ORM Mapped Class Overview

Overview of ORM class mapping configuration.

For readers new to the SQLAlchemy ORM and/or new to Python in general, it’s recommended to browse through the ORM Quick Start and preferably to work through the SQLAlchemy Unified Tutorial, where ORM configuration is first introduced at Using ORM Declarative Forms to Define Table Metadata.

ORM Mapping Styles

SQLAlchemy features two distinct styles of mapper configuration, which then feature further sub-options for how they are set up. The variability in mapper styles is present to suit a varied list of developer preferences, including the degree of abstraction of a user-defined class from how it is to be mapped to relational schema tables and columns, what kinds of class hierarchies are in use, including whether or not custom metaclass schemes are present, and finally if there are other class-instrumentation approaches present such as if Python dataclasses are in use simultaneously.

In modern SQLAlchemy, the difference between these styles is mostly superficial; when a particular SQLAlchemy configurational style is used to express the intent to map a class, the internal process of mapping the class proceeds in mostly the same way for each, where the end result is always a user-defined class that has a Mapper configured against a selectable unit, typically represented by a Table object, and the class itself has been instrumented to include behaviors linked to relational operations both at the level of the class as well as on instances of that class. As the process is basically the same in all cases, classes mapped from different styles are always fully interoperable with each other. The protocol MappedClassProtocol can be used to indicate a mapped class when using type checkers such as mypy.

The original mapping API is commonly referred to as “classical” style, whereas the more automated style of mapping is known as “declarative” style. SQLAlchemy now refers to these two mapping styles as imperative mapping and declarative mapping.

Regardless of what style of mapping used, all ORM mappings as of SQLAlchemy 1.4 originate from a single object known as registry, which is a registry of mapped classes. Using this registry, a set of mapper configurations can be finalized as a group, and classes within a particular registry may refer to each other by name within the configurational process.

Changed in version 1.4: Declarative and classical mapping are now referred to as “declarative” and “imperative” mapping, and are unified internally, all originating from the registry construct that represents a collection of related mappings.

Declarative Mapping

The Declarative Mapping is the typical way that mappings are constructed in modern SQLAlchemy. The most common pattern is to first construct a base class using the DeclarativeBase superclass. The resulting base class, when subclassed will apply the declarative mapping process to all subclasses that derive from it, relative to a particular registry that is local to the new base by default. The example below illustrates the use of a declarative base which is then used in a declarative table mapping:

from sqlalchemy import Integer, String, ForeignKey
from sqlalchemy.orm import DeclarativeBase
from sqlalchemy.orm import Mapped
from sqlalchemy.orm import mapped_column

# declarative base class
class Base(DeclarativeBase):

# an example mapping using the base
class User(Base):
    __tablename__ = "user"

    id: Mapped[int] = mapped_column(primary_key=True)
    name: Mapped[str]
    fullname: Mapped[str] = mapped_column(String(30))
    nickname: Mapped[Optional[str]]

Above, the DeclarativeBase class is used to generate a new base class (within SQLAlchemy’s documentation it’s typically referred to as Base, however can have any desired name) from which new classes to be mapped may inherit from, as above a new mapped class User is constructed.

Changed in version 2.0: The DeclarativeBase superclass supersedes the use of the declarative_base() function and registry.generate_base() methods; the superclass approach integrates with PEP 484 tools without the use of plugins. See ORM Declarative Models for migration notes.

The base class refers to a registry object that maintains a collection of related mapped classes. as well as to a MetaData object that retains a collection of Table objects to which the classes are mapped.

The major Declarative mapping styles are further detailed in the following sections:

Within the scope of a Declarative mapped class, there are also two varieties of how the Table metadata may be declared. These include:

  • Declarative Table with mapped_column() - table columns are declared inline within the mapped class using the mapped_column() directive (or in legacy form, using the Column object directly). The mapped_column() directive may also be optionally combined with type annotations using the Mapped class which can provide some details about the mapped columns directly. The column directives, in combination with the __tablename__ and optional __table_args__ class level directives will allow the Declarative mapping process to construct a Table object to be mapped.

  • Declarative with Imperative Table (a.k.a. Hybrid Declarative) - Instead of specifying table name and attributes separately, an explicitly constructed Table object is associated with a class that is otherwise mapped declaratively. This style of mapping is a hybrid of “declarative” and “imperative” mapping, and applies to techniques such as mapping classes to reflected Table objects, as well as mapping classes to existing Core constructs such as joins and subqueries.

Documentation for Declarative mapping continues at Mapping Classes with Declarative.

