[alpha_api]: javascript:alphaApi()
[alpha_implementation]: javascript:alphaImplementation()

Mapper Configuration {@name=advdatamapping}
======================

This section details all the options available to Mappers, as well as advanced patterns.

### API Reference

Important points of reference, detailing the full configurational ORM API:

[ORM Package Documentation](rel:docstrings_sqlalchemy.orm)

[Options for mapper()](rel:docstrings_sqlalchemy.orm_modfunc_mapper)

[Options for relation()](rel:docstrings_sqlalchemy.orm_modfunc_relation)

### Customizing Column Properties {@name=columns}

The default behavior of a `mapper` is to assemble all the columns in the mapped `Table` into mapped object attributes.  This behavior can be modified in several ways, as well as enhanced by SQL expressions.

To load only a part of the columns referenced by a table as attributes, use the `include_properties` and `exclude_properties` arguments:

    {python}
    mapper(User, users_table, include_properties=['user_id', 'user_name'])
    
    mapper(Address, addresses_table, exclude_properties=['street', 'city', 'state', 'zip'])
    
To change the name of the attribute mapped to a particular column, place the `Column` object in the `properties` dictionary with the desired key:

    {python}
    mapper(User, users_table, properties={ 
       'id' : users_table.c.user_id,
       'name' : users_table.c.user_name,
    })

To change the names of all attributes using a prefix, use the `column_prefix` option.  This is useful for classes which wish to add their own `property` accessors:

    {python}
    mapper(User, users_table, column_prefix='_')
    
The above will place attribute names such as `_user_id`, `_user_name`, `_password` etc. on the mapped `User` class.
    
To place multiple columns which are known to be "synonymous" based on foreign key relationship or join condition into the same mapped attribute, put  them together using a list, as below where we map to a `Join`:

    {python}
    # join users and addresses
    usersaddresses = sql.join(users_table, addresses_table, \
        users_table.c.user_id == addresses_table.c.user_id)
        
    mapper(User, usersaddresses, 
       properties = {
               'id':[users_table.c.user_id, addresses_table.c.user_id],
          })

#### Deferred Column Loading {@name=deferred}

This feature allows particular columns of a table to not be loaded by default, instead being loaded later on when first referenced.  It is essentailly "column-level lazy loading".   This feature is useful when one wants to avoid loading a large text or binary field into memory when its not needed.  Individual columns can be lazy loaded by themselves or placed into groups that lazy-load together.

    {python}
    book_excerpts = Table('books', db, 
        Column('book_id', Integer, primary_key=True),
        Column('title', String(200), nullable=False),
        Column('summary', String(2000)),
        Column('excerpt', String),
        Column('photo', Binary)
      )

    class Book(object):
        pass

    # define a mapper that will load each of 'excerpt' and 'photo' in 
    # separate, individual-row SELECT statements when each attribute
    # is first referenced on the individual object instance
    mapper(Book, book_excerpts, properties = {
      'excerpt' : deferred(book_excerpts.c.excerpt),
      'photo' : deferred(book_excerpts.c.photo)
    })

Deferred columns can be placed into groups so that they load together:

    {python}
    book_excerpts = Table('books', db, 
      Column('book_id', Integer, primary_key=True),
      Column('title', String(200), nullable=False),
      Column('summary', String(2000)),
      Column('excerpt', String),
      Column('photo1', Binary),
      Column('photo2', Binary),
      Column('photo3', Binary)
    )

    class Book(object):
      pass

    # define a mapper with a 'photos' deferred group.  when one photo is referenced,
    # all three photos will be loaded in one SELECT statement.  The 'excerpt' will 
    # be loaded separately when it is first referenced.
    mapper(Book, book_excerpts, properties = {
      'excerpt' : deferred(book_excerpts.c.excerpt),
      'photo1' : deferred(book_excerpts.c.photo1, group='photos'),
      'photo2' : deferred(book_excerpts.c.photo2, group='photos'),
      'photo3' : deferred(book_excerpts.c.photo3, group='photos')
    })

You can defer or undefer columns at the `Query` level with the `options` method:

    {python}
    query = session.query(Book)
    query.options(defer('summary')).all()
    query.options(undefer('excerpt')).all()

#### SQL Expressions as Mapped Attributes {@name=expressions}

To add a SQL clause composed of local or external columns as a read-only, mapped column attribute, use the `column_property()` function.  Any scalar-returning `ClauseElement` may be used, as long as it has a `name` attribute; usually, you'll want to call `label()` to give it a specific name:

    {python}
    mapper(User, users, properties={
        'fullname' : column_property(
            (users_table.c.firstname + " " + users_table.c.lastname).label('fullname')
        )
    })

Correlated subqueries may be used as well:
    
    {python}
    mapper(User, users, properties={
        'address_count' : column_property(
                select(
                    [func.count(addresses_table.c.address_id)], 
                    addresses_table.c.user_id==users_table.c.user_id
                ).label('address_count')
            )
    })
    
### Overriding Attribute Behavior {@name=overriding}

A common request is the ability to create custom class properties that override the behavior of setting/getting an attribute.  You accomplish this using normal Python `property` constructs:

    {python}
    class MyAddress(object):
       def _set_email(self, email):
          self._email = email
       def _get_email(self):
          return self._email
       email = property(_get_email, _set_email)
    
    mapper(MyAddress, addresses_table, properties = {
      # map the '_email' attribute to the "email" column
      # on the table
      '_email': addresses_table.c.email
    })

To have your custom `email` property be recognized by keyword-based `Query` functions such as `filter_by()`, place a `synonym` on your mapper:

    {python}
    mapper(MyAddress, addresses_table, properties = {
      '_email': addresses_table.c.email
      
      'email':synonym('_email')
    })
    
    # use the synonym in a query
    result = session.query(MyAddress).filter_by(email='john@smith.com')

Synonym strategies such as the above can be easily automated, such as this example which specifies all columns and synonyms explicitly:

