Integration with dataclasses and attrs

SQLAlchemy as of version 2.0 features “native dataclass” integration where an Annotated Declarative Table mapping may be turned into a Python dataclass by adding a single mixin or decorator to mapped classes.

Added in version 2.0: Integrated dataclass creation with ORM Declarative classes

There are also patterns available that allow existing dataclasses to be mapped, as well as to map classes instrumented by the attrs third party integration library.

Declarative Dataclass Mapping

SQLAlchemy Annotated Declarative Table mappings may be augmented with an additional mixin class or decorator directive, which will add an additional step to the Declarative process after the mapping is complete that will convert the mapped class in-place into a Python dataclass, before completing the mapping process which applies ORM-specific instrumentation to the class. The most prominent behavioral addition this provides is generation of an __init__() method with fine-grained control over positional and keyword arguments with or without defaults, as well as generation of methods like __repr__() and __eq__().

From a PEP 484 typing perspective, the class is recognized as having Dataclass-specific behaviors, most notably by taking advantage of PEP 681 “Dataclass Transforms”, which allows typing tools to consider the class as though it were explicitly decorated using the @dataclasses.dataclass decorator.

Note

Support for PEP 681 in typing tools as of April 4, 2023 is limited and is currently known to be supported by Pyright as well as Mypy as of version 1.2. Note that Mypy 1.1.1 introduced PEP 681 support but did not correctly accommodate Python descriptors which will lead to errors when using SQLAlchemy’s ORM mapping scheme.

See also

https://peps.python.org/pep-0681/#the-dataclass-transform-decorator - background on how libraries like SQLAlchemy enable PEP 681 support

Dataclass conversion may be added to any Declarative class either by adding the MappedAsDataclass mixin to a DeclarativeBase class hierarchy, or for decorator mapping by using the registry.mapped_as_dataclass() class decorator.

The MappedAsDataclass mixin may be applied either to the Declarative Base class or any superclass, as in the example below:

from sqlalchemy.orm import DeclarativeBase from sqlalchemy.orm import Mapped from sqlalchemy.orm import mapped_column from sqlalchemy.orm import MappedAsDataclass class Base(MappedAsDataclass, DeclarativeBase):  """subclasses will be converted to dataclasses""" class User(Base): __tablename__ = "user_account" id: Mapped[int] = mapped_column(init=False, primary_key=True) name: Mapped[str]

Or may be applied directly to classes that extend from the Declarative base:

from sqlalchemy.orm import DeclarativeBase from sqlalchemy.orm import Mapped from sqlalchemy.orm import mapped_column from sqlalchemy.orm import MappedAsDataclass class Base(DeclarativeBase): pass class User(MappedAsDataclass, Base):  """User class will be converted to a dataclass""" __tablename__ = "user_account" id: Mapped[int] = mapped_column(init=False, primary_key=True) name: Mapped[str]

When using the decorator form, the registry.mapped_as_dataclass() decorator is supported:

from sqlalchemy.orm import Mapped from sqlalchemy.orm import mapped_column from sqlalchemy.orm import registry reg = registry() @reg.mapped_as_dataclass class User: __tablename__ = "user_account" id: Mapped[int] = mapped_column(init=False, primary_key=True) name: Mapped[str]

The same method is available in a standalone function form, which may have better compatibility with some versions of the mypy type checker:

from sqlalchemy.orm import Mapped from sqlalchemy.orm import mapped_as_dataclass from sqlalchemy.orm import mapped_column from sqlalchemy.orm import registry reg = registry() @mapped_as_dataclass(reg) class User: __tablename__ = "user_account" id: Mapped[int] = mapped_column(init=False, primary_key=True) name: Mapped[str]

Added in version 2.0.44: Added mapped_as_dataclass() after observing mypy compatibility issues with the method form of the same feature

Class level feature configuration

Support for dataclasses features is partial. Currently supported are the init, repr, eq, order and unsafe_hash features, match_args and kw_only are supported on Python 3.10+. Currently not supported are the frozen and slots features.

