sqlalchemy-1753 / doc / build / changelog / migration_07.rst

Full commit

What's New in SQLAlchemy 0.7?

About this Document

This document describes changes between SQLAlchemy version 0.6, last released May 5, 2012, and SQLAlchemy version 0.7, undergoing maintenance releases as of October, 2012.

Document date: July 27, 2011


This guide introduces what's new in SQLAlchemy version 0.7, and also documents changes which affect users migrating their applications from the 0.6 series of SQLAlchemy to 0.7.

To as great a degree as possible, changes are made in such a way as to not break compatibility with applications built for 0.6. The changes that are necessarily not backwards compatible are very few, and all but one, the change to mutable attribute defaults, should affect an exceedingly small portion of applications - many of the changes regard non-public APIs and undocumented hacks some users may have been attempting to use.

A second, even smaller class of non-backwards-compatible changes is also documented. This class of change regards those features and behaviors that have been deprecated at least since version 0.5 and have been raising warnings since their deprecation. These changes would only affect applications that are still using 0.4- or early 0.5-style APIs. As the project matures, we have fewer and fewer of these kinds of changes with 0.x level releases, which is a product of our API having ever fewer features that are less than ideal for the use cases they were meant to solve.

An array of existing functionalities have been superseded in SQLAlchemy 0.7. There's not much difference between the terms "superseded" and "deprecated", except that the former has a much weaker suggestion of the old feature would ever be removed. In 0.7, features like synonym and comparable_property, as well as all the Extension and other event classes, have been superseded. But these "superseded" features have been re-implemented such that their implementations live mostly outside of core ORM code, so their continued "hanging around" doesn't impact SQLAlchemy's ability to further streamline and refine its internals, and we expect them to remain within the API for the foreseeable future.

New Features

New Event System

SQLAlchemy started early with the MapperExtension class, which provided hooks into the persistence cycle of mappers. As SQLAlchemy quickly became more componentized, pushing mappers into a more focused configurational role, many more "extension", "listener", and "proxy" classes popped up to solve various activity-interception use cases in an ad-hoc fashion. Part of this was driven by the divergence of activities; ConnectionProxy objects wanted to provide a system of rewriting statements and parameters; AttributeExtension provided a system of replacing incoming values, and DDL objects had events that could be switched off of dialect-sensitive callables.

0.7 re-implements virtually all of these plugin points with a new, unified approach, which retains all the functionalities of the different systems, provides more flexibility and less boilerplate, performs better, and eliminates the need to learn radically different APIs for each event subsystem. The pre-existing classes MapperExtension, SessionExtension, AttributeExtension, ConnectionProxy, PoolListener as well as the DDLElement.execute_at method are deprecated and now implemented in terms of the new system - these APIs remain fully functional and are expected to remain in place for the foreseeable future.

The new approach uses named events and user-defined callables to associate activities with events. The API's look and feel was driven by such diverse sources as JQuery, Blinker, and Hibernate, and was also modified further on several occasions during conferences with dozens of users on Twitter, which appears to have a much higher response rate than the mailing list for such questions.

It also features an open-ended system of target specification that allows events to be associated with API classes, such as for all Session or Engine objects, with specific instances of API classes, such as for a specific Pool or Mapper, as well as for related objects like a user- defined class that's mapped, or something as specific as a certain attribute on instances of a particular subclass of a mapped parent class. Individual listener subsystems can apply wrappers to incoming user- defined listener functions which modify how they are called - an mapper event can receive either the instance of the object being operated upon, or its underlying InstanceState object. An attribute event can opt whether or not to have the responsibility of returning a new value.

Several systems now build upon the new event API, including the new "mutable attributes" API as well as composite attributes. The greater emphasis on events has also led to the introduction of a handful of new events, including attribute expiration and refresh operations, pickle loads/dumps operations, completed mapper construction operations.


