SQLAlchemy version 1.4.36 breaks SQLModel relationships

[archydeberker]

First Check

• I added a very descriptive title to this issue.
• I used the GitHub search to find a similar issue and didn't find it.
• I searched the SQLModel documentation, with the integrated search.
• I already searched in Google "How to X in SQLModel" and didn't find any information.
• I already read and followed all the tutorial in the docs and didn't find an answer.
• I already checked if it is not related to SQLModel but to Pydantic.
• I already checked if it is not related to SQLModel but to SQLAlchemy.

👆 Not quite true - this is definitely related to SQLAlchemy!

Commit to Help

• I commit to help with one of those options 👆

Example Code

from typing import Optional

from sqlmodel import Field, Relationship, SQLModel


class City(SQLModel, table=True):
    name: str = Field(primary_key=True)
    heroes: "Hero" = Relationship(back_populates="city")

class Hero(SQLModel, table=True):
    name: str = Field(primary_key=True)
    city_name: Optional[str] = Field(default=None,foreign_key="city.name")
    city: Optional[City] = Relationship(back_populates="heroes",
                              sa_relationship_kwargs=dict(cascade="all,delete")
                              )


if __name__ == "__main__":

    gotham = City(name="Gotham")
    batman = Hero(name="Batman", city=gotham)

    assert batman.name == 'Batman' # This is fine
    assert batman.city == gotham # This now breaks

Description

Our CI suddenly started failing, despite local SQLModel working fine. The issues turns out to be the transitive dependency on SQLAlchemy, which is weakly pinned: Github Actions pulled the latest version (1.4.36) and most of our tests started failing.

https://github.com/sqlalchemy/sqlalchemy/releases/tag/rel_1_4_36

The problem seems to be related to how relationships are defined, but I haven't yet dug into the SQLAlchemy changes enough to understand why that is.

I'm opening this issue chiefly to help anybody else who is confused by why suddenly their tests are failing. I'm happy to help fix it if it's affecting others too.

For the time being we have just pinned SQLAlchemy==1.4.34 in our requirements.txt.

Operating System

Linux, macOS

Operating System Details

Replicated locally and on Github Actions, both running in Docker

SQLModel Version

0.0.6

Python Version

3.9.10

Additional Context

We were previously running SQLAlchemy 1.4.34 locally and that works fine. Pinning to 1.4.36 breaks SQLModel.

[antont]

We pinned to 1.4.35, in our fork of sqlmodel. Happened to have that locally as upgraded recently and hadn't had problems. Ran into this same this afternoon. Thanks for reporting!

[JLHasson]

Same issue here, pinning to 1.4.35 resolved

[byrman]

It seems related to this recent change:

class DeclarativeMeta(
    _DynamicAttributesType, inspection.Inspectable["Mapper[Any]"]
):
    metadata: MetaData
    registry: "RegistryType"

    def __init__(
        cls, classname: Any, bases: Any, dict_: Any, **kw: Any
    ) -> None:
        # use cls.__dict__, which can be modified by an
        # __init_subclass__() method (#7900)
        dict_ = cls.__dict__  # This line is the culprit?!

https://github.com/sqlalchemy/sqlalchemy/blob/main/lib/sqlalchemy/orm/decl_api.py#L114

[jossefaz]

As @byrman wrote, the change of DeclarativeMeta seems to be the breaking change here.

[orm] [declarative] [bug] Modified the DeclarativeMeta metaclass to pass cls.dict
into the declarative scanning process to look for attributes, rather than
the separate dictionary passed to the type’s init() method. This
allows user-defined base classes that add attributes within an
init_subclass() to work as expected, as init_subclass() can
only affect the cls.dict itself and not the other dictionary. This
is technically a regression from 1.3 where dict was being used.

[byrman]

Not sure if this is a sustainable fix, I don't know how to leverage __init_subclass__, but adding 1 line after here makes things work again:

...
dict_used[rel_name] = rel_value
setattr(cls, rel_name, rel_value)  # Quick fix?

[jossefaz]

The SQLAlchemy maintainers confirmed that this is the issue, and also suggested a fix : sqlalchemy/sqlalchemy#7972 (reply in thread)

[byrman]

The fix I suggested is backwards compatible. Shall I make a pull request or is there more to it than setting rel_name on the cls?

[byrman]

Let me know if this can be improved upon: #322.