SQLAlchemy 0.6 Documentation

Release: 0.6.9 | Release Date: May 5, 2012
SQLAlchemy 0.6 Documentation » SQLAlchemy ORM » Object Relational Tutorial

Object Relational Tutorial

Object Relational Tutorial

The SQLAlchemy Object Relational Mapper presents a method of associating user-defined Python classes with database tables, and instances of those classes (objects) with rows in their corresponding tables. It includes a system that transparently synchronizes all changes in state between objects and their related rows, called a unit of work, as well as a system for expressing database queries in terms of the user defined classes and their defined relationships between each other.

The ORM is in contrast to the SQLAlchemy Expression Language, upon which the ORM is constructed. Whereas the SQL Expression Language, introduced in SQL Expression Language Tutorial, presents a system of representing the primitive constructs of the relational database directly without opinion, the ORM presents a high level and abstracted pattern of usage, which itself is an example of applied usage of the Expression Language.

While there is overlap among the usage patterns of the ORM and the Expression Language, the similarities are more superficial than they may at first appear. One approaches the structure and content of data from the perspective of a user-defined domain model which is transparently persisted and refreshed from its underlying storage model. The other approaches it from the perspective of literal schema and SQL expression representations which are explicitly composed into messages consumed individually by the database.

A successful application may be constructed using the Object Relational Mapper exclusively. In advanced situations, an application constructed with the ORM may make occasional usage of the Expression Language directly in certain areas where specific database interactions are required.

The following tutorial is in doctest format, meaning each >>> line represents something you can type at a Python command prompt, and the following text represents the expected return value.

Version Check

A quick check to verify that we are on at least version 0.6 of SQLAlchemy:

>>> import sqlalchemy
>>> sqlalchemy.__version__ 


For this tutorial we will use an in-memory-only SQLite database. To connect we use create_engine():

>>> from sqlalchemy import create_engine
>>> engine = create_engine('sqlite:///:memory:', echo=True)

The echo flag is a shortcut to setting up SQLAlchemy logging, which is accomplished via Python’s standard logging module. With it enabled, we’ll see all the generated SQL produced. If you are working through this tutorial and want less output generated, set it to False. This tutorial will format the SQL behind a popup window so it doesn’t get in our way; just click the “SQL” links to see what’s being generated.

Define and Create a Table

Next we want to tell SQLAlchemy about our tables. We will start with just a single table called users, which will store records for the end-users using our application (lets assume it’s a website). We define our tables within a catalog called MetaData, using the Table construct, which is used in a manner similar to SQL’s CREATE TABLE syntax:

>>> from sqlalchemy import Table, Column, Integer, String, MetaData, ForeignKey
>>> metadata = MetaData()
>>> users_table = Table('users', metadata,
...     Column('id', Integer, primary_key=True),
...     Column('name', String),
...     Column('fullname', String),
...     Column('password', String)
... )

Schema Definition Language covers all about how to define Table objects, as well as how to load their definition from an existing database (known as reflection).

Next, we can issue CREATE TABLE statements derived from our table metadata, by calling create_all() and passing it the engine instance which points to our database. This will check for the presence of a table first before creating, so it’s safe to call multiple times:

sql>>> metadata.create_all(engine) 


Users familiar with the syntax of CREATE TABLE may notice that the VARCHAR columns were generated without a length; on SQLite and Postgresql, this is a valid datatype, but on others, it’s not allowed. So if running this tutorial on one of those databases, and you wish to use SQLAlchemy to issue CREATE TABLE, a “length” may be provided to the String type as below:

Column('name', String(50))

The length field on String, as well as similar precision/scale fields available on Integer, Numeric, etc. are not referenced by SQLAlchemy other than when creating tables.

Additionally, Firebird and Oracle require sequences to generate new primary key identifiers, and SQLAlchemy doesn’t generate or assume these without being instructed. For that, you use the Sequence construct:

from sqlalchemy import Sequence
Column('id', Integer, Sequence('user_id_seq'), primary_key=True)

A full, foolproof Table is therefore:

users_table = Table('users', metadata,
   Column('id', Integer, Sequence('user_id_seq'), primary_key=True),
   Column('name', String(50)),
   Column('fullname', String(50)),
   Column('password', String(12))

We include this more verbose Table construct separately to highlight the difference between a minimal construct geared primarily towards in-Python usage only, versus one that will be used to emit CREATE TABLE statements on a particular set of backends with more stringent requirements.