Hybrid Attributes, implements/supersedes synonym(), comparable_property()

The "derived attributes" example has now been turned into an official extension. The typical use case for synonym() is to provide descriptor access to a mapped column; the use case for comparable_property() is to be able to return a PropComparator from any descriptor. In practice, the approach of "derived" is easier to use, more extensible, is implemented in a few dozen lines of pure Python with almost no imports, and doesn't require the ORM core to even be aware of it. The feature is now known as the "Hybrid Attributes" extension.

synonym() and comparable_property() are still part of the ORM, though their implementations have been moved outwards, building on an approach that is similar to that of the hybrid extension, so that the core ORM mapper/query/property modules aren't really aware of them otherwise.


Speed Enhancements

As is customary with all major SQLA releases, a wide pass through the internals to reduce overhead and callcounts has been made which further reduces the work needed in common scenarios. Highlights of this release include:

  • The flush process will now bundle INSERT statements into batches fed to cursor.executemany(), for rows where the primary key is already present. In particular this usually applies to the "child" table on a joined table inheritance configuration, meaning the number of calls to cursor.execute for a large bulk insert of joined- table objects can be cut in half, allowing native DBAPI optimizations to take place for those statements passed to cursor.executemany() (such as re-using a prepared statement).
  • The codepath invoked when accessing a many-to-one reference to a related object that's already loaded has been greatly simplified. The identity map is checked directly without the need to generate a new Query object first, which is expensive in the context of thousands of in-memory many-to-ones being accessed. The usage of constructed-per-call "loader" objects is also no longer used for the majority of lazy attribute loads.
  • The rewrite of composites allows a shorter codepath when mapper internals access mapped attributes within a flush.
  • New inlined attribute access functions replace the previous usage of "history" when the "save-update" and other cascade operations need to cascade among the full scope of datamembers associated with an attribute. This reduces the overhead of generating a new History object for this speed-critical operation.
  • The internals of the ExecutionContext, the object corresponding to a statement execution, have been inlined and simplified.
  • The bind_processor() and result_processor() callables generated by types for each statement execution are now cached (carefully, so as to avoid memory leaks for ad-hoc types and dialects) for the lifespan of that type, further reducing per-statement call overhead.
  • The collection of "bind processors" for a particular Compiled instance of a statement is also cached on the Compiled object, taking further advantage of the "compiled cache" used by the flush process to re-use the same compiled form of INSERT, UPDATE, DELETE statements.

A demonstration of callcount reduction including a sample benchmark script is at profiles/

Composites Rewritten

The "composite" feature has been rewritten, like synonym() and comparable_property(), to use a lighter weight implementation based on descriptors and events, rather than building into the ORM internals. This allowed the removal of some latency from the mapper/unit of work internals, and simplifies the workings of composite. The composite attribute now no longer conceals the underlying columns it builds upon, which now remain as regular attributes. Composites can also act as a proxy for relationship() as well as Column() attributes.

The major backwards-incompatible change of composites is that they no longer use the mutable=True system to detect in-place mutations. Please use the Mutation Tracking extension to establish in-place change events to existing composite usage.

:ticket:`2008` :ticket:`2024`

More succinct form of query.join(target, onclause)

The default method of issuing query.join() to a target with an explicit onclause is now:


In 0.6, this usage was considered to be an error, because join() accepts multiple arguments corresponding to multiple JOIN clauses - the two-argument form needed to be in a tuple to disambiguate between single-argument and two- argument join targets. In the middle of 0.6 we added detection and an error message for this specific calling style, since it was so common. In 0.7, since we are detecting the exact pattern anyway, and since having to type out a tuple for no reason is extremely annoying, the non- tuple method now becomes the "normal" way to do it. The "multiple JOIN" use case is exceedingly rare compared to the single join case, and multiple joins these days are more clearly represented by multiple calls to join().

The tuple form will remain for backwards compatibility.

Note that all the other forms of query.join() remain unchanged:

# ... etc

Querying with Joins


Mutation event extension, supersedes "mutable=True"

A new extension, Mutation Tracking, provides a mechanism by which user-defined datatypes can provide change events back to the owning parent or parents. The extension includes an approach for scalar database values, such as those managed by PickleType, postgresql.ARRAY, or other custom MutableType classes, as well as an approach for ORM "composites", those configured using :ref:`composite() <mapper_composite>`_.


