Last modified April 4, 2023

High-availability Kubernetes control plane

Platform support
  • GA in v11.4.0


Kubernetes control plane nodes are the nodes that run the Kubernetes API of a workload cluster, as well as some other important components. In the case of Giant Swarm, the control plane nodes also host the etcd database that keeps all the states that configure the cluster, the workloads, and other resources.

Clusters can run in a fully functional way with one control plane node. However this renders the Kubernetes API unavailable in certain cases like a cluster upgrade or even an outage of the underlying infrastructure.

As of workload cluster release v11.4.0 for AWS, new workload clusters are launched with three control plane nodes by default, spread over multiple availability zones. This results in much higher availability of the API during upgrades or other changes to the cluster configuration, plus enhanced resilience against any data center or hardware failures.


Having multiple control plane nodes in different availability zones has several benefits:

  • API downtimes during upgrades are reduced to a minimum. With a single control plane node, upgrading the cluster requires only the control plane node to be terminated and rebooted with a new configuration. This results in several minutes of downtime. With multiple control plane nodes, the nodes get updated one at a time. The API can only become unreachable in the event that a new etcd leader has to be elected. This usually takes only a few seconds.
  • Resilience in case of infrastructure outages. In the case of a failure of a single control plane node’s EC2 instance, the two remaining control plane nodes take over and a replacement control plane node will be launched automatically. Running multiple control plane nodes of a cluster in different availability zones (AZ) also protects against the risk of losing control of the cluster in case of a single AZ downtime. Refer to below for details.
  • Load balancing of API requests. Read-only requests to the Kubernetes API are distributed over all three control plane nodes, so that latencies stay low in comparison to a single control plane node which can suffer from high loads temporarily.

We recommend that all production clusters on AWS are run with high-availability control planes. As a result, this is the default setting starting with workload cluster release v11.4.0 .

Since these benefits come at the cost of additional EC2 instances and additional network traffic across availability zones, it is still possible to create clusters with a single control plane node. This setting is viable for test purposes or clusters which don’t need high availability and resilience.

Use of availability zones

When high-availability control plane is activated for a cluster, the three control plane nodes are spread over different availability zones (AZ). Depending on the number of AZs in the region, the logic is:

  • In case of three or more AZs in the region, each control plane node is assigned to a different AZ, selected randomly.
  • In case of two AZs in the region, both AZs get used. Two of the control plane nodes will share the same AZ.

When converting a cluster with a single control plane node to high availability, the AZ used by the control plane node before is re-used. Additional AZs are assigned by applying the logic described above. Here, the AZ assignment of existing worker node pools is taken into account.

Upgrades from previous releases

When upgrading a cluster to workload cluster release v11.4.0 or later, the cluster will still only have one control plane node after the upgrade. The API unavailability during the upgrade that is expected for a single control plane node will apply for this upgrade.

A conversion to high availability can be triggered after the upgrade is complete.

Conversion from a single control plane node to high availability

Clusters using workload cluster release v11.4.0 or above on AWS with a single control plane node can be converted via our user interfaces and APIs to run high-availability control planes.

Via the web UI

The web UI presents information on the control plane node in the cluster details page. Next to these details you will find a button Switch to high availability…, unless the cluster is currently undergoing an upgrade. Click this button and follow the instructions in the web UI.

Via gsctl

The gsctl CLI as of v0.23.1 provides the gsctl update cluster to change cluster details. Check the reference for the --master-ha flag.

Via the REST API

Check the v5 cluster modification API reference to find out how to convert a cluster programmatically using the Giant Swarm REST API.

Via the Management API

In order to convert a single node control plane to high availability, the cluster’s G8sControlPlane has to be modified. First you have to find the resource for your cluster ID. The following command helps with that:

kubectl get \

As a result, you should find one G8sControlPlane with the name showing in the first output column. Use that name in the kubectl edit command, or a kubectl patch command alternatively. For interactive editing:

kubectl edit <NAME>

The .spec.replicas attribute, which specifies the number of control plane nodes to run, must be changed from 1 to 3.

For non-interactive patching:

kubectl patch <NAME> \
  --type merge -p '{"spec": {"replicas": 3}}

Once that is done, the operators will reconcile the desired state and create the additional control plane nodes.

Technical details

Why three control plane nodes

At Giant Swarm we operate “stacked” clusters where etcd replicas run on the same machines as the Kubernetes control plane nodes. Each control plane node runs control plane components and a member of an etcd cluster.

In order to achieve high availability in an etcd cluster, a quorum (majority of members) is needed. With three members, one member is allowed to become dysfunctional at any time, e. g. through a machine failure or availability zone outage.

The number of three control plane nodes running three etcd members is a good balance between resilience and cost. While it would be technically possible to run five or more control plane nodes, this is not currently supported with Giant Swarm workload clusters.


  • Currently only supported AWS. Not supported on Azure and KVM.
  • Short API downtimes are still possible during cluster modifications, especially when the leader of the etcd cluster (the member that handles write requests) changes. This happens when the node that hosts the etcd leader has to be modified. Typical cases for this would be an upgrade to a newer workload cluster release or the conversion from a single to multiple control plane nodes. These downtimes are expected to last only a few seconds.
  • Conversion of a cluster from high availability (three control plane nodes) to a single control plane node is not possible.