0.6 Changelog

0.6.9

Released: Sat May 05 2012

general

  • [general]

    Adjusted the “importlater” mechanism, which is used internally to resolve import cycles, such that the usage of __import__ is completed when the import of sqlalchemy or sqlalchemy.orm is done, thereby avoiding any usage of __import__ after the application starts new threads, fixes.

    References: #2279

orm

  • [orm] [bug]

    fixed inappropriate evaluation of user-mapped object in a boolean context within query.get().

    References: #2310

  • [orm] [bug]

    Fixed the error formatting raised when a tuple is inadvertently passed to session.query().

    References: #2297

  • [orm]

    Fixed bug whereby the source clause used by query.join() would be inconsistent if against a column expression that combined multiple entities together.

    References: #2197

  • [orm]

    Fixed bug apparent only in Python 3 whereby sorting of persistent + pending objects during flush would produce an illegal comparison, if the persistent object primary key is not a single integer.

    References: #2228

  • [orm]

    Fixed bug where query.join() + aliased=True from a joined-inh structure to itself on relationship() with join condition on the child table would convert the lead entity into the joined one inappropriately.

    References: #2234

  • [orm]

    Fixed bug whereby mapper.order_by attribute would be ignored in the “inner” query within a subquery eager load. .

    References: #2287

  • [orm]

    Fixed bug whereby if a mapped class redefined __hash__() or __eq__() to something non-standard, which is a supported use case as SQLA should never consult these, the methods would be consulted if the class was part of a “composite” (i.e. non-single-entity) result set.

    References: #2215

  • [orm]

    Fixed subtle bug that caused SQL to blow up if: column_property() against subquery + joinedload + LIMIT + order by the column property() occurred. .

    References: #2188

  • [orm]

    The join condition produced by with_parent as well as when using a “dynamic” relationship against a parent will generate unique bindparams, rather than incorrectly repeating the same bindparam. .

    References: #2207

  • [orm]

    Repaired the “no statement condition” assertion in Query which would attempt to raise if a generative method were called after from_statement() were called..

    References: #2199

  • [orm]

    Cls.column.collate(“some collation”) now works.

    References: #1776

examples

  • [examples]

    Adjusted dictlike-polymorphic.py example to apply the CAST such that it works on PG, other databases.

    References: #2266

engine

  • [engine]

    Backported the fix for introduced in 0.7.4, which ensures that the connection is in a valid state before attempting to call rollback()/prepare()/release() on savepoint and two-phase transactions.

    References: #2317

sql

  • [sql]

    Fixed two subtle bugs involving column correspondence in a selectable, one with the same labeled subquery repeated, the other when the label has been “grouped” and loses itself. Affects.

    References: #2188

  • [sql]

    Fixed bug whereby “warn on unicode” flag would get set for the String type when used with certain dialects. This bug is not in 0.7.

  • [sql]

    Fixed bug whereby with_only_columns() method of Select would fail if a selectable were passed.. However, the FROM behavior is still incorrect here, so you need 0.7 in any case for this use case to be usable.

    References: #2270

schema

  • [schema]

    Added an informative error message when ForeignKeyConstraint refers to a column name in the parent that is not found.

postgresql

  • [postgresql]

    Fixed bug related to whereby the same modified index behavior in PG 9 affected primary key reflection on a renamed column..

    References: #2141, #2291

mysql

  • [mysql]

    Fixed OurSQL dialect to use ansi-neutral quote symbol “’” for XA commands instead of ‘”’. .

    References: #2186

  • [mysql]

    a CREATE TABLE will put the COLLATE option after CHARSET, which appears to be part of MySQL’s arbitrary rules regarding if it will actually work or not.

    References: #2225

mssql

  • [mssql] [bug]

    Decode incoming values when retrieving list of index names and the names of columns within those indexes.

    References: #2269

oracle

  • [oracle]

    Added ORA-00028 to disconnect codes, use cx_oracle _Error.code to get at the code,.

    References: #2200

  • [oracle]

    repaired the oracle.RAW type which did not generate the correct DDL.

    References: #2220

  • [oracle]

    added CURRENT to reserved word list.

    References: #2212

0.6.8

Released: Sun Jun 05 2011

orm

  • [orm]

    Calling query.get() against a column-based entity is invalid, this condition now raises a deprecation warning.

