Mypy / Pep-484 Support for ORM Mappings

    SQLAlchemy Mypy Plugin Status Update

    Updated December 2022

    For SQLAlchemy 2.0, the Mypy plugin continues to work at the level at which it reached in the SQLAlchemy 1.4 release. However, SQLAlchemy 2.0, when released, will feature an all new typing system for ORM Declarative models that removes the need for the Mypy plugin and delivers much more consistent behavior with generally superior capabilities. Note that this new capability is not part of SQLAlchemy 1.4, it is only in SQLAlchemy 2.0, which is out with beta releases as of December 2022.

    The SQLAlchemy Mypy plugin, while it has technically never left the “alpha” stage, should now be considered as deprecated in SQLAlchemy 2.0, even though it is still necessary for full Mypy support when using SQLAlchemy 1.4.

    The Mypy plugin itself does not solve the issue of supplying correct typing with other typing tools such as Pylance/Pyright, Pytype, Pycharm, etc, which cannot make use of Mypy plugins. Additionally, Mypy plugins are extremely difficult to develop, maintain and test, as a Mypy plugin must be deeply integrated with Mypy’s internal datastructures and processes, which itself are not stable within the Mypy project itself. The SQLAlchemy Mypy plugin has lots of limitations when used with code that deviates from very basic patterns which are reported regularly.

    For these reasons, new non-regression issues reported against the Mypy plugin are unlikely to be fixed. When SQLAlchemy 2.0 is released, it will continue to include the plugin, which will have been updated to continue to function as well as it does in SQLAlchemy 1.4, when running under SQLAlchemy 2.0. Existing code that passes Mypy checks using the plugin with SQLAlchemy 1.4 installed will continue to pass all checks in SQLAlchemy 2.0 without any changes required, provided the plugin is still used. The upcoming API to be released with SQLAlchemy 2.0 is fully backwards compatible with the SQLAlchemy 1.4 API and Mypy plugin behavior.

    End-user code that passes all checks under SQLAlchemy 1.4 with the Mypy plugin will be able to incrementally migrate to the new structures, once that code is running exclusively on SQLAlchemy 2.0. See the section for background on how this migration may proceed.

    Code that is running exclusively on SQLAlchemy version 2.0 and has fully migrated to the new declarative constructs will enjoy full compliance with pep-484 as well as working correctly within IDEs and other typing tools, without the need for plugins.

    For SQLAlchemy 2.0 only: No stubs should be installed and packages like and sqlalchemy2-stubs should be fully uninstalled.

    The package itself is a dependency.

    Mypy may be installed using the “mypy” extras hook using pip:

    The plugin itself is configured as described in Configuring mypy to use Plugins, using the module name, such as within setup.cfg:

    1. [mypy]
    2. plugins = sqlalchemy.ext.mypy.plugin

    The primary purpose of the Mypy plugin is to intercept and alter the static definition of SQLAlchemy declarative mappings so that they match up to how they are structured after they have been by their Mapper objects. This allows both the class structure itself as well as code that uses the class to make sense to the Mypy tool, which otherwise would not be the case based on how declarative mappings currently function. The plugin is not unlike similar plugins that are required for libraries like which alter classes dynamically at runtime.

    To cover the major areas where this occurs, consider the following ORM mapping, using the typical example of the User class:

    1. from sqlalchemy import Column, Integer, String, select
    2. from sqlalchemy.orm import declarative_base
    4. # "Base" is a class that is created dynamically from the
    5. # declarative_base() function
    6. Base = declarative_base()
    9. class User(Base):
    10.     __tablename__ = "user"
    12.     id = Column(Integer, primary_key=True)
    13.     name = Column(String)
    16. # "some_user" is an instance of the User class, which
    17. # accepts "id" and "name" kwargs based on the mapping
    18. some_user = User(id=5, name="user")
    20. # it has an attribute called .name that's a string
    21. print(f"Username: {some_user.name}")
    23. # a select() construct makes use of SQL expressions derived from the
    24. # User class itself
    25. select_stmt = select(User).where(User.id.in_([3, 4, 5])).where(User.name.contains("s"))

    Above, the steps that the Mypy extension can take include:

    • Interpretation of the Base dynamic class generated by declarative_base(), so that classes which inherit from it are known to be mapped. It also can accommodate the class decorator approach described at .

    • Application of an __init__() method to mapped classes that do not already include an explicit constructor, which accepts keyword arguments of specific types for all mapped attributes detected.