These are implemented as an extension to the asc() and desc() operators, called nullsfirst() and nullslast().


select.distinct(), query.distinct() accepts *args for Postgresql DISTINCT ON

This was already available by passing a list of expressions to the distinct keyword argument of select(), the distinct() method of select() and Query now accept positional arguments which are rendered as DISTINCT ON when a Postgresql backend is used.




Index() can be placed inline inside of Table, __table_args__

The Index() construct can be created inline with a Table definition, using strings as column names, as an alternative to the creation of the index outside of the Table. That is:

Table('mytable', metadata,
        Column('id',Integer, primary_key=True),
        Column('name', String(50), nullable=False),
        Index('idx_name', 'name')

The primary rationale here is for the benefit of declarative __table_args__, particularly when used with mixins:

class HasNameMixin(object):
    name = Column('name', String(50), nullable=False)
    def __table_args__(cls):
        return (Index('name'), {})

class User(HasNameMixin, Base):
    __tablename__ = 'user'
    id = Column('id', Integer, primary_key=True)


Window Function SQL Construct

A "window function" provides to a statement information about the result set as it's produced. This allows criteria against various things like "row number", "rank" and so forth. They are known to be supported at least by Postgresql, SQL Server and Oracle, possibly others.

The best introduction to window functions is on Postgresql's site, where window functions have been supported since version 8.4: window.html

SQLAlchemy provides a simple construct typically invoked via an existing function clause, using the over() method, which accepts order_by and partition_by keyword arguments. Below we replicate the first example in PG's tutorial:

from sqlalchemy.sql import table, column, select, func

empsalary = table('empsalary',

s = select([

print s


SELECT empsalary.depname, empsalary.empno, empsalary.salary,
avg(empsalary.salary) OVER (PARTITION BY empsalary.depname) AS avg
FROM empsalary



execution_options() on Connection accepts "isolation_level" argument

This sets the transaction isolation level for a single Connection, until that Connection is closed and its underlying DBAPI resource returned to the connection pool, upon which the isolation level is reset back to the default. The default isolation level is set using the isolation_level argument to create_engine().

Transaction isolation support is currently only supported by the Postgresql and SQLite backends.



TypeDecorator works with integer primary key columns

A TypeDecorator which extends the behavior of Integer can be used with a primary key column. The "autoincrement" feature of Column will now recognize that the underlying database column is still an integer so that lastrowid mechanisms continue to function. The TypeDecorator itself will have its result value processor applied to newly generated primary keys, including those received by the DBAPI cursor.lastrowid accessor.

:ticket:`2005` :ticket:`2006`

TypeDecorator is present in the "sqlalchemy" import space

No longer need to import this from sqlalchemy.types, it's now mirrored in sqlalchemy.

New Dialects

Dialects have been added:

  • a MySQLdb driver for the Drizzle database:


  • support for the pymysql DBAPI:

    pymsql Notes

  • psycopg2 now works with Python 3

Behavioral Changes (Backwards Compatible)

C Extensions Build by Default

This is as of 0.7b4. The exts will build if cPython 2.xx is detected. If the build fails, such as on a windows install, that condition is caught and the non-C install proceeds. The C exts won't build if Python 3 or Pypy is used.

Query.count() simplified, should work virtually always

The very old guesswork which occurred within Query.count() has been modernized to use .from_self(). That is, query.count() is now equivalent to:


Previously, internal logic attempted to rewrite the columns clause of the query itself, and upon detection of a "subquery" condition, such as a column-based query that might have aggregates in it, or a query with DISTINCT, would go through a convoluted process of rewriting the columns clause. This logic failed in complex conditions, particularly those involving joined table inheritance, and was long obsolete by the more comprehensive .from_self() call.

The SQL emitted by query.count() is now always of the form:

SELECT count(1) AS count_1 FROM (
    SELECT AS user_id, AS user_name from user
) AS anon_1

that is, the original query is preserved entirely inside of a subquery, with no more guessing as to how count should be applied.


