PEP 634 structural pattern matching adds an auto-generated __match_args__ attribute to classes with the dataclass decorator and to namedtuples.

This change is currently not documented in the dataclass and namedtuple documentation, nor is it mentioned in PEP 557 Data Classes.


I would suggest adding mentions of this behaviour in those three places.

Additionally I think adding a new parameter to switch off generating __match_args__ to dataclass should be considered, as everything else generated by dataclass can be switched off using a parameter.

Turning off the generation of __match_args__ should be a separate issue.

Ok. I created https://bugs.python.org/issue43764 for that.

> This change is currently not documented in the dataclass 
> and namedtuple documentation,

I don't think we need to do that. In the end, __match_args__ will be added in many places.  The right place to document it is in the pattern matching docs.

I agree that __match_args__ shouldn't have to be added to the documentation of any class that supports it, however dataclass and (maybe to a lesser extend) NamedTuple aren't themselves classes, but aid in creating own classes. Their effects on classes generated by them should be documented.
The documentation also mentions other fields/methods generated by dataclass and NamedTyple.

I do think it's worth mentioning that __match_args__ is automatically generated by @dataclass (if the class doesn't already define __match_args__). That seems like behavior that's very specific to @dataclass. Where else would you find out that defining your own __match_args__ prevents @dataclass from adding it?

If we end up adding a “match_args=False” option to the dataclass decorator (issue 43764), then the docs for that option would probably be sufficient for this purpose.

I’m imagining something like the docs for the “init” option, just substituting “__match_args__” for “__init__()”.

For named tuples, there isn't an include/exclude option, so no extra mention is warranted.

I think for namedtuple a short mention in the opening paragraph, where it also mentions the generation of a docstring and __repr__ method should be enough.

Do you need anything from me here?

> Do you need anything from me here?

No, I don't think so. Thanks.

> For named tuples, there isn't an include/exclude option, so no extra mention is warranted.

I think a note in the docs could still be valuable, if only because defining __match_args__ for named tuples changes the inherited tuple behavior for positional matches quite significantly: "case tuple(x)" and "case MyNT(x)" don't mean the same thing.

(Although perhaps this could be inferred from the new constructor signature.)