    References: #2144

  • [orm]

    a non_primary mapper will inherit the _identity_class of the primary mapper. This so that a non_primary established against a class that’s normally in an inheritance mapping will produce results that are identity-map compatible with that of the primary mapper

    References: #2151

  • [orm]

    Backported 0.7’s identity map implementation, which does not use a mutex around removal. This as some users were still getting deadlocks despite the adjustments in 0.6.7; the 0.7 approach that doesn’t use a mutex does not appear to produce “dictionary changed size” issues, the original rationale for the mutex.

    References: #2148

  • [orm]

    Fixed the error message emitted for “can’t execute syncrule for destination column ‘q’; mapper ‘X’ does not map this column” to reference the correct mapper. .

    References: #2163

  • [orm]

    Fixed bug where determination of “self referential” relationship would fail with no workaround for joined-inh subclass related to itself, or joined-inh subclass related to a subclass of that with no cols in the sub-sub class in the join condition.

    References: #2149

  • [orm]

    mapper() will ignore non-configured foreign keys to unrelated tables when determining inherit condition between parent and child class. This is equivalent to behavior already applied to declarative. Note that 0.7 has a more comprehensive solution to this, altering how join() itself determines an FK error.

    References: #2153

  • [orm]

    Fixed bug whereby mapper mapped to an anonymous alias would fail if logging were used, due to unescaped % sign in the alias name.

    References: #2171

  • [orm]

    Modify the text of the message which occurs when the “identity” key isn’t detected on flush, to include the common cause that the Column isn’t set up to detect auto-increment correctly;.

    References: #2170

  • [orm]

    Fixed bug where transaction-level “deleted” collection wouldn’t be cleared of expunged states, raising an error if they later became transient.

    References: #2182

engine

  • [engine]

    Adjusted the __contains__() method of a RowProxy result row such that no exception throw is generated internally; NoSuchColumnError() also will generate its message regardless of whether or not the column construct can be coerced to a string..

    References: #2178

sql

  • [sql]

    Fixed bug whereby if FetchedValue was passed to column server_onupdate, it would not have its parent “column” assigned, added test coverage for all column default assignment patterns.

    References: #2147

  • [sql]

    Fixed bug whereby nesting a label of a select() with another label in it would produce incorrect exported columns. Among other things this would break an ORM column_property() mapping against another column_property(). .

    References: #2167

postgresql

  • [postgresql]

    Fixed bug affecting PG 9 whereby index reflection would fail if against a column whose name had changed. .

    References: #2141

  • [postgresql]

    Some unit test fixes regarding numeric arrays, MATCH operator. A potential floating-point inaccuracy issue was fixed, and certain tests of the MATCH operator only execute within an EN-oriented locale for now. .

    References: #2175

mssql

  • [mssql]

    Fixed bug in MSSQL dialect whereby the aliasing applied to a schema-qualified table would leak into enclosing select statements.

    References: #2169

  • [mssql]

    Fixed bug whereby DATETIME2 type would fail on the “adapt” step when used in result sets or bound parameters. This issue is not in 0.7.

    References: #2159

0.6.7

Released: Wed Apr 13 2011

orm

  • [orm]

    Tightened the iterate vs. remove mutex around the identity map iteration, attempting to reduce the chance of an (extremely rare) reentrant gc operation causing a deadlock. Might remove the mutex in 0.7.

    References: #2087

  • [orm]

    Added a name argument to Query.subquery(), to allow a fixed name to be assigned to the alias object.

    References: #2030

  • [orm]

    A warning is emitted when a joined-table inheriting mapper has no primary keys on the locally mapped table (but has pks on the superclass table).

    References: #2019

  • [orm]

    Fixed bug where “middle” class in a polymorphic hierarchy would have no ‘polymorphic_on’ column if it didn’t also specify a ‘polymorphic_identity’, leading to strange errors upon refresh, wrong class loaded when querying from that target. Also emits the correct WHERE criterion when using single table inheritance.

    References: #2038

  • [orm]

    Fixed bug where a column with a SQL or server side default that was excluded from a mapping with include_properties or exclude_properties would result in UnmappedColumnError.

    References: #1995

  • [orm]

    A warning is emitted in the unusual case that an append or similar event on a collection occurs after the parent object has been dereferenced, which prevents the parent from being marked as “dirty” in the session. This will be an exception in 0.7.

    References: #2046

  • [orm]

    Fixed bug in query.options() whereby a path applied to a lazyload using string keys could overlap a same named attribute on the wrong entity. Note 0.7 has an updated version of this fix.

