What’s new in SQLAlchemy 0.5?

About this Document

This document describes changes between SQLAlchemy version 0.4, last released October 12, 2008, and SQLAlchemy version 0.5, last released January 16, 2010.

Document date: August 4, 2009

This guide documents API changes which affect users migrating their applications from the 0.4 series of SQLAlchemy to 0.5. It’s also recommended for those working from Essential SQLAlchemy, which only covers 0.4 and seems to even have some old 0.3isms in it. Note that SQLAlchemy 0.5 removes many behaviors which were deprecated throughout the span of the 0.4 series, and also deprecates more behaviors specific to 0.4.

Major Documentation Changes

Some sections of the documentation have been completely rewritten and can serve as an introduction to new ORM features. The Query and Session objects in particular have some distinct differences in API and behavior which fundamentally change many of the basic ways things are done, particularly with regards to constructing highly customized ORM queries and dealing with stale session state, commits and rollbacks.

Deprecations Source

Another source of information is documented within a series of unit tests illustrating up to date usages of some common Query patterns; this file can be viewed at [source:sqlalchemy/trunk/test/orm/test_deprecations.py].

Requirements Changes

  • Python 2.4 or higher is required. The SQLAlchemy 0.4 line is the last version with Python 2.3 support.

Object Relational Mapping

  • Column level expressions within Query. - as detailed in the tutorial, Query has the capability to create specific SELECT statements, not just those against full rows:

    session.query(User.name, func.count(Address.id).label("numaddresses")).join(

    The tuples returned by any multi-column/entity query are named’ tuples:

    for row in (
        session.query(User.name, func.count(Address.id).label("numaddresses"))
        print("name", row.name, "number", row.numaddresses)

    Query has a statement accessor, as well as a subquery() method which allow Query to be used to create more complex combinations:

    subq = (
        .filter(Keyword.name.in_(["beans", "carrots"]))
    recipes = session.query(Recipe).filter(
        .where(Recipe.id == recipe_keywords.c.recipe_id)
        .where(recipe_keywords.c.keyword_id == subq.c.keyword_id)
  • Explicit ORM aliases are recommended for aliased joins - The aliased() function produces an “alias” of a class, which allows fine-grained control of aliases in conjunction with ORM queries. While a table-level alias (i.e. table.alias()) is still usable, an ORM level alias retains the semantics of the ORM mapped object which is significant for inheritance mappings, options, and other scenarios. E.g.:

    Friend = aliased(Person)
    session.query(Person, Friend).join((Friend, Person.friends)).all()
  • query.join() greatly enhanced. - You can now specify the target and ON clause for a join in multiple ways. A target class alone can be provided where SQLA will attempt to form a join to it via foreign key in the same way as table.join(someothertable). A target and an explicit ON condition can be provided, where the ON condition can be a relation() name, an actual class descriptor, or a SQL expression. Or the old way of just a relation() name or class descriptor works too. See the ORM tutorial which has several examples.

  • Declarative is recommended for applications which don’t require (and don’t prefer) abstraction between tables and mappers - The [/docs/05/reference/ext/declarative.html Declarative] module, which is used to combine the expression of Table, mapper(), and user defined class objects together, is highly recommended as it simplifies application configuration, ensures the “one mapper per class” pattern, and allows the full range of configuration available to distinct mapper() calls. Separate mapper() and Table usage is now referred to as “classical SQLAlchemy usage” and of course is freely mixable with declarative.

  • The .c. attribute has been removed from classes (i.e. MyClass.c.somecolumn). As is the case in 0.4, class- level properties are usable as query elements, i.e. Class.c.propname is now superseded by Class.propname, and the c attribute continues to remain on Table objects where they indicate the namespace of Column objects present on the table.

    To get at the Table for a mapped class (if you didn’t keep it around already):

    table = class_mapper(someclass).mapped_table

    Iterate through columns:

    for col in table.c:

    Work with a specific column:


    The class-bound descriptors support the full set of Column operators as well as the documented relation-oriented operators like has(), any(), contains(), etc.