    {python}
    mapper(MyAddress, addresses_table, properties = dict(
        [('_'+col.key, col) for col in addresses_table.c] +
        [(col.key, synonym('_'+col.key)) for col in addresses_table.c]
    ))

The `column_prefix` option can also help with the above scenario by setting up the columns automatically with a prefix:

    {python}
    mapper(MyAddress, addresses_table, column_prefix='_', properties = dict(
        [(col.key, synonym('_'+col.key)) for col in addresses_table.c]
    ))


### Alternate Collection Implementations {@name=collections}

Mapping a one-to-many or many-to-many relationship results in a collection of values accessible through an attribute on the parent instance.  By default, this collection is a `list`:

    {python}
    mapper(Parent, properties={
        children = relation(Child)
    })

    parent = Parent()
    parent.children.append(Child())
    print parent.children[0]

Collections are not limited to lists.  Sets, mutable sequences and almost any other Python object that can act as a container can be used in place of the default list.

    {python}
    # use a set
    mapper(Parent, properties={
        children = relation(Child, collection_class=set)
    })

    parent = Parent()
    child = Child()
    parent.children.add(child)
    assert child in parent.children

#### Custom Collection Implementations {@name=custom}

You can use your own types for collections as well.  For most cases, simply inherit from `list` or `set` and add the custom behavior.

Collections in SQLAlchemy are transparently *instrumented*.  Instrumentation means that normal operations on the collection are tracked and result in changes being written to the database at flush time.  Additionally, collection operations can fire *events* which indicate some secondary operation must take place.  Examples of a secondary operation include saving the child item in the parent's `Session` (i.e. the `save-update` cascade), as well as synchronizing the state of a bi-directional relationship (i.e. a `backref`).

The collections package understands the basic interface of lists, sets and dicts and will automatically apply instrumentation to those built-in types and their subclasses.  Object-derived types that implement a basic collection interface are detected and instrumented via duck-typing:

    {python}
    class ListLike(object):
        def __init__(self):
            self.data = []
        def append(self, item):
            self.data.append(item)
        def remove(self, item):
            self.data.remove(item)
        def extend(self, items):
            self.data.extend(items)
        def __iter__(self):
            return iter(self.data)
        def foo(self):
            return 'foo'

`append`, `remove`, and `extend` are known list-like methods, and will be instrumented automatically.  `__iter__` is not a mutator method and won't be instrumented, and `foo` won't be either.

Duck-typing (i.e. guesswork) isn't rock-solid, of course, so you can be explicit about the interface you are implementing by providing an `__emulates__` class attribute:

    {python}
    class SetLike(object):
        __emulates__ = set

        def __init__(self):
            self.data = set()
        def append(self, item):
            self.data.add(item)
        def remove(self, item):
            self.data.remove(item)
        def __iter__(self):
            return iter(self.data)

This class looks list-like because of `append`, but `__emulates__` forces it to set-like.  `remove` is known to be part of the set interface and will be instrumented.

But this class won't work quite yet: a little glue is needed to adapt it for use by SQLAlchemy.  The ORM needs to know which methods to use to append, remove and iterate over members of the collection.  When using a type like `list` or `set`, the appropriate methods are well-known and used automatically when present. This set-like class does not provide the expected `add` method, so we must supply an explicit mapping for the ORM via a decorator.

#### Annotating Custom Collections via Decorators {@name=decorators}

Decorators can be used to tag the individual methods the ORM needs to manage collections.  Use them when your class doesn't quite meet the regular interface for its container type, or you simply would like to use a different method to get the job done.

    {python}
    from sqlalchemy.orm.collections import collection

    class SetLike(object):
        __emulates__ = set

        def __init__(self):
            self.data = set()

        @collection.appender
        def append(self, item):
            self.data.add(item)

        def remove(self, item):
            self.data.remove(item)

        def __iter__(self):
            return iter(self.data)

And that's all that's needed to complete the example.  SQLAlchemy will add instances via the `append` method.  `remove` and `__iter__` are the default methods for sets and will be used for removing and iteration.  Default methods can be changed as well:

    {python}
    from sqlalchemy.orm.collections import collection

    class MyList(list):
        @collection.remover
        def zark(self, item):
            # do something special...

        @collection.iterator
        def hey_use_this_instead_for_iteration(self):
            # ...

There is no requirement to be list-, or set-like at all.  Collection classes can be any shape, so long as they have the append, remove and iterate interface marked for SQLAlchemy's use.  Append and remove methods will be called with a mapped entity as the single argument, and iterator methods are called with no arguments and must return an iterator.

#### Dictionary-Based Collections {@name=dictcollections}

A `dict` can be used as a collection, but a keying strategy is needed to map entities loaded by the ORM to key, value pairs.  The [collections](rel:docstrings_sqlalchemy.orm.collections) package provides several built-in types for dictionary-based collections:

    {python}
    from sqlalchemy.orm.collections import column_mapped_collection, attr_mapped_collection, mapped_collection

    mapper(Item, items_table, properties={
        # key by column
        notes = relation(Note, collection_class=column_mapped_collection(notes_table.c.keyword))
        # or named attribute 
        notes2 = relation(Note, collection_class=attr_mapped_collection('keyword'))
        # or any callable
        notes3 = relation(Note, collection_class=mapped_collection(lambda entity: entity.a + entity.b))
    })

    # ...
    item = Item()
    item.notes['color'] = Note('color', 'blue')
    print item.notes['color']

These functions each provide a `dict` subclass with decorated `set` and `remove` methods and the keying strategy of your choice.