Define a Python Class to be Mapped

While the Table object defines information about our database, it does not say anything about the definition or behavior of the business objects used by our application; SQLAlchemy views this as a separate concern. To correspond to our users table, let’s create a rudimentary User class. It only need subclass Python’s built-in object class (i.e. it’s a new style class):

>>> class User(object):
...     def __init__(self, name, fullname, password):
...         self.name = name
...         self.fullname = fullname
...         self.password = password
...     def __repr__(self):
...        return "<User('%s','%s', '%s')>" % (self.name, self.fullname, self.password)

The class has an __init__() and a __repr__() method for convenience. These methods are both entirely optional, and can be of any form. SQLAlchemy never calls __init__() directly.

Setting up the Mapping

With our users_table and User class, we now want to map the two together. That’s where the SQLAlchemy ORM package comes in. We’ll use the mapper() function to create a mapping between users_table and User:

>>> from sqlalchemy.orm import mapper
>>> mapper(User, users_table) 
<Mapper at 0x...; User>

The mapper() function creates a new Mapper object and stores it away for future reference, associated with our class. Let’s now create and inspect a User object:

>>> ed_user = User('ed', 'Ed Jones', 'edspassword')
>>> ed_user.name
>>> ed_user.password
>>> str(ed_user.id)

The id attribute, which while not defined by our __init__() method, exists due to the id column present within the users_table object. By default, the mapper() creates class attributes for all columns present within the Table. These class attributes exist as Python descriptors, and define instrumentation for the mapped class. The functionality of this instrumentation is very rich and includes the ability to track modifications and automatically load new data from the database when needed.

Since we have not yet told SQLAlchemy to persist Ed Jones within the database, its id is None. When we persist the object later, this attribute will be populated with a newly generated value.

Creating Table, Class and Mapper All at Once Declaratively

The preceding approach to configuration involved a Table, a user-defined class, and a call to mapper(). This illustrates classical SQLAlchemy usage, which values the highest separation of concerns possible. A large number of applications don’t require this degree of separation, and for those SQLAlchemy offers an alternate “shorthand” configurational style called declarative. For many applications, this is the only style of configuration needed. Our above example using this style is as follows:

>>> from sqlalchemy.ext.declarative import declarative_base

>>> Base = declarative_base()
>>> class User(Base):
...     __tablename__ = 'users'
...     id = Column(Integer, primary_key=True)
...     name = Column(String)
...     fullname = Column(String)
...     password = Column(String)
...     def __init__(self, name, fullname, password):
...         self.name = name
...         self.fullname = fullname
...         self.password = password
...     def __repr__(self):
...        return "<User('%s','%s', '%s')>" % (self.name, self.fullname, self.password)

Above, the declarative_base() function defines a new class which we name Base, from which all of our ORM-enabled classes will derive. Note that we define Column objects with no “name” field, since it’s inferred from the given attribute name.

The underlying Table object created by our declarative_base() version of User is accessible via the __table__ attribute:

>>> users_table = User.__table__

The owning MetaData object is available as well:

>>> metadata = Base.metadata

declarative is covered at Declarative as well as throughout Mapper Configuration.

Yet another “declarative” method is available for SQLAlchemy as a third party library called Elixir. This is a full-featured configurational product which also includes many higher level mapping configurations built in. Like declarative, once classes and mappings are defined, ORM usage is the same as with a classical SQLAlchemy configuration.

Creating a Session

We’re now ready to start talking to the database. The ORM’s “handle” to the database is the Session. When we first set up the application, at the same level as our create_engine() statement, we define a Session class which will serve as a factory for new Session objects:

>>> from sqlalchemy.orm import sessionmaker
>>> Session = sessionmaker(bind=engine)

In the case where your application does not yet have an Engine when you define your module-level objects, just set it up like this:

>>> Session = sessionmaker()

Later, when you create your engine with create_engine(), connect it to the Session using configure():

>>> Session.configure(bind=engine)  # once engine is available

This custom-made Session class will create new Session objects which are bound to our database. Other transactional characteristics may be defined when calling sessionmaker() as well; these are described in a later chapter. Then, whenever you need to have a conversation with the database, you instantiate a Session:

>>> session = Session()

The above Session is associated with our SQLite-enabled Engine, but it hasn’t opened any connections yet. When it’s first used, it retrieves a connection from a pool of connections maintained by the Engine, and holds onto it until we commit all changes and/or close the session object.

Adding new Objects

To persist our User object, we add() it to our Session:

>>> ed_user = User('ed', 'Ed Jones', 'edspassword')
>>> session.add(ed_user)

At this point, we say that the instance is pending; no SQL has yet been issued and the object is not yet represented by a row in the database. The Session will issue the SQL to persist Ed Jones as soon as is needed, using a process known as a flush. If we query the database for Ed Jones, all pending information will first be flushed, and the query is issued immediately thereafter.

