GKE Cluster Architecture

This page describes the architecture of Google Kubernetes Engine (GKE) clusters that run your containerized workloads. A GKE cluster consists of a control plane and worker machines called nodes.

Architecture Overview

Mode Comparison

Component	Autopilot	Standard
Control Plane	Managed by GKE	Managed by GKE
System Components	Managed by GKE	Managed by GKE
Nodes	Managed by GKE	Managed by You

About the Control Plane

The control plane runs core Kubernetes processes and is fully managed by Google:

Control Plane Lifecycle

GKE manages:

• Cluster creation to deletion
• Kubernetes version upgrades (automatic or manual)
• High availability and reliability

Kubernetes API Server

The API server is the hub for all cluster communication:

You can make API calls via:

• Direct calls: HTTP/gRPC
• Indirect calls: kubectl, Google Cloud Console

Cluster State Database

Regardless of the database type, every GKE cluster serves the etcd API in the control plane.

Control Plane and Node Interaction

The control plane:

• Schedules workloads to nodes
• Manages workload lifecycle, scaling, and upgrades
• Manages network and storage resources

Artifact Registry Interactions

GKE pulls container images from Artifact Registry for:

• Cluster creation
• Version upgrades

Best Practice: Deploy across multiple regions for availability during regional outages.

About the Nodes

Nodes are Compute Engine VMs that run your containerized applications:

Node Components

Component	Description
Container Runtime	Runs containers (containerd)
kubelet	Node agent, communicates with control plane
DaemonSets	Per-node agents for logging, networking, etc.

Node Management by Mode

Detailed Mode Comparison

Aspect	Autopilot	Standard
Lifecycle	Fully managed (upgrades, repairs, scaling, creation/deletion)	GKE manages upgrades/repairs; you manage scaling, creation/deletion
Visibility	kubectl only; VMs not visible in console	Full visibility in kubectl, gcloud, and console
Connectivity	No direct VM connection	SSH to underlying VMs
Operating System	Container-Optimized OS only	Choose from multiple OS options
Hardware	Request via ComputeClasses	Configure machine types in node pools

Node Lifecycle in Autopilot

Node Lifecycle in Standard

Best Practices

Practice	Description
Use stdout for logs	Let the platform handle application logs
Multi-region deployment	Availability during regional outages
Choose appropriate mode	Autopilot for simplicity, Standard for control
Enable security features	Workload Identity, network policies

---

Last updated: December 2025