Mapping Class Inheritance Hierarchies

SQLAlchemy supports three forms of inheritance: single table inheritance, where several types of classes are represented by a single table, concrete table inheritance, where each type of class is represented by independent tables, and joined table inheritance, where the class hierarchy is broken up among dependent tables, each class represented by its own table that only includes those attributes local to that class.

The most common forms of inheritance are single and joined table, while concrete inheritance presents more configurational challenges.

When mappers are configured in an inheritance relationship, SQLAlchemy has the ability to load elements polymorphically, meaning that a single query can return objects of multiple types.

See also

Inheritance Mapping Recipes - complete examples of joined, single and concrete inheritance

Joined Table Inheritance

In joined table inheritance, each class along a hierarchy of classes is represented by a distinct table. Querying for a particular subclass in the hierarchy will render as a SQL JOIN along all tables in its inheritance path. If the queried class is the base class, the default behavior is to include only the base table in a SELECT statement. In all cases, the ultimate class to instantiate for a given row is determined by a discriminator column or an expression that works against the base table. When a subclass is loaded only against a base table, resulting objects will have base attributes populated at first; attributes that are local to the subclass will lazy load when they are accessed. Alternatively, there are options which can change the default behavior, allowing the query to include columns corresponding to multiple tables/subclasses up front.

The base class in a joined inheritance hierarchy is configured with additional arguments that will refer to the polymorphic discriminator column as well as the identifier for the base class:

  1. class Employee(Base):
  2.     __tablename__ = 'employee'
  3.     id = Column(Integer, primary_key=True)
  4.     name = Column(String(50))
  5.     type = Column(String(50))
  7.     __mapper_args__ = {
  8.         'polymorphic_identity':'employee',
  9.         'polymorphic_on':type
  10.     }

Above, an additional column type is established to act as the discriminator, configured as such using the mapper.polymorphic_on parameter. This column will store a value which indicates the type of object represented within the row. The column may be of any datatype, though string and integer are the most common. The actual data value to be applied to this column for a particular row in the database is specified using the mapper.polymorphic_identity parameter, described below.

While a polymorphic discriminator expression is not strictly necessary, it is required if polymorphic loading is desired. Establishing a simple column on the base table is the easiest way to achieve this, however very sophisticated inheritance mappings may even configure a SQL expression such as a CASE statement as the polymorphic discriminator.

Note

Currently, only one discriminator column or SQL expression may be configured for the entire inheritance hierarchy, typically on the base- most class in the hierarchy. “Cascading” polymorphic discriminator expressions are not yet supported.

We next define Engineer and Manager subclasses of Employee. Each contains columns that represent the attributes unique to the subclass they represent. Each table also must contain a primary key column (or columns), as well as a foreign key reference to the parent table:

  1. class Engineer(Employee):
  2.     __tablename__ = 'engineer'
  3.     id = Column(Integer, ForeignKey('employee.id'), primary_key=True)
  4.     engineer_name = Column(String(30))
  6.     __mapper_args__ = {
  7.         'polymorphic_identity':'engineer',
  8.     }
  10. class Manager(Employee):
  11.     __tablename__ = 'manager'
  12.     id = Column(Integer, ForeignKey('employee.id'), primary_key=True)
  13.     manager_name = Column(String(30))
  15.     __mapper_args__ = {
  16.         'polymorphic_identity':'manager',
  17.     }

In the above example, each mapping specifies the mapper.polymorphic_identity parameter within its mapper arguments. This value populates the column designated by the mapper.polymorphic_on parameter established on the base mapper. The mapper.polymorphic_identity parameter should be unique to each mapped class across the whole hierarchy, and there should only be one “identity” per mapped class; as noted above, “cascading” identities where some subclasses introduce a second identity are not supported.

The ORM uses the value set up by mapper.polymorphic_identity in order to determine which class a row belongs towards when loading rows polymorphically. In the example above, every row which represents an Employee will have the value 'employee' in its type row; similarly, every Engineer will get the value 'engineer', and each Manager will get the value 'manager'. Regardless of whether the inheritance mapping uses distinct joined tables for subclasses as in joined table inheritance, or all one table as in single table inheritance, this value is expected to be persisted and available to the ORM when querying. The mapper.polymorphic_identity parameter also applies to concrete table inheritance, but is not actually persisted; see the later section at Concrete Table Inheritance for details.

