API: what should float(pd.NA) return?

[MarcoGorelli]

From #32265, it seems there was broad agreement that pandas should do

In [4]: pd.Series([1, 2, pd.NA], dtype='Float64').to_numpy()
Out[4]: array([ 1.,  2., nan])

Should float(pd.NA) then return np.nan?

xref matplotlib/matplotlib#23991 (comment)

[phofl]

Just to clarify: The alternative would be to raise for float(pd.NA)? Anything else that should be considered?

[MarcoGorelli]

Yeah that's what currently happens

In [16]: float(pd.NA)
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In [16], line 1
----> 1 float(pd.NA)

TypeError: float() argument must be a string or a number, not 'NAType'

Not sure what other considerations there are, I'll give this a go and see what happens

[MarcoGorelli]

This breaks

pandas/pandas/tests/indexes/test_base.py

Lines 855 to 862 in 1b2cc26

	def test_isin_nan_common_float64(self, nulls_fixture):
	if nulls_fixture is pd.NaT or nulls_fixture is pd.NA:
	# Check 1) that we cannot construct a Float64Index with this value
	# and 2) that with an NaN we do not have .isin(nulls_fixture)
	msg = "data is not compatible with Float64Index"
	with pytest.raises(ValueError, match=msg):
	Float64Index([1.0, nulls_fixture])

, though Float64Index is deprecated anyway

[MarcoGorelli]

from the dev call: it's OK for float(pd.NA) to error, conversion to numpy should be handled in __array__