We have received reports from at least one user that a heavily loaded system reports fewer samples than expected.
This user says that perf also reported incorrect data (similarly to parca-agent).
My suspicion is that under heavy load, the kernel is not actually able to invoke the unwinder 19 times a second (either for us or for perf); we should validate this theory, and if it is true, find a way to detect it and adjust the reported sampling rate.