Reconcile `Series.index` and `DataFrame.index`

[Dr-Irv]

• Series.index
  • def index(self) -> Index | MultiIndex: ...
• DataFrame.index
  • def index(self) -> UnknownIndex: ...

These are inconsistent. I think they should both read def index(self) -> Index: ...

I think this change will allow UnknownIndex to be removed.

[Dr-Irv]

@MarcoGorelli maybe you want to look at this?

[loicdiridollou]

I think we can drop the first one (Index | MultiIndex) that would not change anything in the tests yet last time I checked with the latest of mypy, there is still this issue that it matches to Index[Any] and when you try to checking if it is equal to a timestamp it will complain.

[loicdiridollou]

I found the lines that created an issue:

pandas-stubs/tests/test_scalars.py

Lines 1325 to 1326 in e799ec1

	eq_un1 = check(assert_type(ts == un_idx, np_1darray[np.bool]), np_1darray[np.bool])
	ne_un1 = check(assert_type(ts != un_idx, np_1darray[np.bool]), np_1darray[np.bool])

Turns out if you flip the order (index on the left of the double equal) then it is fine so it is probably in the Timestamp.__eq__ method.

[Dr-Irv]

It's a mypy bug, because reveal_type(ts.__eq__(un_idx)) should reveal the numpy array, not Any. If you comment out def __eq__(self, other: object) -> Literal[False], then it reveals correctly.

I think we should get rid of UnknownIndex, and then on the 2 lines you point out, put in # type: ignore[assert-type] with a comment above those 2 lines indicating that this is a mypy bug, and point to this discussion in the comment.

[Dr-Irv]

The other thing we could do is not allow a comparison between Timestamp and Index and only allow Timestamp.__eq__(self, other: DatetimeIndex) . This would mirror what happens for Timedelta and Period. Having said that, we know it's a mypy bug, so I'm still in favor of just putting in the ignores on the tests and adding a comment. In fact, we probably should change the Timedelta and Period __eq__() and __ne__() methods to take Index and then we'll probably see the same problems if we add tests for unknown indexes of those types.

[Dr-Irv]

The other thing we could do is not allow a comparison between Timestamp and Index and only allow Timestamp.__eq__(self, other: DatetimeIndex) . This would mirror what happens for Timedelta and Period. Having said that, we know it's a mypy bug, so I'm still in favor of just putting in the ignores on the tests and adding a comment. In fact, we probably should change the Timedelta and Period __eq__() and __ne__() methods to take Index and then we'll probably see the same problems if we add tests for unknown indexes of those types.

Thought about this some more, and now I'm leaning to have Timestamp.__eq__(self, other: Index) changed to Timestamp.__eq__(self, other: DatetimeIndex) and removing the test for unknown indexes in test_timestamp_cmp_index()

My reasoning is that you really shouldn't be comparing an Index where you don't know the underlying type to a Timestamp. If you know the index is a DatetimeIndex, then you should do a cast.