GKE on Bare Metal known issues

Installation

Control group v2 incompatibility

Control group v2 (cgroup v2) is incompatible with GKE on Bare Metal 1.6. Kubernetes 1.18 does not support cgroup v2. Also Docker only offers experimental support as of 20.10. systemd switched to cgroup v2 by default in version 247.2-2. The presence of /sys/fs/cgroup/cgroup.controllers indicates that your system uses cgroup v2.

Starting with GKE on Bare Metal 1.6.2, the preflight checks verify that cgroup v2 is not in use on the cluster machine.

Benign error messages during installation

During highly available (HA) cluster installation, you may see errors about etcdserver leader change. These error messages are benign and can be ignored.

When you use bmctl for cluster installation, you may see a Log streamer failed to get BareMetalMachine log message at the very end of the create-cluster.log. This error message is benign and can be ignored.

When examining cluster creation logs, you may notice transient failures about registering clusters or calling webhooks. These errors can be safely ignored, because the installation will retry these operations until they succeed.

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 Platform service account credentials or their associated permissions.

Application default credentials and bmctl

bmctl uses Application Default Credentials (ADC) to validate the cluster operation's location value in the cluster spec when it is not set to global.

For ADC to work, you need to either point the GOOGLE_APPLICATION_CREDENTIALS environment variable to a service account credential file, or run gcloud auth application-default login.

Docker service

On cluster node machines, if the Docker executable is present in the PATH environment variable, but the Docker service is not active, preflight check will fail and report that the Docker service is not active. To fix this error, either remove Docker, or enable the Docker service.

Upgrades and updates

Upgrading is not available in the GKE on Bare Metal 1.6.x releases.

bmctl update cluster fails if the .manifests directory is missing

If the .manifests directory is removed prior to running bmctl update cluster, the command fails with an error similar to the following:

Error updating cluster resources.: failed to get CRD file .manifests/1.9.0/cluster-operator/base/crd/bases/baremetal.cluster.gke.io_clusters.yaml: open .manifests/1.9.0/cluster-operator/base/crd/bases/baremetal.cluster.gke.io_clusters.yaml: no such file or directory

You can fix this issue by running bmctl check cluster first, which will recreate the .manifests directory.

This issue applies to GKE on Bare Metal 1.10 and earlier and is fixed in version 1.11 and later.

bmctl update doesn't remove maintenance blocks

The bmctl update command can't remove or modify the maintenanceBlocks section from the cluster resource configuration. For more information, including instructions for removing nodes from maintenance mode, see Put nodes into maintenance mode.

Operation

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 versions 1.11.1 and earlier.

To determine if a user cluster is affected by this issue, run the following command:

kubectl --kubeconfig ADMIN_KUBECONFIG get secret -n cluster-USER_CLUSTER_NAME \
    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_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.

user-cluster-kubeconfig  -o json  | jq -r '.data.value'  | base64 -d
apiVersion: v1
clusters:
- cluster:
    certificate-authority-data:LS0tLS1CRU...UtLS0tLQo=
    server: https://10.200.0.6:443
  name: ci-aed78cdeca81874
contexts:
- context:
    cluster: ci-aed78cdeca81874
    user: ci-aed78cdeca81874-admin
  name: ci-aed78cdeca81874-admin@ci-aed78cdeca81874
current-context: ci-aed78cdeca81874-admin@ci-aed78cdeca81874
kind: Config
preferences: {}
users:
- name: ci-aed78cdeca81874-admin
  user:
    client-certificate-data: LS0tLS1CRU...UtLS0tLQo=
    client-key-data: LS0tLS1CRU...0tLS0tCg==

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
    

Reset/Deletion

User cluster credentials

The bmctl reset command relies on the top-level credentials section in the cluster configuration file. For user clusters, you will need to manually update the file to add the credentials section.

Mount points and fstab

Reset does not unmount the mount points under /mnt/anthos-system and /mnt/localpv-share/. It also does not clean up the corresponding entries in /etc/fstab.

Namespace deletion

Deleting a namespace will prevent new resources from being created in that namespace, including jobs to reset machines. When deleting a user cluster, you must delete the cluster object first before deleting its namespace. Otherwise, the jobs to reset machines cannot get created, and the deletion process will skip the machine clean-up step.

