Stop 08 · Kubernetes & OrchestrationTopics · 9/12

Horizontal Pod Autoscaling

In short

Traffic doubles every evening when people order dinner, then drops at midnight. Paying for peak capacity all day is wasteful; the Horizontal Pod Autoscaler adds pods when busy and removes them when quiet.

Automatically increasing or decreasing the number of running Pod replicas based on observed metrics like CPU or memory usage — a core mechanism for handling variable traffic without manual intervention.

The mascot opens an extra counter in a bank hall as a long queue of customers grows.

Think of it like this

Extra counters at the bank

On salary day the bank queue stretches out the door, so the manager opens more counters. By afternoon, it's quiet and those staff go back to other work. The HPA is that manager, watching the queue (CPU or requests) and opening or closing counters (pods).

Watch it happen

Replicas through an evening rush

web-1

web-2

1/56 pm. Two pods, average CPU 40% against a 70% target. The HPA leaves things alone.

The big ideas

What you need to know

Target, not threshold

You set a target like 70% CPU. The HPA calculates how many pods would bring the average back to it.

Requests are required

CPU utilisation is measured against each container's CPU request. No requests set, no percentage to scale on.

Metrics come from somewhere

Install metrics-server for CPU and memory. Custom metrics, like queue length, need an adapter or a tool like KEDA.

Scale down slowly

The HPA waits before removing pods, so a short dip doesn't cause flapping up and down.

See it in code

Walk through the code

yaml

1apiVersion: autoscaling/v22kind: HorizontalPodAutoscaler3metadata: { name: web }4spec:5  scaleTargetRef: { apiVersion: apps/v1, kind: Deployment, name: web }6  minReplicas: 27  maxReplicas: 108  metrics:9    - type: Resource10      resource:11        name: cpu12        target: { type: Utilization, averageUtilization: 70 }

1/3This HPA controls the replica count of the web Deployment.

Watch it work with kubectl get hpa web -w while you generate load. The TARGETS column shows current versus target utilisation.

Save yourself the pain

Common mistakes

Oops

Creating an HPA for pods with no CPU requests and seeing <unknown> forever.

Do this instead

Set resources.requests.cpu on the containers, and make sure metrics-server is installed and healthy.

Oops

Setting maxReplicas very high and assuming the cluster will cope.

Do this instead

Pods need nodes. Pair the HPA with a cluster autoscaler, and pick a max your budget and database can handle.

Your mission

Make it scale

~35 min

Quick check

An HPA targets 50% CPU. Three pods are averaging 100%. Roughly how many replicas will it ask for?

Go deeper

1Set up a Horizontal Pod Autoscaler on a Deployment targeting CPU utilization.
2Simulate load and observe the autoscaler adding replicas in response.
3Learn why autoscaling requires resource requests/limits to be set correctly to work well.