get_type_hints fails if there are un-annotated fields in a dataclass

[a-recknagel]

BPO	37948
Nosy	@ericvsmith, @serhiy-storchaka, @ilevkivskyi, @a-recknagel

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = 'https://github.com/ericvsmith'
closed_at = None
created_at = <Date 2019-08-26.06:29:12.163>
labels = ['3.8', 'type-bug', '3.7']
title = 'get_type_hints fails if there are un-annotated fields in a dataclass'
updated_at = <Date 2020-03-08.13:51:23.330>
user = 'https://github.com/a-recknagel'

bugs.python.org fields:

activity = <Date 2020-03-08.13:51:23.330>
actor = 'serhiy.storchaka'
assignee = 'eric.smith'
closed = False
closed_date = None
closer = None
components = []
creation = <Date 2019-08-26.06:29:12.163>
creator = 'arne'
dependencies = []
files = []
hgrepos = []
issue_num = 37948
keywords = []
message_count = 5.0
messages = ['350488', '350973', '355559', '356829', '363662']
nosy_count = 4.0
nosy_names = ['eric.smith', 'serhiy.storchaka', 'levkivskyi', 'arne']
pr_nums = []
priority = 'normal'
resolution = None
stage = None
status = 'open'
superseder = None
type = 'behavior'
url = 'https://bugs.python.org/issue37948'
versions = ['Python 3.7', 'Python 3.8']

Linked PRs

• gh-82129: Fix NameError on get_type_hints in dataclasses #122232
• gh-82129: Provide __annotate__ method for dataclasses from make_dataclass #122262

[a-recknagel]

When declaring a dataclass with make_dataclass, it is valid to omit type information for fields. __annotations__ understands it and just adds typing.Any, but typing.get_type_hints fails with a cryptic error message:

>>> import dataclasses
>>> import typing
>>> A = dataclasses.make_dataclass('A', ['a_var'])
>>> A.__annotations__
{'a_var': 'typing.Any'}
>>> typing.get_type_hints(A)
Traceback (most recent call last):
  File "<input>", line 1, in <module>
  File "/user/venvs/python_3.7/lib/python3.7/typing.py", line 973, in get_type_hints
    value = _eval_type(value, base_globals, localns)
  File "/user/venvs/python_3.7/lib/python3.7/typing.py", line 260, in _eval_type
    return t._evaluate(globalns, localns)
  File "/user/venvs/python_3.7/lib/python3.7/typing.py", line 464, in _evaluate
    eval(self.__forward_code__, globalns, localns),
  File "<string>", line 1, in <module>
NameError: name 'typing' is not defined

Adding typing.Any explicitly is an obvious workaround:

>>> B = dataclasses.make_dataclass('B', [('a_var', typing.Any)])
>>> typing.get_type_hints(B)
{'a_var': typing.Any}

There is already a bug filed regarding datalcasses and get_type_hints which might be related: https://bugs.python.org/issue34776

[ilevkivskyi]

It looks like #9518 will fix also this one.

[ericvsmith]

I'm not sure what can be done with this. The problem is that the decorator doesn't know what's in the caller's namespace. The type being added is "typing.Any". If the caller doesn't import typing, then get_type_hints will fail (as demonstrated here).

The only thing I can think of is using a type that's in builtins. "object" springs to mine, but of course that's semantically incorrect.

Or, maybe I could use "dataclasses.sys.modules['typing'].Any". I don't currently import sys (I don't think), but this should be a cheap import. Then if typing.get_type_hints() is called, we know typing will have already been importing.

But what if "dataclasses" isn't in the caller's namespace? I guess if I could find some way to navigate to sys.modules from __builtins__, that would largely work, absent playing games with builtins.

[ilevkivskyi]

I'm not sure what can be done with this. The problem is that the decorator doesn't know what's in the caller's namespace. The type being added is "typing.Any". If the caller doesn't import typing, then get_type_hints will fail (as demonstrated here).

IIUC the main problem is that get_type_hints() fails even if typing is imported. I would expect this to work (just repeating the original example in a more compact form):

import dataclasses
import typing
A = dataclasses.make_dataclass('A', ['a_var'])
typing.get_type_hints(A)  # This currently crashes

Interestingly, if I use a very similar call that it works:

>>> typing.get_type_hints(A, globalns=globals())
{'a_var': typing.Any}

So the core of the issue is that the globals are identified incorrectly, and indeed if I look at the generated class it looks wrong:

>>> A.__module__
'types'  # Should be '__main__'

I think we should fix the __module__ attribute of the dynamically generated dataclasses (for example the way it is done for named tuples).

Btw, #14166 may potentially fix the __module__ attribute here too.

[serhiy-storchaka]

PR 14166 does not fix this issue.

[AlexWaygood]

Reproduced on 3.11.07a

[DavidCEllis]

I ran into this when trying to make a quick dataclass to demo cattrs code generation.

import dataclasses
import cattrs
A = dataclasses.make_dataclass('A', ['a_var'])
a = A("a_value")
cattrs.unstructure(a)

NameError: name 'typing' is not defined. Did you forget to import 'typing'?

Initially I thought this was a cattrs bug, but on looking at the trace it uses get_type_hints so it's related to this same issue.

In 3.13.0b4, 3.12, 3.11 (and earlier) this still reproduces with:

import dataclasses
from typing import get_type_hints
A = dataclasses.make_dataclass('A', ['a_var'])
print(get_type_hints(A))

However, in 3.13.0b4, and 3.12 (but not 3.11 or earlier) if you plainly import typing this now "works" - I think because __module__ is now set when it was not in 3.11:

import dataclasses
import typing
A = dataclasses.make_dataclass('A', ['a_var'])
print(typing.get_type_hints(A))

I say "works" because it's evaluating the name 'typing' from the script file that defines the dataclass. Unlikely someone would do this, but I'd still consider this incorrect behaviour.

import dataclasses
from typing import get_type_hints
class typing:
    Any = int

A = dataclasses.make_dataclass('A', ['a_var'])
print(get_type_hints(A))

{'a_var': <class 'int'>}

It also fails if A is defined in a file that doesn't import typing and typing.get_type_hints is used from a separate file. Likewise if inspect.get_annotations(..., eval_str=True) is used you get the same error without typing needing to be imported anywhere.

Some possible solutions:

• Use a deferred import of typing only if someone is using make_dataclass with untyped fields and put in the actual typing.Any object.
• Track and remove the untyped fields from __annotations__ after the dataclass has been created
  • I think this is correct, but it might break code if it makes the assumption that a field existing means it's in __annotations__
• Put a dict-like object in __annotations__ that will import and return typing.Any for untyped fields on demand.
  • This is kind of fiddly and adds a whole class just for a presumably lightly used feature

[sobolevn]

Put a dict-like object in annotations that will import and return typing.Any for untyped fields on demand.

I think that this is a proper solution. This will be fixed after __annotate__ will be fully supported.

[DavidCEllis]

If you want to avoid importing typing for the creation of such a class you still have to temporarily put something else in place for dataclasses to use in construction, and then replace it with something that will correctly evaluate afterwards. Otherwise dataclasses will trigger the imports when it inspects the annotations to create the class in the first place.

PEP649/749 could change things, but can/should this also be fixed for 3.13/3.12?