2.4.0: massive test suite failure

[0-wiz-0]

Bug Report

While updating the mypy package in pkgsrc from 2.3.1 to 2.4.0, I ran the test suite.
I see a huge amount of test failures, while the test suite run for 2.3.1 was mostly ok (15 failed, 14149 passed, 403 skipped, 12 xfailed).

To Reproduce

Run pytest after building the package in pkgsrc.

Expected Behavior

2.3.1 output

Actual Behavior

2.4.0 output

Your Environment

• Mypy version used: 2.3.1/2.4.0
• Mypy command-line flags: test suite
• Mypy configuration options from mypy.ini (and other config files): n/a
• Python version used: 3.14.8

This is on NetBSD 11.99.9/x86_64 in case it matters.
For the tests above, I used the exact same versions of the dependencies, just ran the suite once for 2.3.1 and once for 2.4.0:

ast-serialize-0.12.1
attrs-26.1.0
build-1.6.1
coverage-7.16.2
execnet-2.1.2
filelock-4.0.10
iniconfig-2.3.0
installer-1.0.1
librt-0.16.0
mypy_extensions-1.1.0
packaging-26.3
pathspec-1.1.1
pluggy-1.6.0
psutil-7.2.2nb3
pygments-2.21.0
pyproject_hooks-1.3.3
setuptools-84.0.0
test-9.1.1
test-cov-7.1.0
test-xdist-3.8.0
typing-extensions-4.16.0

[ilevkivskyi]

Hm, I guess this may be the problem:

Fatal Python error: _PySemaphore_Init: parking_lot: sem_init failed

Googling this returned

This usually happens because the operating system has run out of available semaphores, file descriptors, or shared memory resources

Most likely this is because we now switched to native parser by default, which parses files in parallel threads. This is probably a NetBSD-specific issue, since everything is fine on other platforms where we run tests.

One possible workaround may be to run tests in small batches.

[ilevkivskyi]

Actually this will probably not work. You should reduce parallelism at the pytest level instead. Right now I see 32 workers in your logs. Can you please try something more modest like pytest -n2?

[0-wiz-0]

I tried with -n2 and it cures the one symptom you mentioned, but there are a lot of problems due to this:

[gw0] netbsd11 -- Python 3.14.8 /usr/pkg/bin/python3.14
data: /tmp/wip/py-mypy/work/mypy-2.4.0/test-data/unit/check-class-namedtuple.test:363:
../wip/py-mypy/work/mypy-2.4.0/mypy/test/testcheck.py:95: in run_case
    self.run_case_once(testcase)
../wip/py-mypy/work/mypy-2.4.0/mypy/test/testcheck.py:200: in run_case_once
    res = build.build(
../wip/py-mypy/work/mypy-2.4.0/mypy/build.py:425: in build
    result = build_inner(
../wip/py-mypy/work/mypy-2.4.0/mypy/build.py:540: in build_inner
    graph = dispatch(sources, manager, stdout, connect_threads)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
../wip/py-mypy/work/mypy-2.4.0/mypy/build.py:4127: in dispatch
    graph = load_graph(sources, manager)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
../wip/py-mypy/work/mypy-2.4.0/mypy/build.py:4375: in load_graph
    manager.parse_all(not_ready)
../wip/py-mypy/work/mypy-2.4.0/mypy/build.py:1049: in parse_all
    parallel_parsed_states, parallel_parsed_states_set = self.parse_files_threaded_raw(
../wip/py-mypy/work/mypy-2.4.0/mypy/build.py:1107: in parse_files_threaded_raw
    available_threads = max(get_available_threads(), self.options.num_workers)
                            ^^^^^^^^^^^^^^^^^^^^^^^
../wip/py-mypy/work/mypy-2.4.0/mypy/util.py:1014: in get_available_threads
    affinity = psutil.Process().cpu_affinity()
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
E   AttributeError: 'Process' object has no attribute 'cpu_affinity'

Most tests still fail due to this:

12135 failed, 1273 passed, 340 skipped, 12 xfailed, 1082 errors

Can the cpu_affinity check be made optional?

There are also a lot of errors about too many open files. I assume a bug with not closing some file descriptors?
The default file descriptor limit on NetBSD is only 1024, but that should be plenty for a test, no?

By terminal backlog was not big enough to capture everything, but the end of the output is here.