For example, below we create a new Query object which loads instances of User. We “filter by” the name attribute of ed, and indicate that we’d like only the first result in the full list of rows. A User instance is returned which is equivalent to that which we’ve added:

sql>>> our_user = session.query(User).filter_by(name='ed').first() 
>>> our_user
<User('ed','Ed Jones', 'edspassword')>

In fact, the Session has identified that the row returned is the same row as one already represented within its internal map of objects, so we actually got back the identical instance as that which we just added:

>>> ed_user is our_user

The ORM concept at work here is known as an identity map and ensures that all operations upon a particular row within a Session operate upon the same set of data. Once an object with a particular primary key is present in the Session, all SQL queries on that Session will always return the same Python object for that particular primary key; it also will raise an error if an attempt is made to place a second, already-persisted object with the same primary key within the session.

We can add more User objects at once using add_all():

>>> session.add_all([
...     User('wendy', 'Wendy Williams', 'foobar'),
...     User('mary', 'Mary Contrary', 'xxg527'),
...     User('fred', 'Fred Flinstone', 'blah')])

Also, Ed has already decided his password isn’t too secure, so lets change it:

>>> ed_user.password = 'f8s7ccs'

The Session is paying attention. It knows, for example, that Ed Jones has been modified:

>>> session.dirty
IdentitySet([<User('ed','Ed Jones', 'f8s7ccs')>])

and that three new User objects are pending:

>>> session.new  
IdentitySet([<User('wendy','Wendy Williams', 'foobar')>,
<User('mary','Mary Contrary', 'xxg527')>,
<User('fred','Fred Flinstone', 'blah')>])

We tell the Session that we’d like to issue all remaining changes to the database and commit the transaction, which has been in progress throughout. We do this via commit():

sql>>> session.commit()

commit() flushes whatever remaining changes remain to the database, and commits the transaction. The connection resources referenced by the session are now returned to the connection pool. Subsequent operations with this session will occur in a new transaction, which will again re-acquire connection resources when first needed.

If we look at Ed’s id attribute, which earlier was None, it now has a value:

sql>>> ed_user.id 

After the Session inserts new rows in the database, all newly generated identifiers and database-generated defaults become available on the instance, either immediately or via load-on-first-access. In this case, the entire row was re-loaded on access because a new transaction was begun after we issued commit(). SQLAlchemy by default refreshes data from a previous transaction the first time it’s accessed within a new transaction, so that the most recent state is available. The level of reloading is configurable as is described in the chapter on Sessions.

Rolling Back

Since the Session works within a transaction, we can roll back changes made too. Let’s make two changes that we’ll revert; ed_user‘s user name gets set to Edwardo:

>>> ed_user.name = 'Edwardo'

and we’ll add another erroneous user, fake_user:

>>> fake_user = User('fakeuser', 'Invalid', '12345')
>>> session.add(fake_user)

Querying the session, we can see that they’re flushed into the current transaction:

sql>>> session.query(User).filter(User.name.in_(['Edwardo', 'fakeuser'])).all() 
[<User('Edwardo','Ed Jones', 'f8s7ccs')>, <User('fakeuser','Invalid', '12345')>]

Rolling back, we can see that ed_user‘s name is back to ed, and fake_user has been kicked out of the session:

sql>>> session.rollback()

sql>>> ed_user.name 
>>> fake_user in session

issuing a SELECT illustrates the changes made to the database:

sql>>> session.query(User).filter(User.name.in_(['ed', 'fakeuser'])).all() 
[<User('ed','Ed Jones', 'f8s7ccs')>]


A Query is created using the query() function on Session. This function takes a variable number of arguments, which can be any combination of classes and class-instrumented descriptors. Below, we indicate a Query which loads User instances. When evaluated in an iterative context, the list of User objects present is returned:

sql>>> for instance in session.query(User).order_by(User.id): 
...     print instance.name, instance.fullname
ed Ed Jones
wendy Wendy Williams
mary Mary Contrary
fred Fred Flinstone

The Query also accepts ORM-instrumented descriptors as arguments. Any time multiple class entities or column-based entities are expressed as arguments to the query() function, the return result is expressed as tuples:

sql>>> for name, fullname in session.query(User.name, User.fullname): 
...     print name, fullname
ed Ed Jones
wendy Wendy Williams
mary Mary Contrary
fred Fred Flinstone

The tuples returned by Query are named tuples, and can be treated much like an ordinary Python object. The names are the same as the attribute’s name for an attribute, and the class name for a class:

sql>>> for row in session.query(User, User.name).all(): 
...    print row.User, row.name
<User('ed','Ed Jones', 'f8s7ccs')> ed
<User('wendy','Wendy Williams', 'foobar')> wendy
<User('mary','Mary Contrary', 'xxg527')> mary
<User('fred','Fred Flinstone', 'blah')> fred