In a polymorphic setup, it is most common that the foreign key constraint is established on the same column or columns as the primary key itself, however this is not required; a column distinct from the primary key may also be made to refer to the parent via foreign key. The way that a JOIN is constructed from the base table to subclasses is also directly customizable, however this is rarely necessary.

Joined inheritance primary keys

One natural effect of the joined table inheritance configuration is that the identity of any mapped object can be determined entirely from rows in the base table alone. This has obvious advantages, so SQLAlchemy always considers the primary key columns of a joined inheritance class to be those of the base table only. In other words, the id columns of both the engineer and manager tables are not used to locate Engineer or Manager objects - only the value in employee.id is considered. engineer.id and manager.id are still of course critical to the proper operation of the pattern overall as they are used to locate the joined row, once the parent row has been determined within a statement.

With the joined inheritance mapping complete, querying against Employee will return a combination of Employee, Engineer and Manager objects. Newly saved Engineer, Manager, and Employee objects will automatically populate the employee.type column with the correct “discriminator” value in this case "engineer", "manager", or "employee", as appropriate.

Relationships with Joined Inheritance

Relationships are fully supported with joined table inheritance. The relationship involving a joined-inheritance class should target the class in the hierarchy that also corresponds to the foreign key constraint; below, as the employee table has a foreign key constraint back to the company table, the relationships are set up between Company and Employee:

  1. class Company(Base):
  2.     __tablename__ = 'company'
  3.     id = Column(Integer, primary_key=True)
  4.     name = Column(String(50))
  5.     employees = relationship("Employee", back_populates="company")
  7. class Employee(Base):
  8.     __tablename__ = 'employee'
  9.     id = Column(Integer, primary_key=True)
  10.     name = Column(String(50))
  11.     type = Column(String(50))
  12.     company_id = Column(ForeignKey('company.id'))
  13.     company = relationship("Company", back_populates="employees")
  15.     __mapper_args__ = {
  16.         'polymorphic_identity':'employee',
  17.         'polymorphic_on':type
  18.     }
  20. class Manager(Employee):
  21.     # ...
  23. class Engineer(Employee):
  24.     # ...

If the foreign key constraint is on a table corresponding to a subclass, the relationship should target that subclass instead. In the example below, there is a foreign key constraint from manager to company, so the relationships are established between the Manager and Company classes:

  1. class Company(Base):
  2.     __tablename__ = 'company'
  3.     id = Column(Integer, primary_key=True)
  4.     name = Column(String(50))
  5.     managers = relationship("Manager", back_populates="company")
  7. class Employee(Base):
  8.     __tablename__ = 'employee'
  9.     id = Column(Integer, primary_key=True)
  10.     name = Column(String(50))
  11.     type = Column(String(50))
  13.     __mapper_args__ = {
  14.         'polymorphic_identity':'employee',
  15.         'polymorphic_on':type
  16.     }
  18. class Manager(Employee):
  19.     __tablename__ = 'manager'
  20.     id = Column(Integer, ForeignKey('employee.id'), primary_key=True)
  21.     manager_name = Column(String(30))
  23.     company_id = Column(ForeignKey('company.id'))
  24.     company = relationship("Company", back_populates="managers")
  26.     __mapper_args__ = {
  27.         'polymorphic_identity':'manager',
  28.     }
  30. class Engineer(Employee):
  31.     # ...

Above, the Manager class will have a Manager.company attribute; Company will have a Company.managers attribute that always loads against a join of the employee and manager tables together.

Loading Joined Inheritance Mappings

See the sections Loading Inheritance Hierarchies and Loading objects with joined table inheritance for background on inheritance loading techniques, including configuration of tables to be queried both at mapper configuration time as well as query time.

Single Table Inheritance

Single table inheritance represents all attributes of all subclasses within a single table. A particular subclass that has attributes unique to that class will persist them within columns in the table that are otherwise NULL if the row refers to a different kind of object.

Querying for a particular subclass in the hierarchy will render as a SELECT against the base table, which will include a WHERE clause that limits rows to those with a particular value or values present in the discriminator column or expression.

Single table inheritance has the advantage of simplicity compared to joined table inheritance; queries are much more efficient as only one table needs to be involved in order to load objects of every represented class.

Single-table inheritance configuration looks much like joined-table inheritance, except only the base class specifies __tablename__. A discriminator column is also required on the base table so that classes can be differentiated from each other.