When using the mixin class form with MappedAsDataclass, class configuration arguments are passed as class-level parameters:

from sqlalchemy.orm import DeclarativeBase from sqlalchemy.orm import Mapped from sqlalchemy.orm import mapped_column from sqlalchemy.orm import MappedAsDataclass class Base(DeclarativeBase): pass class User(MappedAsDataclass, Base, repr=False, unsafe_hash=True):  """User class will be converted to a dataclass""" __tablename__ = "user_account" id: Mapped[int] = mapped_column(init=False, primary_key=True) name: Mapped[str]

When using the decorator form with registry.mapped_as_dataclass() or mapped_as_dataclass(), class configuration arguments are passed to the decorator directly:

from sqlalchemy.orm import registry from sqlalchemy.orm import Mapped from sqlalchemy.orm import mapped_column reg = registry() @reg.mapped_as_dataclass(unsafe_hash=True) class User:  """User class will be converted to a dataclass""" __tablename__ = "user_account" id: Mapped[int] = mapped_column(init=False, primary_key=True) name: Mapped[str]

For background on dataclass class options, see the dataclasses documentation at @dataclasses.dataclass.

Attribute Configuration

SQLAlchemy native dataclasses differ from normal dataclasses in that attributes to be mapped are described using the Mapped generic annotation container in all cases. Mappings follow the same forms as those documented at Declarative Table with mapped_column(), and all features of mapped_column() and Mapped are supported.

Additionally, ORM attribute configuration constructs including mapped_column(), relationship() and composite() support per-attribute field options, including init, default, default_factory and repr. The names of these arguments is fixed as specified in PEP 681. Functionality is equivalent to dataclasses:

Another key difference from dataclasses is that default values for attributes must be configured using the default parameter of the ORM construct, such as mapped_column(default=None). A syntax that resembles dataclass syntax which accepts simple Python values as defaults without using @dataclases.field() is not supported.

As an example using mapped_column(), the mapping below will produce an __init__() method that accepts only the fields name and fullname, where name is required and may be passed positionally, and fullname is optional. The id field, which we expect to be database-generated, is not part of the constructor at all:

from sqlalchemy.orm import Mapped from sqlalchemy.orm import mapped_column from sqlalchemy.orm import registry reg = registry() @reg.mapped_as_dataclass class User: __tablename__ = "user_account" id: Mapped[int] = mapped_column(init=False, primary_key=True) name: Mapped[str] fullname: Mapped[str] = mapped_column(default=None) # 'fullname' is optional keyword argument u1 = User("name")

Column Defaults

In order to accommodate the name overlap of the default argument with the existing Column.default parameter of the Column construct, the mapped_column() construct disambiguates the two names by adding a new parameter mapped_column.insert_default, which will be populated directly into the Column.default parameter of Column, independently of what may be set on mapped_column.default, which is always used for the dataclasses configuration. For example, to configure a datetime column with a Column.default set to the func.utc_timestamp() SQL function, but where the parameter is optional in the constructor:

from datetime import datetime from sqlalchemy import func from sqlalchemy.orm import Mapped from sqlalchemy.orm import mapped_column from sqlalchemy.orm import registry reg = registry() @reg.mapped_as_dataclass class User: __tablename__ = "user_account" id: Mapped[int] = mapped_column(init=False, primary_key=True) created_at: Mapped[datetime] = mapped_column( insert_default=func.utc_timestamp(), default=None )

With the above mapping, an INSERT for a new User object where no parameter for created_at were passed proceeds as:

>>> with Session(e) as session: ... session.add(User()) ... session.commit() 
BEGIN (implicit) INSERT INTO user_account (created_at) VALUES (utc_timestamp()) [generated in 0.00010s] () COMMIT

Integration with Annotated

The approach introduced at Mapping Whole Column Declarations to Python Types illustrates how to use PEP 593 Annotated objects to package whole mapped_column() constructs for re-use. While Annotated objects can be combined with the use of dataclasses, dataclass-specific keyword arguments unfortunately cannot be used within the Annotated construct. This includes PEP 681-specific arguments init, default, repr, and default_factory, which must be present in a mapped_column() or similar construct inline with the class attribute.

