update the allow_unmapped error message to be less confusing with dataclasses; consider adding InitVar to whitelist

Describe the use case

Dataclasses allow for Init-only variables by type hinting using InitVar. This allows for adding __init__ arguments, that are not stored in the dataclass. These are typically used by __post_init__ methods.

Sqlalchemy should should allow one to use InitVar by simply ignore these fields, as it does with ClassVar

It would be useful where one would want to simplify initialization of sqlalchemy dataclass orm objects.

Databases / Backends / Drivers targeted

All

Example Use

@dataclass
class Identity:
    display_name: str
    user_id: str
    principals: List[str] = field(default_factory=list)

@reg.mapped_as_dataclass()
class AuditLog:
    __tablename__ = "audit_log"
    alid: Mapped[int] = mapped_column(primary_key=True, init=False)
    action: Mapped[str] = mapped_column()
    timestamp: Mapped[datetime] = mapped_column(default_factory=datetime.now)
    data: Mapped[Dict] = mapped_column(JSONB, default_factory=dict)

    user_id: Mapped[str] = mapped_column(default_factory=None)
    identity: InitVar[Identity] = field(default=None)

    def __post_init__(self, identity: Identity):
        if identity:
            self.user_id = identity.user_id

identity = Identity(display_name="John Smith", user_id="john")
audit_log = AuditLog(identity=identity, action="delete_object")

Additional context

No response

[zzzeek]

hey thanks for that.

this is still expected behavior, though i guess the ask is that “InitVar” be added to the whitelist. Take a look at what the message says, and from there we can solve the problem by adding __allow_unmapped__ as mentioned:

@reg.mapped_as_dataclass()
class AuditLog:
    __allow_unmapped__ = True
    __tablename__ = "audit_log"
    alid: Mapped[int] = mapped_column(primary_key=True, init=False)
    action: Mapped[str] = mapped_column()
    timestamp: Mapped[datetime] = mapped_column(default_factory=datetime.now)

    user_id: Mapped[Optional[str]] = mapped_column(default=None)
    identity: InitVar[Optional[Identity]] = None

    def __post_init__(self, identity: Optional[Identity]):
        if identity:
            self.user_id = identity.user_id

This is tested as a use case for dataclasses at https://github.com/sqlalchemy/sqlalchemy/blob/e0eea374c2df82f879d69b99ba2230c743bbae27/test/orm/declarative/test_dc_transforms.py#L387 . the goal is to prevent attributes that are ambiguous for their intent to pass silently.

[sqla-tester]

Mike Bayer has proposed a fix for this issue in the main branch:

remove allow_unmapped requirement from dataclasses https://gerrit.sqlalchemy.org/c/sqlalchemy/sqlalchemy/+/4312