REGR: fix return class in _constructor_from_mgr for simple subclasses

[ivirshup]

• closes BUG: pandas 2.1.2 changes how copy works #55763 (Replace xxxx with the GitHub issue number)
• Tests added and passed if fixing a bug or adding a new feature
• All code checks passed.
• Added type annotations to new arguments/methods/functions.
• Added an entry in the latest doc/source/whatsnew/vX.X.X.rst file if fixing a bug or adding a new feature.

This is a draft for fixing #55763. I'm not familiar with this part of the code, but this seems to fix the issue locally.

This is addressing changes brought in by #54922. Instead of checking the class of ._constructor in _constructor_from_mgr, this checks the class being called.

We were relying on most operations on our subclass returning a pandas dataframe. AFAICT, the previous PR changed the behavior so that most operations would return the class of the instance even if _constructor was not set.

Once I see that this doesn't break tests on CI, added tests will be like:

class SimpleSubClass(DataFrame):
    """A subclass of DataFrame that does not define a constructor."""
    pass

class TestSubclassWithoutConstructor():

    def test_copy(self):
        expected = DataFrame({"a": [1, 2, 3]})
        result = SimpleSubClass(expected).copy()

        tm.assert_frame_equal(result, expected)

    def test_groupby(self):
        df = SimpleSubClass(DataFrame({"a": [1, 2, 3]}))

        for _, v in df.groupby("a"):
            assert isinstance(v, DataFrame)

added to test_subclass.py.

[jbrockmendel]

cc @jorisvandenbossche im fine with the change here bc it improves perf the base classes. no strong opinion on the actual behavior.

[jorisvandenbossche]

it improves perf the base classes

Do you mean that the check type(self) is DataFrame is a bit faster than self._constructor is DataFrame? (it's indeed a few nanoseconds faster)

[ivirshup]

I've addressed the comment, added a test for subclasses of Series, and added a release note.

This does throw a depreciation warning:

import pandas as pd

class SimpleDataFrameSubClass(pd.DataFrame):
    """A subclass of DataFrame that does not define a constructor."""

s = SimpleDataFrameSubClass(pd.DataFrame({"a": [1, 2, 3]}))
s.copy()

<ipython-input-2-6e1d2ee6d3ad>:1: DeprecationWarning: Passing a BlockManager to DataFrame is deprecated and will raise in a future version. Use public APIs instead.
  s.copy()

But the previous fix seems to have done this as well.

It would make sense if this didn't warn, since the user code in my example above can't obviously do anything about this warning besides silence it. It may be out of scope for this regression though, since this PR does not introduce this warning.

[jorisvandenbossche]

Yes, we still have to figure out how to deal with this deprecation warning in subclasses. But, that deprecation warning is only for pandas 2.2, and so won't yet be visible in a pandas 2.1.3.

And I think my other comment might solve this warning for the simple case where the subclass doesn't override the constructor.

[jorisvandenbossche]

Assuming the tests pass, this looks good to me!

[jorisvandenbossche]

Thanks @ivirshup !