Changed in version 2.0.14/2.0.22: the Annotated construct when used with an ORM construct like mapped_column() cannot accommodate dataclass field parameters such as init and repr - this use goes against the design of Python dataclasses and is not supported by PEP 681, and therefore is also rejected by the SQLAlchemy ORM at runtime. A deprecation warning is now emitted and the attribute will be ignored.

As an example, the init=False parameter below will be ignored and additionally emit a deprecation warning:

from typing import Annotated from sqlalchemy.orm import Mapped from sqlalchemy.orm import mapped_column from sqlalchemy.orm import registry # typing tools as well as SQLAlchemy will ignore init=False here intpk = Annotated[int, mapped_column(init=False, primary_key=True)] reg = registry() @reg.mapped_as_dataclass class User: __tablename__ = "user_account" id: Mapped[intpk] # typing error as well as runtime error: Argument missing for parameter "id" u1 = User()

Instead, mapped_column() must be present on the right side as well with an explicit setting for mapped_column.init; the other arguments can remain within the Annotated construct:

from typing import Annotated from sqlalchemy.orm import Mapped from sqlalchemy.orm import mapped_column from sqlalchemy.orm import registry intpk = Annotated[int, mapped_column(primary_key=True)] reg = registry() @reg.mapped_as_dataclass class User: __tablename__ = "user_account" # init=False and other pep-681 arguments must be inline id: Mapped[intpk] = mapped_column(init=False) u1 = User()

Using mixins and abstract superclasses

Any mixins or base classes that are used in a MappedAsDataclass mapped class which include Mapped attributes must themselves be part of a MappedAsDataclass hierarchy, such as in the example below using a mixin:

class Mixin(MappedAsDataclass): create_user: Mapped[int] = mapped_column() update_user: Mapped[Optional[int]] = mapped_column(default=None, init=False) class Base(DeclarativeBase, MappedAsDataclass): pass class User(Base, Mixin): __tablename__ = "sys_user" uid: Mapped[str] = mapped_column( String(50), init=False, default_factory=uuid4, primary_key=True ) username: Mapped[str] = mapped_column() email: Mapped[str] = mapped_column()

Python type checkers which support PEP 681 will otherwise not consider attributes from non-dataclass mixins to be part of the dataclass.

Deprecated since version 2.0.8: Using mixins and abstract bases within MappedAsDataclass or registry.mapped_as_dataclass() hierarchies which are not themselves dataclasses is deprecated, as these fields are not supported by PEP 681 as belonging to the dataclass. A warning is emitted for this case which will later be an error.

See also

When transforming <cls> to a dataclass, attribute(s) originate from superclass <cls> which is not a dataclass. - background on rationale

Relationship Configuration

The Mapped annotation in combination with relationship() is used in the same way as described at Basic Relationship Patterns. When specifying a collection-based relationship() as an optional keyword argument, the relationship.default_factory parameter must be passed and it must refer to the collection class that’s to be used. Many-to-one and scalar object references may make use of relationship.default if the default value is to be None:

from typing import List from sqlalchemy import ForeignKey from sqlalchemy.orm import Mapped from sqlalchemy.orm import mapped_column from sqlalchemy.orm import registry from sqlalchemy.orm import relationship reg = registry() @reg.mapped_as_dataclass class Parent: __tablename__ = "parent" id: Mapped[int] = mapped_column(primary_key=True) children: Mapped[List["Child"]] = relationship( default_factory=list, back_populates="parent" ) @reg.mapped_as_dataclass class Child: __tablename__ = "child" id: Mapped[int] = mapped_column(primary_key=True) parent_id: Mapped[int] = mapped_column(ForeignKey("parent.id")) parent: Mapped["Parent"] = relationship(default=None)

The above mapping will generate an empty list for Parent.children when a new Parent() object is constructed without passing children, and similarly a None value for Child.parent when a new Child() object is constructed without passing parent.