Even though subclasses share the base table for all of their attributes, when using Declarative, Column objects may still be specified on subclasses, indicating that the column is to be mapped only to that subclass; the Column will be applied to the same base Table object:

  1. class Employee(Base):
  2.     __tablename__ = 'employee'
  3.     id = Column(Integer, primary_key=True)
  4.     name = Column(String(50))
  5.     type = Column(String(20))
  7.     __mapper_args__ = {
  8.         'polymorphic_on':type,
  9.         'polymorphic_identity':'employee'
  10.     }
  12. class Manager(Employee):
  13.     manager_data = Column(String(50))
  15.     __mapper_args__ = {
  16.         'polymorphic_identity':'manager'
  17.     }
  19. class Engineer(Employee):
  20.     engineer_info = Column(String(50))
  22.     __mapper_args__ = {
  23.         'polymorphic_identity':'engineer'
  24.     }

Note that the mappers for the derived classes Manager and Engineer omit the __tablename__, indicating they do not have a mapped table of their own.

Resolving Column Conflicts

Note in the previous section that the manager_name and engineer_info columns are “moved up” to be applied to Employee.__table__, as a result of their declaration on a subclass that has no table of its own. A tricky case comes up when two subclasses want to specify the same column, as below:

  1. class Employee(Base):
  2.     __tablename__ = 'employee'
  3.     id = Column(Integer, primary_key=True)
  4.     name = Column(String(50))
  5.     type = Column(String(20))
  7.     __mapper_args__ = {
  8.         'polymorphic_on':type,
  9.         'polymorphic_identity':'employee'
  10.     }
  12. class Engineer(Employee):
  13.     __mapper_args__ = {'polymorphic_identity': 'engineer'}
  14.     start_date = Column(DateTime)
  16. class Manager(Employee):
  17.     __mapper_args__ = {'polymorphic_identity': 'manager'}
  18.     start_date = Column(DateTime)

Above, the start_date column declared on both Engineer and Manager will result in an error:

  1. sqlalchemy.exc.ArgumentError: Column 'start_date' on class
  2. <class '__main__.Manager'> conflicts with existing
  3. column 'employee.start_date'

The above scenario presents an ambiguity to the Declarative mapping system that may be resolved by using declared_attr to define the Column conditionally, taking care to return the existing column via the parent __table__ if it already exists:

  1. from sqlalchemy.orm import declared_attr
  3. class Employee(Base):
  4.     __tablename__ = 'employee'
  5.     id = Column(Integer, primary_key=True)
  6.     name = Column(String(50))
  7.     type = Column(String(20))
  9.     __mapper_args__ = {
  10.         'polymorphic_on':type,
  11.         'polymorphic_identity':'employee'
  12.     }
  14. class Engineer(Employee):
  15.     __mapper_args__ = {'polymorphic_identity': 'engineer'}
  17.     @declared_attr
  18.     def start_date(cls):
  19.         "Start date column, if not present already."
  20.         return Employee.__table__.c.get('start_date', Column(DateTime))
  22. class Manager(Employee):
  23.     __mapper_args__ = {'polymorphic_identity': 'manager'}
  25.     @declared_attr
  26.     def start_date(cls):
  27.         "Start date column, if not present already."
  28.         return Employee.__table__.c.get('start_date', Column(DateTime))

Above, when Manager is mapped, the start_date column is already present on the Employee class; by returning the existing Column object, the declarative system recognizes that this is the same column to be mapped to the two different subclasses separately.

A similar concept can be used with mixin classes (see Composing Mapped Hierarchies with Mixins) to define a particular series of columns and/or other mapped attributes from a reusable mixin class:

  1. class Employee(Base):
  2.     __tablename__ = 'employee'
  3.     id = Column(Integer, primary_key=True)
  4.     name = Column(String(50))
  5.     type = Column(String(20))
  7.     __mapper_args__ = {
  8.         'polymorphic_on':type,
  9.         'polymorphic_identity':'employee'
  10.     }
  12. class HasStartDate:
  13.     @declared_attr
  14.     def start_date(cls):
  15.         return cls.__table__.c.get('start_date', Column(DateTime))
  17. class Engineer(HasStartDate, Employee):
  18.     __mapper_args__ = {'polymorphic_identity': 'engineer'}
  20. class Manager(HasStartDate, Employee):
  21.     __mapper_args__ = {'polymorphic_identity': 'manager'}

