ImmutableCollections should have serialized forms which are independent of implementation

[gabbard]

• [x] ImmutableSet
• [ ] ImmutableList
• [x] ImmutableListMultiDict
• [x] ImmutableSetMultiDict

[gabbard]

@nicomitchell : you'll need to override __setstate__ and __getstate__ to control the serialized form. Don't bother with ImmutableList because it is going to die soon. For ImmutableSet, you can just use the ordering tuple as the serialized form. For the multi-dicts, use a tuple of key-value pairs.

[berquist]

Thought I was misunderstanding pickling, but there is in fact a blocker: https://github.com/python-attrs/attrs/issues/512 (with possible PR https://github.com/python-attrs/attrs/pull/513).

[gabbard]

When first implementing these we used attrs for convenience; is it buying us anything at this point or could we easily drop it? IIRC the main thing we get is the frozen=True behavior, which we could just copy.

[berquist]

Writing slots doesn't seem bad. Is making a frozen class as simple as https://github.com/python-attrs/attrs/blob/c2a9dd8e113a0dc72f86490e330f25bc0111971a/src/attr/_make.py#L478? Not like it's hard to get around it if you really want to...

from attr import attrs

class C:
    def __init__(self, val):
        self.val = val

@attrs(auto_attribs=True, frozen=False)
class A:
    m: C

@attrs(auto_attribs=True, frozen=True)
class Af:
    m: C

if __name__ == "__main__":
    c1 = C(1)
    c2 = C(2)
    a = A(c1)
    af = Af(c2)
    assert a.m.val == 1
    assert af.m.val == 2
    a.m.val = 10
    af.m.val = 20
    assert a.m.val == 10
    assert af.m.val == 20

[gabbard]

Yep, that simple.

You can definitely get around frozen if you want to but it prevents getting around it inadvertently. If the user wants to intentionally shoot themselves in the foot, it's their own fault.

[berquist]

In that case it might be worthwhile, since the validation and factory functions should be similarly thin. This would only be for immutablecollections and not other repositories/codebases, right?

[gabbard]

Yes - attrs is in general very useful; we only want to remove it here.

[berquist]

Now that https://github.com/isi-vista/immutablecollections/pull/38 is done, the serialized form can actually be changed.

[berquist]

@nicomitchell I'm working on this now, so we don't have duplicated effort.