The model gets the attention. In my fleet, the serious failures came from schedulers, shell assumptions, expired credentials ...
I noticed some performance issues with my Linux machine. So I used namespaces to shine some light on the problem.