Relationships with Single Table Inheritance

Relationships are fully supported with single table inheritance. Configuration is done in the same manner as that of joined inheritance; a foreign key attribute should be on the same class that’s the “foreign” side of the relationship:

  1. class Company(Base):
  2.     __tablename__ = 'company'
  3.     id = Column(Integer, primary_key=True)
  4.     name = Column(String(50))
  5.     employees = relationship("Employee", back_populates="company")
  7. class Employee(Base):
  8.     __tablename__ = 'employee'
  9.     id = Column(Integer, primary_key=True)
  10.     name = Column(String(50))
  11.     type = Column(String(50))
  12.     company_id = Column(ForeignKey('company.id'))
  13.     company = relationship("Company", back_populates="employees")
  15.     __mapper_args__ = {
  16.         'polymorphic_identity':'employee',
  17.         'polymorphic_on':type
  18.     }
  21. class Manager(Employee):
  22.     manager_data = Column(String(50))
  24.     __mapper_args__ = {
  25.         'polymorphic_identity':'manager'
  26.     }
  28. class Engineer(Employee):
  29.     engineer_info = Column(String(50))
  31.     __mapper_args__ = {
  32.         'polymorphic_identity':'engineer'
  33.     }

Also, like the case of joined inheritance, we can create relationships that involve a specific subclass. When queried, the SELECT statement will include a WHERE clause that limits the class selection to that subclass or subclasses:

  1. class Company(Base):
  2.     __tablename__ = 'company'
  3.     id = Column(Integer, primary_key=True)
  4.     name = Column(String(50))
  5.     managers = relationship("Manager", back_populates="company")
  7. class Employee(Base):
  8.     __tablename__ = 'employee'
  9.     id = Column(Integer, primary_key=True)
  10.     name = Column(String(50))
  11.     type = Column(String(50))
  13.     __mapper_args__ = {
  14.         'polymorphic_identity':'employee',
  15.         'polymorphic_on':type
  16.     }
  19. class Manager(Employee):
  20.     manager_name = Column(String(30))
  22.     company_id = Column(ForeignKey('company.id'))
  23.     company = relationship("Company", back_populates="managers")
  25.     __mapper_args__ = {
  26.         'polymorphic_identity':'manager',
  27.     }
  30. class Engineer(Employee):
  31.     engineer_info = Column(String(50))
  33.     __mapper_args__ = {
  34.         'polymorphic_identity':'engineer'
  35.     }

Above, the Manager class will have a Manager.company attribute; Company will have a Company.managers attribute that always loads against the employee with an additional WHERE clause that limits rows to those with type = 'manager'.

Loading Single Inheritance Mappings

The loading techniques for single-table inheritance are mostly identical to those used for joined-table inheritance, and a high degree of abstraction is provided between these two mapping types such that it is easy to switch between them as well as to intermix them in a single hierarchy (just omit __tablename__ from whichever subclasses are to be single-inheriting). See the sections Loading Inheritance Hierarchies and Loading objects with single table inheritance for documentation on inheritance loading techniques, including configuration of classes to be queried both at mapper configuration time as well as query time.

Concrete Table Inheritance

Concrete inheritance maps each subclass to its own distinct table, each of which contains all columns necessary to produce an instance of that class. A concrete inheritance configuration by default queries non-polymorphically; a query for a particular class will only query that class’ table and only return instances of that class. Polymorphic loading of concrete classes is enabled by configuring within the mapper a special SELECT that typically is produced as a UNION of all the tables.

Warning

Concrete table inheritance is much more complicated than joined or single table inheritance, and is much more limited in functionality especially pertaining to using it with relationships, eager loading, and polymorphic loading. When used polymorphically it produces very large queries with UNIONS that won’t perform as well as simple joins. It is strongly advised that if flexibility in relationship loading and polymorphic loading is required, that joined or single table inheritance be used if at all possible. If polymorphic loading isn’t required, then plain non-inheriting mappings can be used if each class refers to its own table completely.

Whereas joined and single table inheritance are fluent in “polymorphic” loading, it is a more awkward affair in concrete inheritance. For this reason, concrete inheritance is more appropriate when polymorphic loading is not required. Establishing relationships that involve concrete inheritance classes is also more awkward.

