@phofl There's no warning. An exception occurs while processing the warning, and I've included the message in my original post.

The reason for the exception is that left and right might be a Series and they don't have the columns attribute. While _left and _right are the converted DataFrames.

Sorry, did not understand you then.

could you please add a test?

And a whatsnew in 1.4.4, this is a regression

@phofl No worries. I'll add one in a few hours.

@nsarang would you have time to add a test ?

@jorisvandenbossche Forgot about this one!
I've added the test now. The explanation for the test is that using to_frame on a Series with a multi-level name will result in a multi-level column DataFrame (to_frame is used in pd.merge._validate_operand). So it'll create the right condition where the levels of left and right in pd.merge are not equal.

Sure

@phofl I can't reference the PR in the release notes since its path is not added to sphinx.ext.extlinks. Do you want me to add it?

It is sufficient if you add something like Fixed regression in :func:`merge` when...

to https://github.com/pandas-dev/pandas/blob/main/doc/source/whatsnew/v1.4.4.rst

You don't have to do anything else. Maybe you did not merge main in a while?

Fine by me. Just wanted to make sure it's okay that the release note is not having a reference.

Maybe you did not merge main in a while?

I don't follow. Can you explain?

I pushed a small patch to ensure we catch the warning in the test (ideally, our tests don't cause warnings from pandas, if a warning is expected, it should be either asserted or explicitly ignored).

However, doing that also made we wonder why we are actually raising a deprecation warning here. The case from the test:

In [3]: merge(a, b, on=["outer", "inner"])
<ipython-input-3-4caf266e26f3>:1: FutureWarning: merging between different levels is deprecated and will be removed in a future version. (1 levels on the left, 2 on the right)
  merge(a, b, on=["outer", "inner"])
Out[3]: 
             A  (B, C)
outer inner           
a     1      2       1
b     1      4       3

In [4]: a
Out[4]: 
             A
outer inner   
a     0      1
      1      2
b     0      3
      1      4

In [5]: b
Out[5]: 
outer  inner
a      1        1
       2        2
b      1        3
       2        4
Name: (B, C), dtype: int64

So the original a and b don't have different number of levels in the columns. It's only because the Series is converted to a DataFrame that pandas itself introduces that. It's of course quite a corner case ..

@jorisvandenbossche

We wanted to get rid of value dependent behaviour. Your op looks "correct", but if "a" also has a midx, then the tuple name is converted into a MultiIndex itself:

a = DataFrame(
    {("A", "B"): 1},
    index=MultiIndex.from_product([["a"], [0, 1]], names=["o", "i"]),
)
b = Series(
    1,
    index=MultiIndex.from_product([["a"], [1, 2]], names=["o", "i"]),
    name=("B", "C"),
)
pd.merge(b, a, on=["o", "i"])

And the more complicated case:

a = DataFrame(
    {("A", "B"): 1},
    index=MultiIndex.from_product([["a"], [0, 1]], names=["o", "i"]),
)
b = Series(
    1,
    index=MultiIndex.from_product([["a"], [1, 2]], names=["o", "i"]),
    name=("B", "C", "D"),
)
pd.merge(a, b, on=["o", "i"])

raises: ValueError: Length of names must match number of levels in MultiIndex.

If the DataFrame has more levels than our tuple name has elements, we flatten the DataFrame MultiIndex columns to a tuple too:

a = DataFrame(
    {("A", "B", "C"): 1},
    index=MultiIndex.from_product([["a"], [0, 1]], names=["o", "i"]),
)
b = Series(
    1,
    index=MultiIndex.from_product([["a"], [1, 2]], names=["o", "i"]),
    name=("B", "C"),
)
pd.merge(a, b, on=["o", "i"])

     (A, B, C)  (B, C)
o i                   
a 1          1       1

thx @nsarang

And a whatsnew in 1.4.4, this is a regression

and to be clear, this is a regression from 1.0.5?

I've added the deprecation in the 1.3 series, which causes the error

the code sample (test) was raising an exception before that

~/miniconda3/envs/pandas-1.1.5/lib/python3.9/site-packages/pandas/core/reshape/merge.py in __init__(self, left, right, how, on, left_on, right_on, axis, left_index, right_index, sort, suffixes, copy, indicator, validate)
    639                 "merging between different levels can give an unintended "
    640                 f"result ({left.columns.nlevels} levels on the left,"
--> 641                 f"{right.columns.nlevels} on the right)"
    642             )
    643             warnings.warn(msg, UserWarning)

~/miniconda3/envs/pandas-1.1.5/lib/python3.9/site-packages/pandas/core/generic.py in __getattr__(self, name)
   5139             if self._info_axis._can_hold_identifiers_and_holds_name(name):
   5140                 return self[name]
-> 5141             return object.__getattribute__(self, name)
   5142 
   5143     def __setattr__(self, name: str, value) -> None:

AttributeError: 'Series' object has no attribute 'columns'

Ok this is weird. I remember testing this and I thought I installed 1.2.5 for testing purposes. But can not reproduce now. Should we avoid backporting then?

the fix seems low risk so (imo) ok with the backport.

however, if there is an outstanding issue about whether we should be raising a warning here #47946 (comment), then should open an issue about that to ensure it is not forgotten. Whether any further changes get backported would be part of that discussion.

Ok, got it.

I think the warning is oldish, the current behaviour is definitely buggy (#47946 (comment)), so in principal I am ok with he warning, but we could also avoid treating tuples as MultiIndexes. Disallowing this would avoid any potential value dependent bugs

Yes, indeed with the current behaviour in other cases, the warning seems correct.
It's maybe worth a more general question whether we want to continue interpreting a tuple label as MultiIndex (but if you select a Series from a DataFrame with MI columns, you also get such tuple label, so it makes some sense to do that)