Strange "Attributes without a default cannot follow attributes with one" when extending a dataclass from a package

[reinhrst]

Apologies upfront, this must be the worst bug report I've ever made. I can only reproduce it every now and then, and only in a quite complicated code base. Just logging it here in case it rings a bell (or someone else has the same issue).

I have a package (which we distribute through our own, private, pypi), with type-support (py.typed file present). There are no stub files, all types are inline.

The package defines a dataclass, e.g. file cars.py

 1. import dataclasses
 2.
 3.
 4.
 5. # some empty lines to illustrate a point
 6.
 7.
 8.
 9.
10. @dataclasses.dataclass(frozen=True)
11. class Car:
12.    name: str
13.    top_speed: float
14.    brand: str = "Renault"

Now in my main program (main.py) I need to get a dataclass that is very much like a Car, I just need to be able to set the top_speed to "uknown" (a string).

 1. import dataclasses
 2. from mypackage import cars
 3.
 4. @dataclasses.dataclass(frozen=True)
 5. class UnknownCar(cars.Car):
 6.    top_speed: str
 7.
 8. mycar = UnknownCar(name="unknown", top_speed="unkown", brand="unknown")

Now I would expect mypy to give a warning (Subclass not compatible with super class). What happens however is this:

The first run, no errors. Subsequent runs (same command) give the errors below (until .mypy-cache/ is removed, after which the first run again is errorfree)

main.py:6 error: Attributes without a default cannot follow attributes with one
main.py:12 error: Attributes without a default cannot follow attributes with one
main.py:13 error: Attributes without a default cannot follow attributes with one

So what seems to happen is that mypy (incorrectly) assumes that the top_speed for UnknownCar will be added as last argument (and then it complains, correctly, that the argument before it, brand, has a default)

The next 2 error however point to the lines where name and top_speed are defined in cars.py (even though the error says main.py). If I add more newlines to the cars.py file, these line numbers in the errormessage change, and if I give default values to them, the errors disappear. By giving a fake default (of dataclasses.field()) to all entries, all errors disappear. Note that no amount of type: ignore in the main.py file seems to be able to get rid of the errors.

The behaviour was observed (however not always consistently) in mypy 0.670, 0.701 and HEAD.

[ilevkivskyi]

Thanks for reporting!

Hm, interesting, looks like a bug in the caching logic. Could you please try also with --new-semantic-analyzer? (It is available on HEAD.)

[reinhrst]

Yes, --new-semantic-analyzer leads to the same errors on the second run (before crashing)

(the crash is on a different file, AssertionError: fallback can't be filled out until semanal; would you like this submitted as separate ticket?)

[ilevkivskyi]

would you like this submitted as separate ticket?

Yes, please open another issue for this.

[reinhrst]

Yes, please open another issue for this.

Done.

As a side note (may or may not be helpful): I do feel that around the area of code that gives this ttributes without a default cannot follow attributes with one is heavy with circular dependencies, and there were some other issues that felt like they may be due to not all files being processed in the right order by mypy. My first hunch therefore was that somehow the order of reading files changes if the file is in cache -- as I said, this may or may not be useful.

[ilevkivskyi]

is heavy with circular dependencies

This is why I proposed to try --new-semantic-analyzer, that should be much more robust w.r.t. circular dependencies etc. (at least this is our goal for it). Could you please try it on current master with #6972 merged locally?

[Phlogistique]

With mypy 0.720, which according to http://mypy-lang.org/news.html ships with the new semantic analyzer activated by default, the problem still happens to me.

[reinhrst]

Just checked this in 0.770, it seems to work now!