To establish a class as using concrete inheritance, add the mapper.concrete parameter within the __mapper_args__. This indicates to Declarative as well as the mapping that the superclass table should not be considered as part of the mapping:

  1. class Employee(Base):
  2.     __tablename__ = 'employee'
  4.     id = Column(Integer, primary_key=True)
  5.     name = Column(String(50))
  7. class Manager(Employee):
  8.     __tablename__ = 'manager'
  10.     id = Column(Integer, primary_key=True)
  11.     name = Column(String(50))
  12.     manager_data = Column(String(50))
  14.     __mapper_args__ = {
  15.         'concrete': True
  16.     }
  18. class Engineer(Employee):
  19.     __tablename__ = 'engineer'
  21.     id = Column(Integer, primary_key=True)
  22.     name = Column(String(50))
  23.     engineer_info = Column(String(50))
  25.     __mapper_args__ = {
  26.         'concrete': True
  27.     }

Two critical points should be noted:

  • We must define all columns explicitly on each subclass, even those of the same name. A column such as Employee.name here is not copied out to the tables mapped by Manager or Engineer for us.

  • while the Engineer and Manager classes are mapped in an inheritance relationship with Employee, they still do not include polymorphic loading. Meaning, if we query for Employee objects, the manager and engineer tables are not queried at all.

Concrete Polymorphic Loading Configuration

Polymorphic loading with concrete inheritance requires that a specialized SELECT is configured against each base class that should have polymorphic loading. This SELECT needs to be capable of accessing all the mapped tables individually, and is typically a UNION statement that is constructed using a SQLAlchemy helper polymorphic_union().

As discussed in Loading Inheritance Hierarchies, mapper inheritance configurations of any type can be configured to load from a special selectable by default using the mapper.with_polymorphic argument. Current public API requires that this argument is set on a Mapper when it is first constructed.

However, in the case of Declarative, both the mapper and the Table that is mapped are created at once, the moment the mapped class is defined. This means that the mapper.with_polymorphic argument cannot be provided yet, since the Table objects that correspond to the subclasses haven’t yet been defined.

There are a few strategies available to resolve this cycle, however Declarative provides helper classes ConcreteBase and AbstractConcreteBase which handle this issue behind the scenes.

Using ConcreteBase, we can set up our concrete mapping in almost the same way as we do other forms of inheritance mappings:

  1. from sqlalchemy.ext.declarative import ConcreteBase
  3. class Employee(ConcreteBase, Base):
  4.     __tablename__ = 'employee'
  5.     id = Column(Integer, primary_key=True)
  6.     name = Column(String(50))
  8.     __mapper_args__ = {
  9.         'polymorphic_identity': 'employee',
  10.         'concrete': True
  11.     }
  13. class Manager(Employee):
  14.     __tablename__ = 'manager'
  15.     id = Column(Integer, primary_key=True)
  16.     name = Column(String(50))
  17.     manager_data = Column(String(40))
  19.     __mapper_args__ = {
  20.         'polymorphic_identity': 'manager',
  21.         'concrete': True
  22.     }
  24. class Engineer(Employee):
  25.     __tablename__ = 'engineer'
  26.     id = Column(Integer, primary_key=True)
  27.     name = Column(String(50))
  28.     engineer_info = Column(String(40))
  30.     __mapper_args__ = {
  31.         'polymorphic_identity': 'engineer',
  32.         'concrete': True
  33.     }

Above, Declarative sets up the polymorphic selectable for the Employee class at mapper “initialization” time; this is the late-configuration step for mappers that resolves other dependent mappers. The ConcreteBase helper uses the polymorphic_union() function to create a UNION of all concrete-mapped tables after all the other classes are set up, and then configures this statement with the already existing base-class mapper.

Upon select, the polymorphic union produces a query like this:

  1. session.query(Employee).all()
  2. SELECT
  3.     pjoin.id AS pjoin_id,
  4.     pjoin.name AS pjoin_name,
  5.     pjoin.type AS pjoin_type,
  6.     pjoin.manager_data AS pjoin_manager_data,
  7.     pjoin.engineer_info AS pjoin_engineer_info
  8. FROM (
  9.     SELECT
  10.         employee.id AS id,
  11.         employee.name AS name,
  12.         CAST(NULL AS VARCHAR(50)) AS manager_data,
  13.         CAST(NULL AS VARCHAR(50)) AS engineer_info,
  14.         'employee' AS type
  15.     FROM employee
  16.     UNION ALL
  17.     SELECT
  18.         manager.id AS id,
  19.         manager.name AS name,
  20.         manager.manager_data AS manager_data,
  21.         CAST(NULL AS VARCHAR(50)) AS engineer_info,
  22.         'manager' AS type
  23.     FROM manager
  24.     UNION ALL
  25.     SELECT
  26.         engineer.id AS id,
  27.         engineer.name AS name,
  28.         CAST(NULL AS VARCHAR(50)) AS manager_data,
  29.         engineer.engineer_info AS engineer_info,
  30.         'engineer' AS type
  31.     FROM engineer
  32. ) AS pjoin

The above UNION query needs to manufacture “NULL” columns for each subtable in order to accommodate for those columns that aren’t members of that particular subclass.

