This is the implementation of PEP 487.

It adds a metaclass to types that calls a method on a class
once it is subclassed. This way one can customize the creation
of classes without the need to write an own metaclass.

As a second functionality, it calls a method on each descriptor
in a class, such that descriptors know their name.

This is a C implementation of PEP 487, directly in the Python core

I started reviewing Martin's patch, and I initially thought I had found a problem with the way __init_subclass__ is currently defined. It turned out I was wrong about it actually being broken, but I *do* now think it's inherently confusing, and we may be able to do something different that's more obviously correct (or at least easier to document - it was proposing revisions to the documentation that got me thinking along this path).

Specifically, I was thinking using super() in either the zero argument form or the explicit form could create an infinite loop due to the way we're currently proposing to interact with the MRO. Consider:

    class BaseClass:
        @classmethod
        def __init_subclass__(cls):
            super(cls, BaseClass).__init_subclass__()

    class SubClass(BaseClass):
        pass

If the initial call made by type.__new__() is effectively "SubClass.mro()[1].__init_subclass__()", then the super() call is going to call BaseClass.__init_subclass__ again.

However, it turned out I was wrong, as that's not what happens: the call made by the type machinery is instead "super(SubClass, SubClass).__init_subclass__", which gets it to the right place in the MRO and causes further super() calls to do the right thing.

However, the "more obviously correct" signature that occurred to me was to do this instead:

    class BaseClass:
        @classmethod
        def __init_subclass__(cls, subcls):
            super(cls, BaseClass).__init_subclass__(subcls)

    class SubClass(BaseClass):
        pass

Then the invocation from type.__new__ could be defined more simply as:

    SubClass.mro()[1].__init_subclass__(SubClass)

In all cases then (regardless of where you were in the MRO), "cls" would refer to "the class first in the MRO after the class being defined" and "subcls" would refer to "the class currently being defined".

If you consider the plugin example in the PEP, with the revised signature, it would look like:

    class PluginBase:
        subclasses = []

        def __init_subclass__(cls, subcls, **kwargs):
            super().__init_subclass__(**kwargs)
            cls.subclasses.append(subcls)

And *even if the subclass being defined shadowed the "subclasses" attribute*, this initialisation would still work. (You can still get yourself in trouble if a subclass somewhere else in the MRO shadows the attribute, but that's life in complex type hierarchies)

In the version in the PEP, the fact that "cls" is actually a subclass, and we're relying on the MRO to find "subclasses" is a really subtle implementation detail, while having two parameters makes it clear that "the class defining __init_subclass__" is distinct from the "new subclass being defined".

Scratch that, my revised idea is fundamentally broken, as this example illustrates:

>>> class BaseClass: pass
... 
>>> class OtherClass: pass
... 
>>> class SubClass(OtherClass, BaseClass): pass
... 
>>> SubClass.mro()
[<class '__main__.SubClass'>, <class '__main__.OtherClass'>, <class '__main__.BaseClass'>, <class 'object'>]


The PEP as written handles this correctly, while the alternative signature would fail miserably ("OtherClass" wouldn't chain up to "BaseClass" properly). So I'll review the rest of the patch, and we can figure out the documentation problem later.

Martin's current implementation basically looks good to me - I've mainly suggested some changes to the documentation and additional test cases that help stress test some of the more complex inheritance hierarchies described above, although there are a few other other suggestions as well.

Thanks for the nice review!

I made the proposed changes, see attached patch.

I am still waiting for a decision whether type.__setattr__ should call __set_name__ from python-dev, once that's sorted out I'll implement and test the one or the other behavior.

That's a no from me.

On Sunday, July 24, 2016, Martin Teichmann <report@bugs.python.org> wrote:

>
> Martin Teichmann added the comment:
>
> Thanks for the nice review!
>
> I made the proposed changes, see attached patch.
>
> I am still waiting for a decision whether type.__setattr__ should call
> __set_name__ from python-dev, once that's sorted out I'll implement and
> test the one or the other behavior.
>
> ----------
> Added file: http://bugs.python.org/file43854/pep487.patch
>
> _______________________________________
> Python tracker <report@bugs.python.org <javascript:;>>
> <http://bugs.python.org/issue27366>
> _______________________________________
>

I looked over the patch once more and found some places where
I didn't follow normal coding standards. I changed that, namely
now the code returns -1 meaning an exception happened and 0 on
success, which is what many functions in typeobject.c do.

So I think this patch should be ready.

Martin's latest patch looks good to me, so I'm applying it now and will push it once a local run of the test suite gives it the thumbs up :)

New changeset ecc7bff738e0 by Nick Coghlan in branch 'default':
Issue #27366: Implement PEP 487
https://hg.python.org/cpython/rev/ecc7bff738e0

Thanks once again Martin, especially for your patience with the long process in getting this proposal all the way through to resolution :)

I mostly applied your patch as-is, but tweaked a few aspects of the documentation before committing it (mainly expanding the What's New entry, and showing a few more of the moving parts in the  __init_subclass__ example).

Sorry, I'm a bit late to the party. Here are some tweaks to ecc7bff738e0. I've added versionadded directives, updated the tests to use new style classes and removed a duplicate sentence from the __init_subclass__ docstring.

Good catches Berker - go ahead and apply the improvements whenever's convenient :)

New changeset 8747e3455ecb by Berker Peksag in branch 'default':
Issue #27366: Tweak PEP 487 documentation
https://hg.python.org/cpython/rev/8747e3455ecb

Thanks for the review, Nick! (and also thanks to Martin for the great PEP!)

Hello Martin, and thank you for your patch! However, this patch removed the ability to pass keyword arguments to `type`, and it's not documented anywhere. I believe it should be documented at least in e.g. the "Changes in the Python API" of the What's New in Python 3.6 document. Anyone cares to submit a patch?

D'oh, another good catch Emanuel - and I'm even the one that raised the need to mention that in the Porting notes during the python-dev discussion :P

I'll post an update to the What's New shortly.

New changeset 313e8fdb0d0c by Nick Coghlan in branch 'default':
Issue 27366: PEP 487 docs updates
https://hg.python.org/cpython/rev/313e8fdb0d0c

The new porting note doesn't quite capture all the subtleties of the situation, but should be sufficient for folks to get any affected custom metaclasses working again (since the key is to avoid passing those parameters up to type.__new__).

I also added a note about the fact that __init_subclass__ implementations don't have access to the metaclass hint, but can retrieve the actual metaclass based on the passed in class object.

New changeset 545bfa4c20eb by Victor Stinner in branch 'default':
Issue #27366: Fix init_subclass()
https://hg.python.org/cpython/rev/545bfa4c20eb