to version 1.14.x, some resources from the previous version aren't deleted. Specifically, you might see a set of orphaned pods like the following:

capi-webhook-system/capi-controller-manager-xxx
capi-webhook-system/capi-kubeadm-bootstrap-controller-manager-xxx

These orphan objects don't impact cluster operation directly, but as a best practice, we recommend that you remove them.

  • Run the following commands to remove the orphan objects:
    kubectl delete ns capi-webhook-system
    kubectl delete validatingwebhookconfigurations capi-validating-webhook-configuration
    kubectl delete mutatingwebhookconfigurations capi-mutating-webhook-configuration
    

This issue is fixed in GKE on Bare Metal version 1.15.0 and higher.

Installation 1.14

Cluster creation stuck on the machine-init job

If you try to install GKE on Bare Metal version 1.14.x, you might experience a failure due to the machine-init jobs, similar to the following example output:

"kubeadm join" task failed due to:
error execution phase control-plane-join/etcd: error creating local etcd static pod manifest file: etcdserver: re-configuration failed due to not enough started members

"kubeadm reset" task failed due to:
panic: runtime error: invalid memory address or nil pointer dereference

Workaround:

Remove the obsolete etcd member that causes the machine-init job to fail. Complete the following steps on a functioning control plane node:

  1. List the existing etcd members:
    etcdctl --key=/etc/kubernetes/pki/etcd/peer.key \
      --cacert=/etc/kubernetes/pki/etcd/ca.crt \
      --cert=/etc/kubernetes/pki/etcd/peer.crt \
      member list
    
    Look for members with a status of unstarted, as shown in the following example output:
    5feb7ac839625038, started, vm-72fed95a, https://203.0.113.11:2380, https://203.0.113.11:2379, false
    99f09f145a74cb15, started, vm-8a5bc966, https://203.0.113.12:2380, https://203.0.113.12:2379, false
    bd1949bcb70e2cb5, unstarted, , https://203.0.113.10:2380, , false
    
  2. Remove the failed etcd member:
    etcdctl --key=/etc/kubernetes/pki/etcd/peer.key \
      --cacert=/etc/kubernetes/pki/etcd/ca.crt \
      --cert=/etc/kubernetes/pki/etcd/peer.crt \
      member remove MEMBER_ID
    
    Replace MEMBER_ID with the ID of the failed etcd member. In the previous example output, this ID is bd1949bcb70e2cb5.

    The following example output shows that the member has been removed:
    Member bd1949bcb70e2cb5 removed from cluster  9d70e2a69debf2f
    
Networking 1.28.0

Cilium-operator missing Node list and watch permissions

In Cilium 1.13, the cilium-operator ClusterRole permissions are incorrect. The Node list and watch permissions are missing. The cilium-operator fails to start garbage collectors, which results in the following issues:

  • Leakage of Cilium resources.
  • Stale identities aren't removed from BFP policy maps.
  • Policy maps might reach the 16K limit.
    • New entries can't be added.
    • Incorrect NetworkPolicy enforcement.
  • Identities might reach the 64K limit.
    • New Pods can't be created.

An operator that's missing the Node permissions reports the following example log message:

2024-01-02T20:41:37.742276761Z level=error msg=k8sError error="github.com/cilium/cilium/operator/watchers/node.go:83: Failed to watch *v1.Node: failed to list *v1.Node: nodes is forbidden: User \"system:serviceaccount:kube-system:cilium-operator\" cannot list resource \"nodes\" in API group \"\" at the cluster scope" subsys=k8s

The Cilium agent reports an error message when it's unable to insert an entry into a policy map, like the following example:

level=error msg="Failed to add PolicyMap key" bpfMapKey="{6572100 0 0 0}" containerID= datapathPolicyRevision=0 desiredPolicyRevision=7 endpointID=1313 error="Unable to update element for Cilium_policy_01313 map with file descriptor 190: the map is full, please consider resizing it. argument list too long" identity=128 ipv4= ipv6= k8sPodName=/ port=0 subsys=endpoint

Workaround:

Remove the Cilium identities, then add the missing ClusterRole permissions to the operator:

  1. Remove the existing CiliumIdentity objects:
    kubectl delete ciliumid –all
    
  2. Edit the cilium-operator ClusterRole object:
    kubectl edit clusterrole cilium-operator
    
  3. Add a section for nodes that includes the missing permissions, as shown in the following example:
    - apiGroups:
      - ""
      resources:
      - nodes
      verbs:
      - get
      - list
      - watch
    
  4. Save and close the editor. The operator dynamically detects the permission change. You don't need to manually restart the operator.
Upgrades and updates 1.15.0-1.15.7, 1.16.0-1.16.3

Transient issue encountered during the preflight check

One of the kubeadm health check tasks that runs during the upgrade preflight check might fail with the following error message:

[ERROR CreateJob]: could not delete Job \"upgrade-health-check\" in the namespace \"kube-system\": jobs.batch \"upgrade-health-check\" not found

This error can be safely ignored. If you encounter this error that blocks the upgrade, re-run the upgrade command.

If you observe this error when you run the preflight using the bmctl preflightcheck command, nothing is blocked by this failure. You can run the preflight check again to get the accurate preflight information.


Workaround:

Re-run the upgrade command, or if encountered during bmctl preflightcheck, re-run preflightcheck command.

Operation 1.14, 1.15.0-1.15.7, 1.16.0-1.16.3, 1.28.0

Periodic Network health check fails when a node is replaced or removed

This issue affects clusters that perform periodic network health checks after a node has been replaced or removed. If your cluster undergoes periodic health checks, the periodic network health check results in failure following the replacement or removal of a node, because the network inventory ConfigMap doesn't get updated once it's created.


Workaround:

The recommended workaround is to delete the inventory ConfigMap and the periodic network health check. The cluster operator automatically recreates them with the most up-to-date information.

For 1.14.x clusters, run the following commands:

kubectl delete configmap \
    $(kubectl get cronjob CLUSTER_NAME-network -o=jsonpath='{.spec.jobTemplate.spec.template.spec.volumes[?(@name=="inventory-config-volume")].configMap.name}' \
    -n CLUSTER_NAMESPACE) \
    -n CLUSTER_NAMESPACE \
    --kubeconfig ADMIN_KUBECONFIG
kubectl delete healthchecks.baremetal.cluster.gke.io \
    CLUSTER_NAME-network -n CLUSTER_NAMESPACE \
    --kubeconfig ADMIN_KUBECONFIG

For 1.15.0 and later clusters, run the following commands:

kubectl delete configmap \
    $(kubectl get cronjob bm-system-network -o=jsonpath='{.spec.jobTemplate.spec.template.spec.volumes[?(@.name=="inventory-config-volume")]configMap.name}' \
    -n CLUSTER_NAMESPACE) \
    -n CLUSTER_NAMESPACE \
    --kubeconfig ADMIN_KUBECONFIG
kubectl delete healthchecks.baremetal.cluster.gke.io \
    bm-system-network -n CLUSTER_NAMESPACE \
    --kubeconfig ADMIN_KUBECONFIG
Networking 1.14, 1.15, 1.16.0-1.16.2

Network Gateway for GDC can't apply your configuration when the device name contains a period

If you have a network device that includes a period character (.) in the name, such as bond0.2, Network Gateway for GDC treats the period as a path in the directory when it runs sysctl to make changes. When Network Gateway for GDC checks if duplicate address detection (DAD) is enabled, the check might fail and so won't reconcile.

The behavior is different between cluster versions:

  • 1.14 and 1.15: This error only exists when you use IPv6 floating IP addresses. If you don't use IPv6 floating IP addresses, you won't notice this issue when your device names contain a period.
  • 1.16.0 - 1.16.2: This error always exists when your device names contain a period.

Workaround:

Upgrade your cluster to version 1.16.3 or later.

As a workaround until you can upgrade your clusters, remove the period (.) from the name of the device.

Upgrades and updates, Networking, Security 1.16.0

Upgrades to 1.16.0 fail when seccomp is disabled

If seccomp is disabled for your cluster (spec.clusterSecurity.enableSeccomp set to false), then upgrades to version 1.16.0 fail.

GKE on Bare Metal version 1.16 uses Kubernetes version 1.27. In Kubernetes version 1.27.0 and higher, the feature for setting seccomp profiles is GA and no longer uses a feature gate. This Kubernetes change causes upgrades to version 1.16.0 to fail when seccomp is disabled in the cluster configuration. This issue is fixed in version 1.16.1 and higher clusters. If you have the cluster.spec.clusterSecurity.enableSeccomp field set to false, you can upgrade to version 1.16.1 or higher.

Clusters with spec.clusterSecurity.enableSeccomp unset or set to true are not affected.

Installation, Operation 1.11, 1.12, 1.13, 1.14, 1.15.0-1.15.5, 1.16.0-1.16.1

containerd metadata might become corrupt after reboot when /var/lib/containerd is mounted

If you have optionally mounted /var/lib/containerd, the containerd metadata might become corrupt after a reboot. Corrupt metadata might cause Pods to fail, including system-critical Pods.

To check if this issue affects you, see if an optional mount is defined in /etc/fstab for /var/lib/containerd/ and has nofail in the mount options.


Workaround:

Remove the nofail mount option in /etc/fstab, or upgrade your cluster to version 1.15.6 or later.

Operation 1.13, 1.14, 1.15, 1.16, 1.28

Clean up stale Pods in the cluster

You might see Pods managed by a Deployment (ReplicaSet) in a Failed state and with the status of TaintToleration. These Pods don't use cluster resources, but should be deleted.

You can use the following kubectl command to list the Pods that you can clean up:

kubectl get pods –A | grep TaintToleration

The following example output shows a Pod with the TaintToleration status:

kube-system    stackdriver-metadata-agent-[...]    0/1    TaintToleration    0

Workaround:

For each Pod with the described symptoms, check the ReplicaSet that the Pod belongs to. If the ReplicaSet is satisfied, you can delete the Pods:

  1. Get the ReplicaSet that manages the Pod and find the ownerRef.Kind value:
    kubectl get pod POD_NAME -n NAMESPACE -o yaml
    
  2. Get the ReplicaSet and verify that the status.replicas is the same as spec.replicas:
    kubectl get replicaset REPLICA_NAME -n NAMESPACE -o yaml
    
  3. If the names match, delete the Pod:
    kubectl delete pod POD_NAME -n NAMESPACE.
    
Upgrades 1.16.0

etcd-events can stall when upgrade to version 1.16.0

When you upgrade an existing cluster to version 1.16.0, Pod failures related to etcd-events can stall the operation. Specifically, the upgrade-node job fails for the TASK [etcd_events_install : Run etcdevents] step.

If you're affected by this issue, you see Pod failures like the following:

  • The kube-apiserver Pod fails to start with the following error:
    connection error: desc = "transport: Error while dialing dial tcp 127.0.0.1:2382: connect: connection refused"
    
  • The etcd-events pod fails to start with the following error:
    Error: error syncing endpoints with etcd: context deadline exceeded
    

Workaround:

If you can't upgrade your cluster to a version with the fix, use the following temporary workaround to address the errors:

  1. Use SSH to access the control plane node with the reported errors.
  2. Edit the etcd-events manifest file, /etc/kubernetes/manifests/etcd-events.yaml, and remove the initial-cluster-state=existing flag.
  3. Apply the manifest.
  4. Upgrade should continue.
Networking 1.15.0-1.15.2

CoreDNS orderPolicy not recognized

OrderPolicy doesn't get recognized as a parameter and isn't used. Instead, GKE on Bare Metal always uses Random.

This issue occurs because the CoreDNS template was not updated, which causes orderPolicy to be ignored.


Workaround:

Update the CoreDNS template and apply the fix. This fix persists until an upgrade.

  1. Edit the existing template:
    kubectl edit cm -n kube-system coredns-template
    
    Replace the contents of the template with the following:
    coredns-template: |-
      .:53 {
        errors
        health {
          lameduck 5s
        }
        ready
        kubernetes cluster.local in-addr.arpa ip6.arpa {
          pods insecure
          fallthrough in-addr.arpa ip6.arpa
        }
    {{- if .PrivateGoogleAccess }}
        import zones/private.Corefile
    {{- end }}
    {{- if .RestrictedGoogleAccess }}
        import zones/restricted.Corefile
    {{- end }}
        prometheus :9153
        forward . {{ .UpstreamNameservers }} {
          max_concurrent 1000
          {{- if ne .OrderPolicy "" }}
          policy {{ .OrderPolicy }}
          {{- end }}
        }
        cache 30
    {{- if .DefaultDomainQueryLogging }}
        log
    {{- end }}
        loop
        reload
        loadbalance
    }{{ range $i, $stubdomain := .StubDomains }}
    {{ $stubdomain.Domain }}:53 {
      errors
    {{- if $stubdomain.QueryLogging }}
      log
    {{- end }}
      cache 30
      forward . {{ $stubdomain.Nameservers }} {
        max_concurrent 1000
        {{- if ne $.OrderPolicy "" }}
        policy {{ $.OrderPolicy }}
        {{- end }}
      }
    }
    {{- end }}
    
Networking, Operation 1.10, 1.11, 1.12, 1.13, 1.14

Network Gateway for GDC components evicted or pending due to missing priority class

Network gateway Pods in kube-system might show a status of Pending or Evicted, as shown in the following condensed example output:

$ kubectl -n kube-system get pods | grep ang-node
ang-node-bjkkc     2/2     Running     0     5d2h
ang-node-mw8cq     0/2     Evicted     0     6m5s
ang-node-zsmq7     0/2     Pending     0     7h

These errors indicate eviction events or an inability to schedule Pods due to node resources. As Network Gateway for GDC Pods have no PriorityClass, they have the same default priority as other workloads. When nodes are resource-constrained, the network gateway Pods might be evicted. This behavior is particularly bad for the ang-node DaemonSet, as those Pods must be scheduled on a specific node and can't migrate.


