About this Document
This document describes changes between SQLAlchemy version 0.5, last released January 16, 2010, and SQLAlchemy version 0.6, last released May 5, 2012.
Document date: June 6, 2010
This guide documents API changes which affect users migrating their applications from the 0.5 series of SQLAlchemy to 0.6. Note that SQLAlchemy 0.6 removes some behaviors which were deprecated throughout the span of the 0.5 series, and also deprecates more behaviors specific to 0.5.
Dialect modules are now broken up into distinct subcomponents, within the scope of a single database backend. Dialect implementations are now in the sqlalchemy.dialects package. The sqlalchemy.databases package still exists as a placeholder to provide some level of backwards compatibility for simple imports.
For each supported database, a sub-package exists within sqlalchemy.dialects where several files are contained. Each package contains a module called base.py which defines the specific SQL dialect used by that database. It also contains one or more “driver” modules, each one corresponding to a specific DBAPI - these files are named corresponding to the DBAPI itself, such as pysqlite, cx_oracle, or pyodbc. The classes used by SQLAlchemy dialects are first declared in the base.py module, defining all behavioral characteristics defined by the database. These include capability mappings, such as “supports sequences”, “supports returning”, etc., type definitions, and SQL compilation rules. Each “driver” module in turn provides subclasses of those classes as needed which override the default behavior to accommodate the additional features, behaviors, and quirks of that DBAPI. For DBAPIs that support multiple backends (pyodbc, zxJDBC, mxODBC), the dialect module will use mixins from the sqlalchemy.connectors package, which provide functionality common to that DBAPI across all backends, most typically dealing with connect arguments. This means that connecting using pyodbc, zxJDBC or mxODBC (when implemented) is extremely consistent across supported backends.
The URL format used by create_engine() has been enhanced to handle any number of DBAPIs for a particular backend, using a scheme that is inspired by that of JDBC. The previous format still works, and will select a “default” DBAPI implementation, such as the Postgresql URL below that will use psycopg2:
create_engine('postgresql://scott:tiger@localhost/test')However to specify a specific DBAPI backend such as pg8000, add it to the “protocol” section of the URL using a plus sign “+”:
create_engine('postgresql+pg8000://scott:tiger@localhost/test')Important Dialect Links:
Other notes regarding dialects:
The import structure of dialects has changed. Each dialect now exports its base “dialect” class as well as the full set of SQL types supported on that dialect via sqlalchemy.dialects.<name>. For example, to import a set of PG types:
from sqlalchemy.dialects.postgresql import INTEGER, BIGINT, SMALLINT,\
VARCHAR, MACADDR, DATE, BYTEAAbove, INTEGER is actually the plain INTEGER type from sqlalchemy.types, but the PG dialect makes it available in the same way as those types which are specific to PG, such as BYTEA and MACADDR.
There’s one quite significant behavioral change to the expression language which may affect some applications. The boolean value of Python boolean expressions, i.e. ==, !=, and similar, now evaluates accurately with regards to the two clause objects being compared.
As we know, comparing a ClauseElement to any other object returns another ClauseElement:
>>> from sqlalchemy.sql import column
>>> column('foo') == 5
<sqlalchemy.sql.expression._BinaryExpression object at 0x1252490>This so that Python expressions produce SQL expressions when converted to strings:
>>> str(column('foo') == 5)
'foo = :foo_1'But what happens if we say this?
>>> if column('foo') == 5:
... print "yes"
...In previous versions of SQLAlchemy, the returned _BinaryExpression was a plain Python object which evaluated to True. Now it evaluates to whether or not the actual ClauseElement should have the same hash value as to that being compared. Meaning:
>>> bool(column('foo') == 5)
False
>>> bool(column('foo') == column('foo'))
False
>>> c = column('foo')
>>> bool(c == c)
True
>>>That means code such as the following:
if expression:
print "the expression is:", expressionWould not evaluate if expression was a binary clause. Since the above pattern should never be used, the base ClauseElement now raises an exception if called in a boolean context:
>>> bool(c)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
...
raise TypeError("Boolean value of this clause is not defined")
TypeError: Boolean value of this clause is not definedCode that wants to check for the presence of a ClauseElement expression should instead say:
if expression is not None:
print "the expression is:", expressionKeep in mind, this applies to Table and Column objects too.
