Inconsistency in handling None handlers in the Python and C implementations of pickle

In the Python implementation of pickle setting some handlers to None causes a fallback to other methods.

• Setting Pickler.reducer_override to None has the same effect as not setting this attribute -- falling back to other methods.
• Setting the dispatch_table to None leads to falling back to __reduce_ex__.
• Setting the __reduce_ex__ attribute to None leads to falling back to __reduce__.
• Setting the __setstate__ attribute to None leads to falling back to the default implementation of __setstate__ (setting __dict__ and slots).

In the C implementation all this leads to TypeError: 'NoneType' object is not callable.

The copy module follows the Python implementation of pickle, except that it fails for any false value of __reduce__, not just None.

In most other code setting a dunder method to None leads to a TypeError (either raise explicitly or just generic "'NoneType' object is not callable").

We should decide what behavior should be in the pickle and copy module. On one hand, the Python implementation was primary. On other hand, it was inconsistent between the pickle and copy modules and was already changed in the past, the C implementation of pickle is now used by most users, and it is more consistent with other code.

See also:
https://mail.python.org/archives/list/python-dev@python.org/thread/YGAK34DRWJFSIV2VZ4NC2J24XO37GCMM/
https://docs.python.org/3.10/reference/datamodel.html#id2
https://github.com/python/cpython/issues/70146
https://github.com/python/cpython/issues/55781
https://github.com/python/cpython/issues/60755

Linked PRs

• gh-103035
• gh-109498

Commencez par lire les implémentations Python et C de pickle et du module copy, puis examinez les pull requests liées gh-103035 et gh-109498. La finalisation nécessite de s’accorder sur le comportement des gestionnaires None et d’assurer un traitement cohérent dans toutes les implémentations concernées.