API: loc indexing with nested sequence raise on missing

[jbrockmendel]

I think the current behavior is left over from the long long ago.

mi = pd.MultiIndex.from_product([range(3), ["A", "B"]])
ser = pd.Series(len(mi), index=mi)

>>> ser.loc[1, ["A", "B", "not a key"]]
1  A    6
   B    6
dtype: int64

>>> ser.loc[[99999], ["A", "B"]]
Series([], dtype: int64)

>>> ser.loc[[99999]]
KeyError: '[99999] not in index'

I think we should deprecate this and eventually raise like we would with any other missing label.

[attack68]

@Pholf and I have an on going discussion re: this here. For what its worth I typed up how I believe many of the different cases should intersect, and pholf is rightfully challenging a few of those ideas.

Pertinent to your example, I propose:

ser.loc[1, ["A", "B", "not a key"]] would raise since "not a key" is not part of the level1_elements.

but

ser2 = ser.reindex(["A", "B", "not a key"], level=1)
ser2.loc[1, ["A", "B", "not a key"]]

would not raise since "not a key" is now an unused_level1_element, i.e.

>>> ser2.index.levels  
FrozenList([[0, 1, 2], ['A', 'B', 'not a key']])
>>> ser2
0  A    6
   B    6
1  A    6
   B    6
2  A    6
   B    6
dtype: int64

This allows for complex vector product or scalar broadcasting of level slices, without having to explicitly check if all the parts of your level slice will be accepted (and therefore be forced to remove any before using loc to avoid KeyErrors)

[jbrockmendel]

@attack68 thanks for the link, ill work my way through the thread today. I take it this issue is redundant and can be closed?

[attack68]

@attack68 thanks for the link, ill work my way through the thread today. I take it this issue is redundant and can be closed?

yes I believe it to be a duplicate, with this and other cases under consideration elsewhere,