The [collections.MappedCollection](rel:docstrings_sqlalchemy.orm.collections.MappedCollection) class can be used as a base class for your custom types or as a mix-in to quickly add `dict` collection support to other classes.  It uses a keying function to delegate to `__setitem__` and `__delitem__`:

    {python}
    from sqlalchemy.util import OrderedDict
    from sqlalchemy.orm.collections import MappedCollection

    class NodeMap(OrderedDict, MappedCollection):
        """Holds 'Node' objects, keyed by the 'name' attribute with insert order maintained."""

        def __init__(self, *args, **kw):
            MappedCollection.__init__(self, keyfunc=lambda node: node.name)
            OrderedDict.__init__(self, *args, **kw)

The ORM understands the `dict` interface just like lists and sets, and will automatically instrument all dict-like methods if you choose to subclass `dict` or provide dict-like collection behavior in a duck-typed class.  You must decorate appender and remover methods, however- there are no compatible methods in the basic dictionary interface for SQLAlchemy to use by default.  Iteration will go through `itervalues()` unless otherwise decorated.

#### Instrumentation and Custom Types {@name=adv_collections}

Many custom types and existing library classes can be used as a entity collection type as-is without further ado.  However, it is important to note that the instrumentation process _will_ modify the type, adding decorators around methods automatically.

The decorations are lightweight and no-op outside of relations, but they do add unneeded overhead when triggered elsewhere.  When using a library class as a collection, it can be good practice to use the "trivial subclass" trick to restrict the decorations to just your usage in relations.  For example:

    {python}
    class MyAwesomeList(some.great.library.AwesomeList):
        pass

    # ... relation(..., collection_class=MyAwesomeList)

The ORM uses this approach for built-ins, quietly substituting a trivial subclass when a `list`, `set` or `dict` is used directly.

The collections package provides additional decorators and support for authoring custom types.  See the [package documentation](rel:docstrings_sqlalchemy.orm.collections) for more information and discussion of advanced usage and Python 2.3-compatible decoration options.

### Specifying Alternate Join Conditions to relation() {@name=customjoin}

The `relation()` function uses the foreign key relationship between the parent and child tables to formulate the **primary join condition** between parent and child; in the case of a many-to-many relationship it also formulates the **secondary join condition**.  If you are working with a `Table` which has no `ForeignKey` objects on it (which can be the case when using reflected tables with MySQL), or if the join condition cannot be expressed by a simple foreign key relationship, use the `primaryjoin` and possibly `secondaryjoin` conditions to create the appropriate relationship.

In this example we create a relation `boston_addresses` which will only load the user addresses with a city of "Boston":

    {python}
    class User(object):
        pass
    class Address(object):
        pass
    
    mapper(Address, addresses_table)
    mapper(User, users_table, properties={
        'boston_addresses' : relation(Address, primaryjoin=
                    and_(users_table.c.user_id==Address.c.user_id, 
                    Addresses.c.city=='Boston'))
    })
        
Many to many relationships can be customized by one or both of `primaryjoin` and `secondaryjoin`, shown below with just the default many-to-many relationship explicitly set:

    {python}
    class User(object):
        pass
    class Keyword(object):
        pass
    mapper(Keyword, keywords_table)
    mapper(User, users_table, properties={
        'keywords':relation(Keyword, secondary=userkeywords_table,
            primaryjoin=users_table.c.user_id==userkeywords_table.c.user_id,
            secondaryjoin=userkeywords_table.c.keyword_id==keywords_table.c.keyword_id
            )
    })

Very ambitious custom join conditions may fail to be directly persistable, and in some cases may not even load correctly.  To remove the persistence part of the equation, use the flag `viewonly=True` on the `relation()`, which establishes it as a read-only attribute (data written to the collection will be ignored on flush()).  However, in extreme cases, consider using a regular Python property in conjunction with `Query` as follows:

    class User(object):
        def _get_addresses(self):
            return object_session(self).query(Address).with_parent(self).filter(...).all()
        addresses = property(_get_addresses)

#### Multiple Relations against the Same Parent/Child {@name=multiplejoin}

Theres no restriction on how many times you can relate from parent to child.  SQLAlchemy can usually figure out what you want, particularly if the join conditions are straightforward.  Below we add a `newyork_addresses` attribute to complement the `boston_addresses` attribute:

    {python}
    mapper(User, users_table, properties={
        'boston_addresses' : relation(Address, primaryjoin=
                    and_(users_table.c.user_id==Address.c.user_id, 
                    Addresses.c.city=='Boston')),
        'newyork_addresses' : relation(Address, primaryjoin=
                    and_(users_table.c.user_id==Address.c.user_id, 
                    Addresses.c.city=='New York')),
    })


### Working with Large Collections {@name=largecollections}

The default behavior of `relation()` is to fully load the collection of items in, as according to the loading strategy of the relation.  Additionally, the Session by default only knows how to delete objects which are actually present within the session.  When a parent instance is marked for deletion and flushed, the Session loads its full list of child items in so that they may either be deleted as well, or have their foreign key value set to null; this is to avoid constraint violations.  For large collections of child items, there are several strategies to bypass full loading of child items both at load time as well as deletion time.

#### Dynamic Relation Loaders {@name=dynamic}

The most useful by far is the `dynamic_loader()` relation.  This is a variant of `relation()` which returns a `Query` object in place of a collection when accessed.  `filter()` criterion may be applied as well as limits and offsets, either explicitly or via array slices:

    {python}
    mapper(User, users_table, properties={
        'posts':dynamic_loader(Post)
    })
    
    jack = session.query(User).get(id)
    
    # filter Jack's blog posts
    posts = jack.posts.filter(Post.c.headline=='this is a post')
    
    # apply array slices
    posts = jack.posts[5:20]
    
The dynamic relation supports limited write operations, via the `append()` and `remove()` methods.  Since the read side of the dynamic relation always queries the database, changes to the underlying collection will not be visible until the data has been flushed:

    {python}
    oldpost = jack.posts.filter(Post.c.headline=='old post').one()
    jack.posts.remove(oldpost)
    
    jack.posts.append(Post('new post'))
    
To place a dynamic relation on a backref, use `lazy='dynamic'`:

    {python}
    mapper(Post, posts_table, properties={
        'user':relation(User, backref=backref('posts', lazy='dynamic'))
    })
    
Note that eager/lazy loading options cannot be used in conjunction dynamic relations at this time.

