TST, BUG (?): windows failure for integrate/tests/test__quad_vec.py::test_quad_vec_pool

[tylerjereddy]

This failure/permissions error is showing up in the wheels repo for a subset of Windows matrix entries, including backports for 1.7.2 and for master.

I'll put the traceback below, and links to sample PRs:
MacPython/scipy-wheels#135
MacPython/scipy-wheels#138

Can we put a skip marker on this test for Windows temporarily, at least on the maintenance branch, or is this looking real?

================================== FAILURES ===================================
_____________________________ test_quad_vec_pool ______________________________
[gw0] win32 -- Python 3.7.5 C:\Python37-x64\python.exe
C:\Python37-x64\lib\site-packages\scipy\integrate\tests\test__quad_vec.py:98: in test_quad_vec_pool
    res, err = quad_vec(f, -np.inf, np.inf, norm='max', epsabs=1e-4, workers=4)
        Pool       = <function Pool at 0x000002479C210A68>
        f          = <function _lorenzian at 0x0000024787C0FDC8>
C:\Python37-x64\lib\site-packages\scipy\integrate\_quad_vec.py:252: in quad_vec
    res = quad_vec(f2, -1, 1, **kwargs)
        a          = -inf
        b          = inf
        cache_size = 100000000.0
        epsabs     = 0.0001
        epsrel     = 1e-08
        f          = <function _lorenzian at 0x0000024787C0FDC8>
        f2         = <scipy.integrate._quad_vec.DoubleInfiniteFunc object at 0x0000024799FFCA48>
        full_output = False
        kwargs     = {'cache_size': 100000000.0, 'epsabs': 0.0001, 'epsrel': 1e-08, 'full_output': False, ...}
        limit      = 10000
        norm       = 'max'
        points     = None
        quadrature = None
        sgn        = 1
        workers    = 4
C:\Python37-x64\lib\site-packages\scipy\integrate\_quad_vec.py:391: in quad_vec
    break
        CONVERGED  = 0
        NOT_A_NUMBER = 3
        NOT_CONVERGED = 1
        ROUNDING_ERROR = 2
        _quadrature = <function _quadrature_gk15 at 0x0000024787BDB678>
        a          = -1.0
        b          = 0.0
        cache_count = 4166666.0
        cache_size = 100000000.0
        derr       = -0.0001958558976507645
        dint       = -5.188183216375819e-11
        dneval     = 30
        dround_err = 1.7439342490043163e-14
        epsabs     = 0.0001
        epsrel     = 1e-08
        err        = 1.2888960825934786e-10
        err_sum    = 0.00039171231086582484
        f          = <scipy.integrate._quad_vec.DoubleInfiniteFunc object at 0x0000024799FFCA48>
        full_output = False
        global_error = 5.155584007941142e-10
        global_integral = 3.141592653589793
        ier        = 0
        ig         = 0.7853981633974483
        initial_intervals = [(-1.0, 0.0), (0.0, 1.0)]
        interval   = (-0.000195856155429981, -1.0, 0.0)
        interval_cache = LRUDict([((0.0, 0.5), 0.7853981633974485), ((0.5, 1.0), 0.7853981633974483), ((-1.0, -0.5), 0.7853981633974484), ((-0.5, 0.0), 0.7853981633974483)])
        intervals  = [(-1.2888960825934792e-10, -1.0, -0.5), (-1.2888960825934786e-10, -0.5, 0.0), (-1.2888960825934792e-10, 0.5, 1.0), (-1.2888957602994802e-10, 0.0, 0.5)]
        j          = 2
        kwargs     = {'cache_size': 100000000.0, 'epsabs': 0.0001, 'epsrel': 1e-08, 'full_output': False, ...}
        limit      = 10000
        mapwrapper = <scipy._lib._util.MapWrapper object at 0x000002479C20C9C8>
        min_intervals = 2
        neg_old_err = -0.000195856155429981
        neval      = 90
        norm       = 'max'
        norm_func  = <built-in function abs>
        norm_funcs = {None: <function _max_norm at 0x0000024787B5C0D8>, 'max': <function _max_norm at 0x0000024787B5C0D8>, '2': <function norm at 0x00000247FE803E58>}
        old_int    = 1.5707963268467784
        p          = 0.0
        parallel_count = 128
        points     = (0,)
        prev       = 0.0
        quadrature = 'gk15'
        rnd        = 1.743934249061916e-14
        rounding_error = 6.975736996132465e-14
        status_msg = {0: 'Target precision reached.', 1: 'Target precision not reached.', 2: 'Target precision could not be reached due to rounding error.', 3: 'Non-finite values encountered.'}
        subint     = ((-1.0, -0.5, 0.7853981633974484, 1.2888960825934792e-10), (-0.5, 0.0, 0.7853981633974483, 1.2888960825934786e-10))
        to_process = [((0.00019585615543584383, 0.0, 1.0, 1.5707963268467782), <scipy.integrate._quad_vec.DoubleInfiniteFunc object at 0x00...nfiniteFunc object at 0x0000024799FFCA48>, <built-in function abs>, <function _quadrature_gk15 at 0x0000024787BDB678>)]
        tol        = 0.0001
        workers    = 4
        x          = (-0.5, 0.0, 0.7853981633974483, 1.2888960825934786e-10)
        x1         = -0.5
        x2         = 0.0
C:\Python37-x64\lib\site-packages\scipy\_lib\_util.py:466: in __exit__
    self.pool.terminate()
        exc_type   = None
        exc_value  = None
        self       = <scipy._lib._util.MapWrapper object at 0x000002479C20C9C8>
        traceback  = None
C:\Python37-x64\lib\multiprocessing\pool.py:548: in terminate
    self._terminate()
        self       = <multiprocessing.pool.Pool object at 0x000002479C20CBC8>
C:\Python37-x64\lib\multiprocessing\util.py:201: in __call__
    res = self._callback(*self._args, **self._kwargs)
        _finalizer_registry = {(None, 9): <Finalize object, callback=CloseHandle, args=(2044,)>, (None, 10): <Finalize object, callback=CloseHandle,...Finalize object, callback=CloseHandle, args=(2020,)>, (None, 12): <Finalize object, callback=CloseHandle, args=(508,)>}
        getpid     = <built-in function getpid>
        self       = <Finalize object, callback=_terminate_pool, args=(<_queue.SimpleQueue object at 0x000002479C1FB638>, <multiprocessing....emon 2636)>, <Thread(Thread-108, stopped daemon 3004)>, <Thread(Thread-109, stopped daemon 3664)>, {}), exitprority=15>
        sub_debug  = <function sub_debug at 0x00000247FF7A5DC8>
        wr         = None
C:\Python37-x64\lib\multiprocessing\pool.py:601: in _terminate_pool
    p.terminate()
        cache      = {}
        cls        = <class 'multiprocessing.pool.Pool'>
        inqueue    = <multiprocessing.queues.SimpleQueue object at 0x000002479C20CB08>
        outqueue   = <multiprocessing.queues.SimpleQueue object at 0x000002479C2053C8>
        p          = <SpawnProcess(SpawnPoolWorker-8, stopped daemon)>
        pool       = [<SpawnProcess(SpawnPoolWorker-7, stopped[SIGTERM] daemon)>, <SpawnProcess(SpawnPoolWorker-8, stopped daemon)>, <SpawnProcess(SpawnPoolWorker-9, stopped daemon)>, <SpawnProcess(SpawnPoolWorker-10, stopped daemon)>]
        result_handler = <Thread(Thread-109, stopped daemon 3664)>
        task_handler = <Thread(Thread-108, stopped daemon 3004)>
        taskqueue  = <_queue.SimpleQueue object at 0x000002479C1FB638>
        worker_handler = <Thread(Thread-107, stopped daemon 2636)>
C:\Python37-x64\lib\multiprocessing\process.py:124: in terminate
    self._popen.terminate()
        self       = <SpawnProcess(SpawnPoolWorker-8, stopped daemon)>
C:\Python37-x64\lib\multiprocessing\popen_spawn_win32.py:119: in terminate
    _winapi.TerminateProcess(int(self._handle), TERMINATE)
E   PermissionError: [WinError 5] Access is denied
        self       = <multiprocessing.popen_spawn_win32.Popen object at 0x000002479C205748>