Workaround:

Upgrade to 1.15 or later.

As a short-term fix, you can manually assign a PriorityClass to the Network Gateway for GDC components. The GKE on Bare Metal controller overwrites these manual changes during a reconciliation process, such as during a cluster upgrade.

  • Assign the system-cluster-critical PriorityClass to the ang-controller-manager and autoscaler cluster controller Deployments.
  • Assign the system-node-critical PriorityClass to the ang-daemon node DaemonSet.
Installation, Upgrades and updates 1.15.0, 1.15.1, 1.15.2

Cluster creation and upgrades fail due to cluster name length

Creating version 1.15.0, 1.15.1, or 1.15.2 clusters or upgrading clusters to version 1.15.0, 1.15.1, or 1.15.2 fails when the cluster name is longer than 48 characters (version 1.15.0) or 45 characters (version 1.15.1 or 1.15.2). During cluster creation and upgrade operations, GKE on Bare Metal creates a health check resource with a name that incorporates the cluster name and version:

  • For version 1.15.0 clusters, the health check resource name is CLUSTER_NAME-add-ons-CLUSTER_VER.
  • For version 1.15.1 or 1.15.2 clusters, the health check resource name is CLUSTER_NAME-kubernetes-CLUSTER_VER.

For long cluster names, the health check resource name exceeds the Kubernetes 63 character length restriction for label names, which prevents the creation of the health check resource. Without a successful health check, the cluster operation fails.

To see if you are affected by this issue, use kubectl describe to check the failing resource:

kubectl describe healthchecks.baremetal.cluster.gke.io \
    HEALTHCHECK_CR_NAME -n CLUSTER_NAMESPACE \
    --kubeconfig ADMIN_KUBECONFIG

If this issue is affecting you, the response contains a warning for a ReconcileError like the following:

...
Events:
  Type     Reason          Age                   From                    Message
  ----     ------          ----                  ----                    -------
  Warning  ReconcileError  77s (x15 over 2m39s)  healthcheck-controller  Reconcile error, retrying: 1 error occurred:
           * failed to create job for health check
db-uat-mfd7-fic-hybrid-cloud-uk-wdc-cluster-02-kubernetes-1.15.1: Job.batch
"bm-system-db-uat-mfd7-fic-hybrid-cloud-u24d5f180362cffa4a743" is invalid: [metadata.labels: Invalid
value: "db-uat-mfd7-fic-hybrid-cloud-uk-wdc-cluster-02-kubernetes-1.15.1": must be no more than 63
characters, spec.template.labels: Invalid value:
"db-uat-mfd7-fic-hybrid-cloud-uk-wdc-cluster-02-kubernetes-1.15.1": must be no more than 63 characters]

Workaround

To unblock the cluster upgrade or creation, you can bypass the healthcheck. Use the following command to patch the healthcheck custom resource with passing status: (status: {pass: true})

kubectl patch healthchecks.baremetal.cluster.gke.io \
    HEALTHCHECK_CR_NAME -n CLUSTER_NAMESPACE \
    --kubeconfig ADMIN_KUBECONFIG --type=merge \
    --subresource status --patch 'status: {pass: true}'
Upgrades and Updates 1.14, 1.15

Version 1.14.0 and 1.14.1 clusters with preview features can't upgrade to version 1.15.x

If version 1.14.0 and 1.14.1 clusters have a preview feature enabled, they're blocked from successfully upgrading to version 1.15.x. This applies to preview features like the ability to create a cluster without kube-proxy, which is enabled with the following annotation in the cluster configuration file:

preview.baremetal.cluster.gke.io/kube-proxy-free: "enable"

If you're affected by this issue, you get an error like the following during the cluster upgrade:

[2023-06-20 23:37:47+0000] error judging if the cluster is managing itself:
error to parse the target cluster: error parsing cluster config: 1 error
occurred:

Cluster.baremetal.cluster.gke.io "$cluster-name" is invalid:
Annotations[preview.baremetal.cluster.gke.io/$preview-feature-name]:
Forbidden: preview.baremetal.cluster.gke.io/$preview-feature-name feature
isn't supported in 1.15.1 Anthos Bare Metal version

This issue is fixed in version 1.14.2 and higher clusters.


Workaround:

If you're unable to upgrade your clusters to version 1.14.2 or higher before upgrading to version 1.15.x, you can upgrade to version 1.15.x directly by using a bootstrap cluster:

bmctl upgrade cluster --use-bootstrap=true
Operation 1.15

Version 1.15 clusters don't accept duplicate floating IP addresses

Network Gateway for GDC doesn't let you create new NetworkGatewayGroup custom resources that contain IP addresses in spec.floatingIPs that are already used in existing NetworkGatewayGroup custom resources. This rule is enforced by a webhook in GKE on Bare Metal clusters version 1.15.0 and higher. Pre-existing duplicate floating IP addresses don't cause errors. The webhook only prevents the creation of new NetworkGatewayGroups custom resources that contain duplicate IP addresses.

The webhook error message identifies the conflicting IP address and the existing custom resource that is already using it:

IP address exists in other gateway with name default

The initial documentation for advanced networking features, such as the Egress NAT gateway, doesn't caution against duplicate IP addresses. Initially, only the NetworkGatewayGroup resource named default was recognized by the reconciler. Network Gateway for GDC now recognizes all NetworkGatewayGroup custom resources in the system namespace. Existing NetworkGatewayGroup custom resources are honored, as is.


Workaround:

Errors happen for the creation of a new NetworkGatewayGroup custom resource only.

To address the error:

  1. Use the following command to list NetworkGatewayGroups custom resources:
    kubectl get NetworkGatewayGroups --kubeconfig ADMIN_KUBECONFIG \
        -n kube-system -o yaml
    
  2. Open existing NetworkGatewayGroup custom resources and remove any conflicting floating IP addresses (spec.floatingIPs):
    kubectl edit NetworkGatewayGroups --kubeconfig ADMIN_KUBECONFIG \
        -n kube-system RESOURCE_NAME
    
  3. To apply your changes, close and save edited custom resources.
VM Runtime on GDC 1.13.7

VMs might not start on 1.13.7 clusters that use a private registry

When you enable VM Runtime on GDC on a new or upgraded version 1.13.7 cluster that uses a private registry, VMs that connect to the node network or use a GPU might not start properly. This issue is due to some system Pods in the vm-system namespace getting image pull errors. For example, if your VM uses the node network, some Pods might report image pull errors like the following:

macvtap-4x9zp              0/1   Init:ImagePullBackOff  0     70m

This issue is fixed in version 1.14.0 and higher GKE on Bare Metal clusters.

Workaround

If you're unable to upgrade your clusters immediately, you can pull images manually. The following commands pull the macvtap CNI plugin image for your VM and push it to your private registry:

docker pull \
    gcr.io/anthos-baremetal-release/kubevirt/macvtap-cni:v0.5.1-gke.:21
docker tag \
    gcr.io/anthos-baremetal-release/kubevirt/macvtap-cni:v0.5.1-gke.:21 \
    REG_HOST/anthos-baremetal-release/kubevirt/macvtap-cni:v0.5.1-gke.:21
docker push \
    REG_HOST/anthos-baremetal-release/kubevirt/macvtap-cni:v0.5.1-gke.:21

Replace REG_HOST with the domain name of a host that you mirror locally.

Installation 1.11, 1.12

During cluster creation in the kind cluster, the gke-metric-agent pod fails to start

During cluster creation in the kind cluster, the gke-metrics-agent pod fails to start because of an image pulling error as follows:

error="failed to pull and unpack image \"gcr.io/gke-on-prem-staging/gke-metrics-agent:1.8.3-anthos.2\": failed to resolve reference \"gcr.io/gke-on-prem-staging/gke-metrics-agent:1.8.3-anthos.2\": pulling from host gcr.io failed with status code [manifests 1.8.3-anthos.2]: 403 Forbidden"

Also, in the bootstrap cluster's containerd log, you will see the following entry:

Sep 13 23:54:20 bmctl-control-plane containerd[198]: time="2022-09-13T23:54:20.378172743Z" level=info msg="PullImage \"gcr.io/gke-on-prem-staging/gke-metrics-agent:1.8.3-anthos.2\" " Sep 13 23:54:21 bmctl-control-plane containerd[198]: time="2022-09-13T23:54:21.057247258Z" level=error msg="PullImage \"gcr.io/gke-on-prem-staging/gke-metrics-agent:1.8.3-anthos.2\" failed" error="failed to pull and unpack image \"gcr.io/gke-on-prem-staging/gke-metrics-agent:1.8.3-anthos.2\": failed to resolve reference \"gcr.io/gke-on-prem-staging/gke-metrics-agent:1.8.3-anthos.2\": pulling from host gcr.io failed with status code [manifests 1.8.3-anthos.2]: 403 Forbidden"

You will see the following "failing to pull" error in the pod:

gcr.io/gke-on-prem-staging/gke-metrics-agent

Workaround

Despite the errors, the cluster creation process isn't blocked as the purpose of gke-metrics-agent pod in kind cluster is to facilitate the cluster creation success rate and for internal tracking and monitoring. Hence, you can ignore this error.

Workaround

Despite the errors, the cluster creation process isn't blocked as the purpose of gke-metrics-agent pod in kind cluster is to facilitate the cluster creation success rate and for internal tracking and monitoring. Hence, you can ignore this error.

Operation, Networking 1.12, 1.13, 1.14, 1.15, 1.16, 1.28

Accessing an IPv6 Service endpoint crashes the LoadBalancer Node on CentOS or RHEL

When you access a dual-stack Service (a Service that has both IPv4 and IPv6 endpoints) and use the IPv6 endpoint, the LoadBalancer Node that serves the Service might crash. This issue affects customers that use dual-stack services with CentOS or RHEL and kernel version earlier than kernel-4.18.0-372.46.1.el8_6.

If you believe that this issue affects you, check the kernel version on the LoadBalancer Node using the uname -a command.


Workaround:

Update the LoadBalancer Node to kernel version kernel-4.18.0-372.46.1.el8_6 or later. This kernel version is available by default in CentOS and RHEL version 8.6 and later.

Networking 1.11, 1.12, 1.13, 1.14.0

Intermittent connectivity issues after Node reboot

After you restart a Node, you might see intermittent connectivity issues for a NodePort or LoadBalancer Service. For example, you might have intermittent TLS handshake or connection reset errors. This issue is fixed for cluster versions 1.14.1 and higher.

To check if this issue affects you, look at the iptables forward rules on Nodes where the backend Pod for the affected Service is running:

sudo iptables -L FORWARD

If you see the KUBE-FORWARD rule before the CILIUM_FORWARD rule in iptables, you might be affected by this issue. The following example output shows a Node where the problem exists:

Chain FORWARD (policy ACCEPT)
target                  prot opt source   destination
KUBE-FORWARD            all  --  anywhere anywhere                /* kubernetes forwarding rules */
KUBE-SERVICES           all  --  anywhere anywhere    ctstate NEW /* kubernetes service portals */
KUBE-EXTERNAL-SERVICES  all  --  anywhere anywhere    ctstate NEW /* kubernetes externally-visible service portals */
CILIUM_FORWARD          all  --  anywhere anywhere                /* cilium-feeder: CILIUM_FORWARD */

Workaround:

Restart the anetd Pod on the Node that's misconfigured. After you restart the anetd Pod, the forwarding rule in iptables should be configured correctly.

