cpufreq collector is useless

Host operating system: output of uname -a

Linux vivo 5.10.103 #1-NixOS SMP Wed Mar 2 10:42:57 UTC 2022 x86_64 GNU/Linux

node_exporter version: output of node_exporter --version

node_exporter, version 1.2.2 (branch: unknown, revision: v1.2.2)
  build user:       nix@nixpkgs
  build date:       unknown
  go version:       go1.16.13
  platform:         linux/amd64

node_exporter command line flags

  --collector.textfile --collector.textfile.directory /run/prometheus-node-exporter --collector.systemd --collector.systemd.unit-exclude='.+\\.(automount|device|scope|slice)' --collector.diskstats.ignored-devices='^(ram|loop|fd|(h|s|v|xv)d[a-z]|nvme\\d+n\\d+p|mmcblk\\d+p)\\d+$|^zd' \
  --no-collector.ipvs --no-collector.schedstat --no-collector.mdadm --no-collector.nfsd --no-collector.bonding --no-collector.infiniband --no-collector.nfs --no-collector.rapl --no-collector.fibrechannel --no-collector.tapestats --no-collector.nvme \
  --web.listen-address 0.0.0.0:9100

Are you running node_exporter in Docker?

no

What did you do that produced an error?

Looked at the node_cpu_scaling_frequency_hertz metric over time

What did you expect to see?

Representative statistics about CPU frequency

What did you see instead?

Plateaus of high CPU frequency whenever the system is mostly idle.

I suspect that this is because the "boost" frequencies are thermally viable and thus used for the time in which the node exporter is doing its business only when the system is otherwise idle. In other words, the sampling is ineffective because of its own side effects.

Requiring CONFIG_CPU_FREQ_STAT and exporting counters for time_in_state and potentially also trans_table would probably make a lot more sense to see what frequencies the CPU is really running at most of the time. See also the docs for cpufreq-stats: https://www.kernel.org/doc/html/latest/cpu-freq/cpufreq-stats.html

Start with the cpufreq collector and the linked Linux cpufreq-stats documentation. Investigate whether time_in_state and trans_table can provide representative frequency statistics instead of sampling node_exporter's own activity. Done should include a defined metric approach and validation that idle-system observations no longer produce misleading frequency plateaus.