    When the Mypy plugin processes the above file, the resulting static class definition and Python code passed to the Mypy tool is equivalent to the following:

    1. from sqlalchemy import Column, Integer, String, select
    2. from sqlalchemy.orm import Mapped
    3. from sqlalchemy.orm.decl_api import DeclarativeMeta
    6. class Base(metaclass=DeclarativeMeta):
    7.     __abstract__ = True
    10. class User(Base):
    11.     __tablename__ = "user"
    13.     id: Mapped[Optional[int]] = Mapped._special_method(
    14.         Column(Integer, primary_key=True)
    15.     )
    16.     name: Mapped[Optional[str]] = Mapped._special_method(Column(String))
    18.     def __init__(self, id: Optional[int] = ..., name: Optional[str] = ...) -> None:
    19.         ...
    22. some_user = User(id=5, name="user")
    24. print(f"Username: {some_user.name}")
    26. select_stmt = select(User).where(User.id.in_([3, 4, 5])).where(User.name.contains("s"))

    The key steps which have been taken above include:

    • The Base class is now defined in terms of the DeclarativeMeta class explicitly, rather than being a dynamic class.

    • The id and name attributes are defined in terms of the Mapped class, which represents a Python descriptor that exhibits different behaviors at the class vs. instance levels. The class is now the base class for the InstrumentedAttribute class that is used for all ORM mapped attributes.

      is defined as a generic class against arbitrary Python types, meaning specific occurrences of Mapped are associated with a specific Python type, such as Mapped[Optional[int]] and Mapped[Optional[str]] above.

    • The right-hand side of the declarative mapped attribute assignments are removed, as this resembles the operation that the class would normally be doing, which is that it would be replacing these attributes with specific instances of InstrumentedAttribute. The original expression is moved into a function call that will allow it to still be type-checked without conflicting with the left-hand side of the expression. For Mypy purposes, the left-hand typing annotation is sufficient for the attribute’s behavior to be understood.

    • A type stub for the User.__init__() method is added which includes the correct keywords and datatypes.

    The following subsections will address individual uses cases that have so far been considered for pep-484 compliance.

    For mapped columns that include an explicit datatype, when they are mapped as inline attributes, the mapped type will be introspected automatically:

    1. class MyClass(Base):
    2.     # ...
    4.     id = Column(Integer, primary_key=True)
    5.     name = Column("employee_name", String(50), nullable=False)
    6.     other_name = Column(String(50))

    Above, the ultimate class-level datatypes of id, name and other_name will be introspected as Mapped[Optional[int]], Mapped[Optional[str]] and Mapped[Optional[str]]. The types are by default always considered to be Optional, even for the primary key and non-nullable column. The reason is because while the database columns “id” and “name” can’t be NULL, the Python attributes id and name most certainly can be None without an explicit constructor:

    1. >>> m1 = MyClass()
    2. >>> m1.id
    3. None

    The types of the above columns can be stated explicitly, providing the two advantages of clearer self-documentation as well as being able to control which types are optional:

    1. class MyClass(Base):
    2.     # ...
    4.     id: int = Column(Integer, primary_key=True)
    5.     name: str = Column("employee_name", String(50), nullable=False)
    6.     other_name: Optional[str] = Column(String(50))

    The Mypy plugin will accept the above int, str and and convert them to include the Mapped[] type surrounding them. The Mapped[] construct may also be used explicitly:

    1. from sqlalchemy.orm import Mapped
    4.     # ...
    6.     id: Mapped[int] = Column(Integer, primary_key=True)
    7.     name: Mapped[str] = Column("employee_name", String(50), nullable=False)
    8.     other_name: Mapped[Optional[str]] = Column(String(50))

    When the type is non-optional, it simply means that the attribute as accessed from an instance of MyClass will be considered to be non-None:

    For optional attributes, Mypy considers that the type must include None or otherwise be Optional:

    1. mc = MyClass(...)
    3. # will pass mypy --strict
    4. other_name: Optional[str] = mc.name

    Whether or not the mapped attribute is typed as Optional, the generation of the __init__() method will still consider all keywords to be optional. This is again matching what the SQLAlchemy ORM actually does when it creates the constructor, and should not be confused with the behavior of a validating system such as Python dataclasses which will generate a constructor that matches the annotations in terms of optional vs. required attributes.

