Exception that calls super().__init__ breaks copy.copy

[jakkdl]

Bug report

Bug description:

For some reason calling Exception.__init__ breaks copy.copys ability to detect args. The following code

import copy

class MyException(Exception):
    def __init__(self, x):
        super().__init__()
        self.x = x

my_exc = MyException(5)
copy.copy(my_exc)

gives

Traceback (most recent call last):
  File "./foo.py", line 9, in <module>
    copy.copy(my_exc)
  File "/usr/lib/python3.12/copy.py", line 97, in copy
    return _reconstruct(x, None, *rv)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/lib/python3.12/copy.py", line 253, in _reconstruct
    y = func(*args)
        ^^^^^^^^^^^
TypeError: MyException.__init__() missing 1 required positional argument: 'x'

Removing the super().__init__() line makes everything work as expected, but given that it's generally seen as good practice to call super().__init__() in child classes, and you may be working with a more complex class structure, this seems like pretty bad behavior.
I haven't dug deep enough into copy.copy or BaseException.__init__ to decide which one deserves the blame.

For the reason why I want to copy exceptions, see python-trio/flake8-async#298

CPython versions tested on:

3.9, 3.10, 3.11, 3.12, 3.13, 3.14

Operating systems tested on:

Linux

[catmeowjiao]

Your code would have copy.copy() execute MyException() directly without passing in x and therefore report an error, which is not a Python problem.

[jakkdl]

Your code would have copy.copy() execute MyException() directly without passing in x and therefore report an error, which is not a Python problem.

no? copy.copy() is perfectly capable of passing attributes to __init__. This works:

import copy

class MyException(Exception):
    def __init__(self, x):
        # no call to super().__init__()
        self.x = x

my_exc = MyException(5)
my_copy = copy.copy(my_exc)
assert my_copy.x == 5

and it's not just copying the attributes. If you add a print to __init__ it gets printed twice.

[ZeroIntensity]

Yeah, this looks like a bug. Confirmed it on the main branch. Thank you for the report!

[ZeroIntensity]

I suspect this is an issue with copy rather than Exception itself, so I'm marking this as a standard library problem for now. Would you like to try and fix this, @jakkdl?

[oremanj]

copy uses the pickle protocol to reduce the object to a state tuple and then create a new object from that state tuple.

>>> class MyException(Exception):
...     def __init__(self, x):
...         super().__init__()
...         self.x = x
...
>>> MyException(5).__reduce_ex__(4)
(<class '__main__.MyException'>, (), {'x': 5})

Based on https://docs.python.org/3/library/pickle.html#object.__reduce__, this is saying: "call MyException with no arguments, then pass {"x": 5} to its __setstate__ method". Which produces the error shown above. Unpickling produces the same error:

>>> pickle.loads(pickle.dumps(MyException(5)))
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: MyException.__init__() missing 1 required positional argument: 'x'

So I don't think this is copy.copy's fault. BaseException assumes that an exception object can be constructed from its args tuple. That's consistent with how an exception repr is printed out; MyException(5) reprs as MyException(). You will have a better time if you respect the args convention where all arguments to __init__ become self.args. But I think having exception objects not roundtrip through copy/pickle correctly when this convention is violated is a bug that ought to be fixed.

It looks like BaseException accepts args in both __new__ and __init__. So maybe one possible solution direction here is to let exceptions reduce more like any other object: instead of making a new instance via a call to the type object, it should be created by calling __new__ with the exception args, followed by __setstate__.

Unfortunately it seems some exceptions such as StopIteration and SystemExit have additional slots which are set only in __init__. Using __new__ to reconstruct them would create some quite insidious regressions:

Python 3.12.0 (v3.12.0:0fb18b02c8, Oct  2 2023, 09:45:56) [Clang 13.0.0 (clang-1300.0.29.30)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> ex = SystemExit(42)
>>> ex.code
42
>>> ex.args
(42,)
>>> ex.__reduce_ex__(4)
(<class 'SystemExit'>, (42,))
>>> ex2 = SystemExit.__new__(SystemExit, 42)
>>> ex2
SystemExit(42)
>>> ex2.code
>>> raise ex2
shell% echo $?
0

Even if CPython were to fix all examples of this internally, there's no way to know how much user code might be relying on the fact that copying/unpickling an exception calls __init__. And pickling exceptions is quite common in certain workflows, such as multiprocessing. So we may be stuck with the current behavior, in which case this would turn into a documentation issue.

[oremanj]

copy.copy() is perfectly capable of passing attributes to init. This works:

That's a manifestation of the fact that BaseException saves the args on both __new__ and __init__. If you don't override them by calling __init__ with no arguments, then you get self.args == (5,) based on what was passed to the original object's __new__.

[catmeowjiao]

I'm sorry for my mistake.