    The reason for the hard removal of .c. is that in 0.5, class-bound descriptors carry potentially different meaning, as well as information regarding class mappings, versus plain Column objects - and there are use cases where you’d specifically want to use one or the other. Generally, using class-bound descriptors invokes a set of mapping/polymorphic aware translations, and using table- bound columns does not. In 0.4, these translations were applied across the board to all expressions, but 0.5 differentiates completely between columns and mapped descriptors, only applying translations to the latter. So in many cases, particularly when dealing with joined table inheritance configurations as well as when using query(<columns>), Class.propname and table.c.colname are not interchangeable.

    For example, session.query(users.c.id, users.c.name) is different versus session.query(User.id, User.name); in the latter case, the Query is aware of the mapper in use and further mapper-specific operations like query.join(<propname>), query.with_parent() etc. may be used, but in the former case cannot. Additionally, in polymorphic inheritance scenarios, the class-bound descriptors refer to the columns present in the polymorphic selectable in use, not necessarily the table column which directly corresponds to the descriptor. For example, a set of classes related by joined-table inheritance to the person table along the person_id column of each table will all have their Class.person_id attribute mapped to the person_id column in person, and not their subclass table. Version 0.4 would map this behavior onto table-bound Column objects automatically. In 0.5, this automatic conversion has been removed, so that you in fact can use table-bound columns as a means to override the translations which occur with polymorphic querying; this allows Query to be able to create optimized selects among joined-table or concrete-table inheritance setups, as well as portable subqueries, etc.

  • Session Now Synchronizes Automatically with Transactions. Session now synchronizes against the transaction automatically by default, including autoflush and autoexpire. A transaction is present at all times unless disabled using the autocommit option. When all three flags are set to their default, the Session recovers gracefully after rollbacks and it’s very difficult to get stale data into the session. See the new Session documentation for details.

  • Implicit Order By Is Removed. This will impact ORM users who rely upon SA’s “implicit ordering” behavior, which states that all Query objects which don’t have an order_by() will ORDER BY the “id” or “oid” column of the primary mapped table, and all lazy/eagerly loaded collections apply a similar ordering. In 0.5, automatic ordering must be explicitly configured on mapper() and relation() objects (if desired), or otherwise when using Query.

    To convert an 0.4 mapping to 0.5, such that its ordering behavior will be extremely similar to 0.4 or previous, use the order_by setting on mapper() and relation():

        properties={"addresses": relation(Address, order_by=addresses.c.id)},

    To set ordering on a backref, use the backref() function:

    "keywords": relation(
        backref=backref("items", order_by=items.c.id),

    Using declarative ? To help with the new order_by requirement, order_by and friends can now be set using strings which are evaluated in Python later on (this works only with declarative, not plain mappers):

    class MyClass(MyDeclarativeBase):
        "addresses": relation("Address", order_by="Address.id")

    It’s generally a good idea to set order_by on relation()s which load list-based collections of items, since that ordering cannot otherwise be affected. Other than that, the best practice is to use Query.order_by() to control ordering of the primary entities being loaded.

  • Session is now autoflush=True/autoexpire=True/autocommit=False. - To set it up, just call sessionmaker() with no arguments. The name transactional=True is now autocommit=False. Flushes occur upon each query issued (disable with autoflush=False), within each commit() (as always), and before each begin_nested() (so rolling back to the SAVEPOINT is meaningful). All objects are expired after each commit() and after each rollback(). After rollback, pending objects are expunged, deleted objects move back to persistent. These defaults work together very nicely and there’s really no more need for old techniques like clear() (which is renamed to expunge_all() as well).

    P.S.: sessions are now reusable after a rollback(). Scalar and collection attribute changes, adds and deletes are all rolled back.

  • session.add() replaces session.save(), session.update(), session.save_or_update(). - the session.add(someitem) and session.add_all([list of items]) methods replace save(), update(), and save_or_update(). Those methods will remain deprecated throughout 0.5.

