What a Kubernetes CPU value actually measures: CPU-seconds per second on one logical CPU. How a 250m limit becomes cpu.max 25000 100000 and a 100 ms throttle window, why a request is a weight and not a cap, and why 1000m is a hyperthread, not a core.
nice and renice set a CPU scheduling weight and nothing else: the OOM killer reads oom_score, not NI. Why a lower nice cannot lift a cgroup cpu.max quota and why -10 versus +10 is a weight ratio of about 87 to 1 rather than 20 steps, what each field of top's %Cpu(s) line means and why it is every logical CPU summed, and why 717 tasks with 4 running means 713 of them are asleep.