Summary: The descriptor __set_name__ functionality (introduced in Python 3.6) does not seem to be working correctly for dataclass.Field objects with a default pointing to a descriptor. I have attached a file demonstrating the trouble.

Details: If I set a dataclass class object field to a dataclass.field with a descriptor object for the default argument, the descriptor's __set_name__ method is not called during initialization. This is unexpected because descriptors themselves seem to work pretty much flawlessly, otherwise.

(Bravo on that by the way! Working descriptors isn't mentioned at all in the PEP as a feature but I was very pleased to see them working!!)

System details:
Python 3.7b02
Windows 10
PyCharm Community Edition

btw this is my first ever Python bug report; I hope I did a good job.

That's a tough one. Because C.d is not set to a descriptor at type creation time (it's set to a Field object), the __set_name__ behavior is never invoked. It's when the @DataClass decorator is called that C.d is set to D().

https://docs.python.org/3/reference/datamodel.html#creating-the-class-object

I'm not sure it's possible to work around this without duplicating some of the type creation code, and even then I'm not convinced it's doable.

I suppose I could, when overwriting the class member, check for inspect.ismethoddescriptor and call __set_name__ myself.

hmmm... if I check the C.d class attribute it seems to return the
descriptor instance object (not a field object) before any C instances have
been created. i guess this is just a part of how the dataclass
implementation works.

i didn't realize there's nothing "special" going on with descriptors here-
the descriptors "just work" by virtue of being set to the class attribute
at creation time. interesting.

maybe because of this descriptors should short-circuit the field creation
process entirely? that would be a shame though. having the option of
auto-including a descriptor in the class repr turns out to be very useful
and i'm already playing around with it in a project.

one idea: what if it there were a keyword argument to mark a field as a
descriptor, allowing tje descriptor to be set at type creation? this would
need to disallow init=True, i think. and setting a field default to a
descriptor class would then raise a type error.

---
Ricky.

"I've never met a Kentucky man who wasn't either thinking about going home
or actually going home." - Happy Chandler

On Mon, Mar 26, 2018 at 6:47 AM, Eric V. Smith <report@bugs.python.org>
wrote:

Eric V. Smith <eric@trueblade.com> added the comment:

I suppose I could, when overwriting the class member, check for
inspect.ismethoddescriptor and call __set_name__ myself.

----------
components: +Library (Lib)
versions: +Python 3.8

---

Python tracker <report@bugs.python.org>
<https://bugs.python.org/issue33141\>

---

Nick Coghlan suggested (on python-dev) to have Field implement __set_name__ and call through to the default value, if it exists.

That approach seems to work fine. I'll generate a PR with tests sometime before today's release.

Sounds like a relatively easy solution then. Hopefully it makes the beta 3 so I can use it immediately- thanks!

I suppose one downside of that solution is __set_name__ will be called for every Field whether or not it is need. Probably can't be helped without major complications.

I don't expect there to be too much usage of Field objects, so I'm not so worried about it.

If you can, try out the code in the PR. Thanks.

Eric, looking at the PR; note that if you do this for the __set_name__ check:

    if inspect.ismethoddescriptor(self.default):

...an object like the one below will not get its __set_name__ called, even though PEP-487 says it should:

class D:
	def __set_name__(self, o, n):
		self.name = n

class C:
	d: int = D()

if __name__ == "__main__":
	print(f"C.d.name = {C.d.name}") # __set_name__ was called
	assert inspect.ismethoddescriptor(C.d) # Error

Yes, I noticed the same thing. I'll remove the test.

Yeah and I think lines 2709-2712 of TestDescriptors also needs removed.

I've updated the PR.

I left those lines in, and added a different test. After all, it does need to work with real descriptors.

Looks great to me. Thanks!

New changeset de7a2f0 by Eric V. Smith in branch 'master':
bpo-33141: Have dataclasses.Field pass through __set_name__ to any default argument. (GH-6260)
de7a2f0

New changeset c6147ac by Miss Islington (bot) in branch '3.7':
bpo-33141: Have dataclasses.Field pass through __set_name__ to any default argument. (GH-6260)
c6147ac

Thanks for the bug report.