Seems like a reasonable change, but I'd like to understand how you ran into it. Does this only happen if on_cpu_only is true? Can you give any details or repro steps?

I found this in pyroscope where rbspy is used as a library. There it manifests as just no stacktraces to the user of the library. There I start a puma server, it becomes idle (no requests), all samplers are exiting due to incorrect break, later when puma is not idle, no stacktraces is sent.

I guess I can reproduce with the rbspy binary as well

sudo ./target/debug/rbspy record --pid=642619 -s -c

before

Time since start: 1s. Press Ctrl+C to stop.
Summary of profiling data so far:
% self  % total  name
Warning: no profile samples were collected
Wrote raw data to /root/.cache/rbspy/2025-06-03-iIvBIWeeW0.raw.gz
Wrote formatted output to /root/.cache/rbspy/2025-06-03-onWD4497lK.flamegraph.svg

after:

Time since start: 10s. Press Ctrl+C to stop.
Summary of profiling data so far:
% self  % total  name
100.00   100.00  fibonacci - /home/korniltsev/p/ruby_app/app.rb:17
...

Does this only happen if on_cpu_only is true

I have not tried on_cpu_only=false

Reading the code i suspect only oncpu is affected

Out of curiosity: is there a reason to use continue and skip the sleeping ?

No, that was my mistake. I spent some time trying to develop a test for this scenario and misremembered where the sleep code was. Thanks for asking 👍