The following example output shows that the CILIUM_FORWARD rule is now correctly configured before the KUBE-FORWARD rule:

Chain FORWARD (policy ACCEPT)
target                  prot opt source   destination
CILIUM_FORWARD          all  --  anywhere anywhere                /* cilium-feeder: CILIUM_FORWARD */
KUBE-FORWARD            all  --  anywhere anywhere                /* kubernetes forwarding rules */
KUBE-SERVICES           all  --  anywhere anywhere    ctstate NEW /* kubernetes service portals */
KUBE-EXTERNAL-SERVICES  all  --  anywhere anywhere    ctstate NEW /* kubernetes externally-visible service portals */
Upgrades and updates 1.9, 1.10

The preview feature does not retain the original permission and owner information

The preview feature of 1.9.x cluster using bmctl 1.9.x does not retain the original permission and owner information. To verify if you are affected by this feature, extract the backed-up file using the following command:

tar -xzvf BACKUP_FILE

Workaround

Verify if the metadata.json is present and if the bmctlVersion is 1.9.x. If the metadata.json isn't present, upgrade to 1.10.x cluster and use bmctl 1.10.x to backup/restore.

Upgrades and creates 1.14.2

clientconfig-operator stuck in pending state with CreateContainerConfigError

If you've upgraded to or created a version 1.14.2 cluster with an OIDC/LDAP configuration, you may see the clientconfig-operator Pod stuck in a pending state. With this issue, there are two clientconfig-operator Pods, with one in a running state and the other in a pending state.

This issue applies to version 1.14.2 clusters only. Earlier cluster versions such as 1.14.0 and 1.14.1 aren't affected. This issue is fixed in version 1.14.3 and all subsequent releases, including 1.15.0 and later.


Workaround:

As a workaround, you can patch the clientconfig-operator deployment to add additional security context and ensure that the deployment is ready.

Use the following command to patch clientconfig-operator in the target cluster:

kubectl patch deployment clientconfig-operator -n kube-system \
    -p '{"spec":{"template":{"spec":{"containers": [{"name":"oidcproxy","securityContext":{"runAsGroup":2038,"runAsUser":2038}}]}}}}' \
    --kubeconfig CLUSTER_KUBECONFIG

Replace the following:

  • CLUSTER_KUBECONFIG: the path of the kubeconfig file for the target cluster.
Operation 1.11, 1.12, 1.13, 1.14, 1.15

Certificate authority rotation fails for clusters without bundled load balancing

For clusters without bundled load balancing (spec.loadBalancer.mode set to manual), the bmctl update credentials certificate-authorities rotate command can become unresponsive and fail with the following error: x509: certificate signed by unknown authority.

If you're affected by this issue, the bmctl command might output the following message before becoming unresponsive:

Signing CA completed in 3/0 control-plane nodes

In this case, the command eventually fails. The rotate certificate-authority log for a cluster with three control planes may include entries like the following:

[2023-06-14 22:33:17+0000] waiting for all nodes to trust CA bundle OK
[2023-06-14 22:41:27+0000] waiting for first round of pod restart to complete OK
Signing CA completed in 0/0 control-plane nodes
Signing CA completed in 1/0 control-plane nodes
Signing CA completed in 2/0 control-plane nodes
Signing CA completed in 3/0 control-plane nodes
...
Unable to connect to the server: x509: certificate signed by unknown
authority (possibly because of "crypto/rsa: verification error" while
trying to verify candidate authority certificate "kubernetes")

Workaround

If you need additional assistance, contact Google Support.

Installation, Networking 1.11, 1.12, 1.13, 1.14.0-1.14.1

ipam-controller-manager crashloops in dual-stack clusters

When you deploy a dual-stack cluster (a cluster with both IPv4 and IPv6 addresses), the ipam-controller-manager Pod(s) might crashloop. This behavior causes the Nodes to cycle between Ready and NotReady states, and might cause the cluster installation to fail. This problem can occur when the API server is under high load.

To see if this issue affects you, check if the ipam-controller-manager Pod(s) are failing with CrashLoopBackOff errors:

kubectl -n kube-system get pods | grep  ipam-controller-manager

The following example output shows Pods in a CrashLoopBackOff state:

ipam-controller-manager-h7xb8   0/1  CrashLoopBackOff   3 (19s ago)   2m ipam-controller-manager-vzrrf   0/1  CrashLoopBackOff   3 (19s ago)   2m1s
ipam-controller-manager-z8bdw   0/1  CrashLoopBackOff   3 (31s ago)   2m2s

Get details for the Node that's in a NotReady state:

kubectl describe node <node-name> | grep PodCIDRs

In a cluster with this issue, a Node has no PodCIDRs assigned to it, as shown in the following example output:

PodCIDRs:

In a healthy cluster, all the Nodes should have dual-stack PodCIDRs assigned to it, as shown in the following example output:

PodCIDRs:    192.168.6.0/24,222:333:444:555:5:4:7:0/120

Workaround:

Restart the ipam-controller-manager Pod(s):

kubectl -n kube-system rollout restart ds ipam-controller-manager
Operation 1.6, 1.7, 1.8, 1.9, 1.10, 1.11, 1.12, 1.13, and 1.14

etcd watch starvation

Clusters running etcd version 3.4.13 or earlier may experience watch starvation and non-operational resource watches, which can lead to the following problems:

  • Pod scheduling is disrupted
  • Nodes are unable to register
  • kubelet doesn't observe pod changes

These problems can make the cluster non-functional.

This issue is fixed in GKE on Bare Metal version 1.12.9, 1.13.6, 1.14.3, and subsequent releases. These newer releases use etcd version 3.4.21. All prior versions of GKE on Bare Metal are affected by this issue.

Workaround

If you can't upgrade immediately, you can mitigate the risk of cluster failure by reducing the number of nodes in your cluster. Remove nodes until the etcd_network_client_grpc_sent_bytes_total metric is less than 300 MBps.

To view this metric in Metrics Explorer:

  1. Go to the Metrics Explorer in the Google Cloud console:

    Go to Metrics Explorer

  2. Select the Configuration tab.
  3. Expand the Select a metric, enter Kubernetes Container in the filter bar, and then use the submenus to select the metric:
    1. In the Active resources menu, select Kubernetes Container.
    2. In the Active metric categories menu, select Anthos.
    3. In the Active metrics menu, select etcd_network_client_grpc_sent_bytes_total.
    4. Click Apply.
Networking 1.11.6, 1.12.3

SR-IOV operator's vfio-pci mode "Failed" state

The SriovNetworkNodeState object's syncStatus can report the "Failed" value for a configured node. To view the status of a node and determine if the problem affects you, run the following command:

kubectl -n gke-operators get \
    sriovnetworknodestates.sriovnetwork.k8s.cni.cncf.io NODE_NAME \
    -o jsonpath='{.status.syncStatus}'

Replace NODE_NAME with the name of the node to check.


Workaround:

If the SriovNetworkNodeState object status is "Failed", upgrade your cluster to version 1.11.7 or later or version 1.12.4 or later.

Upgrades and updates 1.10, 1.11, 1.12, 1.13, 1.14.0, 1.14.1

Some worker nodes aren't in a Ready state after upgrade

Once upgrade is finished, some worker nodes may have their Ready condition set to false. On the Node resource, you will see an error next to the Ready condition similar to the following example:

container runtime network not ready: NetworkReady=false reason:NetworkPluginNotReady message:Network plugin returns error: cni plugin not initialized

When you log into the stalled machine, the CNI configuration on the machine is empty:

sudo ls /etc/cni/net.d/

Workaround

Restart the node's anetd pod by deleting it.

Upgrades and updates, Security 1.10

Multiple certificate rotations from cert-manager result in inconsistency

After multiple manual or auto certificate rotations, the webhook pod, such as anthos-cluster-operator isn't updated with the new certificates issued by cert-manager. Any update to the cluster custom resource fails and results in an error similar as follows:

Internal error occurred: failed calling
webhook "vcluster.kb.io": failed to call webhook: Post "https://webhook-service.kube-system.svc:443/validate-baremetal-cluster-gke-io-v1-cluster?timeout=10s": x509: certificate signed by unknown authority (possibly because of "x509:
invalid signature: parent certificate cannot sign this kind of certificate"
while trying to verify candidate authority certificate
"webhook-service.kube-system.svc")

This issue might occur in the following circumstances:

  • If you have done two manual cert-manager issued certificate rotations on a cluster older than 180 days or more and never restarted the anthos-cluster-operator.
  • If you have done a manual cert-manager issued certificate rotations on a cluster older than 90 days or more and never restarted the anthos-cluster-operator.

Workaround

Restart the pod by terminating the anthos-cluster-operator.

Upgrades and updates 1.14.0

Outdated lifecycle controller deployer pods created during user cluster upgrade

In version 1.14.0 admin clusters, one or more outdated lifecycle controller deployer pods might be created during user cluster upgrades. This issue applies for user clusters that were initially created at versions lower than 1.12. The unintentionally created pods don't impede upgrade operations, but they might be found in an unexpected state. We recommend that you remove the outdated pods.

This issue is fixed in release 1.14.1.

Workaround:

To remove the outdated lifecycle controller deployer pods:

  1. List preflight check resources:
    kubectl get preflightchecks --kubeconfig ADMIN_KUBECONFIG -A
    

    The output looks like this:

    NAMESPACE                    NAME                      PASS   AGE
    cluster-ci-87a021b9dcbb31c   ci-87a021b9dcbb31c        true   20d
    cluster-ci-87a021b9dcbb31c   ci-87a021b9dcbb31cd6jv6   false  20d
    

    where ci-87a021b9dcbb31c is the cluster name.

  2. Delete resources whose value in the PASS column is either true or false.

    For example, to delete the resources in the preceding sample output, use the following commands:

    kubectl delete preflightchecks ci-87a021b9dcbb31c \
        -n cluster-ci-87a021b9dcbb31c \
        --kubeconfig ADMIN_KUBECONFIG 
    kubectl delete preflightchecks ci-87a021b9dcbb31cd6jv6 \
        -n cluster-ci-87a021b9dcbb31c \
        --kubeconfig ADMIN_KUBECONFIG
    
Networking 1.9, 1.10, 1.11, 1.12, 1.13, 1.14, 1.15, 1.16, 1.28

BGPSession state constantly changing due to large number of incoming routes

GKE on Bare Metal advanced networking fails to manage BGP sessions correctly when external peers advertise a high number of routes (about 100 or more). With a large number of incoming routes, the node-local BGP controller takes too long to reconcile BGP sessions and fails to update the status. The lack of status updates, or a health check, causes the session to be deleted for being stale.

Undesirable behavior on BGP sessions that you might notice and indicate a problem include the following:

  • Continuous bgpsession deletion and recreation.
  • bgpsession.status.state never becomes Established
  • Routes failing to advertise or being repeatedly advertised and withdrawn.

BGP load balancing problems might be noticeable with connectivity issues to LoadBalancer services.

BGP FlatIP issue might be noticeable with connectivity issues to Pods.

To determine if your BGP issues are caused by the remote peers advertising too many routes, use the following commands to review the associated statuses and output:

  • Use kubectl get bgpsessions on the affected cluster. The output shows bgpsessions with state "Not Established" and the last report time continuously counts up to about 10-12 seconds before it appears to reset to zero.
  • The output of kubectl get bgpsessions shows that the affected sessions are being repeatedly recreated:
    kubectl get bgpsessions \
        -o jsonpath="{.items[*]['metadata.name', 'metadata.creationTimestamp']}"
    
  • Log messages indicate that stale BGP sessions are being deleted:
    kubectl logs ang-controller-manager-POD_NUMBER
    

    Replace POD_NUMBER with the leader pod in your cluster.


Workaround:

Reduce or eliminate the number of routes advertised from the remote peer to the cluster with an export policy.

In GKE on Bare Metal cluster version 1.14.2 and later, you can also disable the feature that processes received routes by using an AddOnConfiguration. Add the --disable-received-routes argument to the ang-daemon daemonset's bgpd container.

Networking 1.14, 1.15, 1.16, 1.28

Application timeouts caused by conntrack table insertion failures

Clusters running on an Ubuntu OS that uses kernel 5.15 or higher are susceptible to netfilter connection tracking (conntrack) table insertion failures. Insertion failures can occur even when the conntrack table has room for new entries. The failures are caused by changes in kernel 5.15 and higher that restrict table insertions based on chain length.

To see if you are affected by this issue, you can check the in-kernel connection tracking system statistics with the following command:

sudo conntrack -S

The response looks like this:

cpu=0       found=0 invalid=4 insert=0 insert_failed=0 drop=0 early_drop=0 error=0 search_restart=0 clash_resolve=0 chaintoolong=0
cpu=1       found=0 invalid=0 insert=0 insert_failed=0 drop=0 early_drop=0 error=0 search_restart=0 clash_resolve=0 chaintoolong=0
cpu=2       found=0 invalid=16 insert=0 insert_failed=0 drop=0 early_drop=0 error=0 search_restart=0 clash_resolve=0 chaintoolong=0
cpu=3       found=0 invalid=13 insert=0 insert_failed=0 drop=0 early_drop=0 error=0 search_restart=0 clash_resolve=0 chaintoolong=0
cpu=4       found=0 invalid=9 insert=0 insert_failed=0 drop=0 early_drop=0 error=0 search_restart=0 clash_resolve=0 chaintoolong=0
cpu=5       found=0 invalid=1 insert=0 insert_failed=0 drop=0 early_drop=0 error=519 search_restart=0 clash_resolve=126 chaintoolong=0
...