#### Setting Noload {@name=noload}

The opposite of the dynamic relation is simply "noload", specified using `lazy=None`:

    {python}
    mapper(MyClass, table, properties=relation{
        'children':relation(MyOtherClass, lazy=None)
    })

Above, the `children` collection is fully writeable, and changes to it will be persisted to the database as well as locally available for reading at the time they are added.  However when instances of  `MyClass` are freshly loaded from the database, the `children` collection stays empty.

#### Using Passive Deletes {@name=passivedelete}

Use `passive_deletes=True` to disable child object loading on a DELETE operation, in conjunction with "ON DELETE (CASCADE|SET NULL)" on your database to automatically cascade deletes to child objects.   Note that "ON DELETE" is not supported on SQLite, and requires `InnoDB` tables when using MySQL:

        {python}
        mytable = Table('mytable', meta,
            Column('id', Integer, primary_key=True),
            )

        myothertable = Table('myothertable', meta,
            Column('id', Integer, primary_key=True),
            Column('parent_id', Integer),
            ForeignKeyConstraint(['parent_id'],['mytable.id'], ondelete="CASCADE"),
            )

        mmapper(MyOtherClass, myothertable)
        
        mapper(MyClass, mytable, properties={
            'children':relation(MyOtherClass, cascade="all, delete-orphan", passive_deletes=True)
        })

When `passive_deletes` is applied, the `children` relation will not be loaded into memory when an instance of `MyClass` is marked for deletion.  The `cascade="all, delete-orphan"` *will* take effect for instances of `MyOtherClass` which are currently present in the session; however for instances of `MyOtherClass` which are not loaded, SQLAlchemy assumes that "ON DELETE CASCADE" rules will ensure that those rows are deleted by the database and that no foreign key violation will occur.
        
### Controlling Ordering {@name=orderby}

By default, mappers will attempt to ORDER BY the "oid" column of a table, or the first primary key column, when selecting rows.  This can be modified in several ways.

The "order_by" parameter can be sent to a mapper, overriding the per-engine ordering if any.  A value of None means that the mapper should not use any ordering.  A non-None value, which can be a column, an `asc` or `desc` clause, or an array of either one, indicates the ORDER BY clause that should be added to all select queries:

    {python}
    # disable all ordering
    mapper(User, users_table, order_by=None)

    # order by a column
    mapper(User, users_table, order_by=users_tableusers_table.c.user_id)
    
    # order by multiple items
    mapper(User, users_table, order_by=[users_table.c.user_id, desc(users_table.c.user_name)])

"order_by" can also be specified with queries, overriding all other per-engine/per-mapper orderings:

    {python}
    # order by a column
    l = query.filter(User.user_name=='fred').order_by(User.user_id).all()
    
    # order by multiple criterion
    l = query.filter(User.user_name=='fred').order_by([User.user_id, desc(User.user_name)])

The "order_by" property can also be specified on a `relation()` which will control the ordering of the collection:

    {python}
    mapper(Address, addresses_table)
    
    # order address objects by address id
    mapper(User, users_table, properties = {
        'addresses' : relation(Address, order_by=addresses_table.c.address_id)
    })
    
Note that when using eager loaders with relations, the tables used by the eager load's join are anonymously aliased.  You can only order by these columns if you specify it at the `relation()` level.  To control ordering at the query level based on a related table, you `join()` to that relation, then order by it:

    {python}
    session.query(User).join('addresses').order_by(Address.street)

### Mapping Class Inheritance Hierarchies {@name=inheritance}

SQLAlchemy supports three forms of inheritance:  *single table inheritance*, where several types of classes are stored in one table, *concrete table inheritance*, where each type of class is stored in its own table, and *joined table inheritance*, where the parent/child classes are stored in their own tables that are joined together in a select.  Whereas support for single and joined table inheritance is strong, concrete table inheritance is a less common scenario with some particular problems so is not quite as flexible.

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.

For the following sections, assume this class relationship:

    {python}
    class Employee(object):
        def __init__(self, name):
            self.name = name
        def __repr__(self):
            return self.__class__.__name__ + " " + self.name

    class Manager(Employee):
        def __init__(self, name, manager_data):
            self.name = name
            self.manager_data = manager_data
        def __repr__(self):
            return self.__class__.__name__ + " " + self.name + " " +  self.manager_data

    class Engineer(Employee):
        def __init__(self, name, engineer_info):
            self.name = name
            self.engineer_info = engineer_info
        def __repr__(self):
            return self.__class__.__name__ + " " + self.name + " " +  self.engineer_info

#### Joined Table Inheritance {@name=joined}

In joined table inheritance, each class along a particular classes' list of parents is represented by a unique table.  The total set of attributes for a particular instance is represented as a join along all tables in its inheritance path.  Here, we first define a table to represent the `Employee` class.  This table will contain a primary key column (or columns), and a column for each attribute that's represented by `Employee`.  In this case it's just `name`:

    {python}
    employees = Table('employees', metadata, 
       Column('employee_id', Integer, primary_key=True),
       Column('name', String(50)),
       Column('type', String(30), nullable=False)
    )

The table also has a column called `type`.  It is strongly advised in both single- and joined- table inheritance scenarios that the root table contains a column whose sole purpose is that of the **discriminator**; it stores a value which indicates the type of object represented within the row.  The column may be of any desired datatype.  While there are some "tricks" to work around the requirement that there be a discriminator column, they are more complicated to configure when one wishes to load polymorphically.