Imperative Mapping

An imperative or classical mapping refers to the configuration of a mapped class using the registry.map_imperatively() method, where the target class does not include any declarative class attributes.


The imperative mapping form is a lesser-used form of mapping that originates from the very first releases of SQLAlchemy in 2006. It’s essentially a means of bypassing the Declarative system to provide a more “barebones” system of mapping, and does not offer modern features such as PEP 484 support. As such, most documentation examples use Declarative forms, and it’s recommended that new users start with Declarative Table configuration.

Changed in version 2.0: The registry.map_imperatively() method is now used to create classical mappings. The sqlalchemy.orm.mapper() standalone function is effectively removed.

In “classical” form, the table metadata is created separately with the Table construct, then associated with the User class via the registry.map_imperatively() method, after establishing a registry instance. Normally, a single instance of registry shared for all mapped classes that are related to each other:

from sqlalchemy import Table, Column, Integer, String, ForeignKey
from sqlalchemy.orm import registry

mapper_registry = registry()

user_table = Table(
    Column("id", Integer, primary_key=True),
    Column("name", String(50)),
    Column("fullname", String(50)),
    Column("nickname", String(12)),

class User:

mapper_registry.map_imperatively(User, user_table)

Information about mapped attributes, such as relationships to other classes, are provided via the properties dictionary. The example below illustrates a second Table object, mapped to a class called Address, then linked to User via relationship():

address = Table(
    Column("id", Integer, primary_key=True),
    Column("user_id", Integer, ForeignKey("user.id")),
    Column("email_address", String(50)),

        "addresses": relationship(Address, backref="user", order_by=address.c.id)

mapper_registry.map_imperatively(Address, address)

Note that classes which are mapped with the Imperative approach are fully interchangeable with those mapped with the Declarative approach. Both systems ultimately create the same configuration, consisting of a Table, user-defined class, linked together with a Mapper object. When we talk about “the behavior of Mapper”, this includes when using the Declarative system as well - it’s still used, just behind the scenes.

Mapped Class Essential Components

With all mapping forms, the mapping of the class can be configured in many ways by passing construction arguments that ultimately become part of the Mapper object via its constructor. The parameters that are delivered to Mapper originate from the given mapping form, including parameters passed to registry.map_imperatively() for an Imperative mapping, or when using the Declarative system, from a combination of the table columns, SQL expressions and relationships being mapped along with that of attributes such as __mapper_args__.

There are four general classes of configuration information that the Mapper class looks for:

The class to be mapped

This is a class that we construct in our application. There are generally no restrictions on the structure of this class. [1] When a Python class is mapped, there can only be one Mapper object for the class. [2]

When mapping with the declarative mapping style, the class to be mapped is either a subclass of the declarative base class, or is handled by a decorator or function such as registry.mapped().

When mapping with the imperative style, the class is passed directly as the map_imperatively.class_ argument.

The table, or other from clause object

In the vast majority of common cases this is an instance of Table. For more advanced use cases, it may also refer to any kind of FromClause object, the most common alternative objects being the Subquery and Join object.

When mapping with the declarative mapping style, the subject table is either generated by the declarative system based on the __tablename__ attribute and the Column objects presented, or it is established via the __table__ attribute. These two styles of configuration are presented at Declarative Table with mapped_column() and Declarative with Imperative Table (a.k.a. Hybrid Declarative).

When mapping with the imperative style, the subject table is passed positionally as the map_imperatively.local_table argument.

In contrast to the “one mapper per class” requirement of a mapped class, the Table or other FromClause object that is the subject of the mapping may be associated with any number of mappings. The Mapper applies modifications directly to the user-defined class, but does not modify the given Table or other FromClause in any way.

The properties dictionary

This is a dictionary of all of the attributes that will be associated with the mapped class. By default, the Mapper generates entries for this dictionary derived from the given Table, in the form of ColumnProperty objects which each refer to an individual Column of the mapped table. The properties dictionary will also contain all the other kinds of MapperProperty objects to be configured, most commonly instances generated by the relationship() construct.

When mapping with the declarative mapping style, the properties dictionary is generated by the declarative system by scanning the class to be mapped for appropriate attributes. See the section Defining Mapped Properties with Declarative for notes on this process.

When mapping with the imperative style, the properties dictionary is passed directly as the properties parameter to registry.map_imperatively(), which will pass it along to the Mapper.properties parameter.

