Scaling policy of k8s deployment

[Nicola-Sergio]

Hi everyone,

We're using KEDA to autoscale Azure DevOps self-hosted agents that are deployed as standard Kubernetes Deployments (not Jobs).

What we've noticed is that KEDA often scales in (removes pods) while the agent is still actively running a pipeline, which leads to failed builds and interrupted jobs.

We would like to better understand: On what basis does KEDA decide it's safe to scale in a pod in a Deployment?