Skip to main content
Lesson 10 of 11

Autoscaling

Install Metrics Server on kind, then let an HPA add and remove Pods from live CPU usage.

Metrics Server

Metrics Server aggregates CPU and memory usage from each node’s kubelet and exposes it through the Kubernetes API, so kubectl top and the Horizontal Pod Autoscaler can make scaling decisions.

Step 1: install. This creates the Deployment, ServiceAccount, RBAC, and Service in kube-system:

kubectl apply -f https://github.com/kubernetes-sigs/metrics-server/releases/latest/download/components.yaml

Step 2: patch for kind. Local kind nodes use self-signed certificates, so Metrics Server cannot scrape kubelets. Append --kubelet-insecure-tls:

kubectl -n kube-system patch deployment metrics-server --type=json \
  -p '[{"op":"add","path":"/spec/template/spec/containers/0/args/-","value":"--kubelet-insecure-tls"}]'

Production warning: skipping TLS verification is for local labs only. In production, use properly signed and verified node certificates.

Step 3: verify. The Pod rolls out again. Wait 30–60 seconds for the first collection cycle.

kubectl get deployment metrics-server -n kube-system   # expect READY 1/1
kubectl top nodes
kubectl top pods -A
Command Purpose
kubectl apply -f <URL> Install resources from a URL
kubectl patch deployment ... Modify a running Deployment in place
kubectl top nodes / kubectl top pods -A Live CPU (millicores) and memory (MiB) usage

Horizontal Pod Autoscaler

An HPA automatically adds or removes Pod replicas of a Deployment based on live CPU or memory utilization.

  • Horizontal scaling means more or fewer Pods.
  • Vertical scaling means making an existing Pod bigger with more CPU or memory.

Metrics Server collects CPU and memory from kubelets. The HPA compares that usage with the target. Above the target, it scales toward max. On target, it makes no change. Below the target, it scales toward min. Both directions change the Deployment’s replica count.

Average CPU (target 50%) Action
90% Scale up toward max replicas
50% Hold steady
20% Scale down toward min replicas

min and max are guardrails. There is no scaling to zero in quiet hours, and no endless scaling during a spike.

Requirement: CPU requests. The percentage is calculated relative to the Pod’s CPU request. A request of 200m while the Pod uses 100m is 50%. Without resources.requests.cpu, the HPA shows <unknown> and will not scale. Requests and limits are covered in resources and scheduling.

kubectl autoscale deployment web-auto --cpu-percent=50 --min=1 --max=5
kubectl get hpa
kubectl describe hpa web-auto

kubectl get hpa shows current usage against the target and the replica count. kubectl describe hpa web-auto shows events and missing metrics or requests.

The declarative equivalent:

apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: web-auto
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: web-auto
  minReplicas: 1
  maxReplicas: 5
  metrics:
    - type: Resource
      resource:
        name: cpu
        target:
          type: Utilization
          averageUtilization: 50