You can control the names using the label() construct for scalar attributes and aliased for class constructs:

>>> from sqlalchemy.orm import aliased
>>> user_alias = aliased(User, name='user_alias')
sql>>> for row in session.query(user_alias, user_alias.name.label('name_label')).all(): 
...    print row.user_alias, row.name_label
<User('ed','Ed Jones', 'f8s7ccs')> ed
<User('wendy','Wendy Williams', 'foobar')> wendy
<User('mary','Mary Contrary', 'xxg527')> mary
<User('fred','Fred Flinstone', 'blah')> fred

Basic operations with Query include issuing LIMIT and OFFSET, most conveniently using Python array slices and typically in conjunction with ORDER BY:

sql>>> for u in session.query(User).order_by(User.id)[1:3]: 
...    print u
<User('wendy','Wendy Williams', 'foobar')>
<User('mary','Mary Contrary', 'xxg527')>

and filtering results, which is accomplished either with filter_by(), which uses keyword arguments:

sql>>> for name, in session.query(User.name).filter_by(fullname='Ed Jones'): 
...    print name

...or filter(), which uses more flexible SQL expression language constructs. These allow you to use regular Python operators with the class-level attributes on your mapped class:

sql>>> for name, in session.query(User.name).filter(User.fullname=='Ed Jones'): 
...    print name

The Query object is fully generative, meaning that most method calls return a new Query object upon which further criteria may be added. For example, to query for users named “ed” with a full name of “Ed Jones”, you can call filter() twice, which joins criteria using AND:

sql>>> for user in session.query(User).filter(User.name=='ed').filter(User.fullname=='Ed Jones'): 
...    print user
<User('ed','Ed Jones', 'f8s7ccs')>

Common Filter Operators

Here’s a rundown of some of the most common operators used in filter():

  • equals:

    query.filter(User.name == 'ed')
  • not equals:

    query.filter(User.name != 'ed')
  • LIKE:

  • IN:

    query.filter(User.name.in_(['ed', 'wendy', 'jack']))
    # works with query objects too:
  • NOT IN:

    query.filter(~User.name.in_(['ed', 'wendy', 'jack']))
  • IS NULL:

    filter(User.name == None)

    filter(User.name != None)
  • AND:

    from sqlalchemy import and_
    filter(and_(User.name == 'ed', User.fullname == 'Ed Jones'))
    # or call filter()/filter_by() multiple times
    filter(User.name == 'ed').filter(User.fullname == 'Ed Jones')
  • OR:

    from sqlalchemy import or_
    filter(or_(User.name == 'ed', User.name == 'wendy'))
  • match:

The contents of the match parameter are database backend specific.

Returning Lists and Scalars

The all(), one(), and first() methods of Query immediately issue SQL and return a non-iterator value. all() returns a list:

>>> query = session.query(User).filter(User.name.like('%ed')).order_by(User.id)
sql>>> query.all() 
[<User('ed','Ed Jones', 'f8s7ccs')>, <User('fred','Fred Flinstone', 'blah')>]

first() applies a limit of one and returns the first result as a scalar:

sql>>> query.first() 
<User('ed','Ed Jones', 'f8s7ccs')>

one(), fully fetches all rows, and if not exactly one object identity or composite row is present in the result, raises an error:

sql>>> from sqlalchemy.orm.exc import MultipleResultsFound
>>> try: 
...     user = query.one()
... except MultipleResultsFound, e:
...     print e
Multiple rows were found for one()
sql>>> from sqlalchemy.orm.exc import NoResultFound
>>> try: 
...     user = query.filter(User.id == 99).one()
... except NoResultFound, e:
...     print e
No row was found for one()

Using Literal SQL

Literal strings can be used flexibly with Query. Most methods accept strings in addition to SQLAlchemy clause constructs. For example, filter() and order_by():

sql>>> for user in session.query(User).filter("id<224").order_by("id").all(): 
...     print user.name

Bind parameters can be specified with string-based SQL, using a colon. To specify the values, use the params() method:

sql>>> session.query(User).filter("id<:value and name=:name").\
...     params(value=224, name='fred').order_by(User.id).one() 

To use an entirely string-based statement, using from_statement(); just ensure that the columns clause of the statement contains the column names normally used by the mapper (below illustrated using an asterisk):

sql>>> session.query(User).from_statement(
...                     "SELECT * FROM users where name=:name").\
...                     params(name='ed').all()
[<User('ed','Ed Jones', 'f8s7ccs')>]

