In the PR check cluster, we are seeing quite a lot of instabilities, mainly when the worker nodes are used on 80% or more. The pods are appearing in CrashLoopBackOff and may fix for some time then fail again. The issue was that a node was over-occupied, probably due to some workloads that didn't specify CPU or Memory consumption requests/limits.

Go though the containers and try to find out which are the ones without requests/limits. 

The aim of this issue is not to fix those requests/limits, but to identify the problematic parts and create an issue for appropriate teams.