Starting with version v2.15.0, KKP introduced etcd-launcher as an experimental feature. Etcd-launcher is a lightweight wrapper around the etcd binary. It’s responsible for reading information from KKP API and flexibly control how the user cluster etcd ring is started.
Prior to v2.15.0, user cluster etcd ring was based on a static StatefulSet with 3 pods running the etcd ring nodes. With etcd-launcher, the etcd StatefulSet is updated to include:
etcd-launcher is an experimental feature. It should not be enabled unless all users clusters are operating nominally. It’s possible that a user cluster etcd ring could fail to enable the feature if the etcd ring was not in a stable condition during the update.
Since this is an experimental feature, it’s disabled by default. There are two modes to enable etcd-launcher support:
In this mode, the feature is enabled on the seed-cluster level. The cluster feature flag will be added to all user clusters.
To enable etcd-launcher, the related feature should be enabled in the Kubermatic CRD. To do that, edit your your KubermaticConfiguration file to include the featureGate:
# Snippet, not a complete file! apiVersion: operator.kubermatic.io/v1alpha1 kind: KubermaticConfiguration metadata: name: <<mykubermatic>> namespace: kubermatic spec: # FeatureGates are used to optionally enable certain features. featureGates: EtcdLauncher: enabled: true
Next, simply apply the updated CRD:
$ kubectl apply -f kubermatic_config.yaml
This will update the KKP Operator Deployment and enable etcd-launcher for all users clusters.
Once the seed controller manager is reloaded, all users clusters will get upgraded to use etcd-launcher. In seed clusters with a large number of user clusters, this might take some time depending on the applied
-max-parallel-reconcile value (default is 10). Refer to the Prepare for Reconciliation Load section for more information.
In this mode, to disable etcd-launcher, you need to revert the changes to the
values.yaml file, and apply again. Additionally, you will need to disable it in each user cluster flag.
To disable etcd-launcher for a specific cluster you need to set the cluster level feature flag to
false. This can be applied before enabling the feature gate on the seed controller manager to prevent the cluster from being upgraded in the first place.
In this mode, the feature is only enabled for a specific user cluster. This can be done by editing the object cluster and enabling the feature flag for etcd-launcher:
$ kubectl edit cluster 27k9rzjzs7
Edit the cluster spec to add the
etcdLauncher feature flag, and set it to
spec: features: etcdLauncher: true
Once the cluster object is updated, The etcd StatefulSet will start updating the etcd ring pods one by one. Replacing the old etcd run command with the etcd-launcher.
In this mode, to disable etcd-launcher, you can simply remove the feature flag from the cluster spec, or set it to
Enabling etcd-launcher enables cluster operators to perform several operational tasks that were not possible before. With the the v2.15.0 releases, etcd-launcher provides the following capabilities:
Prior to version v2.15.0, the user cluster etcd ring ran as a simple and static 3-node ring. With etcd-launcher enabled, it’s now possible to resize the etcd ring to increase capacity and/or availability for user clusters.
Currently, the supported minimum etcd ring size is 3 nodes. This is required to maintain etcd quorum during operations. The maximum supported size is 9 nodes, as recommended by etcd upstream.
To resize the etcd ring for a user cluster, you simply need to edit the cluster:
$ kubectl edit cluster 2j2q98dkzt
And set the required size in the cluster spec:
spec: componentsOverride: etcd: clusterSize: 5
The resizing process is currently a disruptive process. Nodes are added/removed one by one and the etcd ring is restarted to add or remove new nodes, which could disrupt the user cluster.
Unfortunately, the Kubernetes StatefulSet controller doesn’t automatically handle cases where a PV is unavailable. This can happen for example if the storage backend is having availability issues or the node running the etcd pod is using network-attached storage and becomes unavailable. In such cases, manual intervention is required to remove the related PVC and reset the StatefulSet.
With etcd-launcher enabled for a cluster, the recovery process is fully automatic. Once a pod PV becomes unavailable, a controller will kick-in and remove the PVC and reset the StatefulSet.
Resetting the etcd StatefulSet means that all etcd nodes in the ring will be restarted. This means that the user cluster API will have a momentary downtime until the etcd ring is available again.