The rationale for the change is twofold:
An “executemany” in SQLAlchemy corresponds to a call to execute(), passing along a collection of bind parameter sets:
connection.execute(table.insert(), {'data':'row1'}, {'data':'row2'}, {'data':'row3'})When the Connection object sends off the given insert() construct for compilation, it passes to the compiler the keynames present in the first set of binds passed along to determine the construction of the statement’s VALUES clause. Users familiar with this construct will know that additional keys present in the remaining dictionaries don’t have any impact. What’s different now is that all subsequent dictionaries need to include at least every key that is present in the first dictionary. This means that a call like this no longer works:
connection.execute(table.insert(),
{'timestamp':today, 'data':'row1'},
{'timestamp':today, 'data':'row2'},
{'data':'row3'})Because the third row does not specify the ‘timestamp’ column. Previous versions of SQLAlchemy would simply insert NULL for these missing columns. However, if the timestamp column in the above example contained a Python-side default value or function, it would not be used. This because the “executemany” operation is optimized for maximum performance across huge numbers of parameter sets, and does not attempt to evaluate Python-side defaults for those missing keys. Because defaults are often implemented either as SQL expressions which are embedded inline with the INSERT statement, or are server side expressions which again are triggered based on the structure of the INSERT string, which by definition cannot fire off conditionally based on each parameter set, it would be inconsistent for Python side defaults to behave differently vs. SQL/server side defaults. (SQL expression based defaults are embedded inline as of the 0.5 series, again to minimize the impact of huge numbers of parameter sets).
SQLAlchemy 0.6 therefore establishes predictable consistency by forbidding any subsequent parameter sets from leaving any fields blank. That way, there’s no more silent failure of Python side default values and functions, which additionally are allowed to remain consistent in their behavior versus SQL and server side defaults.
A rule that was designed to help SQLite has been removed, that of the first compound element within another compound (such as, a union() inside of an except_()) wouldn’t be parenthesized. This is inconsistent and produces the wrong results on Postgresql, which has precedence rules regarding INTERSECTION, and its generally a surprise. When using complex composites with SQLite, you now need to turn the first element into a subquery (which is also compatible on PG). A new example is in the SQL expression tutorial at the end of [http://www.sqlalchemy.org/docs/06/sqlexpression.html #unions-and-other-set-operations]. See #1665 and r6690 for more background.
The ResultProxy and related elements, including most common “row processing” functions such as unicode conversion, numerical/boolean conversions and date parsing, have been re-implemented as optional C extensions for the purposes of performance. This represents the beginning of SQLAlchemy’s path to the “dark side” where we hope to continue improving performance by reimplementing critical sections in C. The extensions can be built by specifying --with-cextensions, i.e. python setup.py --with- cextensions install.
The extensions have the most dramatic impact on result fetching using direct ResultProxy access, i.e. that which is returned by engine.execute(), connection.execute(), or session.execute(). Within results returned by an ORM Query object, result fetching is not as high a percentage of overhead, so ORM performance improves more modestly, and mostly in the realm of fetching large result sets. The performance improvements highly depend on the dbapi in use and on the syntax used to access the columns of each row (eg row['name'] is much faster than row.name). The current extensions have no impact on the speed of inserts/updates/deletes, nor do they improve the latency of SQL execution, that is, an application that spends most of its time executing many statements with very small result sets will not see much improvement.
Performance has been improved in 0.6 versus 0.5 regardless of the extensions. A quick overview of what connecting and fetching 50,000 rows looks like with SQLite, using mostly direct SQLite access, a ResultProxy, and a simple mapped ORM object:
sqlite select/native: 0.260s
0.6 / C extension
sqlalchemy.sql select: 0.360s
sqlalchemy.orm fetch: 2.500s
0.6 / Pure Python
sqlalchemy.sql select: 0.600s
sqlalchemy.orm fetch: 3.000s
0.5 / Pure Python
sqlalchemy.sql select: 0.790s
sqlalchemy.orm fetch: 4.030s
Above, the ORM fetches the rows 33% faster than 0.5 due to in-python performance enhancements. With the C extensions we get another 20%. However, ResultProxy fetches improve by 67% with the C extension versus not. Other tests report as much as a 200% speed improvement for some scenarios, such as those where lots of string conversions are occurring.
