The poll_timeout determines how long the kubernetes runner will wait for a worker pod to start running. The default poll_timeout is 3 minutes, which is a bit too short if k8s cluster is maxed out.

xref https://docs.gitlab.com/runner/executors/kubernetes.html

Example job failure due to the short timeout: https://gitlab.wikimedia.org/repos/releng/cli/-/jobs/59291

I propose increasing the timeout to at least 10 minutes, maybe 20. Note that the default job timeout is 20 minutes.