Next we define individual tables for each of `Engineer` and `Manager`, which each contain columns that represent the attributes unique to the subclass they represent.  Each table also must contain a primary key column (or columns), and in most cases a foreign key reference to the parent table.  It is  standard practice that the same column is used for both of these roles, and that the column is also named the same as that of the parent table.  However this is optional in SQLAlchemy; separate columns may be used for primary key and parent-relation, the column may be named differently than that of the parent, and even a custom join condition can be specified between parent and child tables instead of using a foreign key.  In joined table inheritance, the primary key of an instance is always represented by the primary key of the base table only (new in SQLAlchemy 0.4).

    {python}
    engineers = Table('engineers', metadata, 
       Column('employee_id', Integer, ForeignKey('employees.employee_id'), primary_key=True),
       Column('engineer_info', String(50)),
    )

    managers = Table('managers', metadata, 
       Column('employee_id', Integer, ForeignKey('employees.employee_id'), primary_key=True),
       Column('manager_data', String(50)),
    )

We then configure mappers as usual, except we use some additional arguments to indicate the inheritance relationship, the polymorphic discriminator column, and the **polymorphic identity** of each class; this is the value that will be stored in the polymorphic discriminator column.

    {python}
    mapper(Employee, employees, polymorphic_on=employees.c.type, polymorphic_identity='employee')
    mapper(Engineer, engineers, inherits=Employee, polymorphic_identity='engineer')
    mapper(Manager, managers, inherits=Employee, polymorphic_identity='manager')

And that's it.  Querying against `Employee` will return a combination of `Employee`, `Engineer` and `Manager` objects.

##### Optimizing Joined Table Loads {@name=optimizing}

When loading fresh from the database, the joined-table setup above will query from the parent table first, then for each row will issue a second query to the child table.  For example, for a load of five rows with `Employee` id 3, `Manager` ids 1 and 5 and `Engineer` ids 2 and 4, will produce queries along the lines of this example:

    {python}
    session.query(Employee).all()
    {opensql}
    SELECT employees.employee_id AS employees_employee_id, employees.name AS employees_name, employees.type AS employees_type 
    FROM employees ORDER BY employees.oid
    []
    SELECT managers.employee_id AS managers_employee_id, managers.manager_data AS managers_manager_data 
    FROM managers 
    WHERE ? = managers.employee_id
    [5]
    SELECT engineers.employee_id AS engineers_employee_id, engineers.engineer_info AS engineers_engineer_info 
    FROM engineers 
    WHERE ? = engineers.employee_id
    [2]
    SELECT engineers.employee_id AS engineers_employee_id, engineers.engineer_info AS engineers_engineer_info 
    FROM engineers 
    WHERE ? = engineers.employee_id
    [4]
    SELECT managers.employee_id AS managers_employee_id, managers.manager_data AS managers_manager_data 
    FROM managers 
    WHERE ? = managers.employee_id
    [1]

The above query works well for a `get()` operation, since it limits the queries to only the tables directly involved in fetching a single instance.  For instances which are already present in the session, the secondary table load is not needed.  However, the above loading style is not efficient for loading large groups of objects, as it incurs separate queries for each parent row.

One way to reduce the number of "secondary" loads of child rows is to "defer" them, using `polymorphic_fetch='deferred'`:

    {python}
    mapper(Employee, employees, polymorphic_on=employees.c.type, \
        polymorphic_identity='employee', polymorphic_fetch='deferred')
    mapper(Engineer, engineers, inherits=Employee, polymorphic_identity='engineer')
    mapper(Manager, managers, inherits=Employee, polymorphic_identity='manager')

The above configuration queries in the same manner as earlier, except the load of each "secondary" table occurs only when attributes referencing those columns are first referenced on the loaded instance.  This style of loading is very efficient for cases where large selects of items occur, but a detailed "drill down" of extra inherited properties is less common.

More commonly, an all-at-once load may be achieved by constructing a query which combines all three tables together, and adding it to the mapper configuration as its `select_table`, which is an arbitrary selectable which the mapper will use for load operations (it has no impact on save operations).  Any selectableq can be used for this, such as a UNION of tables.  For joined table inheritance, the easiest method is to use OUTER JOIN:

    {python}
    join = employees.outerjoin(engineers).outerjoin(managers)

    mapper(Employee, employees, polymorphic_on=employees.c.type, \
        polymorphic_identity='employee', select_table=join)
    mapper(Engineer, engineers, inherits=Employee, polymorphic_identity='engineer')
    mapper(Manager, managers, inherits=Employee, polymorphic_identity='manager')

Which produces a query like the following:

    {python}
    session.query(Employee).all()
    {opensql}
    SELECT employees.employee_id AS employees_employee_id, engineers.employee_id AS engineers_employee_id, managers.employee_id AS managers_employee_id, employees.name AS employees_name, employees.type AS employees_type, engineers.engineer_info AS engineers_engineer_info, managers.manager_data AS managers_manager_data 
    FROM employees LEFT OUTER JOIN engineers ON employees.employee_id = engineers.employee_id LEFT OUTER JOIN managers ON employees.employee_id = managers.employee_id ORDER BY employees.oid
    []    

#### Single Table Inheritance

Single table inheritance is where the attributes of the base class as well as all subclasses are represented within a single table.  A column is present in the table for every attribute mapped to the base class and all subclasses; the columns which correspond to a single subclass are nullable.  This configuration looks much like joined-table inheritance except there's only one table.  In this case, a `type` column is required, as there would be no other way to discriminate between classes.  The table is specified in the base mapper only; for the inheriting classes, leave their `table` parameter blank:

    {python}
    employees_table = Table('employees', metadata, 
        Column('employee_id', Integer, primary_key=True),
        Column('name', String(50)),
        Column('manager_data', String(50)),
        Column('engineer_info', String(50)),
        Column('type', String(20), nullable=False)
    )
    
    employee_mapper = mapper(Employee, employees_table, \
        polymorphic_on=employees_table.c.type, polymorphic_identity='employee')
    manager_mapper = mapper(Manager, inherits=employee_mapper, polymorphic_identity='manager')
    engineer_mapper = mapper(Engineer, inherits=employee_mapper, polymorphic_identity='engineer')

#### Concrete Table Inheritance

This form of inheritance maps each class to a distinct table, as below:

    {python}
    employees_table = Table('employees', metadata, 
        Column('employee_id', Integer, primary_key=True),
        Column('name', String(50)),
    )

    managers_table = Table('managers', metadata, 
        Column('employee_id', Integer, primary_key=True),
        Column('name', String(50)),
        Column('manager_data', String(50)),
    )

    engineers_table = Table('engineers', metadata, 
        Column('employee_id', Integer, primary_key=True),
        Column('name', String(50)),
        Column('engineer_info', String(50)),
    )

Notice in this case there is no `type` column.  If polymorphic loading is not required, theres no advantage to using `inherits` here; you just define a separate mapper for each class.

