GeezerVison

whatsnew/3.11.rst shoud be updated?

whatsnew/3.11.rst shoud be updated?

Yes, good catch.

Here it is!
I can't quite come up with a way to summarize it for people who haven't read up on the topic. Hopefully it's enough.

Would it make sense to not add it to the limited API in Python 3.11, try to convert a few third party C extensions using it, and if everything goes well, add it to the limited API?

Would it make sense to not add it to the limited API in Python 3.11, try to convert a few third party C extensions using it, and if everything goes well, add it to the limited API?

It would! OTOH, addition to the limited API is also easy to revert, and the function itself already proved useful in stdlib.
Should I remove the stable ABI addition for now?

I think that makes sense. AFAICS, only a handful of devs (Christian Heimes, Hai Shi, Petr, Victor, me) have been using this function in the CPython code base. In this case, maybe it would make sense to try it out in the wild first. Who could we ask?

Should I remove the stable ABI addition for now?

For Py_NewRef/Py_XNewRef, I chose to add it to the stable ABI immediately. For PyType_GetModuleByDef(), I'm more scared because I introduced a bug and noticed it before your recent review. Sadly, only a minority of 3rd party C extensions are using heap types. Moreover, we discovered issues with heap types defined in the stdlib in Python 3.10:

• Some heap types didn't implement fully the GC protocol: no Py_TPFLAGS_HAVE_GC flag, or no traverse function, or objects not tracked by the GC
• Types became mutable: add Py_TPFLAGS_HAVE_GC flag
• tp_new=NULL didn't work anymore: add Py_TPFLAGS_DISALLOW_INSTANTIATION flag

These bugs made me even more careful than I was previously.

Maybe PyType_GetModuleByDef() is fine, but it's just that heap types in general became "a minefield" if you don't pay attention.

I love heap types, I want to use them everywhere. Maybe we need to enhance the documentation, add more warnings in PEP 630, and even implement new runtime checks (in development/debug mode?).

For example, creating a type with Py_TPFLAGS_HAVE_GC but with no traverse function now raises an exception.

Makes sense, but I'm skeptical of actually finding the issues (ones we can't fix in CPython patch release) unless this is in the limited API. Heap types -- and this function -- aren't very useful for existing projects if they aren't switching to the limited API.
Essentially all of the testing will be on that handful of devs anyway.

more warnings in PEP 630

Those are coming up!

Thank you!