BUG: Index.fillna ignores 'downcast'

[jbrockmendel]

Issue Description

def fillna(self, value=None, downcast=None):
    value = self._require_scalar(value)
     if self.hasnans:
        result = self.putmask(self._isnan, value)
        if downcast is None:
             return Index._with_infer(result, name=self.name)
     return self._view()

In the case with downcast not None and self.hasnans, we fall through to return self._view() and ignore result. We don't have any such cases in the test suite.

Expected Behavior

Not ignoring result or downcast.

I'd be OK with deprecating the keyword; i expect it is just there to match NDFrame.

[jorisvandenbossche]

Concrete example of resulting behaviour:

In [14]: idx = pd.Index([1, 2, np.nan])

In [15]: idx
Out[15]: Float64Index([1.0, 2.0, nan], dtype='float64')

In [16]: idx.fillna(0.0)
Out[16]: Float64Index([1.0, 2.0, 0.0], dtype='float64')

In [17]: idx.fillna(0.0, downcast="infer")
Out[17]: Float64Index([1.0, 2.0, nan], dtype='float64')

So with specifying downcast, it's not actually filling any NaNs.