    {python}
    mapper(Employee, employees_table)
    mapper(Manager, managers_table)
    mapper(Engineer, engineers_table)

To load polymorphically, the `select_table` argument is currently required.  In this case we must construct a UNION of all three tables.  SQLAlchemy includes a helper function to create these called `polymorphic_union`, which will map all the different columns into a structure of selects with the same numbers and names of columns, and also generate a virtual `type` column for each subselect:

    {python}
    pjoin = polymorphic_union({
        'employee':employees_table,
        'manager':managers_table,
        'engineer':engineers_table
    }, 'type', 'pjoin')

    employee_mapper = mapper(Employee, employees_table, select_table=pjoin, \
        polymorphic_on=pjoin.c.type, polymorphic_identity='employee')
    manager_mapper = mapper(Manager, managers_table, inherits=employee_mapper, \
        concrete=True, polymorphic_identity='manager')
    engineer_mapper = mapper(Engineer, engineers_table, inherits=employee_mapper, \
        concrete=True, polymorphic_identity='engineer')

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

    {python}
    session.query(Employee).all()
    {opensql}
    SELECT pjoin.type AS pjoin_type, pjoin.manager_data AS pjoin_manager_data, pjoin.employee_id AS pjoin_employee_id, 
    pjoin.name AS pjoin_name, pjoin.engineer_info AS pjoin_engineer_info 
    FROM (
        SELECT employees.employee_id AS employee_id, CAST(NULL AS VARCHAR(50)) AS manager_data, employees.name AS name, 
        CAST(NULL AS VARCHAR(50)) AS engineer_info, 'employee' AS type 
        FROM employees 
    UNION ALL 
        SELECT managers.employee_id AS employee_id, managers.manager_data AS manager_data, managers.name AS name, 
        CAST(NULL AS VARCHAR(50)) AS engineer_info, 'manager' AS type 
        FROM managers 
    UNION ALL 
        SELECT engineers.employee_id AS employee_id, CAST(NULL AS VARCHAR(50)) AS manager_data, engineers.name AS name, 
        engineers.engineer_info AS engineer_info, 'engineer' AS type 
        FROM engineers
    ) AS pjoin ORDER BY pjoin.oid
    []

#### Using Relations with Inheritance {@name=relations}

Both joined-table and single table inheritance scenarios produce mappings which are usable in relation() functions; that is, it's possible to map a parent object to a child object which is polymorphic.  Similiarly, inheriting mappers can have `relation()`s of their own at any level, which are inherited to each child class.  The only requirement for relations is that there is a table relationship between parent and child.  An example is the following modification to the joined table inheritance example, which sets a bi-directional relationship between `Employee` and `Company`:

    {python}
    employees_table = Table('employees', metadata, 
        Column('employee_id', Integer, primary_key=True),
        Column('name', String(50)),
        Column('company_id', Integer, ForeignKey('companies.company_id'))
    )

    companies = Table('companies', metadata, 
       Column('company_id', Integer, primary_key=True),
       Column('name', String(50)))

    class Company(object):
        pass
    
    mapper(Company, companies, properties={
       'employees': relation(Employee, backref='company')
    })
       
SQLAlchemy has a lot of experience in this area; the optimized "outer join" approach can be used freely for parent and child relationships, eager loads are fully useable, query aliasing and other tricks are fully supported as well.

In a concrete inheritance scenario, mapping `relation()`s is more difficult since the distinct classes do not share a table.  In this case, you *can* establish a relationship from parent to child if a join condition can be constructed from parent to child, if each child table contains a foreign key to the parent:

    {python}
    companies = Table('companies', metadata, 
       Column('id', Integer, primary_key=True),
       Column('name', String(50)))

    employees_table = Table('employees', metadata, 
        Column('employee_id', Integer, primary_key=True),
        Column('name', String(50)),
        Column('company_id', Integer, ForeignKey('companies.id'))
    )

    managers_table = Table('managers', metadata, 
        Column('employee_id', Integer, primary_key=True),
        Column('name', String(50)),
        Column('manager_data', String(50)),
        Column('company_id', Integer, ForeignKey('companies.id'))
    )

    engineers_table = Table('engineers', metadata, 
        Column('employee_id', Integer, primary_key=True),
        Column('name', String(50)),
        Column('engineer_info', String(50)),
        Column('company_id', Integer, ForeignKey('companies.id'))
    )

    mapper(Employee, employees_table, select_table=pjoin, polymorphic_on=pjoin.c.type, polymorphic_identity='employee')
    mapper(Manager, managers_table, inherits=employee_mapper, concrete=True, polymorphic_identity='manager')
    mapper(Engineer, engineers_table, inherits=employee_mapper, concrete=True, polymorphic_identity='engineer')
    mapper(Company, companies, properties={
        'employees':relation(Employee)
    })

Let's crank it up and try loading with an eager load:

