Karpenter on EKS Auto Mode: A New Spot Node in 19 Seconds (Real Demo)

Описание: Six pods stuck in Pending, and no room left on your nodes. On a real EKS Auto Mode cluster we add our own Karpenter NodePool that allows Spot and Graviton, scale a Deployment beyond capacity, and watch Karpenter launch a c6gd.2xlarge Spot node that is Ready 19 seconds later. Then we scale back to zero and watch consolidation delete the empty node.

In this episode you will:
• Compare the Cluster Autoscaler (resizes Auto Scaling groups) with Karpenter (launches the instance that fits)
• See Karpenter's loop: Pending pods → NodePool rules → cheapest instance → node joins → consolidation
• Read EKS Auto Mode's built-in NodePools (general-purpose and system) and why you can't edit them
• Write your own NodePool: Spot + On-Demand, arm64 + amd64, a CPU limit and consolidation, on Auto Mode's default NodeClass
• Watch kubectl get nodeclaims -w go ADDED → Ready, then DELETED after scale-down

Timestamps:
0:00 - Intro
0:13 - Cluster Autoscaler vs Karpenter
0:34 - How Karpenter works
0:58 - Step 1: Built-in NodePools
1:20 - Step 2: Your own NodePool
1:42 - Step 3: Scale out
2:20 - Step 4: Consolidation
2:45 - Recap

🔧 What you saw (eu-north-1, October 2026)
Cluster: EKS Auto Mode, Kubernetes 1.36 · kubectl 1.37.1
Built-in NodePool general-purpose: On-Demand only, amd64 only, C/M/R generation 5+, consolidationPolicy WhenEmptyOrUnderutilized, consolidateAfter 30s, disruption budget 10% of nodes
Custom NodePool flex: karpenter.sh/v1, nodeClassRef group eks.amazonaws.com / kind NodeClass / name default, capacity spot + on-demand, arch arm64 + amd64, categories c/m/r, limits cpu 16
Deployment burst: 6 replicas × 1 CPU request, nodeSelector karpenter.sh/nodepool: flex
Karpenter chose a c6gd.2xlarge (Graviton, 8 vCPU) Spot instance within 6 seconds; the node was Ready at 19 seconds and all 6 pods were Running about 30 seconds after the scale-up
After scaling to 0: disruption started about 50 seconds later (reason: Empty) and the NodeClaim was deleted at about 90 seconds

💰 Cost: one Spot node for about four minutes, a few cents including the Auto Mode fee. The cluster is $0.10 per hour plus nodes; delete it when you're done.

✅ Production notes
1. Give Karpenter flexibility: many instance types and sizes, not one family
2. Use Spot and Graviton where you can, with multi-architecture images and apps that tolerate interruptions
3. Set limits on every NodePool so a runaway Deployment can't launch a fleet
4. Keep CPU and memory requests accurate: Karpenter sizes nodes from them
5. Add Pod Disruption Budgets so consolidation never takes down too many replicas at once
6. On self-managed Karpenter the node class is an EC2NodeClass (karpenter.k8s.aws); on Auto Mode it's eks.amazonaws.com NodeClass

📺 Kubernetes on AWS (EKS) playlist (Episode 17 of 25)
⬅ Previous: Ep 16 · Autoscale Pods With the Horizontal Pod Autoscaler →    • Kubernetes HPA on EKS: Autoscale Pods on C...  
➡ Next: Ep 18 · Helm on EKS: Install Apps With Charts →    • Helm 4 on EKS: Install, Upgrade & Roll Bac...  
▶ Full playlist:    • Kubernetes on AWS (EKS)  

All demos are recorded in a dedicated sandbox AWS account. Account identifiers are blurred.

ClouViQ: cloud engineering, clearly explained.

#AWS #EKS #Kubernetes