A pending ML training job needing 8 GPUs is a classic Karpenter interview scenario — here's the exact four-step lifecycle an interviewer expects you to walk through.
You'll learn:
- Why the K8s scheduler marks pods unschedulable and how Karpenter's controller watches for that signal
- How Karpenter evaluates all pod constraints at once — resource requests, nodeSelector, nodeAffinity, tolerations, and topology spread
- How it calls the EC2 API to select the right instance (p3.16xlarge for 8 GPUs) in the correct availability zone
- Why Karpenter provisions the node but the K8s scheduler still does the final pod binding — a gotcha that trips up a lot of candidates
Keywords: Karpenter node provisioning, Kubernetes GPU scheduling, pending pods interview question, Karpenter vs cluster autoscaler, K8s scheduler lifecycle
🎧 Listen, then go deeper — DevOps & Cloud interview-prep ebooks at DevOpsInterview.Cloud