If a chaintoolong value in the response is a non-zero number, you're affected by this issue.

Workaround

The short term mitigation is to increase the size of both the netfiler hash table (nf_conntrack_buckets) and the netfilter connection tracking table (nf_conntrack_max). Use the following commands on each cluster node to increase the size of the tables:

sysctl -w net.netfilter.nf_conntrack_buckets=TABLE_SIZE

sysctl -w net.netfilter.nf_conntrack_max=TABLE_SIZE

Replace TABLE_SIZE with new size in bytes. The default table size value is 262144. We suggest that you set a value equal to 65,536 times the number of cores on the node. For example, if your node has eight cores, set the table size to 524288.

Upgrades and updates 1.11.3, 1.11.4, 1.11.5, 1.11.6, 1.11.7, 1.11.8, 1.12.4, 1.12.5, 1.12.6, 1.12.7, 1.12.8, 1.13.4, 1.13.5

Can't restore cluster backups with bmctl for some versions

We recommend that you back up your clusters before you upgrade so that you can restore the earlier version if the upgrade doesn't succeed. A problem with the bmctl restore cluster command causes it to fail to restore backups of clusters with the identified versions. This issue is specific to upgrades, where you're restoring a backup of an earlier version.

If your cluster is affected, the bmctl restore cluster log contains the following error:

Error: failed to extract image paths from profile: anthos version VERSION not supported

Workaround:

Until this issue is fixed, we recommend that you use the instructions in Back up and restore clusters to back up your clusters manually and restore them manually, if necessary.
Networking 1.10, 1.11, 1.12, 1.13, 1.14.0-1.14.2

NetworkGatewayGroup crashes if there's no IP address on the interface

NetworkGatewayGroup fails to create daemons for nodes that don't have both IPv4 and IPv6 interfaces on them. This causes features like BGP LB and EgressNAT to fail. If you check the logs of the failing ang-node Pod in the kube-system namespace, errors similar to the following example are displayed when an IPv6 address is missing:

ANGd.Setup    Failed to create ANG daemon    {"nodeName": "bm-node-1", "error":
"creating NDP client failed: ndp: address \"linklocal\" not found on interface \"ens192\""}

In the previous example, there's no IPv6 address on the ens192 interface. Similar ARP errors are displayed if the node is missing an IPv4 address.

NetworkGatewayGroup tries to establish an ARP connection and an NDP connection to the link local IP address. If the IP address doesn't exist (IPv4 for ARP, IPv6 for NDP) then the connection fails and the daemon doesn't continue.

This issue is fixed in release 1.14.3.


Workaround:

Connect to the node using SSH and add an IPv4 or IPv6 address to the link that contains the node IP. In the previous example log entry, this interface was ens192:

ip address add dev INTERFACE scope link ADDRESS

Replace the following:

  • INTERFACE: The interface for your node, such as ens192.
  • ADDRESS: The IP address and subnet mask to apply to the interface.
Reset/Deletion 1.10, 1.11, 1.12, 1.13.0-1.13.2

anthos-cluster-operator crash loop when removing a control plane node

When you try to remove a control plane node by removing the IP address from the Cluster.Spec, the anthos-cluster-operator enters into a crash loop state that blocks any other operations.


Workaround:

Issue is fixed in 1.13.3 and 1.14.0 and later. All other versions are affected. Upgrade to one of the fixed versions

As a workaround, run the following command:

kubectl label baremetalmachine IP_ADDRESS \
    -n CLUSTER_NAMESPACE baremetal.cluster.gke.io/upgrade-apply-

Replace the following:

  • IP_ADDRESS: The IP address of the node in a crash loop state.
  • CLUSTER_NAMESPACE: The cluster namespace.
Installation 1.13.1, 1.13.2 and 1.13.3

kubeadm join fails in large clusters due to token mismatch

When you install GKE on Bare Metal clusters with a large number of nodes, you might see a kubeadmin join error message similar to the following example:

TASK [kubeadm : kubeadm join --config /dev/stdin --ignore-preflight-errors=all] ***
fatal: [10.200.0.138]: FAILED! => {"changed": true, "cmd": "kubeadm join
--config /dev/stdin --ignore-preflight-errors=all", "delta": "0:05:00.140669", "end": "2022-11-01 21:53:15.195648", "msg": "non-zero return code", "rc": 1,
"start": "2022-11-01 21:48:15.054979", "stderr": "W1101 21:48:15.082440   99570 initconfiguration.go:119]
Usage of CRI endpoints without URL scheme is deprecated and can cause kubelet errors in the future.
Automatically prepending scheme \"unix\" to the \"criSocket\" with value \"/run/containerd/containerd.sock\". Please update your configuration!\nerror
execution phase preflight: couldn't validate the identity of the API Server: could not find a JWS signature in the cluster-info ConfigMap for token ID \"yjcik0\"\n
To see the stack trace of this error execute with --v=5 or higher", "stderr_lines":
["W1101 21:48:15.082440   99570 initconfiguration.go:119] Usage of CRI endpoints without URL scheme is deprecated and can cause kubelet errors in the future.
Automatically prepending scheme \"unix\" to the \"criSocket\" with value \"/run/containerd/containerd.sock\".
Please update your configuration!", "error execution phase preflight: couldn't validate the identity of the API Server:
could not find a JWS signature in the cluster-info ConfigMap for token ID \"yjcik0\"",
"To see the stack trace of this error execute with --v=5 or higher"], "stdout": "[preflight]
Running pre-flight checks", "stdout_lines": ["[preflight] Running pre-flight checks"]}

Workaround:

This issue is resolved in GKE on Bare Metal version 1.13.4 and later.

If you need to use an affected version, first create a cluster with less than 20 nodes, and then resize the cluster to add additional nodes after the install is complete.

Logging and monitoring 1.10, 1.11, 1.12, 1.13.0

Low CPU limit for metrics-server in Edge clusters

In GKE on Bare Metal Edge clusters, low CPU limits for metrics-server can cause frequent restarts of metrics-server. Horizontal Pod Autoscaling (HPA) doesn't work due to metrics-server being unhealthy.

If metrics-server CPU limit is less than 40m, your clusters can be affected. To check the metrics-server CPU limits, review one of the following files:

  • GKE on Bare Metal clusters version 1.x-1.12:
    kubectl get deployment metrics-server -n kube-system \
        -o yaml > metrics-server.yaml
    
  • GKE on Bare Metal clusters version 1.13 or later:
    kubectl get deployment metrics-server -n gke-managed-metrics-server \
        -o yaml > metrics-server.yaml
    

Workaround:

This issue is resolved in GKE on Bare Metal clusters version 1.13.1 or later. To fix this issue, upgrade your clusters.

A short-term workaround until you can upgrade clusters is to manually increase the CPU limits for metrics-server as follows:

  1. Scale down metrics-server-operator:
    kubectl scale deploy metrics-server-operator --replicas=0
  2. Update the configuration and increase CPU limits:
    • GKE on Bare Metal clusters version 1.x-1.12:
      kubectl -n kube-system edit deployment metrics-server
    • GKE on Bare Metal clusters version 1.13:
      kubectl -n gke-managed-metrics-server edit deployment metrics-server

    Remove the --config-dir=/etc/config line and increase the CPU limits, as shown in the following example:

    [...]
    - command:
    - /pod_nanny
    # - --config-dir=/etc/config # <--- Remove this line
    - --container=metrics-server
    - --cpu=50m # <--- Increase CPU, such as to 50m
    - --extra-cpu=0.5m
    - --memory=35Mi
    - --extra-memory=4Mi
    - --threshold=5
    - --deployment=metrics-server
    - --poll-period=30000
    - --estimator=exponential
    - --scale-down-delay=24h
    - --minClusterSize=5
    - --use-metrics=true
    [...]
    
  3. Save and close the metrics-server to apply the changes.
Networking 1.14, 1.15, 1.16

Direct NodePort connection to hostNetwork Pod doesn't work

Connection to a Pod enabled with hostNetwork using NodePort Service fails when the backend Pod is on the same node as the targeted NodePort. This issues affects LoadBalancer Services when used with hostNetwork-ed Pods. With multiple backends, there can be a sporadic connection failure.

This issue is caused by a bug in the eBPF program.


Workaround:

When using a Nodeport Service, don't target the node on which any of the backend Pod runs. When using the LoadBalancer Service, make sure the hostNetwork-ed Pods don't run on LoadBalancer nodes.

Upgrades and updates 1.12.3, 1.13.0

1.13.0 admin clusters can't manage 1.12.3 user clusters

Admin clusters that run version 1.13.0 can't manage user clusters that run version 1.12.3. Operations against a version 1.12.3 user cluster fail.


Workaround:

Upgrade your admin cluster to version 1.13.1, or upgrade the user cluster to the same version as the admin cluster.

Upgrades and updates 1.12

Upgrading to 1.13.x is blocked for admin clusters with worker node pools

Version 1.13.0 and higher admin clusters can't contain worker node pools. Upgrades to version 1.13.0 or higher for admin clusters with worker node pools is blocked. If your admin cluster upgrade is stalled, you can confirm if worker node pools are the cause by checking following error in the upgrade-cluster.log file inside the bmctl-workspace folder:

Operation failed, retrying with backoff. Cause: error creating "baremetal.cluster.gke.io/v1, Kind=NodePool" cluster-test-cluster-2023-06-06-140654/np1: admission webhook "vnodepool.kb.io" denied the request: Adding worker nodepool to Admin cluster is disallowed.

Workaround:

Before upgrading, move all worker node pools to user clusters. For instructions to add and remove node pools, see Manage node pools in a cluster.

Upgrades and updates 1.10, 1.11, 1.12, 1.13, 1.14, 1.15, 1.16, 1.28

Errors when updating resources using kubectl apply

If you update existing resources like the ClientConfig or Stackdriver custom resources using kubectl apply, the controller might return an error or revert your input and planned changes.

For example, you might try to edit the Stackdriver custom resource as follows by first getting the resource, and then applying an updated version:

  1. Get the existing YAML definition:
    kubectl get stackdriver -n kube-system stackdriver \
        -o yaml > stackdriver.yaml
    
  2. Enable features or update configuration in the YAML file.
  3. Apply the updated YAML file back:
    kubectl apply -f stackdriver.yaml
    

The final step for kubectl apply is where you might run into problems.


Workaround:

Don't use kubectl apply to make changes to existing resources. Instead, use kubectl edit or kubectl patch as shown in the following examples:

  1. Edit the Stackdriver custom resource:
    kubectl edit stackdriver -n kube-system stackdriver
    
  2. Enable features or update configuration in the YAML file.
  3. Save and exit the editor

Alternate approach using kubectl patch:

  1. Get the existing YAML definition:
    kubectl get stackdriver -n kube-system stackdriver \
        -o yaml > stackdriver.yaml
    
  2. Enable features or update configuration in the YAML file.
  3. Apply the updated YAML file back:
    kubectl patch stackdriver stackdriver --type merge \
        -n kube-system --patch-file stackdriver.yaml
    
Logging and monitoring 1.12, 1.13, 1.14, 1.15, 1.16

Corrupted backlog chunks cause stackdriver-log-forwarder crashloop

The stackdriver-log-forwarder crashloops if it tries to process a corrupted backlog chunk. The following example errors are shown in the container logs:

[2022/09/16 02:05:01] [error] [storage] format check failed: tail.1/1-1659339894.252926599.flb
[2022/09/16 02:05:01] [error] [engine] could not segregate backlog chunks

When this crashloop occurs, you can't see logs in Cloud Logging.


Workaround:

To resolve these errors, complete the following steps:

  1. Identify the corrupted backlog chunks. Review the following example error messages:
    [2022/09/16 02:05:01] [error] [storage] format check failed: tail.1/1-1659339894.252926599.flb
    [2022/09/16 02:05:01] [error] [engine] could not segregate backlog chunks
    
    In this example, the file tail.1/1-1659339894.252926599.flb that's stored in var/log/fluent-bit-buffers/tail.1/ is at fault. Every *.flb file with a format check failed must be removed.
  2. End the running pods for stackdriver-log-forwarder:
    kubectl --kubeconfig KUBECONFIG -n kube-system \
        patch daemonset stackdriver-log-forwarder \
        -p '{"spec": {"template": {"spec": {"nodeSelector": {"non-existing": "true"}}}}}'
    
    Replace KUBECONFIG with the path to your user cluster kubeconfig file.

    Verify that the stackdriver-log-forwarder Pods are deleted before going to the next step.
  3. Connect to the node using SSH where stackdriver-log-forwarder is running.
  4. On the node, delete all corrupted *.flb files in var/log/fluent-bit-buffers/tail.1/.

