I couldn't figure out how to make a crashing test for batch_list_exact. For that case, save() seems to assume that its parameter obj is borrowed, so I think it is safe, but I'm not 100% sure.

The reason that batch_dict_exact crashes is because save(self, key, 0) clears the dict and leaves value as junk, and then save(self, value, 0) crashes, so I think it is only a problem two calls to save() in a row.

Is it worth adding Py_INCREF to make it more obviously correct?

What if persistent_id() clears the list and returns None?

You're right, here's a test case that segfaults:

import pickle
import io

class P(pickle.Pickler):
    def persistent_id(self, obj):
        if obj is a[0]:
            a.clear()
        return None

a = [[[[]]]]
P(io.BytesIO()).dump(a)

Updating the PR soon.

No refleaks

.\python.bat -m test test_pickle -R3:3
Running Debug|x64 interpreter...
0:00:00 Run tests sequentially
0:00:00 [1/1] test_pickle
beginning 6 repetitions
123456
......
test_pickle passed in 3 min 37 sec

Thanks @sweeneyde for the PR 🌮🎉.. I'm working now to backport this PR to: 3.10, 3.11.
🐍🍒⛏🤖

GH-93707 is a backport of this pull request to the 3.11 branch.

GH-93708 is a backport of this pull request to the 3.10 branch.