Consider allowing frozendicts to be assigned to instance dicts

[rhettinger]

Currently, can make an assignment to an instance's __dict__ attribute but the value must be an actual dict. It should allow frozendict as well.

This would be useful for making immutable instances without the awkward work arounds used in frozen dataclasses.

>>> class Point:
... def __init__(self, x, y):
... self.__dict__ = frozendict(x=x, y=y)
...
 
>>> p = Point(10, 20)
Traceback (most recent call last):
 File "<pyshell#53>", line 1, in <module>
 p = Point(10, 20)
 File "<pyshell#52>", line 3, in __init__
 self.__dict__ = frozendict(x=x, y=y)
TypeError: __dict__ must be set to a dictionary, not a 'frozendict'

Linked PRs

• gh-145119: Allow frozendict to be assigned to instance __dict__ #145123

[mpkocher]

Some other considerations:

• frozendict "immutable" style is adding entropy
• creates challenges for linting tools and static analysis tools to understand the "immutable" semantics

Also, to be pedantic, I'm not sure that frozendict fundamentally mean it's immutable.

class Point:
 def __init__(self, x, y):
 self.__dict__ = frozendict(x=x, y=y)
 
 def move_x(self, x):
 self.__dict__ |= frozendict(x=x)

Adding self.__dict__:Final = frozendict(x=x, y=y) might help, but I don't think this is a good pattern to encourage.

I think the current dataclasses model is a reasonable way to have immutability in Python.

[sergey-miryanov]

self.__dict__ |= frozendict(x=x) - you create a new frozendict.

[mpkocher]

self.__dict__ |= frozendict(x=x) - you create a new frozendict.

Correct. That's the point I'm trying to make.

Is the self.__dict__ usage any different than x in the example below?

@dataclass
class A:
 x: frozenset[int]
 
 
a = A(frozenset({1,2}))
a.x |= frozenset({3,4})

[vstinner]

Also, to be pedantic, I'm not sure that frozendict fundamentally mean it's immutable.

If we want to make it more immutable, one option would be to disallow setting/deleting object.__dict__ if it's a frozendict.