Fix flaky RuntimeWarning during array reductions

[jrbourbeau]

This is sort of a shot in the dark. We've got some (pretty rare) flaky RuntimeWarning: invalid value encountered in reduce warnings being emitted in test_reductions_2D (see full traceback below). The warning occurs when we're determining the output meta dtype using np.empty. This PR suggests we use np.ones instead of np.empty to make things more deterministic. Again, not totally sure if this actually solves the flaky warning, but it might, and the change here seems harmless anyways.

I'll run CI here a few times to see if we encounter the test_reductions_2D error

Traceback:

____________________________ test_reductions_2D[f4] ____________________________
[gw3] linux -- Python 3.8.16 /usr/share/miniconda3/envs/test-environment/bin/python3.8

dtype = 'f4'

    @pytest.mark.slow
    @pytest.mark.parametrize("dtype", ["f4", "i4"])
    def test_reductions_2D(dtype):
        x = np.arange(1, 122).reshape((11, 11)).astype(dtype)
        a = da.from_array(x, chunks=(4, 4))
    
        b = a.sum(keepdims=True)
        assert b.__dask_keys__() == [[(b.name, 0, 0)]]
    
        reduction_2d_test(da.sum, a, np.sum, x)
        reduction_2d_test(da.mean, a, np.mean, x)
        reduction_2d_test(da.var, a, np.var, x, False)  # Difference in dtype algo
        reduction_2d_test(da.std, a, np.std, x, False)  # Difference in dtype algo
        reduction_2d_test(da.min, a, np.min, x, False)
        reduction_2d_test(da.max, a, np.max, x, False)
        reduction_2d_test(da.any, a, np.any, x, False)
        reduction_2d_test(da.all, a, np.all, x, False)
    
        reduction_2d_test(da.nansum, a, np.nansum, x)
>       reduction_2d_test(da.nanmean, a, np.mean, x)

dask/array/tests/test_reductions.py:219: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
dask/array/tests/test_reductions.py:158: in reduction_2d_test
    assert same_keys(da_func(darr, axis=1), da_func(darr, axis=1))
dask/array/reductions.py:734: in nanmean
    dt = getattr(np.mean(np.empty(shape=(1,), dtype=a.dtype)), "dtype", object)
<__array_function__ internals>:5: in mean
    ???
/usr/share/miniconda3/envs/test-environment/lib/python3.8/site-packages/numpy/core/fromnumeric.py:3440: in mean
    return _methods._mean(a, axis=axis, dtype=dtype,
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

a = array([nan], dtype=float32), axis = None, dtype = None, out = None
keepdims = False

    def _mean(a, axis=None, dtype=None, out=None, keepdims=False, *, where=True):
        arr = asanyarray(a)
    
        is_float16_result = False
    
        rcount = _count_reduce_items(arr, axis, keepdims=keepdims, where=where)
        if rcount == 0 if where is True else umr_any(rcount == 0, axis=None):
            warnings.warn("Mean of empty slice.", RuntimeWarning, stacklevel=2)
    
        # Cast bool, unsigned int, and int to float64 by default
        if dtype is None:
            if issubclass(arr.dtype.type, (nt.integer, nt.bool_)):
                dtype = mu.dtype('f8')
            elif issubclass(arr.dtype.type, nt.float16):
                dtype = mu.dtype('f4')
                is_float16_result = True
    
>       ret = umr_sum(arr, axis, dtype, out, keepdims, where=where)
E       RuntimeWarning: invalid value encountered in reduce

/usr/share/miniconda3/envs/test-environment/lib/python3.8/site-packages/numpy/core/_methods.py:179: RuntimeWarning

EDIT: Also xref #8892, which is related

[j-bennet]

Looks good. Still makes me wonder why the test is only failing intermittently.

[jrbourbeau]

np.empty can sometimes give arrays with interesting values

In [1]: import numpy as np

In [2]: np.empty([2, 2], dtype="float32")
Out[2]:
array([[1.67e-43, 1.36e-43],
       [1.60e-43, 1.54e-43]], dtype=float32)

In [10]: np.empty([1, 2], dtype="float64")[0][0]
Out[10]: 2.058335917824e-312

My hunch (this is really just a guess) is that some downstream operations might not like some of the values empty can produce and using np.ones would provide the same dtype-determining functionality, but on more predictable data.

[j-bennet]

👍

[jrbourbeau]

Okay, so I've run CI 6 times here and haven't seen the test_reductions_2D failure. It's a pretty rare failure, so maybe 6 times isn't enough, but the changes here seem harmless, regardless. Let's give it a try and I'll keep an eye out for test_reductions_2D failures