CI/BLD: don't use strict xfail for '%m.%Y' format in test_hypothesis_delimited_date

[jorisvandenbossche]

Closes #62093

The last days, some linux wheel builds have been failing while testing with:

   ________________ test_hypothesis_delimited_date[%m %Y-False-.] _________________
  [XPASS(strict)] parse_datetime_string cannot reliably tell whether e.g. %m.%Y is a float or a date

I am not directly sure what is going on (it also only happens on wheel builds, and it is a hypothesis tests, so I am not even sure when this happens (i.e. with which input data). Do we print somewhere some hash so we can reproduce hypothesis failures?).
And it is also not happening consistently on the same builds (in one nightly build, it failed for cp312-musllinux_aarch64, but then in another for cp313, in another for manylinux, etc).

So for now just making this a non-strict xfail to get the nightly wheels working

[jbrockmendel]

i think ive seen this locally too a couple times.

@mroeschke didn't you have a recent PR that tinkered with hypothesis settings?

[mroeschke]

Yeah, I tweaked hypothesis to run less max_examples, but I've seen this flakily fail over the years so OK to set strict=False here.

(We should probably increase the hypothesis verbosity since I've had trouble reproducing this locally in the past)

[mroeschke]

Thanks @jorisvandenbossche

[jbrockmendel]

do we have any other ideas to troubleshoot before the cement sets on this? strict=False makes a test pretty useless

[mroeschke]

do we have any other ideas to troubleshoot before the cement sets on this?

Probably just run this test without the xfail marker (in CI) and increased hypothesis verbosity to actually see what input data is failing. I've not been able to reproduce this locally in the past

[jorisvandenbossche]

Trying to understand what the test was doing, I just hardcoded it with a random datetime, and with the first one I tried the test is already failing for me .. (I am better than hypothesis ;))

--- a/pandas/tests/tslibs/test_parsing.py
+++ b/pandas/tests/tslibs/test_parsing.py
@@ -398,12 +398,12 @@ def _helper_hypothesis_delimited_date(call, date_string, **kwargs):
 def test_hypothesis_delimited_date(
     request, date_format, dayfirst, delimiter, test_datetime
 ):
+    test_datetime = datetime(2000, 2, 1, 9)
     if date_format == "%m %Y" and delimiter == ".":
         request.applymarker(
             pytest.mark.xfail(
                 reason="parse_datetime_string cannot reliably tell whether "
                 "e.g. %m.%Y is a float or a date",
-                strict=False,
             )
         )
     date_string = test_datetime.strftime(date_format.replace(" ", delimiter))

and then running this gives two XPASS(strict) failures.

So in practice, I think this test only fails if the month is larger than 10, because for formatted date strings like '02.2000', the parsing function works "fine" (it gives complete garbage as far as I can see, but it gives the same as dateutil, and that is what is being tested):

>>> from pandas._libs.tslibs import parsing
>>> parsing.py_parse_datetime_string("02.2000")  # was formatted as %m.%Y, so should be 2000, 2, 1
datetime.datetime(1, 1, 2, 0, 0)
>>> parsing.py_parse_datetime_string("10.2000")
...
ValueError: Given date string "10.2000" not likely a datetime

So in summary, we can probably tune that xfail check to allow it be strict, but I am also not entirely sure what is actually being tested for this case of format string ...