test_concurrent_futures/test_shutdown.py: test_cancel_futures_wait_false flaky

[colesbury]

The test looks like:

cpython/Lib/test/test_concurrent_futures/test_shutdown.py

Lines 243 to 250 in bb66600

	t = ThreadPoolExecutor()
	t.submit(sleep_and_print, .1, "apple")
	t.shutdown(wait=False, cancel_futures=True)
	""")
	# Errors in atexit hooks don't change the process exit code, check
	# stderr manually.
	self.assertFalse(err)
	self.assertEqual(out.strip(), b"apple")

The problem is that if the worker thread starts executing after the shutdown() call, then the task will be canceled (because cancel_futures=True) and nothing will be printed. This happens sometimes with the GIL and more frequently in the free-threaded build with the GIL disabled.

You can reliably trigger the problematic behavior by adding a short sleep at the beginning of the worker thread:

cpython/Lib/concurrent/futures/thread.py

Lines 69 to 70 in bb66600

	def _worker(executor_reference, work_queue, initializer, initargs):
	if initializer is not None:

i.e., add a time.sleep(0.2) before the if statement.

This test has been reported failing previously, but those issues were closed:

• test_concurrent_futures: test_cancel_futures_wait_false() and test_interpreter_shutdown() failed on GHA Windows x64 #89099
• test_concurrent_futures: failure on Windows (x64) #88650 (not entirely clear if it's the same issue)

Linked PRs

• gh-116682: stdout may be empty in test_cancel_futures_wait_false #116683
• [3.12] gh-116682: stdout may be empty in test_cancel_futures_wait_false (GH-116683) #116692
• [3.11] gh-116682: stdout may be empty in test_cancel_futures_wait_false (GH-116683) #116693

[sobolevn]

This might be related? https://github.com/python/cpython/actions/runs/8261299493/job/22598342550

 ======================================================================
FAIL: test_check_output_timeout (test.test_subprocess.ProcessTestCaseNoPoll.test_check_output_timeout)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/Users/runner/work/cpython/cpython/Lib/test/test_subprocess.py", line 286, in test_check_output_timeout
    self.assertEqual(c.exception.output, b'BDFL')
    ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError: None != b'BDFL'

======================================================================
FAIL: test_check_output_timeout (test.test_subprocess.RunFuncTestCase.test_check_output_timeout)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/Users/runner/work/cpython/cpython/Lib/test/test_subprocess.py", line 1694, in test_check_output_timeout
    self.assertEqual(c.exception.output, b'BDFL')
    ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError: None != b'BDFL'

----------------------------------------------------------------------

[colesbury]

@sobolevn I think the underlying cause is different:

• test_check_output_timeout failure is due to starting a subprocess taking too long (>3 seconds) on a heavily loaded machine
• test_cancel_futures_wait_false failure was due to a race condition involving threads

They are similar in that sometimes the output is empty and the tests results are inconsistent.

[colesbury]

Let's open a separate issue for test_check_output_timeout