We should discuss this first, @mroeschke and I agreed that we should only add a note to the docs

@phofl

We should discuss this first, @mroeschke and I agreed that we should only add a note to the docs

By all means! This arose due to @rhshadrach's efforts to break my #44267. I found that a trivial and sensible change to frame.py fixed the problem, and also fixed #44410. I know that you decided to document rather than fix. But since the bug was a regression, and the fix seems simple, I suggest to accept my fix. It will also be a prerequisite for my #44267.

Imo this change is neither trivial nor simple. Accidentially ending up with duplictated column names is not desireable. This is a real PITA in indexing among others. For example:

df = pd.DataFrame([[1, 2, 3]], columns=["a", "a", "b"])

df["b"] returns a Series while df["a"] returns a DataFrame.
IMO this is more of an api change instead of a bugfix. Also I think you can make a point that this should be changed inside of insert, not on a higher level.

Accidentially ending up with duplictated column names is not desireable.
IMO this is more of an api change instead of a bugfix.

Agreed on both counts. For #44267, it is the case that the user called the function with duplicate columns, the internals of groupby moved them into the index, and we want to restore them as columns. For this, it would suffice to have an internal version of reset_index where we can override allow_duplicates, i.e. this pattern:

def foo(x):
    return _foo(x, hidden_arg=False)

def _foo(x, hidden_arg=False):
    pass

where #44267 could then call _foo(x, hidden_arg=True) (where hidden_arg is allow_duplicates).

Also I think you can make a point that this should be changed inside of insert, not on a higher level.

I'm not sure what this means (in other words, what should be changed inside of insert?). In any case, for #44267, I don't think we should be reimplementing reset_index via calling insert directly (to be clear, I'm not suggesting you're saying that).

What I meant with respect to insert:

If we want to respect the flag, we should change the keyword allow_duplicates to None or lib.NO_DEFAULT and figure out the correct default inside of insert instead of the caller.

But I would prefer not changing the public behavior, so I like your idea with a private implementation better.

There is a whole chapter on this: https://pandas.pydata.org/pandas-docs/stable/user_guide/duplicates.html.
It implies that the Pandas policy is to default to allowing duplicates:

Both Series and DataFrame disallow duplicate labels by calling .set_flags(allows_duplicate_labels=False). (the default is to allow them).

It seems to me that if you have a legal Series (with duplicates), then it is OK for reset_index() to translate it to a legal DataFrame (with duplicates).
My change just takes allow_duplicates from the frame's own allows_duplicate_labels flag.
But it is not for me to say. If you want reset_index() to have an allow_duplicates flag (over-riding the frame's flag) then let me know.
I can't see anything gained by hiding it in a private method. SInce #44410 is easily resolved then why not let it be resolved?

SInce #44410 is easily resolved then why not let it be resolved?

I personally am not a fan of this resolution. My understanding is that flags.allow_duplicate_labels will not allow duplicates in both the index and the columns. In using pandas, I very often find use in having duplicates in the index, but never want them in the columns. In other words, I want reset_index to throw an error when the result would have duplicate columns but I am not able to set flags.allow_duplicate_labels to be True.

In any case, this an API change that will need to have a FutureWarning before it can be implemented in 2.0 if we go this route.

Hello @johnzangwill! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2022-01-28 22:44:47 UTC

I have added allow_duplicates to DataFrame.reset_index() and Series.reset_index(). I have reverted and added to the tests accordingly, and changed the whatsnew from bug to enhancement.
The defaults are both False, so the current (pre-PR) behaviour is unchanged: duplicates are not created, regardless of the frame flag.
#44410 still raises ValueError. But adding allow_duplicates=None or allow_duplicates=True makes reset_index() work as the OP proposed.