Creating infrastructure MachineSets on Google Compute Platform

On OpenShift, nodes can be created that are dedicated for infrastructure workload. Creating infrastructure MachineSets of the OpenShift documentation explains how to do so. For GCP, this doesn’t work as documented and or the documentation is missing essential details. This how to tries to fill the gaps but must leave open some questions.

This document discusses how to set up infrastructure nodes on a cluster set up using IPI on GCP. The same procedure can be used to create any type of custom node groups. At least in theory …


Please take note of the Open questions.
  1. Create user data secret

    # Get the base url for the machine config endpoint.
    SOURCE=$(oc -n openshift-machine-api get secret worker-user-data -o jsonpath='{.data.userData}' | base64 --decode | jq -r '.ignition.config.append[0].source' | sed 's|/worker||')
    # Get the CA certificate of the machine config endpoint.
    CA=$(oc -n openshift-machine-api get secret worker-user-data -o jsonpath='{.data.userData}' | base64 --decode | jq -r '[0].source')
    # Create ignition config used when creating <role> machines.
    read -r -d '' USER_DATA <<EOF
      "ignition": {
        "config": {
          "append": [
              "source": "${SOURCE}/<role>",
              "verification": {}
        "security": {
          "tls": {
            "certificateAuthorities": [
                "source": "${CA}",
                "verification": {}
        "timeouts": {},
        "version": "2.2.0"
      "networkd": {},
      "passwd": {},
      "storage": {},
      "systemd": {}
    # Place the ignition config into a secret.
    oc -n openshift-machine-api create secret generic <role>-user-data --from-literal=disableTemplating=true --from-literal=userData=${USER_DATA}
  2. Create MachineConfigs and MachineConfigPool

    This is highly speculative and probably not meant to be done like this. See also Open questions.

    Make a copy of all MachineConfigs found with oc get machineconfig -l Rename them accordingly by replacing worker with <role>. Also set the value of the label to <role>. Within 01-<role>-kubelet the kublet Systemd unit is configured. Replace to<role>.

  3. Create a MachineSet for the custom node group

    The following example deviates at several places from what’s documented at Creating infrastructure MachineSets.

    oc apply -f - <<EOF
    kind: MachineSet
      labels: <infrastructureID>
      name: <infrastructureID>-i-a (1)
      namespace: openshift-machine-api
      replicas: 1
          creationTimestamp: null
    <role>: ""
              canIPForward: false
                name: gcp-cloud-credentials
              deletionProtection: false
              - autoDelete: true
                boot: true
                image: <infrastructureID>-rhcos-image
                labels: null
                sizeGb: 128
                type: pd-ssd
              kind: GCPMachineProviderSpec
              machineType: n1-standard-4
                creationTimestamp: null
              - network: <infrastructureID>-network
                subnetwork: <infrastructureID>-worker-subnet (2)
              projectID: <project_name>
              region: us-central1
              - email: <infrastructureID>-w@<project_name>
              - <infrastructureID>-worker (3)
                name: <role>-user-data (4)
              zone: us-central1-a
    1 Use a naming that’s different from workers.
    2 Reuse the worker network. Using a separated network might be desirable. However, the subnetwork doesn’t get created automatically and the worker one is chosen for simplicity.
    3 Use the same tag as worker. This tag is used in a firewall rule giving access to the machine config endpoint. When using <infrastructureID>-<role>, the machine will fail to access the ignition configuration and thus fails to provision. This can be seen within the VMs console output.
    4 Use the custom user data created earlier. Using the worker-user-data will work but then the resulting node will have both the<role> and the` label.

    Once the MachineSet is created, the machines will be provisioned and the nodes will come up with the desired labels.

Open questions

Who’s/what’s supposed to create the subnetwork?

Creating infrastructure MachineSets suggests to use <infrastructureID>-<role>-subnet as the subnetwork within the MachineSet. When doing so, the machine will fail to provision because that subnetwork doesn’t exist. The documentation fails to explain how this subnetwork is supposed to come into existence.

Who’s/what’s supposed to change the firewall rules?

Creating infrastructure MachineSets suggests to use <infrastructureID>-<role> as the machines tag. Those tags are targeted in certain firewall rules. For this case, the rule granting access to the machine config endpoint is relevant. The firewall rules aren’t altered when creating machines with custom tags. As a result, the created VM isn’t allowed to access the machine config endpoint and thus fails to load its ignition configuration.

Who’s/what’s supposed to create the user data secret?

Creating infrastructure MachineSets suggests to use worker-user-data as the user data secret. When doing so, the resulting node, will come up with both the<role> and the label. Reason being, the node role isn’t only set by the MachineSet but also by the MachineConfig. Using the worker-user-data user data secret, will result in the MachineConfigs of the worker MachineConfigPool to be applied. In the MachineConfig 01-worker-kubelet, the label is hardcoded into the arguments to the kublets Systemd unit. Only when creating a custom MachineConfigPool, the nodes won’t receive the worker role.

Who’s/what’s supposed to create the MachineConfig and MachineConfigPool?

Except for the 99-<master|worker>-ssh MachineConfigs, all MachineConfigs do have an owner reference. Their content is fairly specific to a specific cluster and contains things that needs to be updated under certain conditions (for example the cluster CA certificate). This suggests, that creating them manually doesn’t seemed to be intended. The documentation doesn’t hint how those MachineConfig objects come into being.

How can we change MachineConfigs created for master and worker?

As figured above, several MachineConfigs have an owner reference. As a consequence, they can not be changed. Any changes made to them get reverted immediately. Managing Nodes seems to explain this. Machine APIs also has hints on this.