REGR: Adding a row via DataFrame.at

[rhshadrach]

df = pd.DataFrame({'a': [1, 2], 'b': [3, 4]})
df.at[2] = [5, 6]

On 1.4.x, this gives

     a    b
0  1.0  3.0
1  2.0  4.0
2  5.0  6.0

whereas on 1.5.0rc0 it gives pandas.errors.InvalidIndexError: slice(None, None, None). I didn't any mention of this in the whatsnew, and a git bisect points at #47074; cc @jorisvandenbossche.

The docs of DataFrame.at explicitly say a KeyError is raised when the label is not in the frame. It seems to me this operation should not be supported, and can be considered a bugfix, but would be better to raise KeyError: label 2 not found.

[phofl]

Hm I think InvalidIndexError is ok too, since at is for one value only and this tries to update 2

although the message is not optimal

[rhshadrach]

I see; the slice(None, None, None) is referring to the columns - thanks. That makes sense, but this doesn't raise on main:

df = pd.DataFrame({'a': [1, 2]})
df.at[3, 'a'] = 7

and instead adds a row. I believe this should raise a KeyError, and think it would be preferable for the OP example to also raise a KeyError. If the key were to exist, InvalidIndexError in the OP makes sense to me (perhaps with a better message if possible).

I plan to add a note in the whatsnew in the indexing bugfixes section to close this issue and open a new issue on changing the behavior of .at, but I think the behavior change should wait until after the 1.5. release.

[phofl]

This is probably related to #48224 somehow. Maybe you could combine them when opening a new issue