The sqlalchemy.schema package has received some long- needed attention. The most visible change is the newly expanded DDL system. In SQLAlchemy, it was possible since version 0.5 to create custom DDL strings and associate them with tables or metadata objects:
from sqlalchemy.schema import DDL
DDL('CREATE TRIGGER users_trigger ...').execute_at('after-create', metadata)Now the full suite of DDL constructs are available under the same system, including those for CREATE TABLE, ADD CONSTRAINT, etc.:
from sqlalchemy.schema import Constraint, AddConstraint
AddContraint(CheckConstraint("value > 5")).execute_at('after-create', mytable)Additionally, all the DDL objects are now regular ClauseElement objects just like any other SQLAlchemy expression object:
from sqlalchemy.schema import CreateTable
create = CreateTable(mytable)
# dumps the CREATE TABLE as a string
print create
# executes the CREATE TABLE statement
engine.execute(create)and using the sqlalchemy.ext.compiler extension you can make your own:
from sqlalchemy.schema import DDLElement
from sqlalchemy.ext.compiler import compiles
class AlterColumn(DDLElement):
def __init__(self, column, cmd):
self.column = column
self.cmd = cmd
@compiles(AlterColumn)
def visit_alter_column(element, compiler, **kw):
return "ALTER TABLE %s ALTER COLUMN %s %s ..." % (
element.column.table.name,
element.column.name,
element.cmd
)
engine.execute(AlterColumn(table.c.mycolumn, "SET DEFAULT 'test'"))The schema package has also been greatly streamlined. Many options and methods which were deprecated throughout 0.5 have been removed. Other little known accessors and methods have also been removed.
These should be constructed declaratively (i.e. in one construction).
At the expense of a few extra method calls here and there, you can set log levels for INFO and DEBUG after an engine, pool, or mapper has been created, and logging will commence. The isEnabledFor(INFO) method is now called per-Connection and isEnabledFor(DEBUG) per-ResultProxy if already enabled on the parent connection. Pool logging sends to log.info() and log.debug() with no check - note that pool checkout/checkin is typically once per transaction.
The reflection system, which allows reflection of table columns via Table('sometable', metadata, autoload=True) has been opened up into its own fine-grained API, which allows direct inspection of database elements such as tables, columns, constraints, indexes, and more. This API expresses return values as simple lists of strings, dictionaries, and TypeEngine objects. The internals of autoload=True now build upon this system such that the translation of raw database information into sqlalchemy.schema constructs is centralized and the contract of individual dialects greatly simplified, vastly reducing bugs and inconsistencies across different backends.
To use an inspector:
from sqlalchemy.engine.reflection import Inspector
insp = Inspector.from_engine(my_engine)
print insp.get_schema_names()the from_engine() method will in some cases provide a backend-specific inspector with additional capabilities, such as that of Postgresql which provides a get_table_oid() method:
my_engine = create_engine('postgresql://...')
pg_insp = Inspector.from_engine(my_engine)
print pg_insp.get_table_oid('my_table')The insert(), update() and delete() constructs now support a returning() method, which corresponds to the SQL RETURNING clause as supported by Postgresql, Oracle, MS-SQL, and Firebird. It is not supported for any other backend at this time.
Given a list of column expressions in the same manner as that of a select() construct, the values of these columns will be returned as a regular result set:
result = connection.execute(
table.insert().values(data='some data').returning(table.c.id, table.c.timestamp)
)
row = result.first()
print "ID:", row['id'], "Timestamp:", row['timestamp']The implementation of RETURNING across the four supported backends varies wildly, in the case of Oracle requiring an intricate usage of OUT parameters which are re-routed into a “mock” result set, and in the case of MS-SQL using an awkward SQL syntax. The usage of RETURNING is subject to limitations:
RETURNING is also used automatically by SQLAlchemy, when available and when not otherwise specified by an explicit returning() call, to fetch the value of newly generated primary key values for single-row INSERT statements. This means there’s no more “SELECT nextval(sequence)” pre- execution for insert statements where the primary key value is required. Truth be told, implicit RETURNING feature does incur more method overhead than the old “select nextval()” system, which used a quick and dirty cursor.execute() to get at the sequence value, and in the case of Oracle requires additional binding of out parameters. So if method/protocol overhead is proving to be more expensive than additional database round trips, the feature can be disabled by specifying implicit_returning=False to create_engine().
The type system has been completely reworked behind the scenes to provide two goals:
Highlights of these changes include:
As more DBAPIs support returning Python unicode objects directly, the base dialect now performs a check upon the first connection which establishes whether or not the DBAPI returns a Python unicode object for a basic select of a VARCHAR value. If so, the String type and all subclasses (i.e. Text, Unicode, etc.) will skip the “unicode” check/conversion step when result rows are received. This offers a dramatic performance increase for large result sets. The “unicode mode” currently is known to work with:
Other types may choose to disable unicode processing as needed, such as the NVARCHAR type when used with MS-SQL.
In particular, if porting an application based on a DBAPI that formerly returned non-unicode strings, the “native unicode” mode has a plainly different default behavior - columns that are declared as String or VARCHAR now return unicode by default whereas they would return strings before. This can break code which expects non-unicode strings. The psycopg2 “native unicode” mode can be disabled by passing use_native_unicode=False to create_engine().
