CpuPressure

Metrics on the host load. CPU and Network metrics are for the entire host, whereas the scheduler utilization is just for the erlang instance reporting the metric Kernel pressure-stall information for CPU: the share of wall-clock time in which at least one runnable task was WAITING for a CPU rather than running.

This answers a question utilisation cannot. A box can sit at 75% "busy" and still be full, because utilisation counts a thread stalled on DRAM as busy, and because logical CPUs on an SMT core are not independent throughput. It can equally sit at 75% with capacity to spare, when the workload simply has no more parallelism to offer. Pressure separates the two: it counts waiting, not busy-ness.

Read from the process’s own cgroup where available (so a container reports ITS contention, not the host’s), falling back to /proc/pressure/cpu.