You can use from_statement() to go completely “raw”, using string names to identify desired columns:

sql>>> session.query("id", "name", "thenumber12").\
...         from_statement("SELECT id, name, 12 as "
...                 "thenumber12 FROM users where name=:name").\
...                 params(name='ed').all()
[(1, u'ed', 12)]


Query includes a convenience method for counting called count():

sql>>> session.query(User).filter(User.name.like('%ed')).count() 

The count() method is used to determine how many rows the SQL statement would return, and is mainly intended to return a simple count of a single type of entity, in this case User. For more complicated sets of columns or entities where the “thing to be counted” needs to be indicated more specifically, count() is probably not what you want. Below, a query for individual columns does return the expected result:

sql>>> session.query(User.id, User.name).filter(User.name.like('%ed')).count() 

...but if you look at the generated SQL, SQLAlchemy saw that we were placing individual column expressions and decided to wrap whatever it was we were doing in a subquery, so as to be assured that it returns the “number of rows”. This defensive behavior is not really needed here and in other cases is not what we want at all, such as if we wanted a grouping of counts per name:

sql>>> session.query(User.name).group_by(User.name).count()  

We don’t want the number 4, we wanted some rows back. So for detailed queries where you need to count something specific, use the func.count() function as a column expression:

>>> from sqlalchemy import func
sql>>> session.query(func.count(User.name), User.name).group_by(User.name).all()  
[(1, u'ed'), (1, u'fred'), (1, u'mary'), (1, u'wendy')]

Building a Relationship

Now let’s consider a second table to be dealt with. Users in our system also can store any number of email addresses associated with their username. This implies a basic one to many association from the users_table to a new table which stores email addresses, which we will call addresses. Using declarative, we define this table along with its mapped class, Address:

>>> from sqlalchemy import ForeignKey
>>> from sqlalchemy.orm import relationship, backref
>>> class Address(Base):
...     __tablename__ = 'addresses'
...     id = Column(Integer, primary_key=True)
...     email_address = Column(String, nullable=False)
...     user_id = Column(Integer, ForeignKey('users.id'))
...     user = relationship(User, backref=backref('addresses', order_by=id))
...     def __init__(self, email_address):
...         self.email_address = email_address
...     def __repr__(self):
...         return "<Address('%s')>" % self.email_address

The above class introduces a foreign key constraint which references the users table. This defines for SQLAlchemy the relationship between the two tables at the database level. The relationship between the User and Address classes is defined separately using the relationship() function, which defines an attribute user to be placed on the Address class, as well as an addresses collection to be placed on the User class. Such a relationship is known as a bidirectional relationship. Because of the placement of the foreign key, from Address to User it is many to one, and from User to Address it is one to many. SQLAlchemy is automatically aware of many-to-one/one-to-many based on foreign keys.


The relationship() function has historically been known as relation(), which is the name that’s available in all versions of SQLAlchemy prior to 0.6beta2, including the 0.5 and 0.4 series. relationship() is only available starting with SQLAlchemy 0.6beta2. relation() will remain available in SQLAlchemy for the foreseeable future to enable cross-compatibility.

The relationship() function is extremely flexible, and could just have easily been defined on the User class:

class User(Base):
    # ....
    addresses = relationship(Address, order_by=Address.id, backref="user")

We are also free to not define a backref, and to define the relationship() only on one class and not the other. It is also possible to define two separate relationship() constructs for either direction, which is generally safe for many-to-one and one-to-many relationships, but not for many-to-many relationships.

When using the declarative extension, relationship() gives us the option to use strings for most arguments that concern the target class, in the case that the target class has not yet been defined. This only works in conjunction with declarative:

class User(Base):
    addresses = relationship("Address", order_by="Address.id", backref="user")

When declarative is not in use, you typically define your mapper() well after the target classes and Table objects have been defined, so string expressions are not needed.

We’ll need to create the addresses table in the database, so we will issue another CREATE from our metadata, which will skip over tables which have already been created:

sql>>> metadata.create_all(engine) 

Querying with Joins

While joinedload() created an anonymous, non-accessible JOIN specifically to populate a collection, we can also work explicitly with joins in many ways. For example, to construct a simple inner join between User and Address, we can just filter() their related columns together. Below we load the User and Address entities at once using this method:

sql>>> for u, a in session.query(User, Address).filter(User.id==Address.user_id).\
...         filter(Address.email_address=='jack@google.com').all():   
...     print u, a
<User('jack','Jack Bean', 'gjffdd')> <Address('jack@google.com')>