A more general solution for string columns that explicitly do not want a unicode object is to use a TypeDecorator that converts unicode back to utf-8, or whatever is desired:
class UTF8Encoded(TypeDecorator):
"""Unicode type which coerces to utf-8."""
impl = sa.VARCHAR
def process_result_value(self, value, dialect):
if isinstance(value, unicode):
value = value.encode('utf-8')
return valueNote that the assert_unicode flag is now deprecated. SQLAlchemy allows the DBAPI and backend database in use to handle Unicode parameters when available, and does not add operational overhead by checking the incoming type; modern systems like sqlite and Postgresql will raise an encoding error on their end if invalid data is passed. In those cases where SQLAlchemy does need to coerce a bind parameter from Python Unicode to an encoded string, or when the Unicode type is used explicitly, a warning is raised if the object is a bytestring. This warning can be suppressed or converted to an exception using the Python warnings filter documented at: http://docs.python.org/library/warnings.html
We now have an Enum in the types module. This is a string type that is given a collection of “labels” which constrain the possible values given to those labels. By default, this type generates a VARCHAR using the size of the largest label, and applies a CHECK constraint to the table within the CREATE TABLE statement. When using MySQL, the type by default uses MySQL’s ENUM type, and when using Postgresql the type will generate a user defined type using CREATE TYPE <mytype> AS ENUM. In order to create the type using Postgresql, the name parameter must be specified to the constructor. The type also accepts a native_enum=False option which will issue the VARCHAR/CHECK strategy for all databases. Note that Postgresql ENUM types currently don’t work with pg8000 or zxjdbc.
Reflection now returns the most specific type possible from the database. That is, if you create a table using String, then reflect it back, the reflected column will likely be VARCHAR. For dialects that support a more specific form of the type, that’s what you’ll get. So a Text type would come back as oracle.CLOB on Oracle, a LargeBinary might be an mysql.MEDIUMBLOB etc. The obvious advantage here is that reflection preserves as much information possible from what the database had to say.
Some applications that deal heavily in table metadata may wish to compare types across reflected tables and/or non- reflected tables. There’s a semi-private accessor available on TypeEngine called _type_affinity and an associated comparison helper _compare_type_affinity. This accessor returns the “generic” types class which the type corresponds to:
>>> String(50)._compare_type_affinity(postgresql.VARCHAR(50))
True
>>> Integer()._compare_type_affinity(mysql.REAL)
FalseThe usual “generic” types are still the general system in use, i.e. String, Float, DateTime. There’s a few changes there:
Upgrading an ORM application from 0.5 to 0.6 should require little to no changes, as the ORM’s behavior remains almost identical. There are some default argument and name changes, and some loading behaviors have been improved.
The internals for the unit of work, primarily topological.py and unitofwork.py, have been completely rewritten and are vastly simplified. This should have no impact on usage, as all existing behavior during flush has been maintained exactly (or at least, as far as it is exercised by our testsuite and the handful of production environments which have tested it heavily). The performance of flush() now uses 20-30% fewer method calls and should also use less memory. The intent and flow of the source code should now be reasonably easy to follow, and the architecture of the flush is fairly open-ended at this point, creating room for potential new areas of sophistication. The flush process no longer has any reliance on recursion so flush plans of arbitrary size and complexity can be flushed. Additionally, the mapper’s “save” process, which issues INSERT and UPDATE statements, now caches the “compiled” form of the two statements so that callcounts are further dramatically reduced with very large flushes.
Any changes in behavior observed with flush versus earlier versions of 0.6 or 0.5 should be reported to us ASAP - we’ll make sure no functionality is lost.
This to solve the long running issue that “relation” means a “table or derived table” in relational algebra terms. The relation() name, which is less typing, will hang around for the foreseeable future so this change should be entirely painless.
A new kind of eager loading is added called “subquery” loading. This is a load that emits a second SQL query immediately after the first which loads full collections for all the parents in the first query, joining upwards to the parent using INNER JOIN. Subquery loading is used simlarly to the current joined-eager loading, using the `subqueryload()`` and ``subqueryload_all()`` options as well as the ``lazy='subquery'`` setting on ``relationship()`. The subquery load is usually much more efficient for loading many larger collections as it uses INNER JOIN unconditionally and also doesn’t re-load parent rows.
To make room for the new subquery load feature, the existing `eagerload()``/``eagerload_all()`` options are now superseded by ``joinedload()`` and ``joinedload_all()``. The old names will hang around for the foreseeable future just like ``relation()`.
