test_backwards_fork flaky test (plus deadlock after) #5834
Description
Here's what a failure looks like:
07:50:54 ======================================================================
07:50:54 FAIL: test_backwards_fork (__main__.TestMultiprocessing)
07:50:54 backwards() should succeed when called before and after a fork
07:50:54 ----------------------------------------------------------------------
07:50:54 Traceback (most recent call last):
07:50:54 File "test_multiprocessing.py", line 436, in test_backwards_fork
07:50:54 self.assertFalse(p.is_alive())
07:50:54 AssertionError: True is not false
07:50:54
07:50:54 ----------------------------------------------------------------------
07:50:54 Ran 19 tests in 23.804s
07:50:54
07:50:54 FAILED (failures=1, skipped=3)
08:03:07 Build timed out (after 20 minutes). Marking the build as failed.
08:03:07 Build was aborted
First test_backward_fork fails (not good), and then the exit from Python hangs, so that the build is only terminated when timeout happens.
CC @colesbury who originally added the test in b79d74a
Known failure occurrences:
On master:
- https://ci.pytorch.org/jenkins/job/pytorch-builds/job/pytorch-win-ws2016-cuda9-cudnn7-py3-test/2927//console (on master; tests before and after pass)
PRs:
- https://ci.pytorch.org/jenkins/job/pytorch-builds/job/pytorch-linux-xenial-cuda9-cudnn7-py2-test/4624//console
- https://ci.pytorch.org/jenkins/job/pytorch-builds/job/pytorch-linux-xenial-cuda9-cudnn7-py2-test/4623//console
- https://ci.pytorch.org/jenkins/job/pytorch-builds/job/pytorch-linux-xenial-cuda8-cudnn6-py3-test/4685//console