patch does not correctly restore state on objects without __dict__

[c00w]

Bug report

Bug description:

from unittest.mock import patch
class Proxy(object):
    config = {}
    def __getattr__(self, name):
        return self.config[name]
    def __setattr__(self, name, value):
        if name == 'config':
            object.__setattr__(self, name, value)
            return
        self.config[name] = value
    def __delattr__(self, name):
        self.config[name] = 'DEFAULT'

p = Proxy()
p.a = True
print(p.a)
with patch('__main__.p.a', False):
    print(p.a)
print(p.a)

Expected

True
False
True

Actual

True
False
DEFAULT

Note that this is a relatively niche issue, but for example in a config system (i.e. pytorch/pytorch#140779), if you don't want to have a delete work, (i.e. have it set it back to a default value), the code in

cpython/Lib/unittest/mock.py

Line 1637 in 94a7a4e

if not self.create and (not hasattr(self.target, self.attribute) or

will call delattr, and then not call setattr, because it still has the attr.

Possible fixes:
If delattr throws (maybe NotImplementederror), call setattr instead (this is nice and explicit + removes the hacky workaround I have to do).

If the value after it's been deleted doesn't equal the original value, reassign it? (this might require objects to be comparable, not sure if that is already an issue here).

Either way I wanted to see if there was any interest in fixing this, and if anyone saw a solution that seemed reasonable enough.

CPython versions tested on:

3.12

Operating systems tested on:

Linux

[c00w]

Except we're not a proxy class. In particular it makes no logical sense to delete these configs. Making delattr throw on all calls probably makes more sense for the usecase, then defining it to either reset to default, or to allow deletion.

[ZeroIntensity]

@bhagwant4040 Don't use LLMs to help with issues--they're not particularly helpful. If you want to help triage issues, see the devguide :)

[c00w]

@picnixz - No pressure, but curious if patches would be welcome here, or if the risk of breakage is high enough that pateches are unlikely to be accepted :).

[picnixz]

The issue I see here is that the object being patched has custom special setters and deleters.

The current library seems to special case some attributes so it's already a bit fragile. Ideally I would also have expected the original value to be set (via some setattr, possibly suppressing AttributeError raised by custom setattr) but this may break some code out there though I don't know how much.

I think users expect the patch to restore the old value even if you have custom __delattr__. But I don't have an immediate and generic fix for this one. There are many other cases where you create dictlike objects for configuration with default values.

By the way, from a OOP PoV, I wouldn't have changed __delattr__ but rather returned DEFAULT in __getattr__ instead. I don't think __delattr__ should be responsible for setting a default value but that is a matter of personal taste.

Now, before opening a PR, I think this should be discussed on https://discuss.python.org/c/ideas/6.