User-placeholders prevents scale-down in AKS

I just discovered that user-placeholders do not get migrated away from underutilized nodes. Is this expected behavior? I think they used to do that earlier (I might be mistaken as user activities vary over time), but we have upgraded both jupyterhub and AKS itself over the last few months.

If I just kill the pods they get rescheduled and the nodepool scales down as expected.

Regards
Roy

Changing the container from pause to busybox and running “sleep 3600” fixed the issue. The pod then restarts on busy nodes.