[rgommers]

Looks familiar, duplicate of gh-14797 and occurred before in gh-11835. I'll have a closer look.

[rgommers]

Can we put a skip marker on this test for Windows temporarily, at least on the maintenance branch, or is this looking real?

Yes, please do - something is wrong here, but it shouldn't be blocking for adding the Python 3.10 support.

[rgommers]

It's only happening on Python 3.8 builds on Appveyor. Reading the discussion on gh-11835, the solution was import multiprocessing within runtests.py (long story, has to do with a Python 3.8 bug, plus how multiprocessing gets initialized, plus whether it runs in a subprocess).

This seems to be something similar, because this is the only Windows build where we use yet another script to kick off the test suite: https://github.com/MacPython/scipy-wheels/blob/master/run_scipy_tests.py. I bet that if we used pytest, python -c "import scipy; scipy.test()" or runtests.py`, this would not be failing.

[rgommers]

@tylerjereddy you could try to add import multiprocessing to run_scipy_tests.py and see if that fixes it. If not, then I'd say just skip it and only worry about it again if it's reported for a more normal way of running the tests.

[tylerjereddy]

@rgommers thanks, I've tried that first--the import has been pushed into the wheels backport: MacPython/scipy-wheels#135

[mdhaber]

@tylerjereddy was this resolved? I just noticed that the linked issues were closed.

[tylerjereddy]

@mdhaber I think I'm going to close this, given:

• seems to be focused on Python 3.7/3.8, and 3.8 support is getting dropped in NumPy soon-ish anyway I think
• largely related to the old wheels repo/Appveyor stuff, and a weird way of running the tests that I don't think we care about much now that we use modern cibiuldwheel and we have a different build system in place, etc.

Anyone should feel free to reopen if there's a concrete reproducer outside of that weird/older machinery.

[tylerjereddy]

I should mention, this particular test does already have another more recent open issue; #17098

That looks like a separate matter on Linux though with segfault though.