    {python}
    session.query(Company).options(eagerload('employees')).all()
    {opensql}
    SELECT anon_1.type AS anon_1_type, anon_1.manager_data AS anon_1_manager_data, anon_1.engineer_info AS anon_1_engineer_info, 
    anon_1.employee_id AS anon_1_employee_id, anon_1.name AS anon_1_name, anon_1.company_id AS anon_1_company_id, 
    companies.id AS companies_id, companies.name AS companies_name 
    FROM companies LEFT OUTER JOIN (SELECT CAST(NULL AS VARCHAR(50)) AS engineer_info, employees.employee_id AS employee_id, 
    CAST(NULL AS VARCHAR(50)) AS manager_data, employees.name AS name, employees.company_id AS company_id, 'employee' AS type 
    FROM employees UNION ALL SELECT CAST(NULL AS VARCHAR(50)) AS engineer_info, managers.employee_id AS employee_id, 
    managers.manager_data AS manager_data, managers.name AS name, managers.company_id AS company_id, 'manager' AS type 
    FROM managers UNION ALL SELECT engineers.engineer_info AS engineer_info, engineers.employee_id AS employee_id, 
    CAST(NULL AS VARCHAR(50)) AS manager_data, engineers.name AS name, engineers.company_id AS company_id, 'engineer' AS type 
    FROM engineers) AS anon_1 ON companies.id = anon_1.company_id ORDER BY companies.oid, anon_1.oid
    []

The big limitation with concrete table inheritance is that relation()s placed on each concrete mapper do **not** propagate to child mappers.  If you want to have the same relation()s set up on all concrete mappers, they must be configured manually on each.

### Mapping a Class against Multiple Tables {@name=joins}

Mappers can be constructed against arbitrary relational units (called `Selectables`) as well as plain `Tables`.  For example, The `join` keyword from the SQL package creates a neat selectable unit comprised of multiple tables, complete with its own composite primary key, which can be passed in to a mapper as the table.

    {python}
    # a class
    class AddressUser(object):
        pass

    # define a Join
    j = join(users_table, addresses_table)
    
    # map to it - the identity of an AddressUser object will be 
    # based on (user_id, address_id) since those are the primary keys involved
    m = mapper(AddressUser, j, properties={
        'user_id':[users_table.c.user_id, addresses_table.c.user_id]
    })

A second example:

    {python}
    # many-to-many join on an association table
    j = join(users_table, userkeywords, 
            users_table.c.user_id==userkeywords.c.user_id).join(keywords, 
               userkeywords.c.keyword_id==keywords.c.keyword_id)
     
    # a class 
    class KeywordUser(object):
        pass

    # map to it - the identity of a KeywordUser object will be
    # (user_id, keyword_id) since those are the primary keys involved
    m = mapper(KeywordUser, j, properties={
        'user_id':[users_table.c.user_id, userkeywords.c.user_id],
        'keyword_id':[userkeywords.c.keyword_id, keywords.c.keyword_id]
    })

In both examples above, "composite" columns were added as properties to the mappers; these are aggregations of multiple columns into one mapper property, which instructs the mapper to keep both of those columns set at the same value.

### Mapping a Class against Arbitrary Selects {@name=selects}

Similar to mapping against a join, a plain select() object can be used with a mapper as well.  Below, an example select which contains two aggregate functions and a group_by is mapped to a class:

    {python}
    s = select([customers, 
                func.count(orders).label('order_count'), 
                func.max(orders.price).label('highest_order')],
                customers.c.customer_id==orders.c.customer_id,
                group_by=[c for c in customers.c]
                ).alias('somealias')
    class Customer(object):
        pass
    
    m = mapper(Customer, s)
    
Above, the "customers" table is joined against the "orders" table to produce a full row for each customer row, the total count of related rows in the "orders" table, and the highest price in the "orders" table, grouped against the full set of columns in the "customers" table.  That query is then mapped against the Customer class.  New instances of Customer will contain attributes for each column in the "customers" table as well as an "order_count" and "highest_order" attribute.  Updates to the Customer object will only be reflected in the "customers" table and not the "orders" table.  This is because the primary keys of the "orders" table are not represented in this mapper and therefore the table is not affected by save or delete operations.

### Multiple Mappers for One Class {@name=multiple}

The first mapper created for a certain class is known as that class's "primary mapper."  Other mappers can be created as well, these come in two varieties.

* **secondary mapper** - this is a mapper that must be constructed with the keyword argument `non_primary=True`, and represents a load-only mapper.  Objects that are loaded with a secondary mapper will have their save operation processed by the primary mapper.  It is also invalid to add new `relation()`s to a non-primary mapper. To use this mapper with the Session, specify it to the `query` method:

example:

    {python}
    # primary mapper
    mapper(User, users_table)
    
    # make a secondary mapper to load User against a join
    othermapper = mapper(User, users_table.join(someothertable), non_primary=True)
    
    # select
    result = session.query(othermapper).select()

* **entity name mapper** - this is a mapper that is a fully functioning primary mapper for a class, which is distinguished from the regular primary mapper by an `entity_name` parameter.  Instances loaded with this mapper will be totally managed by this new mapper and have no connection to the original one.  Most methods on `Session` include an optional `entity_name` parameter in order to specify this condition.

example:

    {python}
    # primary mapper
    mapper(User, users_table)
    
    # make an entity name mapper that stores User objects in another table
    mapper(User, alternate_users_table, entity_name='alt')
    
    # make two User objects
    user1 = User()
    user2 = User()
    
    # save one in in the "users" table
    session.save(user1)
    
    # save the other in the "alternate_users_table"
    session.save(user2, entity_name='alt')
    
    session.flush()
    
    # select from the alternate mapper
    session.query(User, entity_name='alt').select()

### Self Referential Mappers {@name=selfreferential}

A self-referential mapper is a mapper that is designed to operate with an *adjacency list* table.  This is a table that contains one or more foreign keys back to itself, and is usually used to create hierarchical tree structures.  SQLAlchemy's default model of saving items based on table dependencies is not sufficient in this case, as an adjacency list table introduces dependencies between individual rows.  Fortunately, SQLAlchemy will automatically detect a self-referential mapper and do the extra lifting to make it work.  

