Cloud provider — GCP: FAQ

How do I create a cluster?

1. Set up a cloud environment.
2. Enable the module or pass the --extra-config-map-data base64_encoding_of_custom_config flag with the module parameters to the install.sh script.
3. Create one or more GCPInstanceClass custom resources.
4. Create one or more NodeGroup custom resources for managing the number and the process of provisioning machines in the cloud.

Adding CloudStatic nodes to a cluster

For the VMs, you want to add to a cluster as nodes, add a Network Tag similar to the cluster prefix.

You can find out prefix using the command:

d8 k get mc global -o jsonpath='{.spec.settings.prefix}{"\n"}'

If the value is empty, get it from the deprecated cloud.prefix parameter of ClusterConfiguration (the prefix has moved to the ModuleConfig global):

d8 k -n kube-system get secret d8-cluster-configuration -o json | jq -r '.data."cluster-configuration.yaml"' \
  | base64 -d | grep prefix

How to create preemptible nodes in GCP?

Preemptible virtual machines in Google Cloud are called preemptible instances. To make a node group order them, set the preemptible parameter of the GCPInstanceClass resource to true.

Example:

apiVersion: deckhouse.io/v1
kind: GCPInstanceClass
metadata:
  name: worker-preempt
spec:
  machineType: n1-standard-4
  diskSizeGb: 30
  preemptible: true
---
apiVersion: deckhouse.io/v1
kind: NodeGroup
metadata:
  name: worker-preempt
spec:
  nodeType: CloudEphemeral
  cloudInstances:
    classReference:
      kind: GCPInstanceClass
      name: worker-preempt
    minPerZone: 2
    maxPerZone: 4
    zones:
      - europe-west4-a

Along with this parameter, Deckhouse Platform (DP) disables the automatic restart of the virtual machine and makes it stop during host maintenance.

Limitations of preemptible nodes in GCP

Consider the following limitations:

• The preemptible parameter applies only to CloudEphemeral node groups, because GCPInstanceClass applies only to them.
• DP does not track Google Cloud notices about an upcoming preemption and does not drain the node in advance. When Google Cloud stops the instance, the node switches to the NotReady state, and the group orders a new node in place of the lost one. Until then, the pods of the stopped node may appear in the Kubernetes API with the Running status even though they are no longer available.
• Place only the workload that tolerates a sudden node loss on preemptible nodes.