While the relationship.default_factory can be automatically derived from the given collection class of the relationship() itself, this would break compatibility with dataclasses, as the presence of relationship.default_factory or relationship.default is what determines if the parameter is to be required or optional when rendered into the __init__() method.

Using Non-Mapped Dataclass Fields

When using Declarative dataclasses, non-mapped fields may be used on the class as well, which will be part of the dataclass construction process but will not be mapped. Any field that does not use Mapped will be ignored by the mapping process. In the example below, the fields ctrl_one and ctrl_two will be part of the instance-level state of the object, but will not be persisted by the ORM:

from sqlalchemy.orm import Mapped from sqlalchemy.orm import mapped_column from sqlalchemy.orm import registry reg = registry() @reg.mapped_as_dataclass class Data: __tablename__ = "data" id: Mapped[int] = mapped_column(init=False, primary_key=True) status: Mapped[str] ctrl_one: Optional[str] = None ctrl_two: Optional[str] = None

Instance of Data above can be created as:

d1 = Data(status="s1", ctrl_one="ctrl1", ctrl_two="ctrl2")

A more real world example might be to make use of the Dataclasses InitVar feature in conjunction with the __post_init__() feature to receive init-only fields that can be used to compose persisted data. In the example below, the User class is declared using id, name and password_hash as mapped features, but makes use of init-only password and repeat_password fields to represent the user creation process (note: to run this example, replace the function your_crypt_function_here() with a third party crypt function, such as bcrypt or argon2-cffi):

from dataclasses import InitVar from typing import Optional from sqlalchemy.orm import Mapped from sqlalchemy.orm import mapped_column from sqlalchemy.orm import registry reg = registry() @reg.mapped_as_dataclass class User: __tablename__ = "user_account" id: Mapped[int] = mapped_column(init=False, primary_key=True) name: Mapped[str] password: InitVar[str] repeat_password: InitVar[str] password_hash: Mapped[str] = mapped_column(init=False, nullable=False) def __post_init__(self, password: str, repeat_password: str): if password != repeat_password: raise ValueError("passwords do not match") self.password_hash = your_crypt_function_here(password)

The above object is created with parameters password and repeat_password, which are consumed up front so that the password_hash variable may be generated:

>>> u1 = User(name="some_user", password="xyz", repeat_password="xyz") >>> u1.password_hash '$6$9ppc... (example crypted string....)'

Changed in version 2.0.0rc1: When using registry.mapped_as_dataclass() or MappedAsDataclass, fields that do not include the Mapped annotation may be included, which will be treated as part of the resulting dataclass but not be mapped, without the need to also indicate the __allow_unmapped__ class attribute. Previous 2.0 beta releases would require this attribute to be explicitly present, even though the purpose of this attribute was only to allow legacy ORM typed mappings to continue to function.

Integrating with Alternate Dataclass Providers such as Pydantic

Warning

The dataclass layer of Pydantic is not fully compatible with SQLAlchemy’s class instrumentation without additional internal changes, and many features such as related collections may not work correctly.

For Pydantic compatibility, please consider the SQLModel ORM which is built with Pydantic on top of SQLAlchemy ORM, which includes special implementation details which explicitly resolve these incompatibilities.

