issue_comments

Thanks @shoyer. Cool, then this was easier than I expected. I added the patch and nanargmax/min to the nputils in #3244. What do you think?

Thanks for sharing the patch! I dropped into a debugger by adding --pdb to the pytest command, which revealed what is going on here: ```

PDB post_mortem (IO-capturing turned off) >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> /Users/shoyer/dev/xarray/xarray/core/duck_array_ops.py(295)f() -> raise NotImplementedError(msg) (Pdb) values array(['2000-01-01T00:00:00.000000000', '2000-01-02T00:00:00.000000000', '2000-01-03T00:00:00.000000000'], dtype='datetime64[ns]') (Pdb) func <function nanargmax at 0x1156ff730> (Pdb) func(values, axis=axis, **kwargs) *** AttributeError: module 'xarray.core.nputils' has no attribute 'nanargmax' ```

So it looks like nputils doesn't have nanargmax defined. Instead we need to use nanargmax from NumPy.

Those little changes do solve the MCVE, but break at least one test. I don't have enough of an understanding of the (nan)ops logic in xarray to get around the issue. But may be this helps:

The change

``` diff --git a/xarray/core/nanops.py b/xarray/core/nanops.py index 9ba4eae2..784a1d01 100644 --- a/xarray/core/nanops.py +++ b/xarray/core/nanops.py @@ -91,17 +91,9 @@ def nanargmin(a, axis=None): fill_value = dtypes.get_pos_infinity(a.dtype) if a.dtype.kind == "O": return _nan_argminmax_object("argmin", fill_value, a, axis=axis) - a, mask = _replace_nan(a, fill_value) - if isinstance(a, dask_array_type): - res = dask_array.argmin(a, axis=axis) - else: - res = np.argmin(a, axis=axis)

• if mask is not None:
• mask = mask.all(axis=axis)
• if mask.any():
• raise ValueError("All-NaN slice encountered")
• return res
• module = dask_array if isinstance(a, dask_array_type) else nputils
• return module.nanargmin(a, axis=axis)

def nanargmax(a, axis=None): @@ -109,17 +101,8 @@ def nanargmax(a, axis=None): if a.dtype.kind == "O": return _nan_argminmax_object("argmax", fill_value, a, axis=axis)

• a, mask = _replace_nan(a, fill_value)
• if isinstance(a, dask_array_type):
• res = dask_array.argmax(a, axis=axis)
• else:

• res = np.argmax(a, axis=axis)

• if mask is not None:
• mask = mask.all(axis=axis)
• if mask.any():
• raise ValueError("All-NaN slice encountered")
• return res
• module = dask_array if isinstance(a, dask_array_type) else nputils
• return module.nanargmax(a, axis=axis)

def nansum(a, axis=None, dtype=None, out=None, min_count=None):

```

The failing test

``` python
...
__ TestVariable.testreduce __
...

def f(values, axis=None, skipna=None, **kwargs):
    if kwargs.pop("out", None) is not None:
        raise TypeError("`out` is not valid for {}".format(name))

    values = asarray(values)

    if coerce_strings and values.dtype.kind in "SU":
        values = values.astype(object)

    func = None
    if skipna or (skipna is None and values.dtype.kind in "cfO"):
        nanname = "nan" + name
        func = getattr(nanops, nanname)
    else:
        func = _dask_or_eager_func(name)

    try:
        return func(values, axis=axis, **kwargs)
    except AttributeError:
        if isinstance(values, dask_array_type):
            try:  # dask/dask#3133 dask sometimes needs dtype argument
                # if func does not accept dtype, then raises TypeError
                return func(values, axis=axis, dtype=values.dtype, **kwargs)
            except (AttributeError, TypeError):
                msg = "%s is not yet implemented on dask arrays" % name
        else:
            msg = (
                "%s is not available with skipna=False with the "
                "installed version of numpy; upgrade to numpy 1.12 "
                "or newer to use skipna=True or skipna=None" % name
            )

      raise NotImplementedError(msg)

E NotImplementedError: argmax is not available with skipna=False with the installed version of numpy; upgrade to numpy 1.12 or newer to use skipna=True or skipna=None

... ``` Note: I habe numpy 1.17 instaleed so the error msg here seems missleading.

Yes, this is definitely a bug -- thanks for clear example to reproduce it!

These helper functions were originally added back in https://github.com/pydata/xarray/pull/1883 to handle object dtype arrays properly.

So it would be nice to fix this for object arrays in dask, but for the much more common non-object dtype arrays we should really just be using dask.array.nnargmax.