To emit a non-subquery form of count()

MySQL users have already reported that the MyISAM engine not surprisingly falls over completely with this simple change. Note that for a simple count() that optimizes for DBs that can't handle simple subqueries, func.count() should be used:

from sqlalchemy import func

or for count(*):

from sqlalchemy import func, literal_column

LIMIT/OFFSET clauses now use bind parameters

The LIMIT and OFFSET clauses, or their backend equivalents (i.e. TOP, ROW NUMBER OVER, etc.), use bind parameters for the actual values, for all backends which support it (most except for Sybase). This allows better query optimizer performance as the textual string for multiple statements with differing LIMIT/OFFSET are now identical.


Logging enhancements

Vinay Sajip has provided a patch to our logging system such that the "hex string" embedded in logging statements for engines and pools is no longer needed to allow the echo flag to work correctly. A new system that uses filtered logging objects allows us to maintain our current behavior of echo being local to individual engines without the need for additional identifying strings local to those engines.


Simplified polymorphic_on assignment

The population of the polymorphic_on column-mapped attribute, when used in an inheritance scenario, now occurs when the object is constructed, i.e. its __init__ method is called, using the init event. The attribute then behaves the same as any other column-mapped attribute. Previously, special logic would fire off during flush to populate this column, which prevented any user code from modifying its behavior. The new approach improves upon this in three ways: 1. the polymorphic identity is now present on the object as soon as its constructed; 2. the polymorphic identity can be changed by user code without any difference in behavior from any other column-mapped attribute; 3. the internals of the mapper during flush are simplified and no longer need to make special checks for this column.


contains_eager() chains across multiple paths (i.e. "all()")

The `contains_eager()`` modifier now will chain itself for a longer path without the need to emit individual ``contains_eager()` calls. Instead of:

session.query(A).options(contains_eager(A.b), contains_eager(A.b, B.c))

you can say:

session.query(A).options(contains_eager(A.b, B.c))


Flushing of orphans that have no parent is allowed

We've had a long standing behavior that checks for a so- called "orphan" during flush, that is, an object which is associated with a relationship() that specifies "delete- orphan" cascade, has been newly added to the session for an INSERT, and no parent relationship has been established. This check was added years ago to accommodate some test cases which tested the orphan behavior for consistency. In modern SQLA, this check is no longer needed on the Python side. The equivalent behavior of the "orphan check" is accomplished by making the foreign key reference to the object's parent row NOT NULL, where the database does its job of establishing data consistency in the same way SQLA allows most other operations to do. If the object's parent foreign key is nullable, then the row can be inserted. The "orphan" behavior runs when the object was persisted with a particular parent, and is then disassociated with that parent, leading to a DELETE statement emitted for it.


Warnings generated when collection members, scalar referents not part of the flush

Warnings are now emitted when related objects referenced via a loaded relationship() on a parent object marked as "dirty" are not present in the current Session.

The save-update cascade takes effect when objects are added to the Session, or when objects are first associated with a parent, so that an object and everything related to it are usually all present in the same Session. However, if save-update cascade is disabled for a particular relationship(), then this behavior does not occur, and the flush process does not try to correct for it, instead staying consistent to the configured cascade behavior. Previously, when such objects were detected during the flush, they were silently skipped. The new behavior is that a warning is emitted, for the purposes of alerting to a situation that more often than not is the source of unexpected behavior.


Setup no longer installs a Nose plugin

Since we moved to nose we've used a plugin that installs via setuptools, so that the nosetests script would automatically run SQLA's plugin code, necessary for our tests to have a full environment. In the middle of 0.6, we realized that the import pattern here meant that Nose's "coverage" plugin would break, since "coverage" requires that it be started before any modules to be covered are imported; so in the middle of 0.6 we made the situation worse by adding a separate sqlalchemy-nose package to the build to overcome this.

In 0.7 we've done away with trying to get nosetests to work automatically, since the SQLAlchemy module would produce a large number of nose configuration options for all usages of nosetests, not just the SQLAlchemy unit tests themselves, and the additional sqlalchemy-nose install was an even worse idea, producing an extra package in Python environments. The script in 0.7 is now the only way to run the tests with nose.


