After this lesson, you will be able to: Compare AWS EKS, Google GKE, and Azure AKS, what's managed for you, what's not, and how managed differs from self-hosted.
Almost every production K8s in 2026 is managed. Knowing what 'managed' covers (and doesn't) shapes how you operate the cluster.
Provider runs the control plane (API server, scheduler, controller-manager, etcd), this is the hard, stateful piece. You manage: worker nodes (or use Fargate / GKE Autopilot for serverless nodes), workloads (Deployments etc.), networking integration, IAM, secrets. Provider pricing: AWS EKS $0.10/hr/cluster (~$73/mo) + node cost; GKE free control plane + node cost; AKS free control plane + node cost. Self-hosted is rare outside large platform teams or air-gapped environments.
Tightest IAM integration (IRSA, IAM Roles for Service Accounts let Pods assume AWS roles without long-lived creds). Networking: AWS VPC CNI (each Pod gets a VPC IP, simple security groups). Add-ons: AWS Load Balancer Controller, EBS CSI driver, Cluster Autoscaler, Karpenter (modern node autoscaler). Pricing: $0.10/hr control plane + EC2 / Fargate for workers. Most common in enterprise; ecosystem of vendor integrations is the deepest.
Google invented K8s; GKE is widely considered the most user-friendly managed offering. Autopilot mode: fully serverless, you don't manage nodes at all, pay per Pod CPU/memory. Effective for unpredictable workloads. Built-in monitoring (Cloud Operations), networking (GKE Dataplane V2 with Cilium). Pricing: $0.10/hr per cluster (free for Autopilot management), nodes separately. Best choice if you're starting fresh on GCP.
Free control plane (no hourly cost), nodes you pay for. Tight integration with Entra ID (Azure AD) for cluster auth, same identity layer as the rest of your Azure environment. Azure Container Insights for monitoring, Azure CNI for networking. Often the chosen platform at Microsoft-heavy enterprises. Karpenter and GitOps tools work the same as on EKS.
Treating managed K8s as 'set and forget'. The control plane is managed; everything above is your job. Picking a cloud because of K8s. The cluster will be 5% of your bill; pick the cloud for everything else. Self-hosting K8s as a learning project. Fine for learning; never for prod unless you have a dedicated platform team. Skipping IRSA / Workload Identity. Long-lived AWS / GCP keys mounted into Pods are a leak waiting to happen. Going to multi-region multi-cluster too early. Most apps don't need it; the operational overhead is enormous.
Pick the modern pattern.
Sign in and purchase access to unlock this lesson.