Change __getstate__ and __setstate__ to use a dict (#1004)

[djipko]

Summary

We use a dict instead of a tuple, so that we assign only existing attributes on unpickle, if the class
definition has changed in the meantime.

Resolves #1004

[hynek]

Looks like we're missing a branch: https://github.com/python-attrs/attrs/runs/7863540189?check_suite_focus=true#step:6:22

[hynek]

I think it’s actually an interesting case missing: what are we gonna do, if there’s extra names? By your motivation, we should actually blow up, no?

[djipko]

I think it’s actually an interesting case missing: what are we gonna do, if there’s extra names? By your motivation, we should actually blow up, no?

So the case missing is when what we re-hydrate has less attributes than our current class implementation declares, and we have to leave some of them empty. So that's an interesting one. I don't think we should necessarily blow up in every case, as there's some chance calling code can handle this (say, check for None or hasattr). The difference with the reported case and this is that in the former, the calling code stands no chance, it gets a completely wrong value (and why I thought we should fix it).

I think failing early (on unpickling) is tempting here, but might be too aggressive. It is likely that the calling code that doesn't handle this well and tries to access a non existing attribute will blow up anyway. Another potential solution would be to check if we have an .init on the attribute and assign that but even that seems over-reaching as it re-implements parts of initialisation of the class which is clearly skipped when unpickling by design.

At the end of the day, the point of this is to avoid a clear (if a bit obscure :) ) foot gun, and not re-implement versioning on top of pickle. People who actually need and want that can use protobuf or a number of actual solutions to this problem. So I would go with - just skip it (as it would have before). I will add a test

[djipko]

$ tox -e coverage-report
coverage-report installed: coverage==6.4.4,tomli==2.0.1
coverage-report run-test-pre: PYTHONHASHSEED='1249099045'
coverage-report run-test: commands[0] | coverage combine
Combined data file .coverage.dubndipanovmacbook1.87338.100213
Combined data file .coverage.dubndipanovmacbook1.87526.556153
coverage-report run-test: commands[1] | coverage report
Name    Stmts   Miss Branch BrPart  Cover   Missing
---------------------------------------------------
---------------------------------------------------
TOTAL    1494      0    723      0   100%

19 files skipped due to complete coverage.
______________________________________________________________________ summary _______________________________________________________________________
  coverage-report: commands succeeded
  congratulations :)

Got this locally now so feeling hopeful 🤞 😅

[hynek]

Cool, thanks! I'll write a news fragment.

[Mirandatz]

I feel like this may be a breaking change. Do objects serialized with the previous version can be deserialized with the new version?

[hynek]

They cannot. A fix has been added in #1085 that will be in 23.1, but (de)serializing classes across versions is risky (in general, not just the attrs case) so we'd encourage to look into avoiding that.