Scaling & HPA

Let's think about it this way for a second

When traffic spikes, you want more pods — you can manually scale with `kubectl scale deployment web-deployment --replicas=10`, but if nobody scales it back down when traffic drops at midnight, you're just wasting resources. HPA (Horizontal Pod Autoscaler) monitors CPU/memory usage (or a custom metric) and automatically scales pods up when a target threshold is crossed, and back down when usage falls below it — so 'you don't have to babysit scaling yourself.'

Let's connect this to a real scenario

If you create an HPA with something like `kubectl autoscale deployment web-deployment --min=2 --max=10 --cpu-percent=70`, then once average CPU usage crosses 70%, it'll scale pods up to a max of 10, and scale back down to a minimum of 2 once load drops. To run HPA, you need the metrics-server add-on installed on the cluster (it's the component that supplies CPU/memory data).

kubectl scale = setting the replica count manually
HPA (Horizontal Pod Autoscaler) = auto-scaling based on CPU/memory usage
You need the metrics-server add-on to run HPA

Let's look at it together

bash

# Manual scale
kubectl scale deployment web-deployment --replicas=10

# Create an HPA: 2-10 pods, target 70% CPU
kubectl autoscale deployment web-deployment \
  --min=2 --max=10 --cpu-percent=70

# Watch HPA in action
kubectl get hpa web-deployment --watch

You should see

NAME              REFERENCE                    TARGETS   MINPODS   MAXPODS   REPLICAS
web-deployment    Deployment/web-deployment    45%/70%   2         10        3

5-Minute Try-It

Create an HPA and watch the target/replicas with `kubectl get hpa --watch` (with no traffic, it'll just sit at the minimum replica count — that's normal).

A Quick Word of Caution

HPA calculates its percentage based on the CPU/memory 'requests' value — if you haven't set resource requests in the Pod spec, HPA won't be able to work properly.