Fix test_task_groups_update_start_stop, again

[crusaderky]

• Closes Flaky test_task_groups_update_start_stop #8084
• See also Fix flakiness in tests caused by WindowsTime #8087

[github-actions]

Unit Test Results

See test report for an extended history of previous test failures. This is useful for diagnosing flaky tests.

       18 files  ±0         18 suites  ±0   9h 25m 2s ⏱️ - 56m 11s
  3 767 tests ±0    3 656 ✔️  - 1     107 💤 ±0  4 ❌ +1 
32 648 runs  ±0  31 063 ✔️  - 2  1 581 💤 +1  4 ❌ +1 

For more details on these failures, see this check.

Results for commit f500fdb. ± Comparison against base commit ac5ddc3.

♻️ This comment has been updated with latest results.

[crusaderky]

Tested 100 times per environment. With anything but the sleep I was seeing time() going 1ms back into the past - on Linux and MacOSX too. Typically I was observing tg.start, which is sampled below, to be marginally less than t1 or t2.

distributed/distributed/worker.py

Line 3008 in ac5ddc3

with context_meter.meter("thread-noncpu", func=time) as m:

The only logical explanation I can come up with is that

• on Windows, time.perf_counter() (which is used by _WindowsTime) is not perfectly synchronized between threads on Python 3.9 - or at least that's my interpretation of https://docs.python.org/3/library/time.html#time.perf_counter
• on Linux and MacOSX, a frequent NTP sync may send time.time() back 1ms occasionally?