Continuing on the theme of loader strategies opened up, the standard keywords for the `lazy`` option on ``relationship()`` are now ``select`` for lazy loading (via a SELECT issued on attribute access), ``joined`` for joined-eager loading, ``subquery`` for subquery-eager loading, ``noload`` for no loading should occur, and ``dynamic`` for a “dynamic” relationship. The old ``True``, ``False``, ``None` arguments are still accepted with the identical behavior as before.
Joined-eagerly loaded scalars and collections can now be instructed to use INNER JOIN instead of OUTER JOIN. On Postgresql this is observed to provide a 300-600% speedup on some queries. Set this flag for any many-to-one which is on a NOT NULLable foreign key, and similarly for any collection where related items are guaranteed to exist.
At mapper level:
mapper(Child, child)
mapper(Parent, parent, properties={
'child':relationship(Child, lazy='joined', innerjoin=True)
})At query time level:
session.query(Parent).options(joinedload(Parent.child, innerjoin=True)).all()The innerjoin=True flag at the relationship() level will also take effect for any joinedload() option which does not override the value.
many-to-one relations now fire off a lazyload in fewer cases, including in most cases will not fetch the “old” value when a new one is replaced.
many-to-one relation to a joined-table subclass now uses get() for a simple load (known as the “use_get” condition), i.e. Related->``Sub(Base)``, without the need to redefine the primaryjoin condition in terms of the base table. [ticket:1186]
specifying a foreign key with a declarative column, i.e. ForeignKey(MyRelatedClass.id) doesn’t break the “use_get” condition from taking place [ticket:1492]
relationship(), joinedload(), and joinedload_all() now feature an option called “innerjoin”. Specify True or False to control whether an eager join is constructed as an INNER or OUTER join. Default is False as always. The mapper options will override whichever setting is specified on relationship(). Should generally be set for many-to-one, not nullable foreign key relations to allow improved join performance. [ticket:1544]
the behavior of joined eager loading such that the main query is wrapped in a subquery when LIMIT/OFFSET are present now makes an exception for the case when all eager loads are many-to-one joins. In those cases, the eager joins are against the parent table directly along with the limit/offset without the extra overhead of a subquery, since a many-to-one join does not add rows to the result.
For example, in 0.5 this query:
session.query(Address).options(eagerload(Address.user)).limit(10)would produce SQL like:
SELECT * FROM
(SELECT * FROM addresses LIMIT 10) AS anon_1
LEFT OUTER JOIN users AS users_1 ON users_1.id = anon_1.addresses_user_id
This because the presence of any eager loaders suggests that some or all of them may relate to multi-row collections, which would necessitate wrapping any kind of rowcount-sensitive modifiers like LIMIT inside of a subquery.
In 0.6, that logic is more sensitive and can detect if all eager loaders represent many-to-ones, in which case the eager joins don’t affect the rowcount:
SELECT * FROM addresses LEFT OUTER JOIN users AS users_1 ON users_1.id = addresses.user_id LIMIT 10
A joined table inheritance config where the child table has a PK that foreign keys to the parent PK can now be updated on a CASCADE-capable database like Postgresql. mapper() now has an option passive_updates=True which indicates this foreign key is updated automatically. If on a non-cascading database like SQLite or MySQL/MyISAM, set this flag to False. A future feature enhancement will try to get this flag to be auto-configuring based on dialect/table style in use.
A promising new example of Beaker integration is in examples/beaker_caching. This is a straightforward recipe which applies a Beaker cache within the result- generation engine of Query. Cache parameters are provided via query.options(), and allows full control over the contents of the cache. SQLAlchemy 0.6 includes improvements to the Session.merge() method to support this and similar recipes, as well as to provide significantly improved performance in most scenarios.
Most elements that were deprecated throughout 0.5 and raised deprecation warnings have been removed (with a few exceptions). All elements that were marked “pending deprecation” are now deprecated and will raise a warning upon use.
SQLSoup has been modernized and updated to reflect common 0.5/0.6 capabilities, including well defined session integration. Please read the new docs at [http://www.sqlalc hemy.org/docs/06/reference/ext/sqlsoup.html].
The DeclarativeMeta (default metaclass for declarative_base) previously allowed subclasses to modify dict_ to add class attributes (e.g. columns). This no longer works, the DeclarativeMeta constructor now ignores dict_. Instead, the class attributes should be assigned directly, e.g. cls.id=Column(...), or the MixIn class approach should be used instead of the metaclass approach.