Seems broken on PyPy + Windows

[oprypin]

I tried to apply "watchdog" to a project here: mkdocs/mkdocs#2385

It works fine for everything, but on Windows PyPy3.6 the watches seem to silently produce no events.

Then I saw that dbc5689 kinda just removed PyPy from CI.

I launched a CI run with PyPy added back: https://github.com/oprypin/watchdog/actions/runs/825872726
And indeed, watchdog's own tests are failing just a well.

So, is PyPy supposed to be supported still?

[BoboTiG]

Well, PyPy is incredibly slow on Windows 😮
I added back PyPy CI in #794, and tests failures seems related that is is very slow. Compare a normal run (45 sec) to a PyPy run (15 minutes), that's horrible.

PyPy is known to be slow with ctypes. There may be room for improvements I guess.

To be complete, the old CI was testing PyPy on GNU/Linux only, it was never tested on Windows.

[oprypin]

I don't think it's right to assume that the tests fail because they are slow. I would assume the opposite - that they are slow only because they fail (with the symptom being a timeout). Some tests succeed only occasionally, and when they do, they're very fast then.

[oprypin]

So I added a sleep here

diff --git a/tests/test_emitter.py b/tests/test_emitter.py
index 708a0378b..f44aff602 100644
--- a/tests/test_emitter.py
+++ b/tests/test_emitter.py
@@ -88,6 +88,7 @@ def start_watching(path=None, use_full_emitter=False, recursive=True):
         emitter.suppress_history = True
 
     emitter.start()
+    time.sleep(0.01)
 
 
 def rerun_filter(exc, *args):

And it makes the tests pass just fine and quickly.

What this shows us is that

1. The problem seems to be that modifications that happen instantly after starting a watch don't get detected.
2. My comment (with the argument against your comment) was fully correct.

[oprypin]

Forgot CI link with passing tests: https://github.com/oprypin/watchdog/runs/2549604801?check_suite_focus=true

[BoboTiG]

I stand corrected, thanks for the investigation.
Do you mind opening a PR with your fix + mine + changelog entry?

[oprypin]

Uh sorry, what I have there is not a fix. The library really has a flaw, and adding a sleep into tests would only conceal it.

[BoboTiG]

The test fix may still be revelant for later though.
Let's see if anyone has an idea how to deal with that then.

[BoboTiG]

Done with 50af6eb.

[oprypin]

I found that this delay is a general issue on Windows. There is always a chance that events right after start will be missed, just that on PyPy that chance was very high.

I am presenting 2 runs of unittests (not of watchdog but of some other project using watchdog)
where the test suite is repeated 50 times per platform

The only difference between the two runs is the added sleeps:
https://github.com/oprypin/mkdocs/compare/ac66e7de5ced56007b346ff18b7709746780970b..c6923117a39cb2a5ed335e709960997866c1300d
-- The sleeps are analogous to those that I added to watchdog.

Without sleeps, on Windows it fails 13 out of 200 times (total over all versions): https://github.com/oprypin/mkdocs/actions/runs/857713565
With sleeps passes everything: https://github.com/oprypin/mkdocs/actions/runs/857723657
(Of course, now PyPy succeeds consistently in either case :D)