  • backref configuration made less verbose. - The backref() function now uses the primaryjoin and secondaryjoin arguments of the forwards-facing relation() when they are not explicitly stated. It’s no longer necessary to specify primaryjoin/secondaryjoin in both directions separately.

  • Simplified polymorphic options. - The ORM’s “polymorphic load” behavior has been simplified. In 0.4, mapper() had an argument called polymorphic_fetch which could be configured as select or deferred. This option is removed; the mapper will now just defer any columns which were not present in the SELECT statement. The actual SELECT statement used is controlled by the with_polymorphic mapper argument (which is also in 0.4 and replaces select_table), as well as the with_polymorphic() method on Query (also in 0.4).

    An improvement to the deferred loading of inheriting classes is that the mapper now produces the “optimized” version of the SELECT statement in all cases; that is, if class B inherits from A, and several attributes only present on class B have been expired, the refresh operation will only include B’s table in the SELECT statement and will not JOIN to A.

  • The execute() method on Session converts plain strings into text() constructs, so that bind parameters may all be specified as “:bindname” without needing to call text() explicitly. If “raw” SQL is desired here, use session.connection().execute("raw text").

  • session.Query().iterate_instances() has been renamed to just instances(). The old instances() method returning a list instead of an iterator no longer exists. If you were relying on that behavior, you should use list(your_query.instances()).

Extending the ORM

In 0.5 we’re moving forward with more ways to modify and extend the ORM. Heres a summary:

  • MapperExtension. - This is the classic extension class, which remains. Methods which should rarely be needed are create_instance() and populate_instance(). To control the initialization of an object when it’s loaded from the database, use the reconstruct_instance() method, or more easily the @reconstructor decorator described in the documentation.

  • SessionExtension. - This is an easy to use extension class for session events. In particular, it provides before_flush(), after_flush() and after_flush_postexec() methods. This usage is recommended over MapperExtension.before_XXX in many cases since within before_flush() you can modify the flush plan of the session freely, something which cannot be done from within MapperExtension.

  • AttributeExtension. - This class is now part of the public API, and allows the interception of userland events on attributes, including attribute set and delete operations, and collection appends and removes. It also allows the value to be set or appended to be modified. The @validates decorator, described in the documentation, provides a quick way to mark any mapped attributes as being “validated” by a particular class method.

  • Attribute Instrumentation Customization. - An API is provided for ambitious efforts to entirely replace SQLAlchemy’s attribute instrumentation, or just to augment it in some cases. This API was produced for the purposes of the Trellis toolkit, but is available as a public API. Some examples are provided in the distribution in the /examples/custom_attributes directory.


  • String with no length no longer generates TEXT, it generates VARCHAR - The String type no longer magically converts into a Text type when specified with no length. This only has an effect when CREATE TABLE is issued, as it will issue VARCHAR with no length parameter, which is not valid on many (but not all) databases. To create a TEXT (or CLOB, i.e. unbounded string) column, use the Text type.

  • PickleType() with mutable=True requires an __eq__() method - The PickleType type needs to compare values when mutable=True. The method of comparing pickle.dumps() is inefficient and unreliable. If an incoming object does not implement __eq__() and is also not None, the dumps() comparison is used but a warning is raised. For types which implement __eq__() which includes all dictionaries, lists, etc., comparison will use == and is now reliable by default.

  • convert_bind_param() and convert_result_value() methods of TypeEngine/TypeDecorator are removed. - The O’Reilly book unfortunately documented these methods even though they were deprecated post 0.3. For a user-defined type which subclasses TypeEngine, the bind_processor() and result_processor() methods should be used for bind/result processing. Any user defined type, whether extending TypeEngine or TypeDecorator, which uses the old 0.3 style can be easily adapted to the new style using the following adapter:

    class AdaptOldConvertMethods(object):
        """A mixin which adapts 0.3-style convert_bind_param and
        convert_result_value methods
        def bind_processor(self, dialect):
            def convert(value):
                return self.convert_bind_param(value, dialect)
            return convert
        def result_processor(self, dialect):
            def convert(value):
                return self.convert_result_value(value, dialect)
            return convert
        def convert_result_value(self, value, dialect):
            return value
        def convert_bind_param(self, value, dialect):
            return value

    To use the above mixin:

    class MyType(AdaptOldConvertMethods, TypeEngine): ...
  • The quote flag on Column and Table as well as the quote_schema flag on Table now control quoting both positively and negatively. The default is None, meaning let regular quoting rules take effect. When True, quoting is forced on. When False, quoting is forced off.