Or we can make a real JOIN construct; the most common way is to use join():

sql>>> session.query(User).join(Address).\
...         filter(Address.email_address=='jack@google.com').all() 
[<User('jack','Jack Bean', 'gjffdd')>]

join() knows how to join between User and Address because there’s only one foreign key between them. If there were no foreign keys, or several, join() works better when one of the following forms are used:

query.join((Address, User.id==Address.user_id))  # explicit condition (note the tuple)
query.join(User.addresses)                       # specify relationship from left to right
query.join((Address, User.addresses))            # same, with explicit target
query.join('addresses')                          # same, using a string

Note that when join() is called with an explicit target as well as an ON clause, we use a tuple as the argument. This is so that multiple joins can be chained together, as in:

                        (Bat, bar.bats),
                        (Widget, Bat.widget_id==Widget.id)

The above would produce SQL something like foo JOIN bars ON <onclause> JOIN bats ON <onclause> JOIN widgets ON <onclause>.

The general functionality of join() is also available as a standalone function join(), which is an ORM-enabled version of the same function present in the SQL expression language. This function accepts two or three arguments (left side, right side, optional ON clause) and can be used in conjunction with the select_from() method to set an explicit FROM clause:

>>> from sqlalchemy.orm import join
sql>>> session.query(User).\
...                select_from(join(User, Address, User.addresses)).\
...                filter(Address.email_address=='jack@google.com').all() 
[<User('jack','Jack Bean', 'gjffdd')>]

Using join() to Eagerly Load Collections/Attributes

The “eager loading” capabilities of the joinedload() function and the join-construction capabilities of join() or an equivalent can be combined together using the contains_eager() option. This is typically used for a query that is already joining to some related entity (more often than not via many-to-one), and you’d like the related entity to also be loaded onto the resulting objects in one step without the need for additional queries and without the “automatic” join embedded by the joinedload() function:

>>> from sqlalchemy.orm import contains_eager
sql>>> for address in session.query(Address).\
...                join(Address.user).\
...                filter(User.name=='jack').\
...                options(contains_eager(Address.user)): 
...         print address, address.user
<Address('jack@google.com')> <User('jack','Jack Bean', 'gjffdd')>
<Address('j25@yahoo.com')> <User('jack','Jack Bean', 'gjffdd')>

Note that above the join was used both to limit the rows to just those Address objects which had a related User object with the name “jack”. It’s safe to have the Address.user attribute populated with this user using an inner join. However, when filtering on a join that is filtering on a particular member of a collection, using contains_eager() to populate a related collection may populate the collection with only part of what it actually references, since the collection itself is filtered.

Using Aliases

When querying across multiple tables, if the same table needs to be referenced more than once, SQL typically requires that the table be aliased with another name, so that it can be distinguished against other occurrences of that table. The Query supports this most explicitly using the aliased construct. Below we join to the Address entity twice, to locate a user who has two distinct email addresses at the same time:

>>> from sqlalchemy.orm import aliased
>>> adalias1 = aliased(Address)
>>> adalias2 = aliased(Address)
sql>>> for username, email1, email2 in \
...     session.query(User.name, adalias1.email_address, adalias2.email_address).\
...     join((adalias1, User.addresses), (adalias2, User.addresses)).\
...     filter(adalias1.email_address=='jack@google.com').\
...     filter(adalias2.email_address=='j25@yahoo.com'):
...     print username, email1, email2      
jack jack@google.com j25@yahoo.com

Using Subqueries

The Query is suitable for generating statements which can be used as subqueries. Suppose we wanted to load User objects along with a count of how many Address records each user has. The best way to generate SQL like this is to get the count of addresses grouped by user ids, and JOIN to the parent. In this case we use a LEFT OUTER JOIN so that we get rows back for those users who don’t have any addresses, e.g.:

SELECT users.*, adr_count.address_count FROM users LEFT OUTER JOIN
    (SELECT user_id, count(*) AS address_count FROM addresses GROUP BY user_id) AS adr_count
    ON users.id=adr_count.user_id

Using the Query, we build a statement like this from the inside out. The statement accessor returns a SQL expression representing the statement generated by a particular Query - this is an instance of a select() construct, which are described in SQL Expression Language Tutorial:

>>> from sqlalchemy.sql import func
>>> stmt = session.query(Address.user_id, func.count('*').label('address_count')).group_by(Address.user_id).subquery()

The func keyword generates SQL functions, and the subquery() method on Query produces a SQL expression construct representing a SELECT statement embedded within an alias (it’s actually shorthand for query.statement.alias()).

