Gcp Gke

GKE is Google's managed Kubernetes — technically the most mature managed K8s offering (Google invented Kubernetes). GKE Autopilot is particularly compelling: fully managed nodes (no node pool management), per-pod billing (only pay for running pods), and built-in security hardening. Standard mode gives full node control. Workload Identity maps K8s service accounts to GCP service accounts — enabling pods to access GCP services without stored credentials. For agents: create clusters, manage node pools, retrieve credentials, and deploy workloads. Free tier (1 zonal cluster). Confidence is docs-derived.

GKE API via google-cloud-container Python client. container_client = container_v1.ClusterManagerClient(credentials=credentials). cluster = container_v1.Cluster(name="my-cluster", initial_node_count=3, node_config=container_v1.NodeConfig(machine_type="e2-standard-4"), autopilot=container_v1.Autopilot(enabled=True)). operation = container_client.create_cluster(parent=f"projects/{project}/locations/{region}", cluster=cluster). container_client.get_operation(name=operation.name) polls. Get credentials: subprocess.run(["gcloud","container","clusters","get-credentials",name,"--region",region]) or use kubernetes client with token from credentials. Node pools: container_client.create_node_pool(); container_client.set_node_pool_size() for scaling.

Service Account auth: google.oauth2.service_account.Credentials with container.googleapis.com/clusters.create etc. scopes. Application Default Credentials: container_v1.ClusterManagerClient() (auto-picks ADC). IAM roles: roles/container.clusterAdmin for cluster management. HTTPS enforced. Workload Identity: bind K8s service account to GCP service account via iam.gserviceaccount.com/allowedTo annotation. Container Registry / Artifact Registry: configure imagePullSecrets or Workload Identity for pod image access.

Cluster creation: async — poll operation until done (5-10 minutes for Autopilot, 3-5 minutes for Standard with pre-warmed). Node pool scale: 2-3 minutes. GKE Autopilot: nodes provisioned per-workload — pods may take 60-90s to schedule on new nodes. Cluster upgrade: gradual node upgrade with PodDisruptionBudget respect. GKE uptime at status.cloud.google.com. Resource quotas: project-level CPU/memory limits — request increases for large clusters. Spot VMs (Autopilot): significant cost reduction with 30-second eviction warning. Maintenance windows: configure to avoid upgrades during peak traffic.

cloud.google.com/kubernetes-engine/docs covers cluster creation guide, Autopilot documentation, Workload Identity setup, node pool management, and GKE Gateway (ingress). Getting started: GCP free tier (1 zonal cluster free), gcloud container clusters create — cluster ready in under 5 minutes (Autopilot). google-cloud-container: pip install google-cloud-container. Terraform: google_container_cluster. Community via Google Cloud Community and Stack Overflow (google-kubernetes-engine tag). Comprehensive documentation; GKE Autopilot documentation is particularly good.