SQLAlchemy’s MappedAsDataclass class and registry.mapped_as_dataclass() method call directly into the Python standard library dataclasses.dataclass class decorator, after the declarative mapping process has been applied to the class. This function call may be swapped out for alternateive dataclasses providers, such as that of Pydantic, using the dataclass_callable parameter accepted by MappedAsDataclass as a class keyword argument as well as by registry.mapped_as_dataclass():

from sqlalchemy.orm import DeclarativeBase from sqlalchemy.orm import Mapped from sqlalchemy.orm import mapped_column from sqlalchemy.orm import MappedAsDataclass from sqlalchemy.orm import registry class Base( MappedAsDataclass, DeclarativeBase, dataclass_callable=pydantic.dataclasses.dataclass, ): pass class User(Base): __tablename__ = "user" id: Mapped[int] = mapped_column(primary_key=True) name: Mapped[str]

The above User class will be applied as a dataclass, using Pydantic’s pydantic.dataclasses.dataclasses callable. The process is available both for mapped classes as well as mixins that extend from MappedAsDataclass or which have registry.mapped_as_dataclass() applied directly.

Added in version 2.0.4: Added the dataclass_callable class and method parameters for MappedAsDataclass and registry.mapped_as_dataclass(), and adjusted some of the dataclass internals to accommodate more strict dataclass functions such as that of Pydantic.

Applying ORM Mappings to an existing dataclass (legacy dataclass use)

Legacy Feature

The approaches described here are superseded by the Declarative Dataclass Mapping feature new in the 2.0 series of SQLAlchemy. This newer version of the feature builds upon the dataclass support first added in version 1.4, which is described in this section.

To map an existing dataclass, SQLAlchemy’s “inline” declarative directives cannot be used directly; ORM directives are assigned using one of three techniques:

The general process by which SQLAlchemy applies mappings to a dataclass is the same as that of an ordinary class, but also includes that SQLAlchemy will detect class-level attributes that were part of the dataclasses declaration process and replace them at runtime with the usual SQLAlchemy ORM mapped attributes. The __init__ method that would have been generated by dataclasses is left intact, as is the same for all the other methods that dataclasses generates such as __eq__(), __repr__(), etc.

Mapping pre-existing dataclasses using Declarative With Imperative Table

An example of a mapping using @dataclass using Declarative with Imperative Table (a.k.a. Hybrid Declarative) is below. A complete Table object is constructed explicitly and assigned to the __table__ attribute. Instance fields are defined using normal dataclass syntaxes. Additional MapperProperty definitions such as relationship(), are placed in the __mapper_args__ class-level dictionary underneath the properties key, corresponding to the Mapper.properties parameter:

from __future__ import annotations from dataclasses import dataclass, field from typing import List, Optional from sqlalchemy import Column, ForeignKey, Integer, String, Table from sqlalchemy.orm import registry, relationship mapper_registry = registry() @mapper_registry.mapped @dataclass class User: __table__ = Table( "user", mapper_registry.metadata, Column("id", Integer, primary_key=True), Column("name", String(50)), Column("fullname", String(50)), Column("nickname", String(12)), ) id: int = field(init=False) name: Optional[str] = None fullname: Optional[str] = None nickname: Optional[str] = None addresses: List[Address] = field(default_factory=list) __mapper_args__ = { # type: ignore "properties": { "addresses": relationship("Address"), } } @mapper_registry.mapped @dataclass class Address: __table__ = Table( "address", mapper_registry.metadata, Column("id", Integer, primary_key=True), Column("user_id", Integer, ForeignKey("user.id")), Column("email_address", String(50)), ) id: int = field(init=False) user_id: int = field(init=False) email_address: Optional[str] = None

In the above example, the User.id, Address.id, and Address.user_id attributes are defined as field(init=False). This means that parameters for these won’t be added to __init__() methods, but Session will still be able to set them after getting their values during flush from autoincrement or other default value generator. To allow them to be specified in the constructor explicitly, they would instead be given a default value of None.

For a relationship() to be declared separately, it needs to be specified directly within the Mapper.properties dictionary which itself is specified within the __mapper_args__ dictionary, so that it is passed to the constructor for Mapper. An alternative to this approach is in the next example.

Warning

Declaring a dataclass field() setting a default together with init=False will not work as would be expected with a totally plain dataclass, since the SQLAlchemy class instrumentation will replace the default value set on the class by the dataclass creation process. Use default_factory instead. This adaptation is done automatically when making use of Declarative Dataclass Mapping.