    If there are too many corrupted files and you want to apply a script to clean up all backlog chunks, use the following scripts:
    1. Deploy a DaemonSet to clean up all the dirty data in buffers in fluent-bit:
      kubectl --kubeconfig KUBECONFIG -n kube-system apply -f - << EOF
      apiVersion: apps/v1
      kind: DaemonSet
      metadata:
        name: fluent-bit-cleanup
        namespace: kube-system
      spec:
        selector:
          matchLabels:
            app: fluent-bit-cleanup
        template:
          metadata:
            labels:
              app: fluent-bit-cleanup
          spec:
            containers:
            - name: fluent-bit-cleanup
              image: debian:10-slim
              command: ["bash", "-c"]
              args:
              - |
                rm -rf /var/log/fluent-bit-buffers/
                echo "Fluent Bit local buffer is cleaned up."
                sleep 3600
              volumeMounts:
              - name: varlog
                mountPath: /var/log
              securityContext:
                privileged: true
            tolerations:
            - key: "CriticalAddonsOnly"
              operator: "Exists"
            - key: node-role.kubernetes.io/master
              effect: NoSchedule
            - key: node-role.gke.io/observability
              effect: NoSchedule
            volumes:
            - name: varlog
              hostPath:
                path: /var/log
      EOF
      
    2. Make sure that the DaemonSet has cleaned up all the nodes. The output of the following two commands should be equal to the number of nodes in the cluster:
      kubectl --kubeconfig KUBECONFIG logs \
          -n kube-system -l app=fluent-bit-cleanup | grep "cleaned up" | wc -l
      kubectl --kubeconfig KUBECONFIG \
          -n kube-system get pods -l app=fluent-bit-cleanup --no-headers | wc -l
      
    3. Delete the cleanup DaemonSet:
      kubectl --kubeconfig KUBECONFIG -n kube-system delete ds \
          fluent-bit-cleanup
      
    4. Restart the stackdriver-log-forwarder Pods:
      kubectl --kubeconfig KUBECONFIG \
          -n kube-system patch daemonset stackdriver-log-forwarder --type json \
          -p='[{"op": "remove", "path": "/spec/template/spec/nodeSelector/non-existing"}]'
      
Networking, VM Runtime on GDC 1.14.0

Restarting Dataplane V2 (anetd) on clusters can result in existing VMs unable to attach to non-pod-network

On multi-nic clusters, restarting Dataplane V2 (anetd) can result in virtual machines being unable to attach to networks. An error similar to the following might be observed in the anetd pod logs:

could not find an allocator to allocate the IP of the multi-nic endpoint

Workaround:

You can restart the VM as a quick fix. To avoid a recurrence of the issue, upgrade your cluster to version 1.14.1 or a later.

Operation 1.13, 1.14.0, 1.14.1

gke-metrics-agent has no memory limit on Edge profile clusters

Depending on the cluster's workload, the gke-metrics-agent might use greater than 4608 MiB of memory. This issue only affects GKE on Bare Metal Edge profile clusters. Default profile clusters aren't impacted.


Workaround:

Upgrade your cluster to version 1.14.2 or later.

Installation 1.12, 1.13

Cluster creation might fail due to race conditions

When you create clusters using kubectl, due to race conditions preflight check may never finish. As a result, cluster creation may fail in certain cases.

The preflight check reconciler creates a SecretForwarder to copy the default ssh-key secret to the target namespace. Typically, preflight check leverages on the owner references and reconciles once the SecretForwarder is complete. However, in rare cases the owner references of the SecretForwarder can lose the reference to the preflight check, causing the preflight check to get stuck. As a result, cluster creation fails. In order to continue the reconciliation for the controller-driven preflight check, delete the cluster-operator pod or delete the preflight-check resource. When you delete the preflight-check resource, it creates another one and continues the reconciliation. Alternately, you can upgrade your existing clusters (that were created with an earlier version) to a fixed version.

Networking 1.9, 1.10, 1.11, 1.12, 1.13, 1.14, 1.15

Reserved IP addresses aren't released when using whereabouts plugin with the multi-NIC feature

In the multi-Nic feature, if you're using the CNI whereabouts plugin and you use the CNI DEL operation to delete a network interface for a Pod, some reserved IP addresses might not be released properly. This happens when the CNI DEL operation is interrupted.

You can verify the unused IP address reservations of the Pods by running the following command:

kubectl get ippools -A --kubeconfig KUBECONFIG_PATH

Workaround:

Manually delete the IP addresses (ippools) that aren't used.

Installation 1.10, 1.11.0, 1.11.1, 1.11.2

Node Problem Detector fails in 1.10.4 user cluster

The Node Problem Detector might fail in version 1.10.x user clusters, when version 1.11.0, 1.11.1, or 1.11.2 admin clusters manage 1.10.x user clusters. When the Node Problem Detector fails, the log gets updated with the following error message:

Error - NPD not supported for anthos baremetal version 1.10.4:
anthos version 1.10.4 not supported.

Workaround

Upgrade the admin cluster to 1.11.3 to resolve the issue.

Operation 1.14

1.14 island mode IPv4 cluster nodes have a pod CIDR mask size of 24

In release 1.14, the maxPodsPerNode setting isn't taken into account for island mode clusters, so the nodes are assigned a pod CIDR mask size of 24 (256 IP addresses).nThis might cause the cluster to run out of pod IP addresses earlier than expected. For example, if your cluster has a pod CIDR mask size of 22; each node will be assigned a pod CIDR mask of 24 and the cluster will only be able to support up to 4 nodes. Your cluster may also experience network instability in a period of high pod churn when maxPodsPerNode is set to 129 or higher and there isn't enough overhead in the pod CIDR for each node.

If your cluster is affected, the anetd pod reports the following error when you add a new node to the cluster and there's no podCIDR available:

error="required IPv4 PodCIDR not available"

Workaround

Use the following steps to resolve the issue:

  1. Upgrade to 1.14.1 or a later version.
  2. Remove the worker nodes and add them back.
  3. Remove the control plane nodes and add them back, preferably one by one to avoid cluster downtime.
Upgrades and updates 1.14.0, 1.14.1

Cluster upgrade rollback failure

An upgrade rollback might fail for version 1.14.0 or 1.14.1 clusters. If you upgrade a cluster from 1.14.0 to 1.14.1 and then try to rollback to 1.14.0 by using bmctl restore cluster command, an error like the following example might be returned:

I0119 22:11:49.705596  107905 client.go:48] Operation failed, retrying with backoff.
Cause: error updating "baremetal.cluster.gke.io/v1, Kind=HealthCheck" cluster-user-ci-f3a04dc1b0d2ac8/user-ci-f3a04dc1b0d2ac8-network: admission webhook "vhealthcheck.kb.io"
denied the request: HealthCheck.baremetal.cluster.gke.io "user-ci-f3a04dc1b0d2ac8-network" is invalid:
Spec: Invalid value: v1.HealthCheckSpec{ClusterName:(*string)(0xc0003096e0), AnthosBareMetalVersion:(*string)(0xc000309690),
Type:(*v1.CheckType)(0xc000309710), NodePoolNames:[]string(nil), NodeAddresses:[]string(nil), ConfigYAML:(*string)(nil),
CheckImageVersion:(*string)(nil), IntervalInSeconds:(*int64)(0xc0015c29f8)}: Field is immutable

Workaround:

Delete all healthchecks.baremetal.cluster.gke.io resources under the cluster namespace and then rerun the bmctl restore cluster command:

  1. List all healthchecks.baremetal.cluster.gke.io resources:
    kubectl get healthchecks.baremetal.cluster.gke.io \
        --namespace=CLUSTER_NAMESPACE \
        --kubeconfig=ADMIN_KUBECONFIG
    

    Replace the following:

    • CLUSTER_NAMESPACE: the namespace for the cluster.
    • ADMIN_KUBECONFIG: the path to the admin cluster kubeconfig file.
  2. Delete all healthchecks.baremetal.cluster.gke.io resources listed in the previous step:
    kubectl delete healthchecks.baremetal.cluster.gke.io \
        HEALTHCHECK_RESOURCE_NAME \
        --namespace=CLUSTER_NAMESPACE \
        --kubeconfig=ADMIN_KUBECONFIG
    
    Replace HEALTHCHECK_RESOURCE_NAME with the name of the healthcheck resources.
  3. Rerun the bmctl restore cluster command.
Networking 1.12.0

Service external IP address does not work in flat mode

In a cluster that has flatIPv4 set to true, Services of type LoadBalancer are not accessible by their external IP addresses.

This issue is fixed in version 1.12.1.


Workaround:

In the cilium-config ConfigMap, set enable-415 to "true", and then restart the anetd Pods.

Upgrades and updates 1.13.0, 1.14

In-place upgrades from 1.13.0 to 1.14.x never finish

When you try to do an in-place upgrade from 1.13.0 to 1.14.x using bmctl 1.14.0 and the --use-bootstrap=false flag, the upgrade never finishes.

An error with the preflight-check operator causes the cluster to never schedule the required checks, which means the preflight check never finishes.


Workaround:

Upgrade to 1.13.1 first before you upgrade to 1.14.x. An in-place upgrade from 1.13.0 to 1.13.1 should work. Or, upgrade from 1.13.0 to 1.14.x without the --use-bootstrap=false flag.

Upgrades and updates, Security 1.13 and 1.14

Clusters upgraded to 1.14.0 lose master taints

The control plane nodes require one of two specific taints to prevent workload pods from being scheduled on them. When you upgrade version 1.13 GKE clusters to version 1.14.0, the control plane nodes lose the following required taints:

  • node-role.kubernetes.io/master:NoSchedule
  • node-role.kubernetes.io/master:PreferNoSchedule

This problem doesn't cause upgrade failures, but pods that aren't supposed to run on the control plane nodes may start doing so. These workload pods can overwhelm control plane nodes and lead to cluster instability.

Determine if you're affected

  1. Find control plane nodes, use the following command:
    kubectl get node -l 'node-role.kubernetes.io/control-plane' \
        -o name --kubeconfig KUBECONFIG_PATH
    
  2. To check the list of taints on a node, use the following command:
    kubectl describe node NODE_NAME \
        --kubeconfig KUBECONFIG_PATH
    

    If neither of the required taints is listed, then you're affected.

Workaround

Use the following steps for each control plane node of your affected version 1.14.0 cluster to restore proper function. These steps are for the node-role.kubernetes.io/master:NoSchedule taint and related pods. If you intend for the control plane nodes to use the PreferNoSchedule taint, then adjust the steps accordingly.

Operation, VM Runtime on GDC 1.11, 1.12, 1.13, 1.14, 1.15, 1.16, 1.28

VM creation fails intermittently with upload errors

Creating a new Virtual Machine (VM) with the kubectl virt create vm command fails infrequently during image upload. This issue applies for both Linux and Windows VMs. The error looks something like the following example:

PVC default/heritage-linux-vm-boot-dv not found DataVolume default/heritage-linux-vm-boot-dv created
Waiting for PVC heritage-linux-vm-boot-dv upload pod to be ready... Pod now ready
Uploading data to https://10.200.0.51

 2.38 MiB / 570.75 MiB [>----------------------------------------------------------------------------------]   0.42% 0s

fail to upload image: unexpected return value 500,  ...

Workaround

Retry the kubectl virt create vm command to create your VM.

Upgrades and updates, Logging and monitoring 1.11

Managed collection components in 1.11 clusters aren't preserved in upgrades to 1.12

Managed collection components are part of Managed Service for Prometheus. If you manually deployed managed collection components in the gmp-system namespace of your version 1.11 clusters, the associated resources aren't preserved when you upgrade to version 1.12.

Starting with version 1.12.0 clusters, Managed Service for Prometheus components in the gmp-system namespace and related custom resource definitions are managed by stackdriver-operator with the enableGMPForApplications field. The enableGMPForApplications field defaults to true, so if you manually deploy Managed Service for Prometheus components in the namespace before upgrading to version 1.12, the resources are deleted by stackdriver-operator.

Workaround

To preserve manually managed collection resources:

  1. Backup all existing PodMonitoring custom resources.
  2. Upgrade the cluster to version 1.12 and enable Managed Service for Prometheus.
  3. Redeploy the PodMonitoring custom resources on your upgraded cluster.
Upgrades and updates 1.13

Some version 1.12 clusters with the Docker container runtime can't upgrade to version 1.13

If a version 1.12 cluster that uses the Docker container runtime is missing the following annotation, it can't upgrade to version 1.13:

baremetal.cluster.gke.io/allow-docker-container-runtime:  "true"

If you're affected by this issue, bmctl writes the following error in the upgrade-cluster.log file inside the bmctl-workspace folder:

Operation failed, retrying with backoff. Cause: error creating "baremetal.cluster.gke.io/v1, Kind=Cluster": admission webhook
"vcluster.kb.io" denied the request: Spec.NodeConfig.ContainerRuntime: Forbidden: Starting with Anthos Bare Metal version 1.13 Docker container
runtime will not be supported. Before 1.13 please set the containerRuntime to containerd in your cluster resources.