Other mapper configuration parameters

When mapping with the declarative mapping style, additional mapper configuration arguments are configured via the __mapper_args__ class attribute. Examples of use are available at Mapper Configuration Options with Declarative.

When mapping with the imperative style, keyword arguments are passed to the to registry.map_imperatively() method which passes them along to the Mapper class.

The full range of parameters accepted are documented at Mapper.

Mapped Class Behavior

Across all styles of mapping using the registry object, the following behaviors are common:

Default Constructor

The registry applies a default constructor, i.e. __init__ method, to all mapped classes that don’t explicitly have their own __init__ method. The behavior of this method is such that it provides a convenient keyword constructor that will accept as optional keyword arguments all the attributes that are named. E.g.:

from sqlalchemy.orm import DeclarativeBase
from sqlalchemy.orm import Mapped
from sqlalchemy.orm import mapped_column

class Base(DeclarativeBase):

class User(Base):
    __tablename__ = "user"

    id: Mapped[int] = mapped_column(primary_key=True)
    name: Mapped[str]
    fullname: Mapped[str]

An object of type User above will have a constructor which allows User objects to be created as:

u1 = User(name="some name", fullname="some fullname")


The Declarative Dataclass Mapping feature provides an alternate means of generating a default __init__() method by using Python dataclasses, and allows for a highly configurable constructor form.


The __init__() method of the class is called only when the object is constructed in Python code, and not when an object is loaded or refreshed from the database. See the next section Maintaining Non-Mapped State Across Loads for a primer on how to invoke special logic when objects are loaded.

A class that includes an explicit __init__() method will maintain that method, and no default constructor will be applied.

To change the default constructor used, a user-defined Python callable may be provided to the registry.constructor parameter which will be used as the default constructor.

The constructor also applies to imperative mappings:

from sqlalchemy.orm import registry

mapper_registry = registry()

user_table = Table(
    Column("id", Integer, primary_key=True),
    Column("name", String(50)),

class User:

mapper_registry.map_imperatively(User, user_table)

The above class, mapped imperatively as described at Imperative Mapping, will also feature the default constructor associated with the registry.

New in version 1.4: classical mappings now support a standard configuration-level constructor when they are mapped via the registry.map_imperatively() method.

Maintaining Non-Mapped State Across Loads

The __init__() method of the mapped class is invoked when the object is constructed directly in Python code:

u1 = User(name="some name", fullname="some fullname")

However, when an object is loaded using the ORM Session, the __init__() method is not called:

u1 = session.scalars(select(User).where(User.name == "some name")).first()

The reason for this is that when loaded from the database, the operation used to construct the object, in the above example the User, is more analogous to deserialization, such as unpickling, rather than initial construction. The majority of the object’s important state is not being assembled for the first time, it’s being re-loaded from database rows.

Therefore to maintain state within the object that is not part of the data that’s stored to the database, such that this state is present when objects are loaded as well as constructed, there are two general approaches detailed below.

  1. Use Python descriptors like @property, rather than state, to dynamically compute attributes as needed.

    For simple attributes, this is the simplest approach and the least error prone. For example if an object Point with Point.x and Point.y wanted an attribute with the sum of these attributes:

    class Point(Base):
        __tablename__ = "point"
        id: Mapped[int] = mapped_column(primary_key=True)
        x: Mapped[int]
        y: Mapped[int]
        def x_plus_y(self):
            return self.x + self.y

    An advantage of using dynamic descriptors is that the value is computed every time, meaning it maintains the correct value as the underlying attributes (x and y in this case) might change.

    Other forms of the above pattern include Python standard library cached_property decorator (which is cached, and not re-computed each time), as well as SQLAlchemy’s hybrid_property decorator which allows for attributes that can work for SQL querying as well.

  2. Establish state on-load using InstanceEvents.load(), and optionally supplemental methods InstanceEvents.refresh() and InstanceEvents.refresh_flush().

    These are event hooks that are invoked whenever the object is loaded from the database, or when it is refreshed after being expired. Typically only the InstanceEvents.load() is needed, since non-mapped local object state is not affected by expiration operations. To revise the Point example above looks like:

    from sqlalchemy import event
    class Point(Base):
        __tablename__ = "point"
        id: Mapped[int] = mapped_column(primary_key=True)
        x: Mapped[int]
        y: Mapped[int]
        def __init__(self, x, y, **kw):
            super().__init__(x=x, y=y, **kw)
            self.x_plus_y = x + y
    @event.listens_for(Point, "load")
    def receive_load(target, context):
        target.x_plus_y = target.x + target.y

    If using the refresh events as well, the event hooks can be stacked on top of one callable if needed, as:

    @event.listens_for(Point, "load")
    @event.listens_for(Point, "refresh")
    @event.listens_for(Point, "refresh_flush")
    def receive_load(target, context, attrs=None):
        target.x_plus_y = target.x + target.y

    Above, the attrs attribute will be present for the refresh and refresh_flush events and indicate a list of attribute names that are being refreshed.

