World K5: Lunch Rush

Scale out at lunch rush: HPA

You will learn: how an HPA adds trays when they get busy, and removes them later.

Checked with Kubernetes 1.37 on 31 Oct 2026

The problem

At noon the queue triples. At 3 pm the food court is empty. How many trays do you want?

Watch

You are here: cluster kind-dev namespace default

~/shop $ kubectl get hpa shopNAME   REFERENCE         TARGETS        MINPODS   MAXPODS   REPLICAS   AGEshop   Deployment/shop   cpu: 81%/50%   3         10        10         72s~/shop $ kubectl get deployment shopNAME   READY   UP-TO-DATE   AVAILABLE   AGEshop   10/10   10           10          72s

A (HPA) is a helper manager. It checks every 15 seconds how busy the trays are. Busy? It orders more. Lunch took this shop from 3 to 10.

Predict

“Busy” means CPU use as a share of each tray’s request. The target is 50%. Now 3 trays run at 100%. How many trays does the HPA want?

Try it

An HPA is YAML too. This one will not apply. Read kubectl’s error, then fix the one line.

Fix the manifest, then press Apply.

hpa.yaml12 lines

Try it

Now do it with one command. Make the HPA, then press Lunch rush. Watch the trays grow.

kubectl autoscale deployment shop --cpu=50% --min=3 --max=10

You are here: cluster kind-dev namespace default

shop3/3 ready

dev-control-planeOpenManager

No trays here.

dev-workerOpen

shop·6l2p4

dev-worker2Open

shop·5gkrw
shop·nlvkf

Time: +0s (sped up)

Watch

You are here: cluster kind-dev namespace default

~/shop $ kubectl get hpa shopNAME   REFERENCE         TARGETS              MINPODS   MAXPODS   REPLICAS   AGEshop   Deployment/shop   cpu: <unknown>/50%   3         10        3          32s

cpu: <unknown>/50% means: these trays have no CPU request. Without a request, “50% busy” has no meaning. Add resources.requests.cpu to the container.

Predict

Lunch ends. CPU drops to 5%. When do the extra trays go away?

You did it!

You can now let an HPA add and remove trays for you.

Go deeper (optional)

Scaling down waits 5 minutes by default. Scaling up may double, or add 4 trays, per 15 seconds. Tune both with behavior in the YAML. In Kubernetes 1.37, minReplicas: 0 (scale to zero) is beta and on by default. It needs an object or external metric, not CPU.

Real rep (optional): try the HPA walkthrough on kubernetes.io. Use a throwaway kind cluster.