Security

The cluster CA/certificate will be rotated during upgrade. On-demand rotation support is not currently available.

GKE on Bare Metal rotates kubelet serving certificates automatically. Each kubelet node agent can send out a Certificate Signing Request (CSR) when a certificate nears expiration. A controller in your admin clusters validates and approves the CSR.

Networking

Client source IP with bundled Layer 2 load balancing

Setting the external traffic policy to Local can cause routing errors, such as No route to host, for bundled Layer 2 load balancing. The external traffic policy is set to Cluster (externalTrafficPolicy: Cluster), by default. With this setting, Kubernetes handles cluster-wide traffic. Services of type LoadBalancer or NodePort can use externalTrafficPolicy: Local to preserve the client source IP address. With this setting, however, Kubernetes only handles node-local traffic.

If you want to preserve the client source IP address, additional configuration may be required to ensure service IPs are reachable. For configuration details, see Preserving client source IP address in Configure bundled load balancing.

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 therp_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.

To restore Pod connectivity, either set net.ipv4.conf.all.rp_filter back to 0 manually, 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.

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

Overlapping IP addresses across different clusters

There is no preflight check to validate overlapping IP addresses across different clusters.

hostport feature in GKE on Bare Metal

The hostport feature in ContainerPort is not currently supported.

Operating system

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.

As a workaround, run the following commands to have your CentOS use an archive feed:

sed -i 's/mirrorlist/#mirrorlist/g' /etc/yum.repos.d/CentOS-Linux-*
sed -i 's|#baseurl=http://mirror.centos.org|baseurl=http://vault.centos.org|g' \
    /etc/yum.repos.d/CentOS-Linux-*

As a long-term solution, consider migrating to another supported operating system, such as Ubuntu or RHEL.

Operating system endpoint limitations

On RHEL and CentOS, there is a cluster level limitation of 100,000 endpoints. Kubernetes service. If 2 services reference the same set of pods, this counts as 2 separate sets of endpoints. The underlying nftable implementation on RHEL and CentOS causes this limitation; it is not an intrinsic limitation of GKE on Bare Metal.

Configuration

Control plane and load balancer specifications

The control plane and load balancer node pool specifications are special. These specifications declare and control critical cluster resources. The canonical source for these resources is their respective sections in the cluster config file:

  • spec.controlPlane.nodePoolSpec
  • spec.LoadBalancer.nodePoolSpec

Consequently, do not modify the top-level control plane and load balancer node pool resources directly. Modify the associated sections in the cluster config file instead.

Mutable fields in the cluster and node pool specification

Currently, only the following cluster and node pool specification fields in the cluster config file can be updated after the cluster is created (they are mutable fields):

  • For the Cluster object (kind: Cluster), the following fields are mutable:

    • spec.anthosBareMetalVersion
    • spec.bypassPreflightCheck
    • spec.controlPlane.nodePoolSpec.nodes
    • spec.loadBalancer.nodePoolSpec.nodes
    • spec.maintenanceBlocks
    • spec.nodeAccess.loginUser
  • For the NodePool object (kind: NodePool), the following fields are mutable:

    • spec.nodes

Node shows NotReady status

Under certain load conditions, GKE on Bare Metal 1.6.x nodes may display a NotReady status due to Pod Lifecycle Event Generator (PLEG) being unhealthy. The node status will contain the following entry:

PLEG is not healthy: pleg was last seen active XXXmXXXs ago; threshold is 3m0s

How do I know if I'm affected?

A likely cause of this issue is the runc binary version. To confirm if you have the problematic version installed, connect to one of the cluster machines using SSH and run:

/usr/bin/runc -v

If the output is 1.0.0-rc93, then you have the problematic version installed.

Possible workarounds

To resolve this issue, we recommend upgrading to GKE on Bare Metal 1.7.0 or a later version.

If upgrading is not an option, you can revert the containerd.io package to an earlier version on the problematic node machines. To do this, connect to the node machine using SSH and run:

Ubuntu

apt install containerd.io=1.4.3-1

CentOS/RHEL

dnf install containerd.io-1.3.9-3.1.el8.