BUG: quantile works for all-nan Series but not all-nan Pint ExtensionArray

[MichaelTiemannOSC]

Pandas version checks

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of pandas.

• I have confirmed this bug exists on the main branch of pandas.

Reproducible Example

import pandas as pd
import numpy as np
import pint_pandas
print(pd.Series([1,2,3]).quantile())
# 2.0                                                                                                                                                                                                                                                    
print(pd.Series([np.nan, np.nan, np.nan]).quantile())
# nan                                                                                                                                                                                                                                                    
print(pd.Series([1., np.nan, 3.], dtype='pint[m]').quantile())
# /Users/michael/Documents/GitHub/pint-pandas/pint_pandas/pint_array.py:658: UnitStrippedWarning: The unit of the quantity is stripped when downcasting to ndarray.                                                                                      
#   return np.array(qtys, dtype="object", copy=copy)                                                                                                                                                                                                     
# <Quantity(2.0, 'meter')>                                                                                                                                                                                                                               
print(pd.Series([np.nan, np.nan, np.nan], dtype='pint[m]').quantile())
# ValueError: Cannot infer dtype. No dtype specified and empty array

Issue Description

pd.quantile works as expected for pd.Series of floats, whether some, all, or none of the values are np.nan.

pd.quantile does not work as expected when the pd.Series elements are ExtensionArray elements, such as Pint Quantities. This test case shows it working properly for the case where there's a mixture of np.nan and non-NaN values, but when all values are np.nan it raises a ValueExrror

Expected Behavior

I expect it would return np.nan as the quantile value for an all-nan input under all circumstances.

Installed Versions

I just installed Pandas 1.4.1 this morning (using conda-forge).

No idea why this stopped working:

pd.show_versions()
Traceback (most recent call last):
File "", line 1, in
File "/Users/michael/opt/miniconda3/envs/spyder-env/lib/python3.9/site-packages/pandas/util/_print_versions.py", line 109, in show_versions
deps = _get_dependency_info()
File "/Users/michael/opt/miniconda3/envs/spyder-env/lib/python3.9/site-packages/pandas/util/_print_versions.py", line 88, in _get_dependency_info
mod = import_optional_dependency(modname, errors="ignore")
File "/Users/michael/opt/miniconda3/envs/spyder-env/lib/python3.9/site-packages/pandas/compat/_optional.py", line 126, in import_optional_dependency
module = importlib.import_module(name)
File "/Users/michael/opt/miniconda3/envs/spyder-env/lib/python3.9/importlib/init.py", line 127, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "", line 1030, in _gcd_import
File "", line 1007, in _find_and_load
File "", line 986, in _find_and_load_unlocked
File "", line 680, in _load_unlocked
File "", line 850, in exec_module
File "", line 228, in _call_with_frames_removed
File "/Users/michael/opt/miniconda3/envs/spyder-env/lib/python3.9/site-packages/setuptools/init.py", line 8, in
import _distutils_hack.override # noqa: F401
File "/Users/michael/opt/miniconda3/envs/spyder-env/lib/python3.9/site-packages/_distutils_hack/override.py", line 1, in
import('_distutils_hack').do_override()
File "/Users/michael/opt/miniconda3/envs/spyder-env/lib/python3.9/site-packages/_distutils_hack/init.py", line 71, in do_override
ensure_local_distutils()
File "/Users/michael/opt/miniconda3/envs/spyder-env/lib/python3.9/site-packages/_distutils_hack/init.py", line 59, in ensure_local_distutils
assert '_distutils' in core.file, core.file
AssertionError: /Users/michael/opt/miniconda3/envs/spyder-env/lib/python3.9/distutils/core.py

[MichaelTiemannOSC]

Suggest this implementation for core/arrays/base.py _quantile (note GH reference):

    def _quantile(
        self: ExtensionArrayT, qs: npt.NDArray[np.float64], interpolation: str
    ) -> ExtensionArrayT:
        """
        Compute the quantiles of self for each quantile in `qs`.

        Parameters
        ----------
        qs : np.ndarray[float64]
        interpolation: str

        Returns
        -------
        same type as self
        """
        # asarray needed for Sparse, see GH#24600
        mask = np.asarray(self.isna())
        mask = np.atleast_2d(mask)

        arr = np.atleast_2d(np.asarray(self))
        fill_value = np.nan

        res_values = quantile_with_mask(arr, mask, fill_value, qs, interpolation)

        if self.ndim == 2:
            # i.e. DatetimeArray
            result = type(self)._from_sequence(res_values)

        else:
            # shape[0] should be 1 as long as EAs are 1D
            assert res_values.shape == (1, len(qs)), res_values.shape
            if mask.all():
                # All nan array, so arbitrarily pick first GH #45968
                result = type(self)._from_sequence([self[0]])
            else:
                result = type(self)._from_sequence(res_values[0])

        return result

[phofl]

I tried pd.Series([pd.NA, pd.NA, pd.NA], dtype="Int64").quantile() on master which works as expected. Is the bug you found specific to pint_pandas dtypes or can you also reproduce with one of the pandas extension dtypes?

[MichaelTiemannOSC]

As reported, it failed using a specific pint_pandas dtype (which I canonicalized to pint[m]). I just tried using complex numbers, and it works as expected. I don't know what datatypes sit between complex numbers and pint's units.

[phofl]

Was just asking becasue the post sounded to me like this was relevant to all EAs

[jbrockmendel]

            if mask.all():
                # All nan array, so arbitrarily pick first GH #45968
                result = type(self)._from_sequence([self[0]])
            else:
                result = type(self)._from_sequence(res_values[0])

Instead of doing a mask.all() check, I think the underlying issue is in this case _from_sequence needs dtype=self.dtype. IIRC passing dtype=self.dtype breaks things so that won't quite do it. Maybe something like _from_factorized (which has its own set of problems...)

[MichaelTiemannOSC]

I've just hit this bug again, through a new pathway. I'm pretty sure it fails for all ExtensionArrays. The documentation only promises that for numeric_only, If False, the quantile of datetime and timedelta data will be computed as well.. Which is notoriously non-committal as to supporting ExtensionArrays.