Non-Table-derived constructs can be mapped

A construct that isn't against any Table at all, like a function, can be mapped.

from sqlalchemy import select, func
from sqlalchemy.orm import mapper

class Subset(object):
selectable = select(["x", "y", "z"]).select_from(func.some_db_function()).alias()
mapper(Subset, selectable, primary_key=[selectable.c.x])


aliased() accepts FromClause elements

This is a convenience helper such that in the case a plain FromClause, such as a select, Table or join is passed to the orm.aliased() construct, it passes through to the .alias() method of that from construct rather than constructing an ORM level AliasedClass.


Session.connection(), Session.execute() accept 'bind'

This is to allow execute/connection operations to participate in the open transaction of an engine explicitly. It also allows custom subclasses of Session that implement their own get_bind() method and arguments to use those custom arguments with both the execute() and connection() methods equally.

Session.connection Session.execute


Standalone bind parameters in columns clause auto-labeled.

Bind parameters present in the "columns clause" of a select are now auto-labeled like other "anonymous" clauses, which among other things allows their "type" to be meaningful when the row is fetched, as in result row processors.

SQLite - relative file paths are normalized through os.path.abspath()

This so that a script that changes the current directory will continue to target the same location as subsequent SQLite connections are established.


MS-SQL - String/Unicode/VARCHAR/NVARCHAR/VARBINARY emit "max" for no length

On the MS-SQL backend, the String/Unicode types, and their counterparts VARCHAR/ NVARCHAR, as well as VARBINARY (:ticket:`1833`) emit "max" as the length when no length is specified. This makes it more compatible with Postgresql's VARCHAR type which is similarly unbounded when no length specified. SQL Server defaults the length on these types to '1' when no length is specified.

Behavioral Changes (Backwards Incompatible)

Note again, aside from the default mutability change, most of these changes are *extremely minor* and will not affect most users.

PickleType and ARRAY mutability turned off by default

This change refers to the default behavior of the ORM when mapping columns that have either the PickleType or postgresql.ARRAY datatypes. The mutable flag is now set to False by default. If an existing application uses these types and depends upon detection of in-place mutations, the type object must be constructed with mutable=True to restore the 0.6 behavior:

Table('mytable', metadata,
    # ....

    Column('pickled_data', PickleType(mutable=True))

The mutable=True flag is being phased out, in favor of the new Mutation Tracking extension. This extension provides a mechanism by which user-defined datatypes can provide change events back to the owning parent or parents.

The previous approach of using mutable=True does not provide for change events - instead, the ORM must scan through all mutable values present in a session and compare them against their original value for changes every time flush() is called, which is a very time consuming event. This is a holdover from the very early days of SQLAlchemy when flush() was not automatic and the history tracking system was not nearly as sophisticated as it is now.

Existing applications which use PickleType, postgresql.ARRAY or other MutableType subclasses, and require in-place mutation detection, should migrate to the new mutation tracking system, as mutable=True is likely to be deprecated in the future.


Mutability detection of composite() requires the Mutation Tracking Extension

So-called "composite" mapped attributes, those configured using the technique described at Composite Column Types, have been re-implemented such that the ORM internals are no longer aware of them (leading to shorter and more efficient codepaths in critical sections). While composite types are generally intended to be treated as immutable value objects, this was never enforced. For applications that use composites with mutability, the Mutation Tracking extension offers a base class which establishes a mechanism for user-defined composite types to send change event messages back to the owning parent or parents of each object.

Applications which use composite types and rely upon in- place mutation detection of these objects should either migrate to the "mutation tracking" extension, or change the usage of the composite types such that in-place changes are no longer needed (i.e., treat them as immutable value objects).

SQLite - the SQLite dialect now uses NullPool for file-based databases

This change is 99.999% backwards compatible, unless you are using temporary tables across connection pool connections.

A file-based SQLite connection is blazingly fast, and using NullPool means that each call to Engine.connect creates a new pysqlite connection.

Previously, the SingletonThreadPool was used, which meant that all connections to a certain engine in a thread would be the same connection. It's intended that the new approach is more intuitive, particularly when multiple connections are used.

SingletonThreadPool is still the default engine when a :memory: database is used.

Note that this change breaks temporary tables used across Session commits, due to the way SQLite handles temp tables. See the note at temporary-tables-with-sqlite if temporary tables beyond the scope of one pool connection are desired.


