DEPR: require _constructor/_constructor_sliced to return a class

[jbrockmendel]

(DataFrame|Series)Subclass._constructor(_sliced|_expanddim)? does not currently need to return a class. It can return a callable. This means that we cannot write obj._constructor_sliced.some_method, which has caused bugs/confusion at times. Most recently this surprised a contributor in #51765 and makes the fix there less straightforward.

AFAIK the only/main subclass that relies on this is geopandas, which inspects the arguments to decide which subclass to use. IIUC this could be accomplished with a __new__ method just as easily. cc @jorisvandenbossche

We could even go as far as deprecating _constructor entirely in favor of type(self) and just having _constructor_sliced/_constructor_expanddim.

[jbrockmendel]

@phofl since passing Managers to the constructors came up in another issue: my preferred solution is to do this deprecation, then implement something like NDFrame._from_mgr to pass Manager objects to instead of the constructors

[phofl]

+1

[jorisvandenbossche]

With the idea that _from_mgr is also available to override for subclasses? (it would be needed to preserve the current subclassing functionality)

I think for internal purposes, more consistently using something like _from_mgr instead of _constructor makes sense for code clarity. But I am not sure it solves any of the issues from the top post about interaction with subclasses?

[jbrockmendel]

Good catch on an implicit assumption worth making explicit! I was thinking we'd call self._constructor._from_mgr where relevant. For that usage an overridden self._from_mgr would work fine. For self._constructor_sliced._from_mgr (which im not sure how often it would be needed) we'd either need a _from_mgr_sliced or for _constructor_sliced to be a class.

[jorisvandenbossche]

I was thinking we'd call self._constructor._from_mgr where relevant.

But then you can't deprecate _constructor? (and what would be the benefit of doing self._constructor._from_mgr compared to calling self._from_mgr?)

[jbrockmendel]

_constructor._from_mgr would only work if _constructor reliably returned a class. So the deprecation is for allowing non-class _constructor

[jorisvandenbossche]

Ah, yes, of course. Still, the second part of my question still holds I think? I.e. what would be the benefit of doing self._constructor._from_mgr compared to calling self._from_mgr?