    {python}
    # define a self-referential table
    trees = Table('treenodes', engine,
        Column('node_id', Integer, primary_key=True),
        Column('parent_node_id', Integer, ForeignKey('treenodes.node_id'), nullable=True),
        Column('node_name', String(50), nullable=False),
        )

    # treenode class
    class TreeNode(object):
        pass

    # mapper defines "children" property, pointing back to TreeNode class,
    # with the mapper unspecified.  it will point back to the primary 
    # mapper on the TreeNode class.
    TreeNode.mapper = mapper(TreeNode, trees, properties={
            'children' : relation(
                            TreeNode, 
                            cascade="all"
                         ),
            }
        )
        
This kind of mapper goes through a lot of extra effort when saving and deleting items, to determine the correct dependency graph of nodes within the tree.
    
A self-referential mapper where there is more than one relationship on the table requires that all join conditions be explicitly spelled out.  Below is a self-referring table that contains a "parent_node_id" column to reference parent/child relationships, and a "root_node_id" column which points child nodes back to the ultimate root node:

    {python}
    # define a self-referential table with several relations
    trees = Table('treenodes', engine,
        Column('node_id', Integer, primary_key=True),
        Column('parent_node_id', Integer, ForeignKey('treenodes.node_id'), nullable=True),
        Column('root_node_id', Integer, ForeignKey('treenodes.node_id'), nullable=True),
        Column('node_name', String(50), nullable=False),
        )

    # treenode class
    class TreeNode(object):
        pass

    # define the "children" property as well as the "root" property
    mapper(TreeNode, trees, properties={
            'children' : relation(
                            TreeNode, 
                            primaryjoin=trees.c.parent_node_id == trees.c.node_id
                            cascade="all",
                            backref=backref("parent", remote_side=[trees.c.node_id])
                         ),
            'root' : relation(
                    TreeNode,
                    primaryjoin=trees.c.root_node_id == trees.c.node_id, 
                    remote_side=[trees.c.node_id],
                    uselist=False
                )
            }
        )
        
The "root" property on a TreeNode is a many-to-one relationship.  By default, a self-referential mapper declares relationships as one-to-many, so the extra parameter `remote_side`, pointing to a column or list of columns on the remote side of a relationship, is needed to indicate a "many-to-one" self-referring relationship (note the previous keyword argument `foreignkey` is deprecated).
Both TreeNode examples above are available in functional form in the `examples/adjacencytree` directory of the distribution.    

#### Self-Referential Query Strategies {@name=query}

todo

### Combining Eager Loads with Statement/Result Set Queries

When full statement/result loads are used with `Query`, SQLAlchemy does not affect the SQL query itself, and therefore has no way of tacking on its own `LEFT [OUTER] JOIN` conditions that are normally used to eager load relationships.  If the query being constructed is created in such a way that it returns rows not just from a parent table (or tables) but also returns rows from child tables, the result-set mapping can be notified as to which additional properties are contained within the result set.  This is done using the `contains_eager()` query option, which specifies the name of the relationship to be eagerly loaded.

    {python}
    # mapping is the users->addresses mapping
    mapper(User, users_table, properties={
        'addresses':relation(Address, addresses_table)
    })
    
    # define a query on USERS with an outer join to ADDRESSES
    statement = users_table.outerjoin(addresses_table).select(use_labels=True)
    
    # construct a Query object which expects the "addresses" results 
    query = session.query(User).options(contains_eager('addresses'))
    
    # get results normally
    r = query.from_statement(statement)

If the "eager" portion of the statement is "alisaed", the `alias` keyword argument to `contains_eager()` may be used to indicate it.  This is a string alias name or reference to an actual `Alias` object:

    {python}
    # use an alias of the addresses table
    adalias = addresses_table.alias('adalias')
    
    # define a query on USERS with an outer join to adalias
    statement = users_table.outerjoin(adalias).select(use_labels=True)

    # construct a Query object which expects the "addresses" results 
    query = session.query(User).options(contains_eager('addresses', alias=adalias))

    # get results normally
    {sql}r = query.from_statement(query).all()
    SELECT users.user_id AS users_user_id, users.user_name AS users_user_name, adalias.address_id AS adalias_address_id, 
    adalias.user_id AS adalias_user_id, adalias.email_address AS adalias_email_address, (...other columns...)
    FROM users LEFT OUTER JOIN email_addresses AS adalias ON users.user_id = adalias.user_id

In the case that the main table itself is also aliased, the `contains_alias()` option can be used:

    {python}
    # define an aliased UNION called 'ulist'
    statement = users.select(users.c.user_id==7).union(users.select(users.c.user_id>7)).alias('ulist')

    # add on an eager load of "addresses"
    statement = statement.outerjoin(addresses).select(use_labels=True)
    
    # create query, indicating "ulist" is an alias for the main table, "addresses" property should
    # be eager loaded
    query = create_session().query(User).options(contains_alias('ulist'), contains_eager('addresses'))
    
    # results
    r = query.from_statement(statement)

### Extending Mapper {@name=extending}

Mappers can have functionality augmented or replaced at many points in its execution via the usage of the MapperExtension class.  This class is just a series of "hooks" where various functionality takes place.  An application can make its own MapperExtension objects, overriding only the methods it needs.  Methods that are not overridden return the special value `sqlalchemy.orm.EXT_CONTINUE` to allow processing to continue to the next MapperExtension or simply proceed normally if there are no more extensions.

API documentation for MapperExtension: [docstrings_sqlalchemy.orm_MapperExtension](rel:docstrings_sqlalchemy.orm_MapperExtension)

To use MapperExtension, make your own subclass of it and just send it off to a mapper:

    {python}
    m = mapper(User, users_table, extension=MyExtension())

Multiple extensions will be chained together and processed in order; they are specified as a list:

    {python}
    m = mapper(User, users_table, extension=[ext1, ext2, ext3])
    