Session.merge() checks version ids for versioned mappers

Session.merge() will check the version id of the incoming state against that of the database, assuming the mapping uses version ids and incoming state has a version_id assigned, and raise StaleDataError if they don't match. This is the correct behavior, in that if incoming state contains a stale version id, it should be assumed the state is stale.

If merging data into a versioned state, the version id attribute can be left undefined, and no version check will take place.

This check was confirmed by examining what Hibernate does - both the merge() and the versioning features were originally adapted from Hibernate.


Tuple label names in Query Improved

This improvement is potentially slightly backwards incompatible for an application that relied upon the old behavior.

Given two mapped classes Foo and Bar each with a column spam:

qa = session.query(Foo.spam)
qb = session.query(Bar.spam)

qu = qa.union(qb)

The name given to the single column yielded by qu will be spam. Previously it would be something like foo_spam due to the way the union would combine things, which is inconsistent with the name spam in the case of a non-unioned query.


Mapped column attributes reference the most specific column first

This is a change to the behavior involved when a mapped column attribute references multiple columns, specifically when dealing with an attribute on a joined-table subclass that has the same name as that of an attribute on the superclass.

Using declarative, the scenario is this:

class Parent(Base):
    __tablename__ = 'parent'
    id = Column(Integer, primary_key=True)

class Child(Parent):
   __tablename__ = 'child'
    id = Column(Integer, ForeignKey(''), primary_key=True)

Above, the attribute refers to both the column as well as - this due to the name of the attribute. If it were named differently on the class, such as Child.child_id, it then maps distinctly to, with being the same attribute as

When the id attribute is made to reference both and, it stores them in an ordered list. An expression such as then refers to just one of those columns when rendered. Up until 0.6, this column would be In 0.7, it is the less surprising

The legacy of this behavior deals with behaviors and restrictions of the ORM that don't really apply anymore; all that was needed was to reverse the order.

A primary advantage of this approach is that it's now easier to construct primaryjoin expressions that refer to the local column:

class Child(Parent):
   __tablename__ = 'child'
    id = Column(Integer, ForeignKey(''), primary_key=True)
    some_related = relationship("SomeRelated",

class SomeRelated(Base):
   __tablename__ = 'some_related'
    id = Column(Integer, primary_key=True)
    child_id = Column(Integer, ForeignKey(''))

Prior to 0.7 the expression would reference, and it would be necessary to map to a distinct attribute.

It also means that a query like this one changes its behavior:

session.query(Parent).filter( > 7)

In 0.6, this would render:

SELECT AS parent_id
FROM parent
WHERE > :id_1

in 0.7, you get:

SELECT AS parent_id
FROM parent, child
WHERE > :id_1

which you'll note is a cartesian product - this behavior is now equivalent to that of any other attribute that is local to Child. The with_polymorphic() method, or a similar strategy of explicitly joining the underlying Table objects, is used to render a query against all Parent objects with criteria against Child, in the same manner as that of 0.5 and 0.6:

print s.query(Parent).with_polymorphic([Child]).filter( > 7)

Which on both 0.6 and 0.7 renders:

SELECT AS parent_id, AS child_id
FROM parent LEFT OUTER JOIN child ON =
WHERE > :id_1

Another effect of this change is that a joined-inheritance load across two tables will populate from the child table's value, not that of the parent table. An unusual case is that a query against "Parent" using with_polymorphic="*" issues a query against "parent", with a LEFT OUTER JOIN to "child". The row is located in "Parent", sees the polymorphic identity corresponds to "Child", but suppose the actual row in "child" has been deleted. Due to this corruption, the row comes in with all the columns corresponding to "child" set to NULL - this is now the value that gets populated, not the one in the parent table.


Mapping to joins with two or more same-named columns requires explicit declaration

This is somewhat related to the previous change in :ticket:`1892`. When mapping to a join, same-named columns must be explicitly linked to mapped attributes, i.e. as described in Mapping a Class Against Multiple Tables.

