test_exception_handling_no_traceback leaking FDs (LsofFdLeakChecker plugin)

[nicoddemus]

While implementing #2292, it was discovered that LsofFdLeakChecker plugin declares a hook wrapper incorrectly as pytest_runtest_item instead of pytest_runtest_protocol.

AFAICT, the hook was introduced in db5649ec6a8fcd9ef148b9f77797a22cb9f3cda1 already using the incorrect name, so it seems it never checked for leaked file descriptors. @hpk42 could you comment here?

Fixing the hook name enabled the plugin and it is now crashing the test suite because of some leaked fds (example log).

In order to move #2292 forward, the plugin was changed to emit a warning instead of failing the suite.

We need to decide if we want to keep the plugin and fix the tests, or just remove it altogether.

[nicoddemus]

Update: after turning the failure into a warning, it seems only test_exception_handling_no_traceback is leaking fds, and on this Python versions:

• py26: https://travis-ci.org/pytest-dev/pytest/jobs/221875491
• py34: https://travis-ci.org/pytest-dev/pytest/jobs/221875494
• py35: https://travis-ci.org/pytest-dev/pytest/jobs/221875495
• py36: https://travis-ci.org/pytest-dev/pytest/jobs/221875507

[hpk42]

FWIW and just for quick comment (i leave tomorrow for a couple of days) the leak checker did work some time ago as i used it to fix several leaks. The warnings do sound legit so probably should be fixed and the plugin kept ... unless i am missing something.

[nicoddemus]

Thanks @hpk42! Now that it is clear that a single test is failing, I agree it should be fixed.

[The-Compiler]

Two things I noticed:

• The affected file descriptors point to /dev/shm/... (shared memory), so this is probably something in multiprocessing and not pytest
• They show up as DEL whatever that means. The leak checker already filters out lines with (deleted), I'm wondering what the difference is... The manpage says "or ``DEL'' for a Linux map file that has been deleted;".

[blueyed]

JFI: it was introduced in d4fe273b initially (not db5649e like stated initially).

[blueyed]

Tried it with https://github.com/pytest-dev/pytest/pull/5006, but there are still issues. Just wondered how it should be enabled / used to fail single test only - using pytest.fail in pytest_runtest_protocol makes pytest throw an internal error. But using pytest_runtest_setup / pytest_runtest_teardown is not early enough I assume.