Scale replicas with a HorizontalPodAutoscaler, based on requests

A HorizontalPodAutoscaler (HPA) adds and removes replicas to keep a metric near a target. For CPU and memory, the target is a percentage of the request, so an HPA on a Deployment with no requests cannot compute utilisation and will not scale.

Practical rules:

• Set requests first, then pick a utilisation target that leaves headroom for spikes while new Pods start.
• Set minReplicas to at least 2 for anything user-facing so one Pod restart is not an outage.
• Set maxReplicas from what your database and downstream services can take, not from what the cluster could fit.
• Do not set spec.replicas in the Deployment you apply from Git when an HPA owns the count, or every deploy resets it.

For queue workers, scale on queue depth through a custom or external metric rather than CPU.

To check, run kubectl get hpa and confirm the TARGETS column shows real numbers. <unknown> usually means missing requests or a missing metrics server.

apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: checkout
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: checkout
  minReplicas: 2
  maxReplicas: 10
  metrics:
    - type: Resource
      resource:
        name: cpu
        target:
          type: Utilization
          averageUtilization: 70