Given two tables foo and bar, each with a primary key column id, the following now produces an error:

foobar = foo.join(bar,
mapper(FooBar, foobar)

This because the mapper() refuses to guess what column is the primary representation of - is it or is it ? The attribute must be explicit:

foobar = foo.join(bar,
mapper(FooBar, foobar, properties={


Mapper requires that polymorphic_on column be present in the mapped selectable

This is a warning in 0.6, now an error in 0.7. The column given for polymorphic_on must be in the mapped selectable. This to prevent some occasional user errors such as:

mapper(SomeClass, sometable,

where above the polymorphic_on needs to be on a sometable column, in this case perhaps sometable.c.some_lookup_id. There are also some "polymorphic union" scenarios where similar mistakes sometimes occur.

Such a configuration error has always been "wrong", and the above mapping doesn't work as specified - the column would be ignored. It is however potentially backwards incompatible in the rare case that an application has been unknowingly relying upon this behavior.


DDL() constructs now escape percent signs

Previously, percent signs in DDL() strings would have to be escaped, i.e. %% depending on DBAPI, for those DBAPIs that accept pyformat or format binds (i.e. psycopg2, mysql-python), which was inconsistent versus text() constructs which did this automatically. The same escaping now occurs for DDL() as for text().


Table.c / MetaData.tables refined a bit, don't allow direct mutation

Another area where some users were tinkering around in such a way that doesn't actually work as expected, but still left an exceedingly small chance that some application was relying upon this behavior, the construct returned by the .c attribute on Table and the .tables attribute on MetaData is explicitly non-mutable. The "mutable" version of the construct is now private. Adding columns to .c involves using the append_column() method of Table, which ensures things are associated with the parent Table in the appropriate way; similarly, MetaData.tables has a contract with the Table objects stored in this dictionary, as well as a little bit of new bookkeeping in that a set() of all schema names is tracked, which is satisfied only by using the public Table constructor as well as Table.tometadata().

It is of course possible that the ColumnCollection and dict collections consulted by these attributes could someday implement events on all of their mutational methods such that the appropriate bookkeeping occurred upon direct mutation of the collections, but until someone has the motivation to implement all that along with dozens of new unit tests, narrowing the paths to mutation of these collections will ensure no application is attempting to rely upon usages that are currently not supported.

:ticket:`1893` :ticket:`1917`

server_default consistently returns None for all inserted_primary_key values

Established consistency when server_default is present on an Integer PK column. SQLA doesn't pre-fetch these, nor do they come back in cursor.lastrowid (DBAPI). Ensured all backends consistently return None in result.inserted_primary_key for these - some backends may have returned a value previously. Using a server_default on a primary key column is extremely unusual. If a special function or SQL expression is used to generate primary key defaults, this should be established as a Python-side "default" instead of server_default.

Regarding reflection for this case, reflection of an int PK col with a server_default sets the "autoincrement" flag to False, except in the case of a PG SERIAL col where we detected a sequence default.

:ticket:`2020` :ticket:`2021`

The sqlalchemy.exceptions alias in sys.modules is removed

For a few years we've added the string sqlalchemy.exceptions to sys.modules, so that a statement like "import sqlalchemy.exceptions" would work. The name of the core exceptions module has been exc for a long time now, so the recommended import for this module is:

from sqlalchemy import exc

The exceptions name is still present in "sqlalchemy" for applications which might have said from sqlalchemy import exceptions, but they should also start using the exc name.

Query Timing Recipe Changes

While not part of SQLAlchemy itself, it's worth mentioning that the rework of the ConnectionProxy into the new event system means it is no longer appropriate for the "Timing all Queries" recipe. Please adjust query-timers to use the before_cursor_execute() and after_cursor_execute() events, demonstrated in the updated recipe UsageRecipes/Profiling.

Deprecated API

Default constructor on types will not accept arguments

Simple types like Integer, Date etc. in the core types module don't accept arguments. The default constructor that accepts/ignores a catchall \*args, \**kwargs is restored as of 0.7b4/0.7.0, but emits a deprecation warning.