  • Column DEFAULT value DDL can now be more conveniently specified with Column(..., server_default='val'), deprecating Column(..., PassiveDefault('val')). default= is now exclusively for Python-initiated default values, and can coexist with server_default. A new server_default=FetchedValue() replaces the PassiveDefault('') idiom for marking columns as subject to influence from external triggers and has no DDL side effects.

  • SQLite’s DateTime, Time and Date types now only accept datetime objects, not strings as bind parameter input. If you’d like to create your own “hybrid” type which accepts strings and returns results as date objects (from whatever format you’d like), create a TypeDecorator that builds on String. If you only want string-based dates, just use String.

  • Additionally, the DateTime and Time types, when used with SQLite, now represent the “microseconds” field of the Python datetime.datetime object in the same manner as str(datetime) - as fractional seconds, not a count of microseconds. That is:

    dt = datetime.datetime(2008, 6, 27, 12, 0, 0, 125)  # 125 usec
    # old way
    "2008-06-27 12:00:00.125"
    # new way
    "2008-06-27 12:00:00.000125"

    So if an existing SQLite file-based database intends to be used across 0.4 and 0.5, you either have to upgrade the datetime columns to store the new format (NOTE: please test this, I’m pretty sure its correct):

    UPDATE mytable SET somedatecol =
      substr(somedatecol, 0, 19) || '.' || substr((substr(somedatecol, 21, -1) / 1000000), 3, -1);

    or, enable “legacy” mode as follows:

    from sqlalchemy.databases.sqlite import DateTimeMixin
    DateTimeMixin.__legacy_microseconds__ = True

Connection Pool no longer threadlocal by default

0.4 has an unfortunate default setting of “pool_threadlocal=True”, leading to surprise behavior when, for example, using multiple Sessions within a single thread. This flag is now off in 0.5. To re-enable 0.4’s behavior, specify pool_threadlocal=True to create_engine(), or alternatively use the “threadlocal” strategy via strategy="threadlocal".

*args Accepted, *args No Longer Accepted

The policy with method(\*args) vs. method([args]) is, if the method accepts a variable-length set of items which represent a fixed structure, it takes \*args. If the method accepts a variable-length set of items that are data-driven, it takes [args].

  • The various Query.options() functions eagerload(), eagerload_all(), lazyload(), contains_eager(), defer(), undefer() all accept variable-length \*keys as their argument now, which allows a path to be formulated using descriptors, ie.:

    query.options(eagerload_all(User.orders, Order.items, Item.keywords))

    A single array argument is still accepted for backwards compatibility.

  • Similarly, the Query.join() and Query.outerjoin() methods accept a variable length *args, with a single array accepted for backwards compatibility:

    query.join("orders", "items")
    query.join(User.orders, Order.items)
  • the in_() method on columns and similar only accepts a list argument now. It no longer accepts \*args.


  • entity_name - This feature was always problematic and rarely used. 0.5’s more deeply fleshed out use cases revealed further issues with entity_name which led to its removal. If different mappings are required for a single class, break the class into separate subclasses and map them separately. An example of this is at [wiki:UsageRecipes/EntityName]. More information regarding rationale is described at https://groups.google.c om/group/sqlalchemy/browse_thread/thread/9e23a0641a88b96d? hl=en .

  • get()/load() cleanup

    The load() method has been removed. Its functionality was kind of arbitrary and basically copied from Hibernate, where it’s also not a particularly meaningful method.

