Changing cluster-id for existing vSphere user clusters

Some vSphere user clusters have the cluster-id value in the CSI config set to the vSphere Compute Cluster Name. However, this value is supposed to be a unique identifier of the cluster. If there are multiple vSphere user clusters, those clusters might be using the same cluster-id value, causing users to experience issues with volumes (e.g. the CSI driver failing to find or mount a volume).

As a resolution, you should follow the steps described in this document to change the cluster-id value to a unique value for all affected vSphere user clusters.

Glossary

• User cluster name (name of the user cluster): the value of the NAME column in output of kubectl get clusters (e.g. s8kkpcccfq)
• Seed cluster kubeconfig: kubeconfig file used to access the Seed cluster
• User cluster kubeconfig: kubeconfig file used to access the specific user cluster
• User cluster namespace (in the seed cluster): each KKP user cluster has a namespace in the seed cluster named cluster-<name>, where <name> is the user cluster name as defined above (e.g. cluster-s8kkpcccfq)

Affected KKP versions and vSphere User Clusters

The following vSphere user clusters are affected by this issue:

• Clusters created with the following KKP versions:
  • 2.18.0 - 2.18.9
  • 2.19.0 - 2.19.4
• Clusters migrated to the external CCM/CSI with the following KKP versions:
  • 2.18.0 - 2.18.9
  • 2.19.0 - 2.19.4
  • 2.20.0 - 2.20.1

You can confirm if your user cluster is affected by taking the following steps. The following steps should be done in the seed cluster for each vSphere user cluster.

First, get all user clusters and filter vSphere user clusters using grep:

kubectl --kubeconfig=<seed-cluster-kubeconfig> get clusters | grep vsphere

You should get output similar to the following:

NAME         HUMANREADABLENAME         OWNER                    VERSION   PROVIDER   DATACENTER          PHASE         PAUSED   AGE
s8kkpcccfq   focused-spence            test@kubermatic.com      1.23.8    vsphere    your-dc             Running       false    16h

For each user cluster in the list, note down the cluster name (in this case s8kkpcccfq) and inspect the vSphere CSI cloud-config to check value of the cluster-id field.

kubectl --kubeconfig=<seed-cluster-kubeconfig> get configmap -n cluster-<cluster-name> cloud-config-csi -o yaml

The following excerpt shows the most important part of the output. You need to locate the cluster-id field under the [Global] group.

apiVersion: v1
data:
  config: |+
    [Global]
    user              = "username"
    ...
    cluster-id        = "gph75zqs2q"
...

Finally, verify value of the cluster-id field:

1. If it’s set to the cluster name (e.g. s8kkpcccfq) you’re not affected by this issue and you don’t need to take any action
2. If it’s set to the vSphere Compute Cluster Name, you’re affected and you should follow this guide
3. If the ConfigMap doesn’t exist, you’re still running the in-tree cloud provider, and therefore you’re not affected. Make sure that you don’t use an affected KKP version when migrating the cluster to the external CCM/CSI

Changing the cluster-id value

This guide documents two different approaches to change the cluster-id value.

The first approach is recommended by VMware, however, it requires pausing affected KKP user clusters and stopping the CSI driver for about one hour. For that time, KKP will not reconcile your affected user clusters and you’ll not be able to work with vSphere volumes (attach/detach volumes, create new volumes, delete volumes…).

The second approach assumes changing cluster-id without stopping the CSI driver. This approach is not documented by VMware, however, it worked in our environment. In this case, there’s no significant downtime. Since this approach is not documented by VMware, we heavily advise that you:

• follow the first approach
• if you decide to follow this approach, make sure to extensively test it in a staging/testing environment before applying it in the production

Approach 1 (recommended)

Introduction and Warnings

This approach assumes doing this procedure for each affected user cluster. You can optionally choose one cluster and leave it with the old cluster-id value, but we strongly recommend migrating all your affected user clusters. This approach is based on the official recommendation from VMware.

Pausing User Clusters and Enabling vsphereCSIClusterID Feature Flag

Before starting, make sure to download kubeconfig files for each affected user cluster.

First, pause affected user clusters by running the following command in the seed cluster for each affected user cluster:

clusterPatch='{"spec":{"pause":true,"features":{"vsphereCSIClusterID":true}}}'
kubectl --kubeconfig=<seed-cluster> patch cluster <cluster-name-1> --type=merge -p $clusterPatch
...
kubectl --kubeconfig=<seed-cluster> patch cluster <cluster-name-n> --type=merge -p $clusterPatch

Once done, scale down the vSphere CSI driver deployment in each affected user cluster:

kubectl --kubeconfig=<user-cluster-1> scale deployment -n kube-system vsphere-csi-controller --replicas=0
...
kubectl --kubeconfig=<user-cluster-n> scale deployment -n kube-system vsphere-csi-controller --replicas=0

Wait a minute or two to give time for the CSI controller pods to get scaled down and terminated, and then proceed to change the cluster-id value which you need to do in two places:

1. The cloud-config-csi Secret in the user cluster (in the kube-system namespace)
2. The cloud-config-csi ConfigMap in the user cluster namespace in the seed cluster

Changing the cloud-config-csi Secret in the user cluster

In this section, you’ll change the cloud-config-csi Secret in to set the cluster-id value to the name of the user cluster. The name of the user cluster is value of the NAME column in output of kubectl get clusters for that user cluster. For example, it looks something like s8kkpcccfq.

Since the values of Secrets are base64-encoded, you need to take the config stored in the Secret, decode it, change the cluster-id value, then encode the config and update the Secret.

