Which component are you using?:

Horizontal workload autoscaler.

What version of the component are you using?:

Not relevant.

What k8s version are you using (kubectl version)?:

$ kubectl version
Client Version: v1.30.6-dispatcher
Kustomize Version: v5.0.4-0.20230601165947-6ce0bf390ce3
Server Version: v1.30.5-gke.1443001

What environment is this in?:

Not relevant.

What did you expect to happen?:

When the horizontal autoscaler computes the expected number of replicas, it uses the following formula:

usageRatio := float64(usage) / (float64(targetUsagePerPod) * float64(statusReplicas))

(This formula (or a variation of it) appears a couple of times in this file.)

It's not impossible that statusReplicas (i.e. the status.replicas field of the /scale subresource) equals zero (e.g. if a user kills pods), leading to a division by zero.

The Golang spec allows divisions by zero to trigger traps, although most implementations would return +Inf (which leads to the correct behaviour) or NaN (if usage == 0., in which case the func would return a negative number of replicas).

What happened instead?:

This problem could lead to a panic, or an incorrect behaviour. This could only happen as a result of a race condition, and in the unlikely situation where status.replicas == 0. I'm not sure of its practical significance.

How to reproduce it (as minimally and precisely as possible):

This issue only appears as a result of a race condition. I couldn't produce it.