Once we have our statement, it behaves like a Table construct, such as the one we created for users at the start of this tutorial. The columns on the statement are accessible through an attribute called c:

sql>>> for u, count in session.query(User, stmt.c.address_count).\
...     outerjoin((stmt, User.id==stmt.c.user_id)).order_by(User.id): 
...     print u, count
<User('ed','Ed Jones', 'f8s7ccs')> None
<User('wendy','Wendy Williams', 'foobar')> None
<User('mary','Mary Contrary', 'xxg527')> None
<User('fred','Fred Flinstone', 'blah')> None
<User('jack','Jack Bean', 'gjffdd')> 2

Selecting Entities from Subqueries

Above, we just selected a result that included a column from a subquery. What if we wanted our subquery to map to an entity ? For this we use aliased() to associate an “alias” of a mapped class to a subquery:

sql>>> stmt = session.query(Address).filter(Address.email_address != 'j25@yahoo.com').subquery()
>>> adalias = aliased(Address, stmt)
>>> for user, address in session.query(User, adalias).join((adalias, User.addresses)): 
...     print user, address
<User('jack','Jack Bean', 'gjffdd')> <Address('jack@google.com')>


The EXISTS keyword in SQL is a boolean operator which returns True if the given expression contains any rows. It may be used in many scenarios in place of joins, and is also useful for locating rows which do not have a corresponding row in a related table.

There is an explicit EXISTS construct, which looks like this:

>>> from sqlalchemy.sql import exists
>>> stmt = exists().where(Address.user_id==User.id)
sql>>> for name, in session.query(User.name).filter(stmt):   
...     print name

The Query features several operators which make usage of EXISTS automatically. Above, the statement can be expressed along the User.addresses relationship using any():

sql>>> for name, in session.query(User.name).filter(User.addresses.any()):   
...     print name

any() takes criterion as well, to limit the rows matched:

sql>>> for name, in session.query(User.name).\
...     filter(User.addresses.any(Address.email_address.like('%google%'))):   
...     print name

has() is the same operator as any() for many-to-one relationships (note the ~ operator here too, which means “NOT”):

sql>>> session.query(Address).filter(~Address.user.has(User.name=='jack')).all() 

Common Relationship Operators

Here’s all the operators which build on relationships:

  • equals (used for many-to-one):

    query.filter(Address.user == someuser)
  • not equals (used for many-to-one):

    query.filter(Address.user != someuser)
  • IS NULL (used for many-to-one):

    query.filter(Address.user == None)
  • contains (used for one-to-many and many-to-many collections):

  • any (used for one-to-many and many-to-many collections):

    query.filter(User.addresses.any(Address.email_address == 'bar'))
    # also takes keyword arguments:
  • has (used for many-to-one):

  • with_parent (used for any relationship):

    session.query(Address).with_parent(someuser, 'addresses')


Let’s try to delete jack and see how that goes. We’ll mark as deleted in the session, then we’ll issue a count query to see that no rows remain:

>>> session.delete(jack)
sql>>> session.query(User).filter_by(name='jack').count() 

So far, so good. How about Jack’s Address objects ?

sql>>> session.query(Address).filter(
...     Address.email_address.in_(['jack@google.com', 'j25@yahoo.com'])
...  ).count() 

Uh oh, they’re still there ! Analyzing the flush SQL, we can see that the user_id column of each address was set to NULL, but the rows weren’t deleted. SQLAlchemy doesn’t assume that deletes cascade, you have to tell it to do so.

Configuring delete/delete-orphan Cascade

We will configure cascade options on the User.addresses relationship to change the behavior. While SQLAlchemy allows you to add new attributes and relationships to mappings at any point in time, in this case the existing relationship needs to be removed, so we need to tear down the mappings completely and start again.


Tearing down mappers with clear_mappers() is not a typical operation, and normal applications do not need to use this function. It is here so that the tutorial code can be executed as a whole.

>>> session.close()  # roll back and close the transaction
>>> from sqlalchemy.orm import clear_mappers
>>> clear_mappers() # remove all class mappings

Below, we use mapper to reconfigure an ORM mapping for User and Address, on our existing but currently un-mapped classes. The User.addresses relationship now has delete, delete-orphan cascade on it, which indicates that DELETE operations will cascade to attached Address objects as well as Address objects which are removed from their parent:

>>> users_table = User.__table__
>>> mapper(User, users_table, properties={    
...     'addresses':relationship(Address, backref='user', cascade="all, delete, delete-orphan")
... })
<Mapper at 0x...; User>

>>> addresses_table = Address.__table__
>>> mapper(Address, addresses_table) 
<Mapper at 0x...; Address>

Now when we load Jack (below using get(), which loads by primary key), removing an address from his addresses collection will result in that Address being deleted:

# load Jack by primary key
sql>>> jack = session.query(User).get(5)    

# remove one Address (lazy load fires off)
sql>>> del jack.addresses[1] 

# only one address remains
sql>>> session.query(Address).filter(
...     Address.email_address.in_(['jack@google.com', 'j25@yahoo.com'])
... ).count() 

Deleting Jack will delete both Jack and his remaining Address:

>>> session.delete(jack)

sql>>> session.query(User).filter_by(name='jack').count() 

sql>>> session.query(Address).filter(
...    Address.email_address.in_(['jack@google.com', 'j25@yahoo.com'])
... ).count() 

Building a Many To Many Relationship

We’re moving into the bonus round here, but lets show off a many-to-many relationship. We’ll sneak in some other features too, just to take a tour. We’ll make our application a blog application, where users can write BlogPost items, which have Keyword items associated with them.

