Opscraft
Amazon EKS · Lesson 1 of 11 · 4 min read

EKS architecture and responsibilities

EKS is AWS's managed Kubernetes and the most common way AWS-based companies run Kubernetes. Senior interviews go beyond "AWS runs the control plane": they probe how pods get IAM permissions (IRSA vs Pod Identity), how the VPC CNI assigns IPs and what to do when you run out, how load balancers are provisioned, how Karpenter differs from Cluster Autoscaler, how access is managed with access entries, how you upgrade safely, and how you control cost and multi-tenancy. This topic covers all of it, ending with a model answer describing a Kustomize + Flux setup.

What AWS runs and what you run

Enlarge diagramAWS-MANAGED ACCOUNTYOUR ACCOUNT / VPCAPI serversmulti-AZ, behind a load balanceretcd3 AZs, encryptedscheduler, controllerspatched and scaled by AWSworker nodesEC2, Fargate or Auto Modekubelet, CNI, CoreDNSand your podsLBs, EBS, NAT, endpointsyou pay and designENIsin your subnets
Figure 1. Who runs what in EKS. AWS runs the control plane in its own account; it reaches your nodes through network interfaces placed in your subnets.
Swipe sideways to see all of it, or tap to open it full screen.
  • AWS manages: the control plane — API servers spread across AZs behind a load balancer, etcd across three AZs, the scheduler and controller managers, control plane patching, scaling and availability (99.95% SLA).
  • You manage: worker nodes (unless Fargate or Auto Mode), add-ons and their versions, networking design, IAM, workloads, security policies, upgrades (you trigger them), observability.
  • The control plane connects to your VPC through cross-account ENIs placed in the subnets you choose — that's how the API server reaches kubelets (for logs, exec, webhooks).

Cluster endpoint access

Mode API server reachable from
Public (default) The internet (restrict with allowed CIDRs); nodes reach it via NAT/IGW
Public + private Internet (restricted CIDRs) and privately from within the VPC (nodes use the private path)
Private only Only inside the VPC and connected networks (VPN/Direct Connect/peering/TGW) — kubectl from a bastion, SSM, or a CI runner inside the network

Private (or public with tight CIDRs) is the standard for production.

Enlarge diagramPublicinternet; restrict CIDRsPublic + privatenodes use the private pathPrivate onlyVPC, VPN, Direct Connect
Figure 2. Where the API server can be reached from. Production clusters use private access, or public access locked to a few CIDRs.
Swipe sideways to see all of it, or tap to open it full screen.

Versions, support and pricing

  • EKS supports each Kubernetes minor version for 14 months of standard support, followed by 12 months of extended support at a much higher per-cluster hourly price. Clusters on unsupported versions are auto-upgraded eventually — so plan upgrades as routine work.
  • Pricing: a per-cluster hourly fee (higher in extended support), plus EC2/Fargate, EBS, load balancers, NAT and data transfer.
  • EKS upgrade insights flag deprecated API usage and add-on incompatibilities before upgrades.
Enlarge diagramstandard support: 14 monthsextended: 12 months, higher feethen upgraded automatically by AWS
Figure 3. The EKS support clock for each Kubernetes version. Extended support costs much more per cluster hour; plan upgrades inside standard support.
Swipe sideways to see all of it, or tap to open it full screen.

Useful CLI

aws eks update-kubeconfig --name acme-prod --region ap-south-1   # write kubeconfig context
aws eks describe-cluster --name acme-prod --query 'cluster.{v:version,ep:resourcesVpcConfig}'
aws eks list-addons --cluster-name acme-prod
aws eks describe-addon-versions --addon-name vpc-cni --kubernetes-version 1.31
eksctl get nodegroup --cluster acme-prod

eksctl is a handy CLI for creating and managing clusters (generates CloudFormation); in production most teams use Terraform.