Abstract Concrete Classes

The concrete mappings illustrated thus far show both the subclasses as well as the base class mapped to individual tables. In the concrete inheritance use case, it is common that the base class is not represented within the database, only the subclasses. In other words, the base class is “abstract”.

Normally, when one would like to map two different subclasses to individual tables, and leave the base class unmapped, this can be achieved very easily. When using Declarative, just declare the base class with the __abstract__ indicator:

  1. class Employee(Base):
  2.     __abstract__ = True
  4. class Manager(Employee):
  5.     __tablename__ = 'manager'
  6.     id = Column(Integer, primary_key=True)
  7.     name = Column(String(50))
  8.     manager_data = Column(String(40))
  10.     __mapper_args__ = {
  11.         'polymorphic_identity': 'manager',
  12.     }
  14. class Engineer(Employee):
  15.     __tablename__ = 'engineer'
  16.     id = Column(Integer, primary_key=True)
  17.     name = Column(String(50))
  18.     engineer_info = Column(String(40))
  20.     __mapper_args__ = {
  21.         'polymorphic_identity': 'engineer',
  22.     }

Above, we are not actually making use of SQLAlchemy’s inheritance mapping facilities; we can load and persist instances of Manager and Engineer normally. The situation changes however when we need to query polymorphically, that is, we’d like to emit session.query(Employee) and get back a collection of Manager and Engineer instances. This brings us back into the domain of concrete inheritance, and we must build a special mapper against Employee in order to achieve this.

Mappers can always SELECT

In SQLAlchemy, a mapper for a class always has to refer to some “selectable”, which is normally a Table but may also refer to any select() object as well. While it may appear that a “single table inheritance” mapper does not map to a table, these mappers in fact implicitly refer to the table that is mapped by a superclass.

To modify our concrete inheritance example to illustrate an “abstract” base that is capable of polymorphic loading, we will have only an engineer and a manager table and no employee table, however the Employee mapper will be mapped directly to the “polymorphic union”, rather than specifying it locally to the mapper.with_polymorphic parameter.

To help with this, Declarative offers a variant of the ConcreteBase class called AbstractConcreteBase which achieves this automatically:

  1. from sqlalchemy.ext.declarative import AbstractConcreteBase
  3. class Employee(AbstractConcreteBase, Base):
  4.     pass
  6. class Manager(Employee):
  7.     __tablename__ = 'manager'
  8.     id = Column(Integer, primary_key=True)
  9.     name = Column(String(50))
  10.     manager_data = Column(String(40))
  12.     __mapper_args__ = {
  13.         'polymorphic_identity': 'manager',
  14.         'concrete': True
  15.     }
  17. class Engineer(Employee):
  18.     __tablename__ = 'engineer'
  19.     id = Column(Integer, primary_key=True)
  20.     name = Column(String(50))
  21.     engineer_info = Column(String(40))
  23.     __mapper_args__ = {
  24.         'polymorphic_identity': 'engineer',
  25.         'concrete': True
  26.     }

The AbstractConcreteBase helper class has a more complex internal process than that of ConcreteBase, in that the entire mapping of the base class must be delayed until all the subclasses have been declared. With a mapping like the above, only instances of Manager and Engineer may be persisted; querying against the Employee class will always produce Manager and Engineer objects.

Classical and Semi-Classical Concrete Polymorphic Configuration

The Declarative configurations illustrated with ConcreteBase and AbstractConcreteBase are equivalent to two other forms of configuration that make use of polymorphic_union() explicitly. These configurational forms make use of the Table object explicitly so that the “polymorphic union” can be created first, then applied to the mappings. These are illustrated here to clarify the role of the polymorphic_union() function in terms of mapping.