    Columns that Don’t have an Explicit Type

    1. # .. other imports
    2. from sqlalchemy.sql.schema import ForeignKey
    4. Base = declarative_base()
    7. class User(Base):
    8.     __tablename__ = "user"
    10.     id = Column(Integer, primary_key=True)
    11.     name = Column(String)
    14. class Address(Base):
    15.     __tablename__ = "address"
    17.     id = Column(Integer, primary_key=True)
    18.     user_id = Column(ForeignKey("user.id"))

    The plugin will deliver the message as follows:

    1. $ mypy test3.py --strict
    2. test3.py:20: error: [SQLAlchemy Mypy plugin] Can't infer type from
    3. ORM mapped expression assigned to attribute 'user_id'; please specify a
    4. Python type or Mapped[<python type>] on the left hand side.
    5. Found 1 error in 1 file (checked 1 source file)

    To resolve, apply an explicit type annotation to the Address.user_id column:

    1. class Address(Base):
    2.     __tablename__ = "address"
    4.     id = Column(Integer, primary_key=True)
    5.     user_id: int = Column(ForeignKey("user.id"))

    In , the Column definitions are given inside of a construct which is separate from the mapped attributes themselves. The Mypy plugin does not consider this Table, but instead supports that the attributes can be explicitly stated with a complete annotation that must use the class to identify them as mapped attributes:

    1. class MyClass(Base):
    2.     __table__ = Table(
    3.         "mytable",
    4.         Base.metadata,
    5.         Column(Integer, primary_key=True),
    6.         Column("employee_name", String(50), nullable=False),
    7.         Column(String(50)),
    8.     )
    10.     id: Mapped[int]
    11.     name: Mapped[str]
    12.     other_name: Mapped[Optional[str]]

    The above Mapped annotations are considered as mapped columns and will be included in the default constructor, as well as provide the correct typing profile for MyClass both at the class level and the instance level.

    Mapping Relationships

    The plugin has limited support for using type inference to detect the types for relationships. For all those cases where it can’t detect the type, it will emit an informative error message, and in all cases the appropriate type may be provided explicitly, either with the Mapped class or optionally omitting it for an inline declaration. The plugin also needs to determine whether or not the relationship refers to a collection or a scalar, and for that it relies upon the explicit value of the and/or relationship.collection_class parameters. An explicit type is needed if neither of these parameters are present, as well as if the target type of the is a string or callable, and not a class:

    1. class User(Base):
    2.     __tablename__ = "user"
    4.     id = Column(Integer, primary_key=True)
    5.     name = Column(String)
    8. class Address(Base):
    9.     __tablename__ = "address"
    11.     id = Column(Integer, primary_key=True)
    12.     user_id: int = Column(ForeignKey("user.id"))
    14.     user = relationship(User)

    The above mapping will produce the following error:

    1. test3.py:22: error: [SQLAlchemy Mypy plugin] Can't infer scalar or
    2. collection for ORM mapped expression assigned to attribute 'user'
    3. if both 'uselist' and 'collection_class' arguments are absent from the
    4. relationship(); please specify a type annotation on the left hand side.
    5. Found 1 error in 1 file (checked 1 source file)

    The error can be resolved either by using relationship(User, uselist=False) or by providing the type, in this case the scalar User object:

    For collections, a similar pattern applies, where in the absence of uselist=True or a relationship.collection_class, a collection annotation such as List may be used. It is also fully appropriate to use the string name of the class in the annotation as supported by pep-484, ensuring the class is imported with in the as appropriate:

    1. from typing import TYPE_CHECKING, List
    3. from .mymodel import Base
    5. if TYPE_CHECKING:
    6.     # if the target of the relationship is in another module
    7.     # that cannot normally be imported at runtime
    8.     from .myaddressmodel import Address
    11. class User(Base):
    12.     __tablename__ = "user"
    14.     id = Column(Integer, primary_key=True)
    15.     name = Column(String)
    16.     addresses: List["Address"] = relationship("Address")

    As is the case with columns, the Mapped class may also be applied explicitly:

    1. class User(Base):
    4.     name = Column(String)
    6.     addresses: Mapped[List["Address"]] = relationship("Address", back_populates="user")
    9. class Address(Base):
    10.     __tablename__ = "address"
    12.     id = Column(Integer, primary_key=True)
    13.     user_id: int = Column(ForeignKey("user.id"))
    15.     user: Mapped[User] = relationship(User, back_populates="addresses")