The declarative setup is as follows:

>>> from sqlalchemy import Text

>>> # association table
>>> post_keywords = Table('post_keywords', metadata,
...     Column('post_id', Integer, ForeignKey('posts.id')),
...     Column('keyword_id', Integer, ForeignKey('keywords.id'))
... )

>>> class BlogPost(Base):
...     __tablename__ = 'posts'
...     id = Column(Integer, primary_key=True)
...     user_id = Column(Integer, ForeignKey('users.id'))
...     headline = Column(String(255), nullable=False)
...     body = Column(Text)
...     # many to many BlogPost<->Keyword
...     keywords = relationship('Keyword', secondary=post_keywords, backref='posts')
...     def __init__(self, headline, body, author):
...         self.author = author
...         self.headline = headline
...         self.body = body
...     def __repr__(self):
...         return "BlogPost(%r, %r, %r)" % (self.headline, self.body, self.author)

>>> class Keyword(Base):
...     __tablename__ = 'keywords'
...     id = Column(Integer, primary_key=True)
...     keyword = Column(String(50), nullable=False, unique=True)
...     def __init__(self, keyword):
...         self.keyword = keyword

Above, the many-to-many relationship is BlogPost.keywords. The defining feature of a many-to-many relationship is the secondary keyword argument which references a Table object representing the association table. This table only contains columns which reference the two sides of the relationship; if it has any other columns, such as its own primary key, or foreign keys to other tables, SQLAlchemy requires a different usage pattern called the “association object”, described at Association Object.

The many-to-many relationship is also bi-directional using the backref keyword. This is the one case where usage of backref is generally required, since if a separate posts relationship were added to the Keyword entity, both relationships would independently add and remove rows from the post_keywords table and produce conflicts.

We would also like our BlogPost class to have an author field. We will add this as another bidirectional relationship, except one issue we’ll have is that a single user might have lots of blog posts. When we access User.posts, we’d like to be able to filter results further so as not to load the entire collection. For this we use a setting accepted by relationship() called lazy='dynamic', which configures an alternate loader strategy on the attribute. To use it on the “reverse” side of a relationship(), we use the backref() function:

>>> from sqlalchemy.orm import backref
>>> # "dynamic" loading relationship to User
>>> BlogPost.author = relationship(User, backref=backref('posts', lazy='dynamic'))

Create new tables:

sql>>> metadata.create_all(engine) 

Usage is not too different from what we’ve been doing. Let’s give Wendy some blog posts:

sql>>> wendy = session.query(User).filter_by(name='wendy').one() 
>>> post = BlogPost("Wendy's Blog Post", "This is a test", wendy)
>>> session.add(post)

We’re storing keywords uniquely in the database, but we know that we don’t have any yet, so we can just create them:

>>> post.keywords.append(Keyword('wendy'))
>>> post.keywords.append(Keyword('firstpost'))

We can now look up all blog posts with the keyword ‘firstpost’. We’ll use the any operator to locate “blog posts where any of its keywords has the keyword string ‘firstpost’”:

sql>>> session.query(BlogPost).filter(BlogPost.keywords.any(keyword='firstpost')).all() 
[BlogPost("Wendy's Blog Post", 'This is a test', <User('wendy','Wendy Williams', 'foobar')>)]

If we want to look up just Wendy’s posts, we can tell the query to narrow down to her as a parent:

sql>>> session.query(BlogPost).filter(BlogPost.author==wendy).\
... filter(BlogPost.keywords.any(keyword='firstpost')).all() 
[BlogPost("Wendy's Blog Post", 'This is a test', <User('wendy','Wendy Williams', 'foobar')>)]

Or we can use Wendy’s own posts relationship, which is a “dynamic” relationship, to query straight from there:

sql>>> wendy.posts.filter(BlogPost.keywords.any(keyword='firstpost')).all() 
[BlogPost("Wendy's Blog Post", 'This is a test', <User('wendy','Wendy Williams', 'foobar')>)]

Further Reference

Query Reference: Querying

Mapper Reference: Mapper Configuration

Relationship Reference: Relationship Configuration

Session Reference: Using the Session.