detly
commented
2 years ago I've narrowed it down a bit and found that while some tests (in calloop-subproc) complete in 200s, others take more than an hour (I still haven't found how long):
200 test_chain_single_failure
200 test_chain_single_success
? test_chain_multiple_success_with_failure_and_cleanup
? test_chain_multiple_success_fail
? test_chain_multiple_success_with_failure_and_cleanup_failure
? test_chain_multiple_success_with_cleanup_failure
? test_chain_multiple_success
202 test_listen_one_line
199 test_listen_three_lines
xd009642
I've written a crate calloop-subproc that uses the Calloop event loop and async-process to manage subprocesses in an event loop. It's about 700 lines of code with 9 integration tests. In order to run the tests, the test binary has to build another test binary that can serve as the "thing to run as a subprocess" part of the tests.
Without tarpaulin, the tests take about 5s including the extra binary build. Under tarpaulin, they never finish. They don't seem to hang, however — I can gradually let more and more tests finish by increasing the timeout from 1m to 2m, to 10m, to an hour. So far, however, I have not found a long enough duration for the whole set of (9) tests to finish.
I have tried to create a minimal example, but without much luck. It really seems to be a result of the following factors used together:
...and possibly other factors?
Nonetheless, my tarp-test project sort-of shows the problem, where a usually-20s test runs in 230s under tarpaulin (including the one-off build time). It certainly doesn't show the same magnitude of slowdown that the original project shows though.
I initially thought it was due to signals for subprocess management (kill, sigchld etc.) being blocked, but the tests themselves seem to actually work, so I doubt it's that after all.
I'll keep trying to see if I can identify a particular factor that's having such a disproportionate impact.