Baffling random failures in core Python behaviour in Windows CI

[crusaderky]

I'm seeing very frequent (once every 2-3 runs) random failures in CI, which happen

• exclusively on windows-latest runners, and
• exclusively on Python 3.13 and 3.14 (not 3.14t).
• I cannot yet say conclusively if they only happen in actions/setup-python or also inside actions/cibuildwheel.

The failures are as follows:

C:\hostedtoolcache\windows\Python\3.13.9\x64\Lib\site-packages\hypothesis\statistics.py:90: in describe_statistics
    runtime_ms = format_ms(t["runtime"] for t in cases)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

times = <generator object describe_statistics.<locals>.<genexpr> at 0x00000200EA90CAD0>

    def format_ms(times: Iterable[float]) -> str:
        """Format `times` into a string representing approximate milliseconds.
    
        `times` is a collection of durations in seconds.
        """
        ordered = sorted(times)
        n = len(ordered) - 1
        if n < 0 or any(math.isnan(t) for t in ordered):  # pragma: no cover
            return "NaN ms"
>       lower = int(ordered[math.floor(n * 0.05)] * 1000)
                            ^^^^^^^^^^^^^^^^^^^^
E       OverflowError: cannot convert float infinity to integer

C:\hostedtoolcache\windows\Python\3.13.9\x64\Lib\site-packages\hypothesis\statistics.py:58: OverflowError
============================ slowest 10 durations =============================
1.06s call     tests/test_gemm.py::test_threads_share_input
0.59s call     tests/test_dotv.py::test_threads_share_input
0.42s call     tests/test_dotv.py::test_memoryview_noconj
0.25s call     tests/test_gemm.py::test_memoryview_notrans

times is a collection of runtimes for a test and I expect it to be fairly short - anywhere between 0 and a few thousands elements.
Note the underline, which is saying that n * 0.05 is valued inf.

Either:

• list.__len__ returned float inf instead of the expected (fairly small) integer, or
• a small integer * 0.05 returned inf.

Both of the above are absurd and I have no explanation.

[ngoldbaum]

@Liam-DeVoe have you ever seen anything like this? Somehow in hypothesis internals, pure-python expressions aren't evaluating to the correct result.

[crusaderky]

Example: https://github.com/explosion/cython-blis/actions/runs/20043099915

[Liam-DeVoe]

This is completely cursed and I have have no explanation beyond the two hypotheses you've posed 😄. I've never seen something like this before

[rgommers]

I'm seeing these as well now on a new PR, it seems likely that this is memory corruption from one of the tests, that then manifests itself as impossible pure Python behavior. That it's exclusively on Python 3.13 and 3.14 on Windows points in that direction as well; the switch to mimalloc was made in 3.13 I believe.

[rgommers]

It happens for 3.14t as well: https://github.com/explosion/cython-blis/actions/runs/23814279556/job/69409454711?pr=155