Bug: Jsonschema of DocList not correct #1521

Activity

1. JohannesMessner

   added this to DocArray backlogon May 9, 2023
2. JohannesMessner

   converted this from a draft issue on May 9, 2023
3. JohannesMessner

   self-assigned this

   on May 9, 2023
4. JohannesMessner commented on May 16, 2023

   

   JohannesMessner

   on May 16, 2023

   Last edited by JohannesMessner

   MemberAuthor

   More actions

   After digging into this for a while it seems like we have a bit of a problem.

   tl;dr: We cannot have working validation for DocList and have working jsonschema at the same time. The reason our validation currently works at all is that we have a bug that breaks pydantic in just the right way.

   longer version:

   • I discovered that DocList does not fully behave like a built-in generic: get_origin(DocList[Doc]) returns None instead of DocList. This means that pydantic does not pick it up as a list and does not treat it properly. For most use cases this can be worked around though, by addiing __origin__ and __args__ like below, and pydantic's custom implementation of get_origin will pick it up properly:

   # in AnyDocArray.__class_getitem__
   class _DocArrayTyped(cls):  # type: ignore
       doc_type: Type[BaseDoc] = cast(Type[BaseDoc], item)
       __origin__: Type['AnyDocArray'] = cls  # add this
       __args__: Tuple[Any, ...] = (item,)  # add this

   • The step above fixes Jsonschema and FastAPI, great success!
   • However, then validation does not work anymore:

   class MyDoc(BaseDoc):
       docs: DocList[ImageDoc]
   d = MyDoc(docs=DocList[ImageDoc]([ImageDoc()]))
   print(type(d.docs))  # this will be a `list` instead of `DocList[ImageDoc]`

   This error can be reproduced in "pure" pydantic, showing that it is not per se a docarray issue:

   T_item = TypeVar('T_item')
   
   
   class MyOtherList(List[T_item], typing.Generic[T_item]):
       @classmethod
       def __get_validators__(cls):
           yield cls.validate
   
       @classmethod
       def validate(cls, value, field, config):
           print('validating :)')  # this is never called/printed
           if isinstance(value, MyList):
               print(f'{type(value)=}')
               return value
           raise TypeError(f'Cannot validate')
   
   
   class OtherDoc(BaseModel):
       docs: MyOtherList[int]
   
   
   d = OtherDoc(docs=MyOtherList[int]([1, 2, 3]))
   print(type(d.docs))  # this will be a list instead of MyList

   • Basically, as soon as we actually make DocList a proper python list, validation doesn't work anymore :/
   • It is a known issue (validators in custom subclasses of Mapping are not called pydantic/pydantic#1457) and should be avoidable in pydantic v2: Parametrized generics validation is broken/inconsistent pydantic/pydantic#4161 (comment)

   Conclusion: The reason validation on DocList works right now is that we failed to implement some list-like things, and so pydantic doesn't pick it up as a real list. If we fix this bug, validation will break. Thus, we cannot have working validation and working jsonschema/fastapi at the same time.
5. samsja commented on May 16, 2023

   

   samsja

   on May 16, 2023

   Member

   More actions

   So we wait for Pydantic v2 to fix this ? Or we can undo the list like things and try to fix it (I don't think we should do it )
6. JohannesMessner commented on May 16, 2023

   

   JohannesMessner

   on May 16, 2023

   MemberAuthor

   More actions

   I would wait

   👍React with 👍1Reacted by samsja
7. JoanFM commented on May 16, 2023

   

   JoanFM

   on May 16, 2023

   Member

   More actions

   We wait and document

   👍React with 👍1Reacted by samsja
8. JoanFM

   moved this from In Progress to Blocked in DocArray backlogon May 19, 2023
9. JoanFM

   linked a pull request that will close this issuefix: fix DocList schema when using Pydantic V2 #1876on Mar 12, 2024
10. JoanFM

    closed this as completedin #1876on Mar 21, 2025
11. github-project-automation

    moved this from Blocked to Done in DocArray backlogon Mar 21, 2025