How to Configure GKE ComputeClass for Node Pool Auto-Creation

Set spec.nodePoolAutoCreation.enabled: true in your ComputeClass manifest and use the cloud.google.com/compute-class node selector in your pod specs to enable automatic node pool provisioning without requiring cluster-wide Node Auto Provisioning.

GKE ComputeClasses provide a declarative abstraction for binding node-pool-creation rules to logical class names. When you configure GKE ComputeClass for node pool auto-creation, the system automatically manages node pool lifecycle operations based on workload demands, available starting with GKE version 1.33.3-gke.1136000 according to the google/skills source code.

Enable Node Pool Auto-Creation

To activate automatic node pool management, set the nodePoolAutoCreation field in your ComputeClass specification. According to skills/cloud/gke-compute-classes/SKILL.md (lines 50-54), this feature operates independently of the cluster-wide Node Auto Provisioning setting.

apiVersion: gke.googleapis.com/v1
kind: ComputeClass
metadata:
  name: auto-provisioned-class
spec:
  nodePoolAutoCreation:
    enabled: true
  priorities:
  - machineFamily: n4
    machineType: n4-standard-2

Critical distinction: Do not enable the cluster-wide Node Auto Provisioning flag when using this feature. The ComputeClass handles pool creation autonomously (SKILL.md lines 49-54).

Configure Priority Rules for Machine Selection

Define a priorities[] list to establish a fallback ladder for machine family selection. The auto-scaler attempts each priority in sequence, respecting committed use discounts (CUDs) and reservations. As documented in skills/cloud/gke-compute-classes/SKILL.md (lines 31-34 and 84-88), this configuration ensures cost-effective scaling from Spot instances to On-Demand and finally to reserved capacity.

spec:
  priorities:
  - machineFamily: n4
    machineType: n4-standard-2
    spot: true
  - machineFamily: n2
    machineType: n2-standard-4
  - machineFamily: c3
    machineType: c3-standard-8

Schedule Workloads on Auto-Created Pools

Pods must explicitly target the ComputeClass using node selection criteria. The cloud.google.com/compute-class label directs the scheduler to pools managed by that specific class (SKILL.md lines 46-48).

apiVersion: v1
kind: Pod
metadata:
  name: workload-pod
spec:
  containers:
  - name: app
    image: gcr.io/project/app:latest
  nodeSelector:
    cloud.google.com/compute-class: auto-provisioned-class

Handle Taints and Tolerations Correctly

Auto-created node pools automatically receive the cloud.google.com/compute-class toleration. Do not duplicate this taint in your pod specifications, as this breaks scheduling (SKILL.md lines 55-58). Manual pools still require both label and taint configuration.

For specialized hardware, add specific tolerations:

• GPU nodes: Require nvidia.com/gpu:NoSchedule toleration
• Spot nodes: Require cloud.google.com/gke-spot=true:NoSchedule toleration

These requirements are documented in skills/cloud/gke-compute-classes/SKILL.md (lines 71-82).

tolerations:
- key: cloud.google.com/gke-spot
  operator: Equal
  value: "true"
  effect: NoSchedule

Complete Implementation Example

The following configuration demonstrates a production-ready ComputeClass with node pool auto-creation enabled, referencing the templates in skills/cloud/gke-compute-classes/assets/system-pool-compute-class.yaml.

apiVersion: gke.googleapis.com/v1
kind: ComputeClass
metadata:
  name: spot-fallback
spec:
  nodePoolAutoCreation:
    enabled: true
  priorities:
  - machineFamily: n4
    machineType: n4-standard-2
    spot: true
  - machineFamily: n2
    machineType: n2-standard-4
  - machineFamily: c3
    machineType: c3-standard-8

Target this class with a pod specification that includes the appropriate tolerations for Spot instances:

apiVersion: v1
kind: Pod
metadata:
  name: example-pod
spec:
  containers:
  - name: app
    image: gcr.io/my-project/my-app:latest
    resources:
      requests:
        cpu: "2"
        memory: "4Gi"
  nodeSelector:
    cloud.google.com/compute-class: spot-fallback
  tolerations:
  - key: cloud.google.com/gke-spot
    operator: Equal
    value: "true"
    effect: NoSchedule

Summary

• Enable auto-creation by setting spec.nodePoolAutoCreation.enabled: true in the ComputeClass manifest.
• Avoid cluster-level NAP—Node Auto Provisioning is not required and should remain disabled.
• Define priorities to cascade from Spot → On-Demand → Reservations for cost optimization.
• Use node selectors with the key cloud.google.com/compute-class to bind pods to specific classes.
• Skip redundant taints—auto-created pools already carry the cloud.google.com/compute-class toleration.
• Add hardware tolerations for GPU (nvidia.com/gpu) or Spot (cloud.google.com/gke-spot) workloads only when targeting those specific node types.

Frequently Asked Questions

What GKE version supports ComputeClass node pool auto-creation?

Node pool auto-creation for ComputeClasses requires GKE version 1.33.3-gke.1136000 or later. This version introduces the nodePoolAutoCreation API field that allows ComputeClasses to manage node pools independently of the cluster-wide Node Auto Provisioning feature.

Do I need to enable Node Auto Provisioning on my cluster to use ComputeClass auto-creation?

No. The ComputeClass nodePoolAutoCreation feature operates independently of cluster-wide Node Auto Provisioning. According to skills/cloud/gke-compute-classes/SKILL.md (lines 49-54), enabling the cluster-wide NAP flag is unnecessary and potentially conflicting when using ComputeClass-scoped auto-creation.

Why are my pods stuck pending when targeting an auto-created ComputeClass pool?

Pending pods usually indicate a missing node selector or an incorrect toleration configuration. Ensure your pod spec includes nodeSelector: cloud.google.com/compute-class: <NAME> (SKILL.md lines 46-48). Additionally, verify you have not manually added the cloud.google.com/compute-class taint to your pods, as auto-created pools already include this toleration internally (SKILL.md lines 55-58).

How do I schedule GPU workloads on auto-created node pools?

When your ComputeClass targets GPU machine families, add the nvidia.com/gpu:NoSchedule toleration to your pod spec. The auto-created node pools will automatically apply the corresponding taint. This pattern is documented in skills/cloud/gke-compute-classes/SKILL.md (lines 71-82) alongside Spot instance toleration requirements.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}