    To get equivalent functionality:

    x = session.query(SomeClass).populate_existing().get(7)

    Session.get(cls, id) and Session.load(cls, id) have been removed. Session.get() is redundant vs. session.query(cls).get(id).

    MapperExtension.get() is also removed (as is MapperExtension.load()). To override the functionality of Query.get(), use a subclass:

    class MyQuery(Query):
        def get(self, ident): ...
    session = sessionmaker(query_cls=MyQuery)()
    ad1 = session.query(Address).get(1)
  • sqlalchemy.orm.relation()

    The following deprecated keyword arguments have been removed:

    foreignkey, association, private, attributeext, is_backref

    In particular, attributeext is replaced with extension - the AttributeExtension class is now in the public API.

  • session.Query()

    The following deprecated functions have been removed:

    list, scalar, count_by, select_whereclause, get_by, select_by, join_by, selectfirst, selectone, select, execute, select_statement, select_text, join_to, join_via, selectfirst_by, selectone_by, apply_max, apply_min, apply_avg, apply_sum

    Additionally, the id keyword argument to join(), outerjoin(), add_entity() and add_column() has been removed. To target table aliases in Query to result columns, use the aliased construct:

    from sqlalchemy.orm import aliased
    address_alias = aliased(Address)
    print(session.query(User, address_alias).join((address_alias, User.addresses)).all())
  • sqlalchemy.orm.Mapper

    • instances()

    • get_session() - this method was not very noticeable, but had the effect of associating lazy loads with a particular session even if the parent object was entirely detached, when an extension such as scoped_session() or the old SessionContextExt was used. It’s possible that some applications which relied upon this behavior will no longer work as expected; but the better programming practice here is to always ensure objects are present within sessions if database access from their attributes are required.

  • mapper(MyClass, mytable)

    Mapped classes no are longer instrumented with a “c” class attribute; e.g. MyClass.c

  • sqlalchemy.orm.collections

    The _prepare_instrumentation alias for prepare_instrumentation has been removed.

  • sqlalchemy.orm

    Removed the EXT_PASS alias of EXT_CONTINUE.

  • sqlalchemy.engine

    The alias from DefaultDialect.preexecute_sequences to .preexecute_pk_sequences has been removed.

    The deprecated engine_descriptors() function has been removed.

  • sqlalchemy.ext.activemapper

    Module removed.

  • sqlalchemy.ext.assignmapper

    Module removed.

  • sqlalchemy.ext.associationproxy

    Pass-through of keyword args on the proxy’s .append(item, \**kw) has been removed and is now simply .append(item)

  • sqlalchemy.ext.selectresults, sqlalchemy.mods.selectresults

    Modules removed.

  • sqlalchemy.ext.declarative

    declared_synonym() removed.

  • sqlalchemy.ext.sessioncontext

    Module removed.

  • sqlalchemy.log

    The SADeprecationWarning alias to sqlalchemy.exc.SADeprecationWarning has been removed.

  • sqlalchemy.exc

    exc.AssertionError has been removed and usage replaced by the Python built-in of the same name.

  • sqlalchemy.databases.mysql

    The deprecated get_version_info dialect method has been removed.

Renamed or Moved

  • sqlalchemy.exceptions is now sqlalchemy.exc

    The module may still be imported under the old name until 0.6.

  • FlushError, ConcurrentModificationError, UnmappedColumnError -> sqlalchemy.orm.exc

    These exceptions moved to the orm package. Importing ‘sqlalchemy.orm’ will install aliases in sqlalchemy.exc for compatibility until 0.6.

  • sqlalchemy.logging -> sqlalchemy.log

    This internal module was renamed. No longer needs to be special cased when packaging SA with py2app and similar tools that scan imports.

  • session.Query().iterate_instances() -> session.Query().instances().


  • Session.save(), Session.update(), Session.save_or_update()

    All three replaced by Session.add()

  • sqlalchemy.PassiveDefault

    Use Column(server_default=...) Translates to sqlalchemy.DefaultClause() under the hood.

  • session.Query().iterate_instances(). It has been renamed to instances().