BUG: regression in error raised by idxmin/idxmax for extension dtypes

[jorisvandenbossche]

With pandas 1.0.1:

In [7]: from pandas.tests.extension.decimal import DecimalArray, make_data 

In [8]: s = pd.Series(DecimalArray(make_data()[:5]))  

In [9]: s 
Out[9]: 
0    Decimal: 0.25866656130631138221787068687262944...
1    Decimal: 0.11511905452780368808163302674074657...
2    Decimal: 0.15301679241167220890673661415348760...
3    Decimal: 0.41125672759464626526693109553889371...
4    Decimal: 0.46391316685725048074573351186700165...
dtype: decimal

In [10]: s.idxmin()   
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-10-5060aab05d12> in <module>
----> 1 s.idxmin()

~/miniconda3/envs/pandas10/lib/python3.8/site-packages/pandas/core/series.py in idxmin(self, axis, skipna, *args, **kwargs)
   2037         """
   2038         skipna = nv.validate_argmin_with_skipna(skipna, args, kwargs)
-> 2039         i = nanops.nanargmin(com.values_from_object(self), skipna=skipna)
   2040         if i == -1:
   2041             return np.nan

~/miniconda3/envs/pandas10/lib/python3.8/site-packages/pandas/core/nanops.py in _f(*args, **kwargs)
     62             if any(self.check(obj) for obj in obj_iter):
     63                 f_name = f.__name__.replace("nan", "")
---> 64                 raise TypeError(
     65                     f"reduction operation '{f_name}' not allowed for this dtype"
     66                 )

TypeError: reduction operation 'argmin' not allowed for this dtype

Now with master you get:

In [4]: s.idxmin() 
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-4-7e62303a9d43> in <module>
----> 1 s.idxmin()

~/scipy/pandas/pandas/core/series.py in idxmin(self, axis, skipna, *args, **kwargs)
   1988         """
   1989         skipna = nv.validate_argmin_with_skipna(skipna, args, kwargs)
-> 1990         i = nanops.nanargmin(self._values, skipna=skipna)
   1991         if i == -1:
   1992             return np.nan

~/scipy/pandas/pandas/core/nanops.py in _f(*args, **kwargs)
     69             try:
     70                 with np.errstate(invalid="ignore"):
---> 71                     return f(*args, **kwargs)
     72             except ValueError as e:
     73                 # we want to transform an object array

~/scipy/pandas/pandas/core/nanops.py in nanargmin(values, axis, skipna, mask)
    943     """
    944     values, mask, dtype, _, _ = _get_values(
--> 945         values, True, fill_value_typ="+inf", mask=mask
    946     )
    947     result = values.argmin(axis)

~/scipy/pandas/pandas/core/nanops.py in _get_values(values, skipna, fill_value, fill_value_typ, mask)
    311         # promote if needed
    312         else:
--> 313             values, _ = maybe_upcast_putmask(values, mask, fill_value)
    314 
    315     # return a platform independent precision dtype

~/scipy/pandas/pandas/core/dtypes/cast.py in maybe_upcast_putmask(result, mask, other)
    278     """
    279     if not isinstance(result, np.ndarray):
--> 280         raise ValueError("The result input must be a ndarray.")
    281     if not is_scalar(other):
    282         # We _could_ support non-scalar other, but until we have a compelling

ValueError: The result input must be a ndarray.

@jbrockmendel this is from no longer using values_from_object, but passing _values to the nanops function, while this expects to always receive an ndarray

[jbrockmendel]

sounds like a 2-liner to check for non-ndarray and raise the old exception?

[jorisvandenbossche]

Maybe not necessarily, you might have EAs that can be converted to numerical ndarrays?

[jbrockmendel]

you might have EAs that can be converted to numerical ndarrays?

Should values_for_argsort work for that?

[jorisvandenbossche]

I suppose so? I suppose if argsort works on the values, also argmin/argmax should do the expected thing?

[TomAugspurger]

I'd like to fix this by having Series.idxmin() eventually call .array._reduce("argmin"). However, there's a potential issue with how we've defined argmin on our arrays for missing values.

In [52]: pd.array([1, 2, None], dtype="float64")._reduce("argmin", skipna=False)  # PandasArray
Out[52]: <NA>

In [53]: pd.array([1, 2, None], dtype="Int64")._reduce("argmin", skipna=False)
Out[53]: -1

IMO, _reduce should return the correct result, which in this case is I think NA (under the rule that skipna propagates NA).

[jorisvandenbossche]

However, there's a potential issue with how we've defined argmin on our arrays for missing values.

@TomAugspurger see also #33941, #33942