Mapping pre-existing dataclasses using Declarative-style fields

Legacy Feature

This approach to Declarative mapping with dataclasses should be considered as legacy. It will remain supported however is unlikely to offer any advantages against the new approach detailed at Declarative Dataclass Mapping.

Note that mapped_column() is not supported with this use; the Column construct should continue to be used to declare table metadata within the metadata field of dataclasses.field().

The fully declarative approach requires that Column objects are declared as class attributes, which when using dataclasses would conflict with the dataclass-level attributes. An approach to combine these together is to make use of the metadata attribute on the dataclass.field object, where SQLAlchemy-specific mapping information may be supplied. Declarative supports extraction of these parameters when the class specifies the attribute __sa_dataclass_metadata_key__. This also provides a more succinct method of indicating the relationship() association:

from __future__ import annotations from dataclasses import dataclass, field from typing import List from sqlalchemy import Column, ForeignKey, Integer, String from sqlalchemy.orm import registry, relationship mapper_registry = registry() @mapper_registry.mapped @dataclass class User: __tablename__ = "user" __sa_dataclass_metadata_key__ = "sa" id: int = field(init=False, metadata={"sa": Column(Integer, primary_key=True)}) name: str = field(default=None, metadata={"sa": Column(String(50))}) fullname: str = field(default=None, metadata={"sa": Column(String(50))}) nickname: str = field(default=None, metadata={"sa": Column(String(12))}) addresses: List[Address] = field( default_factory=list, metadata={"sa": relationship("Address")} ) @mapper_registry.mapped @dataclass class Address: __tablename__ = "address" __sa_dataclass_metadata_key__ = "sa" id: int = field(init=False, metadata={"sa": Column(Integer, primary_key=True)}) user_id: int = field(init=False, metadata={"sa": Column(ForeignKey("user.id"))}) email_address: str = field(default=None, metadata={"sa": Column(String(50))})

Using Declarative Mixins with pre-existing dataclasses

In the section Composing Mapped Hierarchies with Mixins, Declarative Mixin classes are introduced. One requirement of declarative mixins is that certain constructs that can’t be easily duplicated must be given as callables, using the declared_attr decorator, such as in the example at Mixing in Relationships:

class RefTargetMixin: @declared_attr def target_id(cls) -> Mapped[int]: return mapped_column("target_id", ForeignKey("target.id")) @declared_attr def target(cls): return relationship("Target")

This form is supported within the Dataclasses field() object by using a lambda to indicate the SQLAlchemy construct inside the field(). Using declared_attr() to surround the lambda is optional. If we wanted to produce our User class above where the ORM fields came from a mixin that is itself a dataclass, the form would be:

@dataclass class UserMixin: __tablename__ = "user" __sa_dataclass_metadata_key__ = "sa" id: int = field(init=False, metadata={"sa": Column(Integer, primary_key=True)}) addresses: List[Address] = field( default_factory=list, metadata={"sa": lambda: relationship("Address")} ) @dataclass class AddressMixin: __tablename__ = "address" __sa_dataclass_metadata_key__ = "sa" id: int = field(init=False, metadata={"sa": Column(Integer, primary_key=True)}) user_id: int = field( init=False, metadata={"sa": lambda: Column(ForeignKey("user.id"))} ) email_address: str = field(default=None, metadata={"sa": Column(String(50))}) @mapper_registry.mapped class User(UserMixin): pass @mapper_registry.mapped class Address(AddressMixin): pass

Added in version 1.4.2: Added support for “declared attr” style mixin attributes, namely relationship() constructs as well as Column objects with foreign key declarations, to be used within “Dataclasses with Declarative Table” style mappings.