Although highly discouraged, you can create a cluster with Docker node pools until 1.13 by passing the flag "--allow-docker-container-runtime" to bmctl
create cluster or add the annotation "baremetal.cluster.gke.io/allow-docker- container-runtime: true" to the cluster configuration file.

This is most likely to occur with version 1.12 Docker clusters that were upgraded from 1.11, as that upgrade doesn't require the annotation to maintain the Docker container runtime. In this case, clusters don't have the annotation when upgrading to 1.13. Note that starting with version 1.13, containerd is the only permitted container runtime.

Workaround:

If you're affected by this problem, update the cluster resource with the missing annotation. You can add the annotation either while the upgrade is running or after canceling and before retrying the upgrade.

Installation 1.11

bmctl exits before cluster creation completes

Cluster creation may fail for GKE on Bare Metal version 1.11.0 (this issue is fixed in GKE on Bare Metal release 1.11.1). In some cases, the bmctl create cluster command exits early and writes errors like the following to the logs:

Error creating cluster: error waiting for applied resources: provider cluster-api watching namespace USER_CLUSTER_NAME not found in the target cluster

Workaround

The failed operation produces artifacts, but the cluster isn't operational. If this issue affects you, use the following steps to clean up artifacts and create a cluster:

Installation, VM Runtime on GDC 1.11, 1.12

Installation reports VM runtime reconciliation error

The cluster creation operation may report an error similar to the following:

I0423 01:17:20.895640 3935589 logs.go:82]  "msg"="Cluster reconciling:" "message"="Internal error occurred: failed calling webhook \"vvmruntime.kb.io\": failed to call webhook: Post \"https://vmruntime-webhook-service.kube-system.svc:443/validate-vm-cluster-gke-io-v1vmruntime?timeout=10s\": dial tcp 10.95.5.151:443: connect: connection refused" "name"="xxx" "reason"="ReconciliationError"

Workaround

This error is benign and you can safely ignore it.

Installation 1.10, 1.11, 1.12

Cluster creation fails when using multi-NIC, containerd, and HTTPS proxy

Cluster creation fails when you have the following combination of conditions:

  • Cluster is configured to use containerd as the container runtime (nodeConfig.containerRuntime set to containerd in the cluster configuration file, the default for GKE on Bare Metal version 1.13 and higher).
  • Cluster is configured to provide multiple network interfaces, multi-NIC, for pods (clusterNetwork.multipleNetworkInterfaces set to true in the cluster configuration file).
  • Cluster is configured to use a proxy (spec.proxy.url is specified in the cluster configuration file). Even though cluster creation fails, this setting is propagated when you attempt to create a cluster. You may see this proxy setting as an HTTPS_PROXY environment variable or in your containerd configuration (/etc/systemd/system/containerd.service.d/09-proxy.conf).

Workaround

Append service CIDRs (clusterNetwork.services.cidrBlocks) to the NO_PROXY environment variable on all node machines.

Installation 1.10, 1.11, 1.12

Failure on systems with restrictive umask setting

GKE on Bare Metal release 1.10.0 introduced a rootless control plane feature that runs all the control plane components as a non-root user. Running all components as a non-root user may cause installation or upgrade failures on systems with a more restrictive umask setting of 0077.


Workaround

Reset the control plane nodes and change the umask setting to 0022 on all the control plane machines. After the machines have been updated, retry the installation.

Alternatively, you can change the directory and file permissions of /etc/kubernetes on the control-plane machines for the installation or upgrade to proceed.

  • Make /etc/kubernetes and all its subdirectories world readable: chmod o+rx.
  • Make all the files owned by root user under the directory (recursively) /etc/kubernetes world readable (chmod o+r). Exclude private key files (.key) from these changes as they are already created with correct ownership and permissions.
  • Make /usr/local/etc/haproxy/haproxy.cfg world readable.
  • Make /usr/local/etc/bgpadvertiser/bgpadvertiser-cfg.yaml world readable.
Installation 1.10, 1.11, 1.12, 1.13

Control group v2 incompatibility

Control group v2 (cgroup v2) isn't supported in GKE on Bare Metal clusters versions 1.13 and earlier of GKE on Bare Metal. However, version 1.14 supports cgroup v2 as a Preview feature. The presence of /sys/fs/cgroup/cgroup.controllers indicates that your system uses cgroup v2.


Workaround

If your system uses cgroup v2, upgrade your cluster to version 1.14.

Installation 1.10, 1.11, 1.12, 1.13, 1.14, 1.15, 1.16, 1.28

Preflight checks and service account credentials

For installations triggered by admin or hybrid clusters (in other words, clusters not created with bmctl, like user clusters), the preflight check does not verify Google Cloud service account credentials or their associated permissions.

Installation 1.10, 1.11, 1.12, 1.13, 1.14, 1.15, 1.16, 1.28

Installing on vSphere

When installing GKE on Bare Metal clusters on vSphere VMs, you must set the tx-udp_tnl-segmentation and tx-udp_tnl-csum-segmentation flags to off. These flags are related to the hardware segmentation offload done by the vSphere driver VMXNET3 and they don't work with the GENEVE tunnel of GKE on Bare Metal clusters.


Workaround

Run the following command on each node to check the current values for these flags:

ethtool -k NET_INTFC | grep segm

Replace NET_INTFC with the network interface associated with the IP address of the node.

The response should have entries like the following:

...
tx-udp_tnl-segmentation: on
tx-udp_tnl-csum-segmentation: on
...
Sometimes in RHEL 8.4, ethtool shows these flags are off while they aren't. To explicitly set these flags to off, toggle the flags on and then off with the following commands:
ethtool -K ens192 tx-udp_tnl-segmentation on ethtool -K ens192 \
    tx-udp_tnl-csum-segmentation on
ethtool -K ens192 tx-udp_tnl-segmentation off ethtool -K ens192 \
    tx-udp_tnl-csum-segmentation off

This flag change does not persist across reboots. Configure the startup scripts to explicitly set these flags when the system boots.

Upgrades and updates 1.10

bmctl can't create, update, or reset lower version user clusters

The bmctl CLI can't create, update, or reset a user cluster with a lower minor version, regardless of the admin cluster version. For example, you can't use bmctl with a version of 1.N.X to reset a user cluster of version 1.N-1.Y, even if the admin cluster is also at version 1.N.X.

If you are affected by this issue, you should see the logs similar to the following when you use bmctl:

[2022-06-02 05:36:03-0500] error judging if the cluster is managing itself: error to parse the target cluster: error parsing cluster config: 1 error occurred:

*   cluster version 1.8.1 isn't supported in bmctl version 1.9.5, only cluster version 1.9.5 is supported

Workaround:

Use kubectl to create, edit, or delete the user cluster custom resource inside the admin cluster.

The ability to upgrade user clusters is unaffected.

Upgrades and updates 1.12

Cluster upgrades to version 1.12.1 may stall

Upgrading clusters to version 1.12.1 sometimes stalls due to the API server becoming unavailable. This issue affects all cluster types and all supported operating systems. When this issue occurs, the bmctl upgrade clustercommand can fail at multiple points, including during the second phase of preflight checks.


Workaround

You can check your upgrade logs to determine if you are affected by this issue. Upgrade logs are located in /baremetal/bmctl-workspace/CLUSTER_NAME/log/upgrade-cluster-TIMESTAMP by default.

The upgrade-cluster.log may contain errors like the following:

Failed to upgrade cluster: preflight checks failed: preflight check failed
The machine log may contain errors like the following (repeated failures indicate that you are affected by this issue):
FAILED - RETRYING: Query CNI health endpoint (30 retries left). FAILED - RETRYING: Query CNI health endpoint (29 retries left).
FAILED - RETRYING: Query CNI health endpoint (28 retries left). ...

HAProxy and Keepalived must be running on each control plane node before you reattempt to upgrade your cluster to version 1.12.1. Use the crictl command-line interface on each node to check to see if the haproxy and keepalived containers are running:

docker/crictl ps | grep haproxy docker/crictl ps | grep keepalived

If either HAProxy or Keepalived isn't running on a node, restart kubelet on the node:

systemctl restart kubelet
Upgrades and updates, VM Runtime on GDC 1.11, 1.12

Upgrading clusters to version 1.12.0 or higher fails when VM Runtime on GDC is enabled

In version 1.12.0 GKE on Bare Metal clusters, all resources related to VM Runtime on GDC are migrated to the vm-system namespace to better support the VM Runtime on GDC GA release. If you have VM Runtime on GDC enabled in a version 1.11.x or lower cluster, upgrading to version 1.12.0 or higher fails unless you first disable VM Runtime on GDC. When you're affected by this issue, the upgrade operation reports the following error:

Failed to upgrade cluster: cluster isn't upgradable with vmruntime enabled from
version 1.11.x to version 1.12.0: please disable VMruntime before upgrade to
1.12.0 and higher version

Workaround

To disable VM Runtime on GDC:

  1. Edit the VMRuntime custom resource:
    kubectl edit vmruntime
    
  2. Set enabled to false in the spec:
    apiVersion: vm.cluster.gke.io/v1
    kind: VMRuntime
    metadata:
      name: vmruntime
    spec:
      enabled: false
      ...
    
  3. Save the custom resource in your editor.
  4. Once the cluster upgrade is complete, re-enable VM Runtime on GDC.

For more information, see Working with VM-based workloads.

Upgrades and updates 1.10, 1.11, 1.12

Upgrade stuck at error during manifests operations

In some situations, cluster upgrades fail to complete and the bmctl CLI becomes unresponsive. This problem can be caused by an incorrectly updated resource. To determine if you're affected by this issue and to correct it, check the anthos-cluster-operator logs and look for errors similar to the following entries:

controllers/Cluster "msg"="error during manifests operations" "error"="1 error occurred: ... {RESOURCE_NAME} is invalid: metadata.resourceVersion: Invalid value: 0x0: must be specified for an update

These entries are a symptom of an incorrectly updated resource, where {RESOURCE_NAME} is the name of the problem resource.


Workaround

If you find these errors in your logs, complete the following steps:

  1. Use kubectl edit to remove the kubectl.kubernetes.io/last-applied-configuration annotation from the resource contained in the log message.
  2. Save and apply your changes to the resource.
  3. Retry the cluster upgrade.
Upgrades and updates 1.10, 1.11, 1.12

Upgrades are blocked for clusters with features that use Network Gateway for GDC

Cluster upgrades from 1.10.x to 1.11.x fail for clusters that use either egress NAT gateway or bundled load-balancing with BGP. These features both use Network Gateway for GDC. Cluster upgrades get stuck at the Waiting for upgrade to complete... command-line message and the anthos-cluster-operator logs errors like the following:

apply run failed ... MatchExpressions:[]v1.LabelSelectorRequirement(nil)}: field
is immutable...

Workaround

To unblock the upgrade, run the following commands against the cluster you are upgrading:

kubectl -n kube-system delete deployment \
    ang-controller-manager-autoscaler
kubectl -n kube-system delete deployment \
    ang-controller-manager
kubectl -n kube-system delete ds ang-node
Upgrades and updates 1.10, 1.11, 1.12, 1.13, 1.14, 1.15

bmctl update doesn't remove maintenance blocks

The bmctl update command can't remove or modify the maintenanceBlocks section from the cluster resource configuration.


Workaround

For more information, including instructions for removing nodes from maintenance mode, see Put nodes into maintenance mode.

Operation 1.10, 1.11, 1.12

Nodes uncordoned if you don't use the maintenance mode procedure

If you runversion 1.12.0 clusters (anthosBareMetalVersion: 1.12.0) or lower and manually use kubectl cordon on a node, GKE on Bare Metal might uncordon the node before you're ready in an effort to reconcile the expected state.


Workaround

For version 1.12.0 and lower clusters, use maintenance mode to cordon and drain nodes safely.

In version 1.12.1 (anthosBareMetalVersion: 1.12.1) or higher, GKE on Bare Metal won't uncordon your nodes unexpectedly when you use kubectl cordon.

Operation 1.11

Version 11 admin clusters using a registry mirror can't manage version 1.10 clusters

If your admin cluster is on version 1.11 and uses a registry mirror, it can't manage user clusters that are on a lower minor version. This issue affects reset, update, and upgrade operations on the user cluster.

To determine whether this issue affects you, check your logs for cluster operations, such as create, upgrade, or reset. These logs are located in the bmctl-workspace/CLUSTER_NAME/ folder by default. If you're affected by the issue, your logs contain the following error message:

flag provided but not defined: -registry-mirror-host-to-endpoints
Operation 1.10, 1.11

kubeconfig Secret overwritten

The bmctl check cluster command, when run on user clusters, overwrites the user cluster kubeconfig Secret with the admin cluster kubeconfig. Overwriting the file causes standard cluster operations, such as updating and upgrading, to fail for affected user clusters. This problem applies to GKE on Bare Metal cluster versions 1.11.1 and earlier.