A semi-classical mapping for example makes use of Declarative, but establishes the Table objects separately:

  1. metadata = Base.metadata
  3. employees_table = Table(
  4.     'employee', metadata,
  5.     Column('id', Integer, primary_key=True),
  6.     Column('name', String(50)),
  7. )
  9. managers_table = Table(
  10.     'manager', metadata,
  11.     Column('id', Integer, primary_key=True),
  12.     Column('name', String(50)),
  13.     Column('manager_data', String(50)),
  14. )
  16. engineers_table = Table(
  17.     'engineer', metadata,
  18.     Column('id', Integer, primary_key=True),
  19.     Column('name', String(50)),
  20.     Column('engineer_info', String(50)),
  21. )

Next, the UNION is produced using polymorphic_union():

  1. from sqlalchemy.orm import polymorphic_union
  3. pjoin = polymorphic_union({
  4.     'employee': employees_table,
  5.     'manager': managers_table,
  6.     'engineer': engineers_table
  7. }, 'type', 'pjoin')

With the above Table objects, the mappings can be produced using “semi-classical” style, where we use Declarative in conjunction with the __table__ argument; our polymorphic union above is passed via __mapper_args__ to the mapper.with_polymorphic parameter:

  1. class Employee(Base):
  2.     __table__ = employee_table
  3.     __mapper_args__ = {
  4.         'polymorphic_on': pjoin.c.type,
  5.         'with_polymorphic': ('*', pjoin),
  6.         'polymorphic_identity': 'employee'
  7.     }
  9. class Engineer(Employee):
  10.     __table__ = engineer_table
  11.     __mapper_args__ = {
  12.         'polymorphic_identity': 'engineer',
  13.         'concrete': True}
  15. class Manager(Employee):
  16.     __table__ = manager_table
  17.     __mapper_args__ = {
  18.         'polymorphic_identity': 'manager',
  19.         'concrete': True}

Alternatively, the same Table objects can be used in fully “classical” style, without using Declarative at all. A constructor similar to that supplied by Declarative is illustrated:

  1. class Employee(object):
  2.     def __init__(self, **kw):
  3.         for k in kw:
  4.             setattr(self, k, kw[k])
  6. class Manager(Employee):
  7.     pass
  9. class Engineer(Employee):
  10.     pass
  12. employee_mapper = mapper_registry.map_imperatively(
  13.     Employee,
  14.     pjoin,
  15.     with_polymorphic=('*', pjoin),
  16.     polymorphic_on=pjoin.c.type,
  17. )
  18. manager_mapper = mapper_registry.map_imperatively(
  19.     Manager,
  20.     managers_table,
  21.     inherits=employee_mapper,
  22.     concrete=True,
  23.     polymorphic_identity='manager',
  24. )
  25. engineer_mapper = mapper_registry.map_imperatively(
  26.     Engineer,
  27.     engineers_table,
  28.     inherits=employee_mapper,
  29.     concrete=True,
  30.     polymorphic_identity='engineer',
  31. )

The “abstract” example can also be mapped using “semi-classical” or “classical” style. The difference is that instead of applying the “polymorphic union” to the mapper.with_polymorphic parameter, we apply it directly as the mapped selectable on our basemost mapper. The semi-classical mapping is illustrated below:

  1. from sqlalchemy.orm import polymorphic_union
  3. pjoin = polymorphic_union({
  4.     'manager': managers_table,
  5.     'engineer': engineers_table
  6. }, 'type', 'pjoin')
  8. class Employee(Base):
  9.     __table__ = pjoin
  10.     __mapper_args__ = {
  11.         'polymorphic_on': pjoin.c.type,
  12.         'with_polymorphic': '*',
  13.         'polymorphic_identity': 'employee'
  14.     }
  16. class Engineer(Employee):
  17.     __table__ = engineer_table
  18.     __mapper_args__ = {
  19.         'polymorphic_identity': 'engineer',
  20.         'concrete': True}
  22. class Manager(Employee):
  23.     __table__ = manager_table
  24.     __mapper_args__ = {
  25.         'polymorphic_identity': 'manager',
  26.         'concrete': True}

Above, we use polymorphic_union() in the same manner as before, except that we omit the employee table.

See also

Imperative (a.k.a. Classical) Mappings - background information on “classical” mappings

Relationships with Concrete Inheritance

In a concrete inheritance scenario, mapping relationships is challenging since the distinct classes do not share a table. If the relationships only involve specific classes, such as a relationship between Company in our previous examples and Manager, special steps aren’t needed as these are just two related tables.