Runtime Introspection of Mapped classes, Instances and Mappers

A class that is mapped using registry will also feature a few attributes that are common to all mappings:

  • The __mapper__ attribute will refer to the Mapper that is associated with the class:

    mapper = User.__mapper__

    This Mapper is also what’s returned when using the inspect() function against the mapped class:

    from sqlalchemy import inspect
    mapper = inspect(User)
  • The __table__ attribute will refer to the Table, or more generically to the FromClause object, to which the class is mapped:

    table = User.__table__

    This FromClause is also what’s returned when using the Mapper.local_table attribute of the Mapper:

    table = inspect(User).local_table

    For a single-table inheritance mapping, where the class is a subclass that does not have a table of its own, the Mapper.local_table attribute as well as the .__table__ attribute will be None. To retrieve the “selectable” that is actually selected from during a query for this class, this is available via the Mapper.selectable attribute:

    table = inspect(User).selectable

Inspection of Mapper objects

As illustrated in the previous section, the Mapper object is available from any mapped class, regardless of method, using the Runtime Inspection API system. Using the inspect() function, one can acquire the Mapper from a mapped class:

>>> from sqlalchemy import inspect
>>> insp = inspect(User)

Detailed information is available including Mapper.columns:

>>> insp.columns
<sqlalchemy.util._collections.OrderedProperties object at 0x102f407f8>

This is a namespace that can be viewed in a list format or via individual names:

>>> list(insp.columns)
[Column('id', Integer(), table=<user>, primary_key=True, nullable=False), Column('name', String(length=50), table=<user>), Column('fullname', String(length=50), table=<user>), Column('nickname', String(length=50), table=<user>)]
>>> insp.columns.name
Column('name', String(length=50), table=<user>)

Other namespaces include Mapper.all_orm_descriptors, which includes all mapped attributes as well as hybrids, association proxies:

>>> insp.all_orm_descriptors
<sqlalchemy.util._collections.ImmutableProperties object at 0x1040e2c68>
>>> insp.all_orm_descriptors.keys()
['fullname', 'nickname', 'name', 'id']

As well as Mapper.column_attrs:

>>> list(insp.column_attrs)
[<ColumnProperty at 0x10403fde0; id>, <ColumnProperty at 0x10403fce8; name>, <ColumnProperty at 0x1040e9050; fullname>, <ColumnProperty at 0x1040e9148; nickname>]
>>> insp.column_attrs.name
<ColumnProperty at 0x10403fce8; name>
>>> insp.column_attrs.name.expression
Column('name', String(length=50), table=<user>)

See also


Inspection of Mapped Instances

The inspect() function also provides information about instances of a mapped class. When applied to an instance of a mapped class, rather than the class itself, the object returned is known as InstanceState, which will provide links to not only the Mapper in use by the class, but also a detailed interface that provides information on the state of individual attributes within the instance including their current value and how this relates to what their database-loaded value is.

Given an instance of the User class loaded from the database:

>>> u1 = session.scalars(select(User)).first()

The inspect() function will return to us an InstanceState object:

>>> insp = inspect(u1)
>>> insp
<sqlalchemy.orm.state.InstanceState object at 0x7f07e5fec2e0>

With this object we can see elements such as the Mapper:

>>> insp.mapper
<Mapper at 0x7f07e614ef50; User>

The Session to which the object is attached, if any:

>>> insp.session
<sqlalchemy.orm.session.Session object at 0x7f07e614f160>

Information about the current persistence state for the object:

>>> insp.persistent
>>> insp.pending

Attribute state information such as attributes that have not been loaded or lazy loaded (assume addresses refers to a relationship() on the mapped class to a related class):

>>> insp.unloaded

Information regarding the current in-Python status of attributes, such as attributes that have not been modified since the last flush:

>>> insp.unmodified
{'nickname', 'name', 'fullname', 'id'}

as well as specific history on modifications to attributes since the last flush:

>>> insp.attrs.nickname.value
>>> u1.nickname = "new nickname"
>>> insp.attrs.nickname.history
History(added=['new nickname'], unchanged=(), deleted=['nickname'])