Modelplane Modelplane docs

GKECluster Custom Resource

This document is for an unreleased version of Modelplane.

This document applies to the Modelplane main branch and not to the latest release v0.5.

A GKECluster provisions a GKE cluster with dedicated node pools for GPU inference and system workloads. It outputs secrets containing the cluster kubeconfig and a GCP service account key that consumers can use to target the cluster.

#Metadata

API version
infrastructure.modelplane.ai/v1alpha1
Kind
GKECluster
Scope
Namespaced
Short names
gke

#Example

Manifest
apiVersion: infrastructure.modelplane.ai/v1alpha1
kind: GKECluster
metadata:
  name: inference-us-central
  namespace: platform
spec:
  region: us-central1
  nodePools:
    - name: system
      role: System
      machineType: e2-standard-4
      nodeCount: 2
    - name: gpu-h100
      role: GPU
      machineType: a3-highgpu-8g
      diskSizeGb: 500
      nodeCount: 0
      maxNodeCount: 4
      zones: [us-central1-a]
      gpu:
        acceleratorType: nvidia-h100-80gb
        acceleratorCount: 8

#Spec

GKEClusterSpec defines the desired state of GKECluster.

# credentials optional object

GCP ProviderConfig or ClusterProviderConfig used to authenticate to the GCP API. Defaults to the ClusterProviderConfig named default.

# name optional string 1–253 chars default: default
# type optional enum: ProviderConfig | ClusterProviderConfig default: ClusterProviderConfig
# kubernetesVersion optional string 1–16 chars default: 1.35

GKE cluster Kubernetes version. Must be a version supported by the REGULAR release channel.

# networking optional object

VPC networking configuration. Defaults are suitable for standalone clusters. Override when VPC-peering multiple clusters to avoid CIDR collisions.

# nodeCidr optional string ≤ 18 chars default: 10.0.0.0/24

Primary IP range for nodes.

# podCidr optional string ≤ 18 chars default: 10.1.0.0/16

Secondary IP range for pods.

# serviceCidr optional string ≤ 18 chars default: 10.2.0.0/16

Secondary IP range for services.

# nodePools required object[] 1–8 items
# diskSizeGb optional integer 10–65536 default: 100

Boot disk size in GB.

# gpu optional object

GPU configuration. Required when role is GPU.

# acceleratorCount optional integer 1–16 default: 1

Number of GPUs to attach per node.

# acceleratorType required string 1–63 chars

GPU accelerator type (e.g. nvidia-tesla-a100, nvidia-h100-80gb, nvidia-l4). Attached to each node along with acceleratorCount.

# machineType required string 1–63 chars

GCE machine type (e.g. e2-standard-4, a2-highgpu-8g, g2-standard-48).

# maxNodeCount optional integer ≤ 1000 default: 8

Maximum number of nodes for autoscaling.

# minNodeCount optional integer ≤ 1000

Minimum number of nodes for autoscaling. Set to 1 or higher for pools that must always be available.

# name required string 1–40 chars

Unique name for this node pool. Used as a suffix in the GKE NodePool resource name.

# nodeCount optional integer ≤ 1000 default: 1

Initial number of nodes.

# role required enum: System | GPU | CPU

Determines what workloads this pool runs. System pools host controllers, gateways, and infrastructure. GPU pools host inference workloads and are tainted to exclude non-GPU pods. CPU pools host CPU inference workloads: a standard node image, no GPU, and the modelplane.ai/cpu label and taint.

# zones optional string[] 1–8 items 1–63 chars
# region required string 1–32 chars

GCP region for the cluster (e.g. us-central1, europe-west4).

#Status

# cache optional object

Observed ModelCache RWX storage state.

# storageClassName optional string ≤ 253 chars

Name of the Modelplane-managed ReadWriteMany StorageClass composed on this cluster for ModelCache PVCs. ModelCache reads this to target the cache PVC.

# nodePools optional object[] ≤ 8 items
# message optional string

The provider’s error, prefixed with the name of the composed resource it’s about. Set only when state is Failed.

# name required string ≤ 40 chars

Name of the node pool in spec.nodePools.

# state required enum: Pending | Provisioning | Ready | Failed

Pending until Modelplane composes the pool’s cloud resources, Provisioning until their provider reports them all Ready, then Ready. Failed while the provider reports that its last create, update or delete of one of them failed.

# secrets optional object[]
# key required string ≤ 253 chars

Key within the Secret that holds the credential data.

# name required string ≤ 253 chars

Name of the Secret.

# type required enum: Kubeconfig | GoogleApplicationCredentials

The type of credential this secret contains. Kubeconfig contains a kubeconfig file with the cluster endpoint and CA certificate. GoogleApplicationCredentials contains a GCP service account JSON key that authenticates to the cluster via GKE IAM.