compatibility with numpy>=2.0

[keewis]

• closes Test for compatibility with Numpy prerelease (2.0 is coming this month) #3950

This also replaces a couple of top-level pytest.warns with ignores, which was the simplest way I found to silence the errors caused by the change in behavior of numpy>=2.

Still missing: a way to actually test this with a nightly build of numpy, but since I don't know your CI well enough I'll need help with that.

With numpy>=2, I'm getting a bunch of other failures locally, mostly concerning the removal of numpy.asfarray.

[Zac-HD]

Thanks @keewis! Both the implementation and the test changes look good to me. The CI setup is a little tricky, but we can imitate the "oldest numpy" job:

1. set up a tox environment, like below, with this install command:

   hypothesis/hypothesis-python/tox.ini

   Lines 46 to 56 in 5578efc

   	# This test job runs on the oldest version of CPython we support, against the minimum
   	# version specified in our runtime dependencies. For now, that's the oldest version
   	# with a wheel for non-EOL Python. In future we might deprecate faster per NEP-29.
   	[testenv:py38-oldestnumpy]
   	deps=
   	-r../requirements/test.txt
   	allowlist_externals =
   	bash
   	commands=
   	bash -c "pip install --only-binary=:all: numpy==$(grep 'numpy>=' setup.py | grep -oE '[0-9.]+')"
   	python -bb -X dev -m pytest tests/numpy/ -n auto

2. re-export that for our tooling (no need for the py= argument in this case):

   hypothesis/tooling/src/hypothesistooling/__main__.py

   Line 504 in 5578efc

   standard_tox_task("py38-oldestnumpy", py="3.8")

3. invoke it for GH actions:

   hypothesis/.github/workflows/main.yml

   Line 79 in 5578efc

   - check-py38-oldestnumpy

and then I think we're done 🙂

[Zac-HD]

woohoo, some Numpy-2-specific CI failures!

• looks like we need to wrap this in a pytest.param("a", marks=pytest.mark.skipif(<is numpy >= 2>) because the "a" alias for "S" has been removed
• update this fails_with decorator to include OverflowError on Numpy >= 2

[keewis]

I've seen a lot more than these locally, but maybe those are not running because of the environment? Edit: nope, that was because I copied the tests/numpy restriction from the oldest-numpy job

Most notably, some ghostwriter tests error at test collection time:

  File ".../hypothesis/hypothesis-python/tests/ghostwriter/test_expected_output.py", line 132, in <module>
    ("magic_gufunc", ghostwriter.magic(numpy.matmul)),
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".../hypothesis/hypothesis-python/src/hypothesis/extra/ghostwriter.py", line 1301, in magic
    make_(_make_ufunc_body, func, annotate=annotate)
  File ".../hypothesis/hypothesis-python/src/hypothesis/extra/ghostwriter.py", line 1202, in make_
    imp, body = how(*args, **kwargs, except_=except_, style=style)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".../hypothesis/hypothesis-python/src/hypothesis/extra/ghostwriter.py", line 1884, in _make_ufunc_body
    call=_write_call(func, *ascii_lowercase[: func.nin], except_=except_),
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".../hypothesis/hypothesis-python/src/hypothesis/extra/ghostwriter.py", line 781, in _write_call
    args = ", ".join(
           ^^^^^^^^^^
  File ".../hypothesis/hypothesis-python/src/hypothesis/extra/ghostwriter.py", line 784, in <genexpr>
    if p.kind is inspect.Parameter.POSITIONAL_ONLY
       ^^^^^^
AttributeError: 'NoneType' object has no attribute 'kind'

and (after commenting out the first one):

File ".../hypothesis/hypothesis-python/tests/ghostwriter/test_expected_output.py", line 191, in <module>
    ("addition_op_multimagic", ghostwriter.magic(add, operator.add, numpy.add)),
                               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".../hypothesis/hypothesis-python/src/hypothesis/extra/ghostwriter.py", line 1301, in magic
    make_(_make_ufunc_body, func, annotate=annotate)
  File ".../hypothesis/hypothesis-python/src/hypothesis/extra/ghostwriter.py", line 1202, in make_
    imp, body = how(*args, **kwargs, except_=except_, style=style)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".../hypothesis/hypothesis-python/src/hypothesis/extra/ghostwriter.py", line 1884, in _make_ufunc_body
    call=_write_call(func, *ascii_lowercase[: func.nin], except_=except_),
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".../hypothesis/hypothesis-python/src/hypothesis/extra/ghostwriter.py", line 781, in _write_call
    args = ", ".join(
           ^^^^^^^^^^
  File ".../hypothesis/hypothesis-python/src/hypothesis/extra/ghostwriter.py", line 784, in <genexpr>
    if p.kind is inspect.Parameter.POSITIONAL_ONLY
       ^^^^^^
AttributeError: 'NoneType' object has no attribute 'kind'

Also, there appears to be an issue with your formatting checks, it complains that ruff is not installed.

[keewis]

in the most recent commit I removed the limitation to just tests/numpy. This means it runs all tests, which is probably not intended. Instead, maybe we just need to enumerate all those directories that make use of numpy?

[Zac-HD]

Most notably, some ghostwriter tests error at test collection time:

here, I guess we need if p is None or p.kind...

in the most recent commit I removed the limitation to just tests/numpy. This means it runs all tests, which is probably not intended. Instead, maybe we just need to enumerate all those directories that make use of numpy?

Ah, for that copy the "niche" tests - that covers everything with optional dependencies.

(and thanks for pursuing this!)

[keewis]

Ah, for that copy the "niche" tests - that covers everything with optional dependencies.

Not sure if that can work, in scripts/other-tests.sh the pinned version of numpy will be enforced, replacing the nightly wheel we just installed.

[keewis]

here, I guess we need if p is None or p.kind...

It appears the error only occurs on python>=3.11. My guess is that this is because inspect.signature stopped raising for certain functions built in C (but the parameters returned are *args, **kwargs so not really useful) – I can't find the entry in the python changelog, though. I can also reproduce this on versions other than numpy>=2.0, so I think this is unrelated to this PR. No idea why CI doesn't fail, though.

[keewis]

the final failure on numpy>=2.0 appears to be caused by numpy.ndarray gaining the __array_namespace__ protocol.

In the most recent commits I've hacked together a wrapping mechanism that allows removing arbitrary attributes on the array returned by any of the functions in the namespace (I don't know if that also works on properties, though), which can then be used to patch out the __array_namespace__. Not sure if that's the best way to resolve this, though.

Edit: not sure what to make of the windows 3.9 cover+rest CI failure... it says "stack overflow", but the traceback looks like a recursion, which is weird. To make sure it's not the issue, we could explicitly call object.__getattribute__ instead of super().__getattribute__, though.

[Zac-HD]

Edit: not sure what to make of the windows 3.9 cover+rest CI failure... it says "stack overflow", but the traceback looks like a recursion, which is weird. To make sure it's not the issue, we could explicitly call object.__getattribute__ instead of super().__getattribute__, though.

Hmm. Might as well try this - it's certainly suspicious that we hit this with mocks around.

[Zac-HD]

tiny nits, not necessary to merge

[keewis]

Hmm. Might as well try this - it's certainly suspicious that we hit this with mocks around.

Didn't help, see the most recent commit. So I don't know what's wrong here.

[keewis]

my last attempt, replacing __getattribute__ with __getattr__, appears to have worked. I have no clue why, though!

[Zac-HD]

Thanks @keewis! Delighted to see this all working 🙂