kubectl scale sets a number. A HorizontalPodAutoscaler (HPA) sets it for you, from what the Pods are actually doing.

Create one

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

NAME         REFERENCE               TARGETS       MINPODS   MAXPODS   REPLICAS   AGE
market-web   Deployment/market-web   cpu: 4%/50%   1         5         1          30s

TARGETS reads current over target. The current figure is the average over the Deployment's Pods of usage divided by request. That ratio is why the HPA needs requests: a Pod that requested nothing has no denominator, and TARGETS shows <unknown>.

The arithmetic

Every 15 seconds the controller reads usage from metrics-server, the same source as kubectl top, and computes:

desired = ceil( current replicas × current utilisation / target )

The result is clamped between min and max, and small differences, within ten percent of the target, are ignored.

Up fast, down slow

Scale-up happens as soon as the numbers say so. Scale-down waits out a stabilization window (five minutes by default on a real cluster, one minute here) so that a short dip does not remove Pods that are needed again a moment later. During the wait kubectl describe hpa says ScaleDownStabilized.

Reading it

kubectl describe hpa market-web

Metrics:                                               ( current / target )
  resource cpu on pods  (as a percentage of request):  41% (41m) / 50%
Min replicas:                                          1
Max replicas:                                          5
Deployment pods:                                       4 current / 4 desired
Events:
  Normal  SuccessfulRescale  New size: 4; reason: cpu resource utilization (percentage of request) above target

An HPA and a hand-set replicas: fight each other; the HPA wins at its next sync. Once a Deployment has an HPA, leave spec.replicas alone.