Xarray: test failure with upstream-dev

[dcherian]

https://dev.azure.com/xarray/xarray/_build/results?buildId=1101&view=logs

=================================== FAILURES ===================================
_________________________ test_datetime_reduce[False] __________________________

dask = False

    @arm_xfail
    @pytest.mark.parametrize("dask", [False, True])
    def test_datetime_reduce(dask):
        time = np.array(pd.date_range("15/12/1999", periods=11))
        time[8:11] = np.nan
        da = DataArray(np.linspace(0, 365, num=11), dims="time", coords={"time": time})

        if dask and has_dask:
            chunks = {"time": 5}
            da = da.chunk(chunks)

        actual = da["time"].mean()
>       assert not pd.isnull(actual)
E       AssertionError: assert not True
E        +  where True = <function isna at 0x7f2475449ae8>(<xarray.DataArray 'time' ()>\narray('NaT', dtype='datetime64[ns]'))
E        +    where <function isna at 0x7f2475449ae8> = pd.isnull

xarray/tests/test_duck_array_ops.py:288: AssertionError
__________________________ test_datetime_reduce[True] __________________________

dask = True

    @arm_xfail
    @pytest.mark.parametrize("dask", [False, True])
    def test_datetime_reduce(dask):
        time = np.array(pd.date_range("15/12/1999", periods=11))
        time[8:11] = np.nan
        da = DataArray(np.linspace(0, 365, num=11), dims="time", coords={"time": time})

        if dask and has_dask:
            chunks = {"time": 5}
            da = da.chunk(chunks)

        actual = da["time"].mean()
>       assert not pd.isnull(actual)
E       AssertionError: assert not True
E        +  where True = <function isna at 0x7f2475449ae8>(<xarray.DataArray 'time' ()>\narray('NaT', dtype='datetime64[ns]'))
E        +    where <function isna at 0x7f2475449ae8> = pd.isnull

xarray/tests/test_duck_array_ops.py:288: AssertionError
=============================== warnings summary 

[crusaderky]

This has now spread to all tests without pinned dependencies

[crusaderky]

conda list (py36) before and after the breakage:

< boto3                     1.9.252                    py_0    conda-forge
> boto3                     1.9.253                    py_0    conda-forge
< botocore                  1.12.252                   py_0    conda-forge
> botocore                  1.12.253                   py_0    conda-forge
< expat                     2.2.5             he1b5a44_1003    conda-forge
> expat                     2.2.5             he1b5a44_1004    conda-forge
< libxkbcommon              0.8.4                h516909a_0    conda-forge
> libxkbcommon              0.9.0                hebb1f50_0    conda-forge
< mypy_extensions           0.4.2                    py36_0    conda-forge
> mypy_extensions           0.4.3                    py36_0    conda-forge
< nss                       3.46                 he751ad9_0    conda-forge
> nss                       3.47                 he751ad9_0    conda-forge
< pandas                    0.25.1           py36hb3f55d8_0    conda-forge
> pandas                    0.25.2           py36hb3f55d8_0    conda-forge
< pip                       19.3                     py36_0    conda-forge
> pip                       19.3.1                   py36_0    conda-forge
< pseudonetcdf              3.0.2                      py_0    conda-forge
> pseudonetcdf              3.1.0                      py_0    conda-forge

[shoyer]

This looks like this upstream pandas issue: https://github.com/pandas-dev/pandas/issues/29053

[crusaderky]

It doesn't add up - the upstream ticket mentions an incompatibility with numpy 1.18. However, our failing tests are all running with numpy 1.17.

[crusaderky]

Looking more closely at the failing tests, there are actually two problems.
One is with pandas/numpy git tip, and it only affects the upstream-dev test suite.
The other is with pseudonetcdf-3.1, and it affects py36, py37, and upstream-dev.

[crusaderky]

Narrowed down:

>>> import numpy as np
>>> a = np.array(['1999-12-15', 'NaT'], dtype='M8[ns]')
>>> np.min(a)

Output:
numpy 1.17: numpy.datetime64('1999-12-15T00:00:00.000000000')
numpy 1.18: numpy.datetime64('NaT')

triggered by:
https://github.com/pydata/xarray/blob/0f7ab0e909d9b6272f734a0b6fa4318e9522d3a2/xarray/core/duck_array_ops.py#L372

[shoyer]

I think we have our own version of the same issue that pandas had.

[crusaderky]

Same problem:

a = xarray.DataArray(['1999-12-15', 'NaT']).astype('M8[ns]')            
a.min(skipna=False)  # np 1.17: 1999-12-15; np 1.18: NaT
a.min(skipna=True)  # np 1.17: crashes; np 1.18: crashes

[dcherian]

Looks like numpy will fix this: https://github.com/numpy/numpy/pull/14841

[crusaderky]

@dcherian I didn't try running the PR code, but I don't think so?
The PR may mean (must test) that nanmin() and nanmax() now work with NaT. However, as highlighted above https://github.com/pydata/xarray/issues/3409#issuecomment-544299242 xarray is invoking min() on an array that contains NaT - which in numpy 1.17 ignores them, while in 1.18 correctly returns NaT.

Does anybody have the time to test it?

[dcherian]

Sigh yes you're right

[mathause]

Unfortunately this does not seem to make np.nanmin and np.nanmax work for datetime arrays (yet), see: https://github.com/numpy/numpy/pull/14841#issuecomment-551824320