    The declared_attr class allows Declarative mapped attributes to be declared in class level functions, and is particularly useful when using . For these functions, the return type of the function should be annotated using either the Mapped[] construct or by indicating the exact kind of object returned by the function. Additionally, “mixin” classes that are not otherwise mapped (i.e. don’t extend from a declarative_base() class nor are they mapped with a method such as ) should be decorated with the declarative_mixin() decorator, which provides a hint to the Mypy plugin that a particular class intends to serve as a declarative mixin:

    1. from sqlalchemy.orm import declarative_mixin, declared_attr
    4. @declarative_mixin
    5. class HasUpdatedAt:
    6.     @declared_attr
    7.     def updated_at(cls) -> Column[DateTime]:  # uses Column
    8.         return Column(DateTime)
    11. @declarative_mixin
    12. class HasCompany:
    13.     @declared_attr
    14.     def company_id(cls) -> Mapped[int]:  # uses Mapped
    15.         return Column(ForeignKey("company.id"))
    17.     @declared_attr
    18.     def company(cls) -> Mapped["Company"]:
    19.         return relationship("Company")
    22. class Employee(HasUpdatedAt, HasCompany, Base):
    23.     __tablename__ = "employee"
    25.     id = Column(Integer, primary_key=True)
    26.     name = Column(String)

    Note the mismatch between the actual return type of a method like HasCompany.company vs. what is annotated. The Mypy plugin converts all @declared_attr functions into simple annotated attributes to avoid this complexity:

    1. # what Mypy sees
    2. class HasCompany:
    3.     company_id: Mapped[int]
    4.     company: Mapped["Company"]

    Combining with Dataclasses or Other Type-Sensitive Attribute Systems

    The examples of Python dataclasses integration at Applying ORM Mappings to an existing dataclass presents a problem; Python dataclasses expect an explicit type that it will use to build the class, and the value given in each assignment statement is significant. That is, a class as follows has to be stated exactly as it is in order to be accepted by dataclasses:

    1. mapper_registry: registry = registry()
    4. @mapper_registry.mapped
    5. @dataclass
    6. class User:
    7.     __table__ = Table(
    8.         "user",
    9.         mapper_registry.metadata,
    10.         Column("id", Integer, primary_key=True),
    11.         Column("name", String(50)),
    12.         Column("fullname", String(50)),
    13.         Column("nickname", String(12)),
    14.     )
    15.     id: int = field(init=False)
    16.     name: Optional[str] = None
    17.     fullname: Optional[str] = None
    18.     nickname: Optional[str] = None
    19.     addresses: List[Address] = field(default_factory=list)
    21.     __mapper_args__ = {  # type: ignore
    22.         "properties": {"addresses": relationship("Address")}
    23.     }

    We can’t apply our Mapped[] types to the attributes id, name, etc. because they will be rejected by the @dataclass decorator. Additionally, Mypy has another plugin for dataclasses explicitly which can also get in the way of what we’re doing.

    The above class will actually pass Mypy’s type checking without issue; the only thing we are missing is the ability for attributes on User to be used in SQL expressions, such as:

    1. stmt = select(User.name).where(User.id.in_([1, 2, 3]))

    To provide a workaround for this, the Mypy plugin has an additional feature whereby we can specify an extra attribute _mypy_mapped_attrs, that is a list that encloses the class-level objects or their string names. This attribute can be conditional within the TYPE_CHECKING variable:

    1. @mapper_registry.mapped
    2. @dataclass
    3. class User:
    4.     __table__ = Table(
    5.         "user",
    6.         mapper_registry.metadata,
    7.         Column("id", Integer, primary_key=True),
    8.         Column("name", String(50)),
    9.         Column("fullname", String(50)),
    10.         Column("nickname", String(12)),
    11.     )
    12.     id: int = field(init=False)
    13.     name: Optional[str] = None
    14.     fullname: Optional[str]
    15.     nickname: Optional[str]
    16.     addresses: List[Address] = field(default_factory=list)
    18.     if TYPE_CHECKING:
    19.         _mypy_mapped_attrs = [id, name, "fullname", "nickname", addresses]
    21.     __mapper_args__ = {  # type: ignore
    22.     }

    With the above recipe, the attributes listed in _mypy_mapped_attrs will be applied with the typing information so that the class will behave as a SQLAlchemy mapped class when used in a class-bound context.