The following command reads the config stored in the Secret, decodes it and then saves it to a file called cloud-config-csi:

kubectl --kubeconfig=<user-cluster-kubeconfig> get secret -n kube-system cloud-config-csi -o=jsonpath='{.data.config}' | base64 -d > cloud-config-csi

Open the cloud-config-csi file in some text editor:

vi cloud-config-csi

The following excerpt shows the most important part of the file. You need to locate the cluster-id field under the [Global] group, and replace <vsphere-compute-cluster> with the name of your user cluster (e.g. s8kkpcccfq).

[Global]
user              = "username"
password          = "password"
...
cluster-id        = "<vsphere-compute-cluster>"

[Disk]
...

Save the file, exit your editor, and then encode the file:

cat cloud-config-csi | base64 -w0

Copy the encoded output and run the following kubectl edit command:

kubectl --kubeconfig=<user-cluster-kubeconfig> edit secret -n kube-system cloud-config-csi

This will open your default text editor and you should see a Secret such as the following one. Replace <base64-encoded-config> with what you have copied, i.e. with the new config, then save the file and close the editor.

apiVersion: v1
data:
  config: <base64-encoded-config>
kind: Secret
metadata:
  creationTimestamp: "2022-07-06T07:39:51Z"
  name: cloud-config-csi
  namespace: kube-system
  resourceVersion: "560"
  uid: c77cf6e9-e69f-4e0d-b69f-f551e5233271
type: Opaque

Before proceeding to the next step, you need to update the cloud-config-csi Secret in other affected user clusters. Once that’s done, proceed to the next section where you’ll update the cloud-config-csi Secret in the user cluster namespaces in the seed cluster.

Changing the cloud-config-csi ConfigMap in the user cluster namespaces

The ConfigMap is changed in the same way as the Secret, i.e. you need to change the cluster-id value to the name of the user cluster. Run the following kubectl edit command. Replace <cluster-name> in the command with the name of user cluster (e.g. s8kkpcccfq).

kubectl --kubeconfig=<seed-cluster-kubeconfig> edit configmap -n cluster-<cluster-name> cloud-config-csi

This will open the default text editor. You should see a ConfigMap like the following one. Replace <vsphere-compute-cluster> with the name of your user cluster (e.g. s8kkpcccfq), then save the file and exit your editor.

apiVersion: v1
data:
  config: |+
    [Global]
    user              = "username"
    password          = "password"
    ...
    cluster-id        = "<vsphere-compute-cluster>"

    [Disk]
    ...    

Before proceeding to the next step, you need to update the cloud-config-csi ConfigMap for other affected user clusters. Once that’s done, proceed to the next section where you’ll finalize the procedure.

Finaling the procedure

Before proceeding with this section, you MUST WAIT FOR AN HOUR to give time to vSphere to de-register all volumes.

After an hour, patch each affected Cluster object to unpause the cluster. The vsphereCSIClusterID feature flag enabled at the beginning ensures that your cluster-id changes are persisted once the clusters are unpaused.

clusterPatch='{"spec":{"pause":false}}'
kubectl patch cluster <cluster-name-1> --type=merge -p $clusterPatch
...
kubectl patch cluster <cluster-name-n> --type=merge -p $clusterPatch

Wait for a few minutes for KKP to reconcile all unpaused user clusters. The vSphere CSI controller deployment should get scaled up in all user clusters after a few minutes (up to 5 minutes). Make sure the vSphere CSI controller pods are running and that volume operations are working (e.g. you can try creating a pod that uses a vSphere volume).

Approach 2

Introduction and Warnings

This approach assumes changing cluster-id without stopping the CSI driver. This approach also assumes doing this procedure for each affected user cluster like the first approach. You can optionally choose one cluster and leave it with the old cluster-id value, but we strongly recommend migrating all your affected user clusters.

Enabling vsphereCSIClusterID Feature Flag

Before starting, make sure to download kubeconfig files for each affected user cluster.

Start with patching the Cluster object for each affected user clusters to enable the vsphereCSIClusterID feature flag. Enabling this feature flag automatically changes the cluster-id value to the cluster name.

clusterPatch='{"spec":{"features":{"vsphereCSIClusterID":true}}}'
kubectl patch cluster <cluster-name-1> --type=merge -p $clusterPatch
...
kubectl patch cluster <cluster-name-n> --type=merge -p $clusterPatch

Wait for a minute or two to give time to KKP to reconcile and apply changes on all user clusters.

Verifying that ConfigMaps and Secrets are updated

Ensure that the cloud-config-csi ConfigMap in the user cluster namespace in the seed cluster AND the cloud-config-csi Secret in the user cluster (kube-system namespace) are updated. The first command reads the config from the ConfigMap in the user cluster namespace in seed cluster, and the second commands reads the config from the Secret in the user cluster.

kubectl --kubeconfig=<seed-cluster-kubeconfig> get configmap -n cluster-<cluster-name> cloud-config-csi
kubectl --kubeconfig=<user-cluster-kubeconfig> get secret -n kube-system cloud-config-csi -o jsonpath='{.data.config}' | base64 -d

Both the Secret and the ConfigMap should have config with cluster-id set to the user cluster name (e.g. s8kkpcccfq).

[Global]
user              = "username"
password          = "password"
...
cluster-id        = "<vsphere-compute-cluster>"

[Disk]
...

Repeat steps in this section for each affected user cluster and then proceed to the next section.

Restarting the CSI controller pods

Finally, restart the vSphere CSI controller pods in the each affected user cluster to put those changes in the effect:

kubectl --kubeconfig=<user-cluster-kubeconfig-1> delete pods -n kube-system -l app=vsphere-csi-controller
...
kubectl --kubeconfig=<user-cluster-kubeconfig-n> delete pods -n kube-system -l app=vsphere-csi-controller

Wait for the new vSphere CSI controller pods to get healthy. Once pods are healthy, make sure that volume operations are working (e.g. you can try creating a pod that uses a vSphere volume).