Mapping pre-existing dataclasses using Imperative Mapping

As described previously, a class which is set up as a dataclass using the @dataclass decorator can then be further decorated using the registry.mapped() decorator in order to apply declarative-style mapping to the class. As an alternative to using the registry.mapped() decorator, we may also pass the class through the registry.map_imperatively() method instead, so that we may pass all Table and Mapper configuration imperatively to the function rather than having them defined on the class itself as class variables:

from __future__ import annotations from dataclasses import dataclass from dataclasses import field from typing import List from sqlalchemy import Column from sqlalchemy import ForeignKey from sqlalchemy import Integer from sqlalchemy import MetaData from sqlalchemy import String from sqlalchemy import Table from sqlalchemy.orm import registry from sqlalchemy.orm import relationship mapper_registry = registry() @dataclass class User: id: int = field(init=False) name: str = None fullname: str = None nickname: str = None addresses: List[Address] = field(default_factory=list) @dataclass class Address: id: int = field(init=False) user_id: int = field(init=False) email_address: str = None metadata_obj = MetaData() user = Table( "user", metadata_obj, Column("id", Integer, primary_key=True), Column("name", String(50)), Column("fullname", String(50)), Column("nickname", String(12)), ) address = Table( "address", metadata_obj, Column("id", Integer, primary_key=True), Column("user_id", Integer, ForeignKey("user.id")), Column("email_address", String(50)), ) mapper_registry.map_imperatively( User, user, properties={ "addresses": relationship(Address, backref="user", order_by=address.c.id), }, ) mapper_registry.map_imperatively(Address, address)

The same warning mentioned in Mapping pre-existing dataclasses using Declarative With Imperative Table applies when using this mapping style.

Applying ORM mappings to an existing attrs class

Warning

The attrs library is not part of SQLAlchemy’s continuous integration testing, and compatibility with this library may change without notice due to incompatibilities introduced by either side.

The attrs library is a popular third party library that provides similar features as dataclasses, with many additional features provided not found in ordinary dataclasses.

A class augmented with attrs uses the @define decorator. This decorator initiates a process to scan the class for attributes that define the class’ behavior, which are then used to generate methods, documentation, and annotations.

The SQLAlchemy ORM supports mapping an attrs class using Imperative mapping. The general form of this style is equivalent to the Mapping pre-existing dataclasses using Imperative Mapping mapping form used with dataclasses, where the class construction uses attrs alone, with ORM mappings applied after the fact without any class attribute scanning.

The @define decorator of attrs by default replaces the annotated class with a new __slots__ based class, which is not supported. When using the old style annotation @attr.s or using define(slots=False), the class does not get replaced. Furthermore attrs removes its own class-bound attributes after the decorator runs, so that SQLAlchemy’s mapping process takes over these attributes without any issue. Both decorators, @attr.s and @define(slots=False) work with SQLAlchemy.

Changed in version 2.0: SQLAlchemy integration with attrs works only with imperative mapping style, that is, not using Declarative. The introduction of ORM Annotated Declarative style is not cross-compatible with attrs.

The attrs class is built first. The SQLAlchemy ORM mapping can be applied after the fact using registry.map_imperatively():

from __future__ import annotations from typing import List from attrs import define from sqlalchemy import Column from sqlalchemy import ForeignKey from sqlalchemy import Integer from sqlalchemy import MetaData from sqlalchemy import String from sqlalchemy import Table from sqlalchemy.orm import registry from sqlalchemy.orm import relationship mapper_registry = registry() @define(slots=False) class User: id: int name: str fullname: str nickname: str addresses: List[Address] @define(slots=False) class Address: id: int user_id: int email_address: Optional[str] metadata_obj = MetaData() user = Table( "user", metadata_obj, Column("id", Integer, primary_key=True), Column("name", String(50)), Column("fullname", String(50)), Column("nickname", String(12)), ) address = Table( "address", metadata_obj, Column("id", Integer, primary_key=True), Column("user_id", Integer, ForeignKey("user.id")), Column("email_address", String(50)), ) mapper_registry.map_imperatively( User, user, properties={ "addresses": relationship(Address, backref="user", order_by=address.c.id), }, ) mapper_registry.map_imperatively(Address, address)