To determine if this issue affects a user cluster, run the following command:

kubectl --kubeconfig ADMIN_KUBECONFIG \
    get secret -n USER_CLUSTER_NAMESPACE \
    USER_CLUSTER_NAME -kubeconfig \
    -o json  | jq -r '.data.value'  | base64 -d

Replace the following:

  • ADMIN_KUBECONFIG: the path to the admin cluster kubeconfig file.
  • USER_CLUSTER_NAMESPACE: the namespace for the cluster. By default, the cluster namespaces for GKE on Bare Metal clusters are the name of the cluster prefaced with cluster-. For example, if you name your cluster test, the default namespace is cluster-test.
  • USER_CLUSTER_NAME: the name of the user cluster to check.

If the cluster name in the output (see contexts.context.cluster in the following sample output) is the admin cluster name, then the specified user cluster is affected.

apiVersion: v1
clusters:
- cluster:
    certificate-authority-data:LS0tLS1CRU...UtLS0tLQo=
    server: https://10.200.0.6:443
  name: ci-aed78cdeca81874
contexts:
- context:
    cluster: ci-aed78cdeca81
    user: ci-aed78cdeca81-admin
  name: ci-aed78cdeca81-admin@ci-aed78cdeca81
current-context: ci-aed78cdeca81-admin@ci-aed78cdeca81
kind: Config
preferences: {}
users:
- name: ci-aed78cdeca81-admin
  user:
    client-certificate-data: LS0tLS1CRU...UtLS0tLQo=
    client-key-data: LS0tLS1CRU...0tLS0tCg==

Workaround

The following steps restore function to an affected user cluster (USER_CLUSTER_NAME):

  1. Locate the user cluster kubeconfig file. GKE on Bare Metal generates the kubeconfig file on the admin workstation when you create a cluster. By default, the file is in the bmctl-workspace/USER_CLUSTER_NAME directory.
  2. Verify the kubeconfig is correct user cluster kubeconfig:
    kubectl get nodes \
        --kubeconfig PATH_TO_GENERATED_FILE
    
    Replace PATH_TO_GENERATED_FILE with the path to the user cluster kubeconfig file. The response returns details about the nodes for the user cluster. Confirm the machine names are correct for your cluster.
  3. Run the following command to delete the corrupted kubeconfig file in the admin cluster:
    kubectl delete secret \
        -n USER_CLUSTER_NAMESPACE \
        USER_CLUSTER_NAME-kubeconfig
    
  4. Run the following command to save the correct kubeconfig secret back to the admin cluster:
    kubectl create secret generic \
        -n USER_CLUSTER_NAMESPACE \
        USER_CLUSTER_NAME-kubeconfig \
        --from-file=value=PATH_TO_GENERATED_FILE
    
Operation 1.10, 1.11, 1.12, 1.13, 1.14, 1.15, 1.16, 1.28

Taking a snapshot as a non-root login user

If you use containerd as the container runtime, running snapshot as non-root user requires /usr/local/bin to be in the user's PATH. Otherwise it will fail with a crictl: command not found error.

When you aren't logged in as the root user, sudo is used to run the snapshot commands. The sudo PATH can differ from the root profile and may not contain /usr/local/bin.


Workaround

Update the secure_path in /etc/sudoers to include /usr/local/bin. Alternatively, create a symbolic link for crictl in another /bin directory.

Logging and monitoring 1.10

stackdriver-log-forwarder has [parser:cri] invalid time format warning logs

If the container runtime interface (CRI) parser uses an incorrect regular expression for parsing time, the logs for the stackdriver-log-forwarder Pod contain errors and warnings like the following:

[2022/03/04 17:47:54] [error] [parser] time string length is too long [2022/03/04 20:16:43] [ warn] [parser:cri] invalid time format %Y-%m-%dT%H:%M:%S.%L%z for '2022-03-04T20:16:43.680484387Z'

Workaround:

Logging and monitoring 1.10, 1.11, 1.12, 1.13, 1.14, 1.15

Unexpected monitoring billing

For GKE on Bare Metal cluster versions 1.10 to 1.15, some customers have found unexpectedly high billing for Metrics volume on the Billing page. This issue affects you only when all of the following circumstances apply:

  • Application monitoring is enabled (enableStackdriverForApplications=true)
  • Managed Service for Prometheus isn't enabled (enableGMPForApplications)
  • Application Pods have the prometheus.io/scrap=true annotation

To confirm whether you are affected by this issue, list your user-defined metrics. If you see billing for unwanted metrics, then this issue applies to you.


Workaround

If you are affected by this issue, we recommend that you upgrade your clusters to version 1.12 and switch to new application monitoring solution managed-service-for-prometheus that address this issue:

  • Separate flags to control the collection of application logs versus application metrics
  • Bundled Google Cloud Managed Service for Prometheus
  • If you can't upgrade to version 1.12, use the following steps:

    1. Find the source Pods and Services that have the unwanted billing:
      kubectl --kubeconfig KUBECONFIG \
          get pods -A -o yaml | grep 'prometheus.io/scrape: "true"'
      kubectl --kubeconfig KUBECONFIG get \
          services -A -o yaml | grep 'prometheus.io/scrape: "true"'
      
    2. Remove the prometheus.io/scrap=true annotation from the Pod or Service.
    Logging and monitoring 1.11, 1.12, 1.13, 1.14, 1.15, 1.16, 1.28

    Edits to metrics-server-config aren't persisted

    High pod density can, in extreme cases, create excessive logging and monitoring overhead, which can cause Metrics Server to stop and restart. You can edit the metrics-server-config ConfigMap to allocate more resources to keep Metrics Server running. However, due to reconciliation, edits made to metrics-server-config can get reverted to the default value during a cluster update or upgrade operation. Metrics Server isn't affected immediately, but the next time it restarts, it picks up the reverted ConfigMap and is vulnerable to excessive overhead, again.


    Workaround

    For 1.11.x, you can script the ConfigMap edit and perform it along with updates or upgrades to the cluster. For 1.12 and onward, contact support.

    Logging and monitoring 1.11, 1.12

    Deprecated metrics affects Cloud Monitoring dashboard

    Several GKE on Bare Metal metrics have been deprecated and, starting with GKE on Bare Metal release 1.11, data is no longer collected for these deprecated metrics. If you use these metrics in any of your alerting policies, there won't be any data to trigger the alerting condition.

    The following table lists the individual metrics that have been deprecated and the metric that replaces them.

    Deprecated metrics Replacement metric
    kube_daemonset_updated_number_scheduled kube_daemonset_status_updated_number_scheduled
    kube_node_status_allocatable_cpu_cores
    kube_node_status_allocatable_memory_bytes
    kube_node_status_allocatable_pods
    kube_node_status_allocatable
    kube_node_status_capacity_cpu_cores
    kube_node_status_capacity_memory_bytes
    kube_node_status_capacity_pods
    kube_node_status_capacity

    In GKE on Bare Metal cluster versions lower than 1.11, the policy definition file for the recommended Anthos on baremetal node cpu usage exceeds 80 percent (critical) alert uses the deprecated metrics. The node-cpu-usage-high.json JSON definition file is updated for releases 1.11.0 and later.


    Workaround

    Use the following steps to migrate to the replacement metrics:

    1. In the Google Cloud console, select Monitoring or click the following button:
      Go to Monitoring
    2. In the navigation pane, select Dashboards, and delete the Anthos cluster node status dashboard.
    3. Click the Sample library tab and reinstall the Anthos cluster node status dashboard.
    4. Follow the instructions in Creating alerting policies to create a policy using the updated node-cpu-usage-high.json policy definition file.
    Logging and monitoring 1.10, 1.11

    stackdriver-log-forwarder has CrashloopBackOff errors

    In some situations, the fluent-bit logging agent can get stuck processing corrupt chunks. When the logging agent is unable to bypass corrupt chunks, you may observe that stackdriver-log-forwarder keeps crashing with a CrashloopBackOff error. If you are having this problem, your logs have entries like the following

    [2022/03/09 02:18:44] [engine] caught signal (SIGSEGV) #0  0x5590aa24bdd5
    in  validate_insert_id() at plugins/out_stackdriver/stackdriver.c:1232
    #1  0x5590aa24c502      in  stackdriver_format() at plugins/out_stackdriver/stackdriver.c:1523
    #2  0x5590aa24e509      in  cb_stackdriver_flush() at plugins/out_stackdriver/stackdriver.c:2105
    #3  0x5590aa19c0de      in  output_pre_cb_flush() at include/fluent-bit/flb_output.h:490
    #4  0x5590aa6889a6      in  co_init() at lib/monkey/deps/flb_libco/amd64.c:117 #5  0xffffffffffffffff  in  ???() at ???:0
    

    Workaround:

    Clean up the buffer chunks for the Stackdriver Log Forwarder.

    Note: In the following commands, replace KUBECONFIG with the path to the admin cluster kubeconfig file.

    1. Terminate all stackdriver-log-forwarder pods:
      kubectl --kubeconfig KUBECONFIG -n kube-system patch daemonset \
          stackdriver-log-forwarder -p \
          '{"spec": {"template": {"spec": {"nodeSelector": {"non-existing": "true"}}}}}'
      
      Verify that the stackdriver-log-forwarder pods are deleted before going to the next step.
    2. Deploy the following DaemonSet to clean up any corrupted data in fluent-bit buffers:
      kubectl --kubeconfig KUBECONFIG -n kube-system apply -f - << EOF
      apiVersion: apps/v1
      kind: DaemonSet
      metadata:
        name: fluent-bit-cleanup
        namespace: kube-system
      spec:
        selector:
          matchLabels:
            app: fluent-bit-cleanup
        template:
          metadata:
            labels:
              app: fluent-bit-cleanup
          spec:
            containers:
            - name: fluent-bit-cleanup
              image: debian:10-slim
              command: ["bash", "-c"]
              args:
              - |
                rm -rf /var/log/fluent-bit-buffers/
                echo "Fluent Bit local buffer is cleaned up."
                sleep 3600
              volumeMounts:
              - name: varlog
                mountPath: /var/log
              securityContext:
                privileged: true
            tolerations:
            - key: "CriticalAddonsOnly"
              operator: "Exists"
            - key: node-role.kubernetes.io/master
              effect: NoSchedule
            - key: node-role.gke.io/observability
              effect: NoSchedule
            volumes:
            - name: varlog
              hostPath:
                path: /var/log
      EOF
      
    3. Use the following commands to verify that the DaemonSet has cleaned up all the nodes:
      kubectl --kubeconfig KUBECONFIG logs \
          -n kube-system -l \
          app=fluent-bit-cleanup | grep "cleaned up" | wc -l
      kubectl --kubeconfig KUBECONFIG -n \
          kube-system get pods -l \
          app=fluent-bit-cleanup --no-headers | wc -l
      
      The output of the two commands should be equal to the number of nodes in your cluster.
    4. Delete the cleanup DaemonSet:
      kubectl --kubeconfig KUBECONFIG -n \
          kube-system delete ds fluent-bit-cleanup
      
    5. Restart the log forwarder pods:
      kubectl --kubeconfig KUBECONFIG \
          -n kube-system patch daemonset \
          stackdriver-log-forwarder --type json \
          -p='[{"op": "remove", "path": "/spec/template/spec/nodeSelector/non-existing"}]'
      
    Logging and monitoring 1.10, 1.11, 1.12, 1.13, 1.14, 1.15, 1.16, 1.28

    Unknown metrics error in gke-metrics-agent log

    gke-metrics-agent is a daemonset that is collecting metrics on each node and forward them to Cloud Monitoring. It may produce logs such as the following:

    Unknown metric: kubernetes.io/anthos/go_gc_duration_seconds_summary_percentile
    

    Similar errors may happen to other metrics types, including (but not limited to):

    • apiserver_admission_step_admission_duration_seconds_summary
    • go_gc_duration_seconds
    • scheduler_scheduling_duration_seconds
    • gkeconnect_http_request_duration_seconds_summary
    • alertmanager_nflog_snapshot_duration_seconds_summary

    These error logs can be safely ignored as the metrics they refer to are not supported and not critical for monitoring purposes.

    Logging and monitoring 1.10, 1.11

    Intermittent metrics export interruptions

    GKE on Bare Metal clusters might experience interruptions in normal, continuous exporting of metrics, or missing metrics on some nodes. If this issue affects your clusters, you may see gaps in data for the following metrics (at a minimum):

    • kubernetes.io/anthos/container_memory_working_set_bytes
    • kubernetes.io/anthos/container_cpu_usage_seconds_total
    • kubernetes.io/anthos/container_network_receive_bytes_total

    Workaround

    Upgrade your clusters to version 1.11.1 or later.

