Using machine-controller

Kubermatic machine-controller is an open-source Cluster API implementation that takes care of:

• creating and managing instances for worker nodes
• joining worker nodes a cluster
• reconciling worker nodes and ensuring they are healthy

Kubermatic machine-controller allows you define all worker nodes as Kubernetes object, more precisely, as MachineDeployments. MachineDeployments work similar to core Deployments. You provide information needed to create instances, while machine-controller creates underlying MachineSet and Machine objects, and based on that, (cloud) provider instances. The (cloud) provider instances are then provisioned and joined the cluster by machine-controller automatically.

machine-controller watches all MachineDeployment, MachineSet, and Machine objects all the time, and if any change happens, it ensures that the actual state matches the desired.

As all worker nodes are defined as Kubernetes objects, you can manage them using kubectl or by interacting with the Kubernetes API directly. This is a powerful mechanism because you can create new worker nodes, delete existing ones, or scale them up and down, using a single kubectl command.

Kubermatic machine-controller works only with natively-supported providers. If your provider is natively-supported, we highly recommend using machine-controller. Otherwise, you can use KubeOne Static Workers.

For the required permissions of the machine-controller checkout the machine-controller requirements.

Creating Initial Worker Nodes

The initial worker nodes (MachineDeployment objects) can be created on the provisioning time by defining them in the KubeOne Configuration Manifest or in the output.tf file if you’re using Terraform.

If you’re using the KubeOne Terraform Integration, you can define initial MachineDeployment objects in the output.tf file under the kubeone_workers section. We already define initial MachineDeployment objects in our example Terraform configs and you can modify them by setting the appropriate variables or by modifying the output.tf file.

If you are not using Terraform, other options is to use a yaml file definition to provide MachineDeployment CRD values. MachineDeployment CRD is part of Kubernetes Cluster API - which is a spec from Kubernetes project itself. Go spec for this CRD can be found here.

If you want to use yaml approach to provide machine-controller deployment, then do not define kubeone_workers object in output.tf of terraform. Instead, provide the values via machines.yaml file as below.

# Create a machines.yaml with MachineDeployment resource definition
# Apply this file directly using kubectl
kubectl apply -f machines.yaml

Some examples of possible machine deployment yamls can be found in Machine-controller examples directory

Otherwise, you can also define MachineDeployment objects directly in the KubeOne Configuration Manifest, under dynamicWorkers key. You can run kubeone config print --full for an example configuration.

Creating Additional Worker Nodes

If you want to create additional worker nodes once the cluster is provisioned, you need to create the appropriate MachineDeployments manifest. You can do that by grabbing the existing MachineDeployment object from the cluster or by using KubeOne, such as:

kubeone config machinedeployments --manifest kubeone.yaml -t tf.json

This command will output MachineDeployments defined in the KubeOne Configuration Manifest and tf.json Terraform state file. You can use that as a template/foundation to create your desired manifest.

Inspecting Worker Nodes

If you already have a provisioned cluster, you can use kubectl to inspect nodes in the cluster.

The following command returns all nodes, including control plane nodes, machine-controller managed nodes, and nodes managed using any other way (if applicable).

kubectl get nodes

All nodes should have status Ready. Additionally, worker nodes have <none> set for roles.

If you want to filter just nodes created by machine-controller, you can utilize the appropriate label selector.

kubectl get nodes -l "machine-controller/owned-by"

You can use the following command to list all MachineDeployment, MachineSet, and Machine objects. KubeOne deploys all those objects in the kube-system namespace. You can include additional details by using the -o wide flag.

kubectl get machinedeployments,machinesets,machines -n kube-system

The output includes various details, such as the number of replicas, cloud provider name, IP addresses, and more. Adding -o wide would also include information about underlying MachineDeployment, MachineSet, and Node objects.

Editing Worker Nodes

You can easily edit existing MachineDeployment objects using the kubectl edit command, for example:

kubectl edit -n kube-system machinedeployment <machinedeployment-name>

This will open a text editor, where you can edit various properties. If you want to change number of replicas, you can also use the scale command.

Scaling Worker Nodes

The MachineDeployment objects can be scaled up and down (including to 0) using the scale command:

# Scaling up
kubectl scale -n kube-system machinedeployment <machinedeployment-name> --replicas=5

# Scalding down
kubectl scale -n kube-system machinedeployment <machinedeployment-name> --replicas=2

Scaling down to zero is useful when you want to “temporarily” delete worker nodes, i.e. have the ability to easily recreate them by scaling up.

# Scalding down
kubectl scale -n kube-system machinedeployment <machinedeployment-name> --replicas=0