If arguments are being used with a core type like Integer, it may be that you intended to use a dialect specific type, such as sqlalchemy.dialects.mysql.INTEGER which does accept a "display_width" argument for example.

compile_mappers() renamed configure_mappers(), simplified configuration internals

This system slowly morphed from something small, implemented local to an individual mapper, and poorly named into something that's more of a global "registry-" level function and poorly named, so we've fixed both by moving the implementation out of Mapper altogether and renaming it to configure_mappers(). It is of course normally not needed for an application to call configure_mappers() as this process occurs on an as-needed basis, as soon as the mappings are needed via attribute or query access.


Core listener/proxy superseded by event listeners

PoolListener, ConnectionProxy, DDLElement.execute_at are superseded by event.listen(), using the PoolEvents, EngineEvents, DDLEvents dispatch targets, respectively.

ORM extensions superseded by event listeners

MapperExtension, AttributeExtension, SessionExtension are superseded by event.listen(), using the MapperEvents/InstanceEvents, AttributeEvents, SessionEvents, dispatch targets, respectively.

Sending a string to 'distinct' in select() for MySQL should be done via prefixes

This obscure feature allows this pattern with the MySQL backend:

select([mytable], distinct='ALL', prefixes=['HIGH_PRIORITY'])

The prefixes keyword or prefix_with() method should be used for non-standard or unusual prefixes:

select([mytable]).prefix_with('HIGH_PRIORITY', 'ALL')

useexisting superseded by extend_existing and keep_existing

The useexisting flag on Table has been superseded by a new pair of flags keep_existing and extend_existing. extend_existing is equivalent to useexisting - the existing Table is returned, and additional constructor elements are added. With keep_existing, the existing Table is returned, but additional constructor elements are not added - these elements are only applied when the Table is newly created.

Backwards Incompatible API Changes

Callables passed to bindparam() don't get evaluated - affects the Beaker example


Note this affects the Beaker caching example, where the workings of the _params_from_query() function needed a slight adjustment. If you're using code from the Beaker example, this change should be applied.

types.type_map is now private, types._type_map

We noticed some users tapping into this dictionary inside of sqlalchemy.types as a shortcut to associating Python types with SQL types. We can't guarantee the contents or format of this dictionary, and additionally the business of associating Python types in a one-to-one fashion has some grey areas that should are best decided by individual applications, so we've underscored this attribute.


Renamed the alias keyword arg of standalone alias() function to name

This so that the keyword argument name matches that of the alias() methods on all FromClause objects as well as the name argument on Query.subquery().

Only code that uses the standalone alias() function, and not the method bound functions, and passes the alias name using the explicit keyword name alias, and not positionally, would need modification here.

Non-public Pool methods underscored

All methods of Pool and subclasses which are not intended for public use have been renamed with underscores. That they were not named this way previously was a bug.

Pooling methods now underscored or removed:

Pool.create_connection() -> Pool._create_connection()

Pool.do_get() -> Pool._do_get()

Pool.do_return_conn() -> Pool._do_return_conn()

Pool.do_return_invalid() -> removed, was not used

Pool.return_conn() -> Pool._return_conn()

Pool.get() -> Pool._get(), public API is Pool.connect()

SingletonThreadPool.cleanup() -> _cleanup()

SingletonThreadPool.dispose_local() -> removed, use conn.invalidate()


Previously Deprecated, Now Removed

Query.join(), Query.outerjoin(), eagerload(), eagerload_all(), others no longer allow lists of attributes as arguments

Passing a list of attributes or attribute names to Query.join, eagerload(), and similar has been deprecated since 0.5:

# old way, deprecated since 0.5
session.query(Houses).join([Houses.rooms, Room.closets])
session.query(Houses).options(eagerload_all([Houses.rooms, Room.closets]))

These methods all accept *args as of the 0.5 series:

# current way, in place since 0.5
session.query(Houses).join(Houses.rooms, Room.closets)
session.query(Houses).options(eagerload_all(Houses.rooms, Room.closets))

ScopedSession.mapper is removed

This feature provided a mapper extension which linked class- based functionality with a particular ScopedSession, in particular providing the behavior such that new object instances would be automatically associated with that session. The feature was overused by tutorials and frameworks which led to great user confusion due to its implicit behavior, and was deprecated in 0.5.5. Techniques for replicating its functionality are at [wiki:UsageRecipes/SessionAwareMapper]