Cluster Autoscaler

Introduction

Cluster Autoscaler is a practical Kubernetes capability, not just a definition to memorize. This lesson explains the problem it solves, why teams use it in production, how it behaves under failure, and how to practice it hands-on.

Purpose of this lesson

By the end, you should be able to decide when Cluster Autoscaler belongs in a design, implement it with a clear manifest or command flow, and troubleshoot the most common failure modes using Kubernetes status, events, logs, metrics, and ownership relationships.

Understanding the topic

Cluster Autoscaler adds nodes when pods cannot schedule and removes underutilized nodes. It connects Kubernetes scheduling pressure to cloud node groups.

Use Cluster Autoscaler when pods are unschedulable because the cluster lacks capacity and nodes can be added or removed automatically.

• Read every object through metadata, spec, and status: who owns it, what should happen, and what the cluster reports actually happened.
• Connect YAML to the responsible component: scheduler, kubelet, controller manager, cloud controller, CSI driver, CNI, CoreDNS, admission webhook, or application runtime.
• Use it only when it improves a real operating concern such as availability, rollout safety, service discovery, isolation, security, cost, or developer workflow.

Visual explanation

Use this mental model when explaining the lesson during design reviews or incidents:

Human / CI / GitOps
  |
  v
kube-apiserver stores desired state
  |
  +-> controller reconciles objects
  +-> scheduler places pods
  +-> kubelet runs containers
  |
  v
events, status, logs, metrics reveal reality

Step-by-step explanation

1. Identify the workload or platform problem first: availability, traffic routing, storage, configuration, identity, policy, scaling, observability, or troubleshooting.
2. Write the smallest useful desired state for Cluster Autoscaler; include labels, namespace, ownership, resource settings, and health checks where relevant.
3. Apply or render the change in a safe environment, then read status and events before assuming the manifest worked.
4. Break one realistic dependency such as a selector, image tag, probe, permission, quota, or endpoint and practice the recovery path.
5. Promote through Git or your release process with a rollback plan, alert coverage, and a short runbook.

Informative example

Inspect unschedulable pods: After applying it, verify both desired and observed state. A production-ready workflow should include kubectl diff, apply, describe, get events, and a rollout or health check when the object supports it.

kubectl get pods --field-selector=status.phase=Pending
kubectl describe pod pending-api
kubectl -n kube-system logs deploy/cluster-autoscaler

Real-world use

HPA creates more replicas than current nodes can fit. Pending pods trigger node group scale-out, then underutilized nodes are removed after traffic falls.

Best practices

• Keep manifests reviewed in Git and treat manual cluster changes as temporary break-glass actions.
• Use standard labels such as app.kubernetes.io/name, part-of, and managed-by so selectors, dashboards, alerts, and cost reports line up.
• Attach ownership, environment, and runbook metadata before resources reach production.

Common mistakes

• Confusing resource exists with resource is healthy. Always inspect status, events, and downstream dependencies.
• Changing selectors, labels, or names casually; these are contracts between controllers, Services, policies, dashboards, and GitOps tools.
• Debugging from memory instead of reading the object: kubectl describe, events, endpoints, and controller logs usually tell the story.

Debugging tips

• Start with kubectl describe and recent events sorted by time; they often identify scheduling, image, probe, volume, or policy failures.
• Compare desired state with live state using kubectl get -o yaml, kubectl diff, and the owning controller's status.
• Follow the traffic or lifecycle path one hop at a time rather than jumping straight to the node or the application code.

Optimization strategies

• Tune requests, limits, probes, and rollout settings from observed production behavior instead of copying defaults.
• Reduce blast radius with namespaces, quotas, PodDisruptionBudgets, topology spread, and progressive delivery.
• Automate validation with CI, policy checks, and GitOps drift detection so correctness is enforced before outages.

Advanced interview questions

Hands-on exercise

Create a workload with high requests, watch pods go Pending, and inspect autoscaler logs for scale-up reasoning.

# Suggested lab loop for Cluster Autoscaler
kubectl create namespace cluster-autoscaler-lab
kubectl -n cluster-autoscaler-lab apply -f lesson.yaml
kubectl -n cluster-autoscaler-lab get all
kubectl -n cluster-autoscaler-lab describe all
kubectl -n cluster-autoscaler-lab get events --sort-by=.lastTimestamp

Summary

Cluster Autoscaler becomes valuable when you connect the API object or command to real operational behavior. Practice the happy path, then deliberately break it so troubleshooting becomes evidence-driven rather than guesswork.