However, if Company is to have a one-to-many relationship to Employee, indicating that the collection may include both Engineer and Manager objects, that implies that Employee must have polymorphic loading capabilities and also that each table to be related must have a foreign key back to the company table. An example of such a configuration is as follows:

  1. from sqlalchemy.ext.declarative import ConcreteBase
  4. class Company(Base):
  5.     __tablename__ = 'company'
  6.     id = Column(Integer, primary_key=True)
  7.     name = Column(String(50))
  8.     employees = relationship("Employee")
  11. class Employee(ConcreteBase, Base):
  12.     __tablename__ = 'employee'
  13.     id = Column(Integer, primary_key=True)
  14.     name = Column(String(50))
  15.     company_id = Column(ForeignKey('company.id'))
  17.     __mapper_args__ = {
  18.         'polymorphic_identity': 'employee',
  19.         'concrete': True
  20.     }
  23. class Manager(Employee):
  24.     __tablename__ = 'manager'
  25.     id = Column(Integer, primary_key=True)
  26.     name = Column(String(50))
  27.     manager_data = Column(String(40))
  28.     company_id = Column(ForeignKey('company.id'))
  30.     __mapper_args__ = {
  31.         'polymorphic_identity': 'manager',
  32.         'concrete': True
  33.     }
  36. class Engineer(Employee):
  37.     __tablename__ = 'engineer'
  38.     id = Column(Integer, primary_key=True)
  39.     name = Column(String(50))
  40.     engineer_info = Column(String(40))
  41.     company_id = Column(ForeignKey('company.id'))
  43.     __mapper_args__ = {
  44.         'polymorphic_identity': 'engineer',
  45.         'concrete': True
  46.     }

The next complexity with concrete inheritance and relationships involves when we’d like one or all of Employee, Manager and Engineer to themselves refer back to Company. For this case, SQLAlchemy has special behavior in that a relationship() placed on Employee which links to Company does not work against the Manager and Engineer classes, when exercised at the instance level. Instead, a distinct relationship() must be applied to each class. In order to achieve bi-directional behavior in terms of three separate relationships which serve as the opposite of Company.employees, the relationship.back_populates parameter is used between each of the relationships:

  1. from sqlalchemy.ext.declarative import ConcreteBase
  4. class Company(Base):
  5.     __tablename__ = 'company'
  6.     id = Column(Integer, primary_key=True)
  7.     name = Column(String(50))
  8.     employees = relationship("Employee", back_populates="company")
  11. class Employee(ConcreteBase, Base):
  12.     __tablename__ = 'employee'
  13.     id = Column(Integer, primary_key=True)
  14.     name = Column(String(50))
  15.     company_id = Column(ForeignKey('company.id'))
  16.     company = relationship("Company", back_populates="employees")
  18.     __mapper_args__ = {
  19.         'polymorphic_identity': 'employee',
  20.         'concrete': True
  21.     }
  24. class Manager(Employee):
  25.     __tablename__ = 'manager'
  26.     id = Column(Integer, primary_key=True)
  27.     name = Column(String(50))
  28.     manager_data = Column(String(40))
  29.     company_id = Column(ForeignKey('company.id'))
  30.     company = relationship("Company", back_populates="employees")
  32.     __mapper_args__ = {
  33.         'polymorphic_identity': 'manager',
  34.         'concrete': True
  35.     }
  38. class Engineer(Employee):
  39.     __tablename__ = 'engineer'
  40.     id = Column(Integer, primary_key=True)
  41.     name = Column(String(50))
  42.     engineer_info = Column(String(40))
  43.     company_id = Column(ForeignKey('company.id'))
  44.     company = relationship("Company", back_populates="employees")
  46.     __mapper_args__ = {
  47.         'polymorphic_identity': 'engineer',
  48.         'concrete': True
  49.     }

The above limitation is related to the current implementation, including that concrete inheriting classes do not share any of the attributes of the superclass and therefore need distinct relationships to be set up.

Loading Concrete Inheritance Mappings

The options for loading with concrete inheritance are limited; generally, if polymorphic loading is configured on the mapper using one of the declarative concrete mixins, it can’t be modified at query time in current SQLAlchemy versions. Normally, the with_polymorphic() function would be able to override the style of loading used by concrete, however due to current limitations this is not yet supported.