High Availability on EKS#

In the HA notes for the kubeadm cluster, HA was a project: stacked etcd members, an external etcd option, a keepalived VIP in front of the apiserver, and a failover test where you pull the power on a VM. On EKS, most of that project is already done for you - but “done for you” is not the same as “you have no homework”.

The control plane: AWS’s job#

When you create patientping-eks, AWS runs the control plane across three Availability Zones, with an etcd cluster that tolerates member loss and a regional API endpoint your kubeconfig already points at. There is no VIP to keep alive, no --control-plane-endpoint to design, no etcd quorum to babysit.

What you give up in exchange: no access. No etcdctl snapshot save, no manual defrag, no kubectl exec into the apiserver. Backups and restores of the control plane are AWS’s responsibility (they exist - you just can’t take them). For most teams that trade is a bargain; if you need etcd-level forensics, you’re on kubeadm for a reason.

Not: ha-control-plane notundaki “stable endpoint” ve “external etcd” bölümleri EKS’te anlamsız - endpoint regional DNS, etcd AWS’in kontrolünde. Senin HA sorumluluğun sadece worker katmanı.

The worker layer: your job#

Control plane HA without worker HA is a fancy car with one wheel. Three things make the worker layer survive an AZ loss:

1. Spread the node group across AZs. A managed node group created with multiple subnets (like the two private subnets in our cluster config) places nodes in each. For real AZ coverage, give it all the private subnets in the VPC.
2. Topology spread constraints - the same tool from your scheduling notes:

      topologySpreadConstraints:
        - maxSkew: 1
          topologyKey: topology.kubernetes.io/zone
          whenUnsatisfiable: ScheduleAnyway
          labelSelector:
            matchLabels: { app: patientping-web }

3. Pod Disruption Budgets - from the pod priority and disruption notes. A PDB is what keeps a node drain (or a node group upgrade) from killing every replica at once:

apiVersion: policy/v1
kind: PodDisruptionBudget
metadata:
  name: patientping-web-pdb
spec:
  minAvailable: 1
  selector:
    matchLabels: { app: patientping-web }

Node replacement#

On kubeadm, replacing a dead node means a new VM and kubeadm join. On EKS, delete the instance (or let the node group’s instance refresh do it) and the group launches a replacement that joins automatically - labels and taints included (see eks-node-affinity-taints). For unplanned node loss, Cluster Autoscaler (or Karpenter) keeps a buffer of capacity so the replacement is fast.

Assignment#

Make the PatientPing deployment survive losing an entire node - and prove it.

Cost check: Spreading nodes across AZs doesn’t cost extra; you’re already paying for the nodes. The drain test below is free, just disruptive.

1. Patch the deployment with the topology spread constraint above and make sure the PDB exists:

kubectl apply -f patientping-web.yaml
kubectl apply -f patientping-web-pdb.yaml
kubectl get pdb

2. Confirm both replicas are on different nodes (and therefore, in our multi-AZ group, different AZs):

kubectl get pods -o wide
kubectl get nodes --show-labels | grep zone

3. Simulate losing a node - the EKS version of the “pull the power” failover test from your HA notes:

NODE=$(kubectl get pods -l app=patientping-web -o wide | awk 'NR==2{print $7}')
kubectl drain $NODE --ignore-daemonsets --delete-emptydir-data

4. Watch the PDB and the pods do their job:

kubectl get pods -w

The evicted replica should be recreated on the surviving node/zone. kubectl get pdb shows ALLOWED DISRUPTIONS: 1 before the drain - that’s the budget letting exactly one replica go down at a time.

5. Bring the node back (uncordon it) and check the node group in the console - or simply delete the drained node’s instance and watch a fresh one join.

kubectl uncordon $NODE

Sanity check: tek node’a kill -9 atmak yerine kubectl drain kullanın - drain, PDB’ye saygılı ve kontrollü bir tahliyedir. Gerçek AZ kaybını test etmek isterseniz node’un instance’ını console’dan terminate edin; node group otomatik yenisini açar.