    If you can't upgrade, perform the following steps as a workaround:

    1. Open your stackdriver resource for editing:
      kubectl -n kube-system edit stackdriver stackdriver
      
    2. To increase the CPU request for gke-metrics-agent from 10m to 50m, add the following resourceAttrOverride section to the stackdriver manifest:
      spec:
        resourceAttrOverride:
          gke-metrics-agent/gke-metrics-agent:
            limits:
              cpu: 100m
              memory: 4608Mi
            requests:
              cpu: 50m
              memory: 200Mi
      
      Your edited resource should look similar to the following:
      spec:
        anthosDistribution: baremetal
        clusterLocation: us-west1-a
        clusterName: my-cluster
        enableStackdriverForApplications: true
        gcpServiceAccountSecretName: ...
        optimizedMetrics: true
        portable: true
        projectID: my-project-191923
        proxyConfigSecretName: ...
        resourceAttrOverride:
          gke-metrics-agent/gke-metrics-agent:
            limits:
              cpu: 100m
              memory: 4608Mi
            requests:
              cpu: 50m
              memory: 200Mi
      
    3. Save your changes and close the text editor.
    4. To verify your changes have taken effect, run the following command:
      kubectl -n kube-system get daemonset \
          gke-metrics-agent -o yaml | grep "cpu: 50m"
      
      The command finds cpu: 50m if your edits have taken effect.
    Networking 1.10

    Multiple default gateways breaks connectivity to external endpoints

    Having multiple default gateways in a node can lead to broken connectivity from within a Pod to external endpoints, such as google.com.

    To determine if you're affected by this issue, run the following command on the node:

    ip route show
    

    Multiple instances of default in the response indicate that you're affected.

    Networking 1.12

    Networking custom resource edits on user clusters get overwritten

    Version 1.12.x GKE on Bare Metal clusters don't prevent you from manually editing networking custom resources in your user cluster. GKE on Bare Metal reconciles custom resources in the user clusters with the custom resources in your admin cluster during cluster upgrades. This reconciliation overwrites any edits made directly to the networking custom resources in the user cluster. The networking custom resources should be modified in the admin cluster only, but version 1.12.x clusters don't enforce this requirement.

    Advanced networking features, such as bundled load balancing with BGP, egress NAT gateway, SR-IOV networking, flat-mode with BGP, and multi-NIC for Pods use the following custom resources:

    • BGPLoadBalancer
    • BGPPeer
    • NetworkGatewayGroup
    • NetworkAttachmentDefinition
    • ClusterCIDRConfig
    • FlatIPMode

    You edit these custom resources in your admin cluster and the reconciliation step applies the changes to your user clusters.


    Workaround

    If you've modified any of the previously mentioned custom resources on a user cluster, modify the corresponding custom resources on your admin cluster to match before upgrading. This step ensures that your configuration changes are preserved. GKE on Bare Metal cluster versions 1.13.0 and higher prevent you from modifying the networking custom resources on your user clusters directly.

    Networking 1.10, 1.11, 1.12, 1.13, 1.14, 1.15, 1.16, 1.28

    Pod connectivity failures and reverse path filtering

    GKE on Bare Metal configures reverse path filtering on nodes to disable source validation (net.ipv4.conf.all.rp_filter=0). If the rp_filter setting is changed to 1 or 2, pods will fail due to out-of-node communication timeouts.

    Reverse path filtering is set with rp_filter files in the IPv4 configuration folder (net/ipv4/conf/all). This value may also be overridden by sysctl, which stores reverse path filtering settings in a network security configuration file, such as /etc/sysctl.d/60-gce-network-security.conf.


    Workaround

    Pod connectivity can be restored by performing either of the following workarounds:

    Set the value for net.ipv4.conf.all.rp_filter back to 0 manually, and then run sudo sysctl -p to apply the change.

    Or

    Restart the anetd Pod to set net.ipv4.conf.all.rp_filter back to 0. To restart the anetd Pod, use the following commands to locate and delete the anetd Pod and a new anetd Pod will start up in its place:

    kubectl get pods -n kube-system kubectl delete pods -n kube-system ANETD_XYZ
    

    Replace ANETD_XYZ with the name of the anetd Pod.

    After performing either of the workarounds verify that the net.ipv4.conf.all.rp_filter value is set to 0 by running sysctl net.ipv4.conf.all.rp_filter on each node.

    Networking 1.10, 1.11, 1.12, 1.13, 1.14, 1.15, 1.16, 1.28

    Bootstrap (kind) cluster IP addresses and cluster node IP addresses overlapping

    192.168.122.0/24 and 10.96.0.0/27 are the default pod and service CIDRs used by the bootstrap (kind) cluster. Preflight checks will fail if they overlap with cluster node machine IP addresses.


    Workaround

    To avoid the conflict, you can pass the --bootstrap-cluster-pod-cidr and --bootstrap-cluster-service-cidr flags to bmctl to specify different values.

    Operating system 1.10, 1.11, 1.12, 1.13, 1.14, 1.15, 1.16, 1.28

    Cluster creation or upgrade fails on CentOS

    In December 2020, the CentOS community and Red Hat announced the sunset of CentOS. On January 31, 2022, CentOS 8 reached its end of life (EOL). As a result of the EOL, yum repositories stopped working for CentOS, which causes cluster creation and cluster upgrade operations to fail. This applies to all supported versions of CentOS and affects all versions of GKE on Bare Metal clusters.


    Workaround

    Security 1.10, 1.11, 1.12, 1.13, 1.14, 1.15, 1.16, 1.28

    Container can't write to VOLUME defined in Dockerfile with containerd and SELinux

    If you use containerd as the container runtime and your operating system has SELinux enabled, the VOLUME defined in the application Dockerfile might not be writable. For example, containers built with the following Dockerfile aren't able to write to the /tmp folder.

    FROM ubuntu:20.04 RUN chmod -R 777 /tmp VOLUME /tmp
    

    To verify if you're affected by this issue, run the following command on the node that hosts the problematic container:

    ausearch -m avc
    

    If you're affected by this issue, you see a denied error like the following:

    time->Mon Apr  4 21:01:32 2022 type=PROCTITLE msg=audit(1649106092.768:10979): proctitle="bash" type=SYSCALL msg=audit(1649106092.768:10979): arch=c000003e syscall=257 success=no exit=-13 a0=ffffff9c a1=55eeba72b320 a2=241 a3=1b6 items=0 ppid=75712 pid=76042 auid=4294967295 uid=0 gid=0 euid=0 suid=0 fsuid=0 egid=0 sgid=0 fsgid=0 tty=pts0 ses=4294967295 comm="bash" exe="/usr/bin/bash" subj=system_u:system_r:container_t:s0:c701,c935 key=(null) type=AVC msg=audit(1649106092.768:10979): avc:  denied { write } for  pid=76042 comm="bash" name="aca03d7bb8de23c725a86cb9f50945664cb338dfe6ac19ed0036c" dev="sda2" ino=369501097 scontext=system_u:system_r: container_t:s0:c701,c935 tcontext=system_u:object_r: container_ro_file_t:s0 tclass=dir permissive=0
    

    Workaround

    To work around this issue, make either of the following changes:

    • Turn off SELinux.
    • Don't use the VOLUME feature inside Dockerfile.
    Upgrades and updates 1.10, 1.11, 1.12

    Node Problem Detector isn't enabled by default after cluster upgrades

    When you upgrade GKE on Bare Metal clusters, Node Problem Detector isn't enabled by default. This issue is applicable for upgrades in release 1.10 to 1.12.1 and has been fixed in release 1.12.2.


    Workaround:

    To enable the Node Problem Detector:

    1. Verify if node-problem-detector systemd service is running on the node.
      1. Use the SSH command and connect to the node.
      2. Check if node-problem-detector systemd service is running on the node:
        systemctl is-active node-problem-detector
        
        If the command result displays inactive, then the node-problem-detector isn't running on the node.
    2. To enable the Node Problem Detector, use the kubectl edit command and edit the node-problem-detector-config ConfigMap. For more information, see Node Problem Detector.
    Operation 1.9, 1.10

    Cluster backup fails when using non-root login

    The bmctl backup cluster command fails if nodeAccess.loginUser is set to a non-root username.]


    Workaround:

    This issue applies to versions 1.9.x, 1.10.0, and 1.10.1 and is fixed in version 1.10.2 and later.

    Networking 1.10, 1.11, 1.12

    Load Balancer Services don't work with containers on the control plane host network

    There is a bug in anetd where packets are dropped for LoadBalancer Services if the backend pods are both running on the control plane node and are using the hostNetwork: true field in the container's spec.

    The bug isn't present in version 1.13 or later.


    Workaround:

    The following workarounds can help if you use a LoadBalancer Service that is backed by hostNetwork Pods:

    1. Run them on worker nodes (not control plane nodes).
    2. Use externalTrafficPolicy: local in the Service spec and ensure your workloads run on load balancer nodes.
    Upgrades and Updates 1.12, 1.13

    Orphaned anthos-version-$version$ pod failing to pull image

    Cluster upgrading from 1.12.x to 1.13.x might observe a failing anthos-version-$version$ pod with ImagePullBackOff error. This happens due to the race condition of anthos-cluster-operator gets upgraded and it shouldn't affect any regular cluster capabilities.

    The bug isn't present after version 1.13 or later.


    Workaround:

    Delete the Job of dynamic-version-installer by kubectl delete job anthos-version-$version$ -n kube-system

    Upgrades and updates 1.13

    1.12 clusters upgraded from 1.11 can't upgrade to 1.13.0

    Version 1.12 clusters that were upgraded from version 1.11 can't be upgraded to version 1.13.0. This upgrade issue doesn't apply to clusters that were created at version 1.12.

    To determine if you're affected, check the logs of the upgrade job that contains the upgrade-first-no* string in the admin cluster. If you see the following error message, you're affected.

    TASK [kubeadm_upgrade_apply : Run kubeadm upgrade apply] *******
    ...
    [upgrade/config] FATAL: featureGates: Invalid value: map[string]bool{\"IPv6DualStack\":false}: IPv6DualStack isn't a valid feature name.
    ...
    

    Workaround:

    To work around this issue:

    1. Run the following commands on your admin workstation:
      echo '[{ "op": "remove", "path": \
          "/spec/clusterConfiguration/featureGates" }]' \
          > remove-feature-gates.patch
      export KUBECONFIG=$ADMIN_KUBECONFIG
      kubectl get kubeadmconfig -A --no-headers | xargs -L1 bash -c \
          'kubectl patch kubeadmconfig $1 -n $0 --type json \
          --patch-file remove-feature-gates.patch'
      
    2. Re-attempt the cluster upgrade.
    Logging and Monitoring 1.16.2, 1.16.3

    High CPU usage for stackdriver-operator

    There's an issue in stackdriver-operator that causes it to consume higher CPU time than normal. Normal CPU usage is less than 50 milliCPU (50m) for stackdriver-operator in idle state. The cause is a mismatch of Certificate resources that stackdriver-operator applies with the expectations from cert-manager. This mismatch causes a race condition between cert-manager and stackdriver-operator in updating those resources.

    This issue may result in reduced performance on clusters with limited CPU availability.


    Workaround:

    Until you can upgrade to a version that fixed this bug, use the following workaround:

    1. To temporarily scale down stackdriver-operator to 0 replicas, apply an AddonConfiguration custom resource:
      kubectl scale deploy stackdriver-operator --replicas=0
      
    2. Once you've upgraded to a version that fixes this issue, scale stackdriver-operator back up again:
      kubectl scale deploy stackdriver-operator --replicas=1
      
    Logging and Monitoring 1.16.0, 1.16.1

    Annotation-based metrics scraping not working

    In the GKE on Bare Metal 1.16 minor release, the enableStackdriverForApplications field in the stackdriver custom resource spec is deprecated. This field is replaced by two fields, enableCloudLoggingForApplications and enableGMPForApplications, in the stackdriver custom resource.

    We recommend that you to use Google Cloud Managed Service for Prometheus for monitoring your workloads. Use the enableGMPForApplications field to enable this feature.

    If you rely on metrics collection triggered by prometheus.io/scrape annotations on your workloads, you can use the annotationBasedApplicationMetrics feature gate flag to keep the old behavior. However, there is an issue that prevents the annotationBasedApplicationMetrics from working properly, preventing metrics collection from your applications into Cloud Monitoring.


    Workaround:

    To resolve this issue, upgrade your cluster to version 1.16.2 or higher.

    The annotation-based workload metrics collection enabled by the annotationBasedApplicationMetrics feature gate collects metrics for objects that have the prometheus.io/scrape annotation. Many software systems with open source origin may use this annotation. If you continue using this method of metrics collection, be aware of this dependency so that you aren't surprised by metrics charges in Cloud Monitoring.

    If you need additional assistance, reach out to