    References: #2098

  • [orm]

    Reworded the exception raised when a flush is attempted of a subclass that is not polymorphic against the supertype.

    References: #2063

  • [orm]

    Some fixes to the state handling regarding backrefs, typically when autoflush=False, where the back-referenced collection wouldn’t properly handle add/removes with no net change. Thanks to Richard Murri for the test case + patch.

    References: #2123

  • [orm]

    a “having” clause would be copied from the inside to the outside query if from_self() were used..

    References: #2130

examples

  • [examples]

    The Beaker caching example allows a “query_cls” argument to the query_callable() function.

    References: #2090

engine

  • [engine]

    Fixed bug in QueuePool, SingletonThreadPool whereby connections that were discarded via overflow or periodic cleanup() were not explicitly closed, leaving garbage collection to the task instead. This generally only affects non-reference-counting backends like Jython and PyPy. Thanks to Jaimy Azle for spotting this.

    References: #2102

sql

  • [sql]

    Column.copy(), as used in table.tometadata(), copies the ‘doc’ attribute.

    References: #2028

  • [sql]

    Added some defs to the resultproxy.c extension so that the extension compiles and runs on Python 2.4.

    References: #2023

  • [sql]

    The compiler extension now supports overriding the default compilation of expression._BindParamClause including that the auto-generated binds within the VALUES/SET clause of an insert()/update() statement will also use the new compilation rules.

    References: #2042

  • [sql]

    Added accessors to ResultProxy “returns_rows”, “is_insert”

    References: #2089

  • [sql]

    The limit/offset keywords to select() as well as the value passed to select.limit()/offset() will be coerced to integer.

    References: #2116

postgresql

  • [postgresql]

    When explicit sequence execution derives the name of the auto-generated sequence of a SERIAL column, which currently only occurs if implicit_returning=False, now accommodates if the table + column name is greater than 63 characters using the same logic PostgreSQL uses.

    References: #1083

  • [postgresql]

    Added an additional libpq message to the list of “disconnect” exceptions, “could not receive data from server”

    References: #2044

  • [postgresql]

    Added RESERVED_WORDS for postgresql dialect.

    References: #2092

  • [postgresql]

    Fixed the BIT type to allow a “length” parameter, “varying” parameter. Reflection also fixed.

    References: #2073

mysql

  • [mysql]

    oursql dialect accepts the same “ssl” arguments in create_engine() as that of MySQLdb.

    References: #2047

sqlite

  • [sqlite]

    Fixed bug where reflection of foreign key created as “REFERENCES <tablename>” without col name would fail.

    References: #2115

mssql

  • [mssql]

    Rewrote the query used to get the definition of a view, typically when using the Inspector interface, to use sys.sql_modules instead of the information schema, thereby allowing views definitions longer than 4000 characters to be fully returned.

    References: #2071

oracle

  • [oracle]

    Using column names that would require quotes for the column itself or for a name-generated bind parameter, such as names with special characters, underscores, non-ascii characters, now properly translate bind parameter keys when talking to cx_oracle.

    References: #2100

  • [oracle]

    Oracle dialect adds use_binds_for_limits=False create_engine() flag, will render the LIMIT/OFFSET values inline instead of as binds, reported to modify the execution plan used by Oracle.

    References: #2116

firebird

  • [firebird]

    The “implicit_returning” flag on create_engine() is honored if set to False.

    References: #2083

misc

  • [informix]

    Added RESERVED_WORDS informix dialect.

    References: #2092

  • [ext]

    The horizontal_shard ShardedSession class accepts the common Session argument “query_cls” as a constructor argument, to enable further subclassing of ShardedQuery.

    References: #2090

  • [declarative]

    Added an explicit check for the case that the name ‘metadata’ is used for a column attribute on a declarative class.

    References: #2050

  • [declarative]

    Fix error message referencing old @classproperty name to reference @declared_attr

    References: #2061

  • [declarative]

    Arguments in __mapper_args__ that aren’t “hashable” aren’t mistaken for always-hashable, possibly-column arguments.

    References: #2091

  • [documentation]

    Documented SQLite DATE/TIME/DATETIME types.

    References: #2029

0.6.6

Released: Sat Jan 08 2011

orm

  • [orm]

    Fixed bug whereby a non-“mutable” attribute modified event which occurred on an object that was clean except for preceding mutable attribute changes would fail to strongly reference itself in the identity map. This would cause the object to be garbage collected, losing track of any changes that weren’t previously saved in the “mutable changes” dictionary.

  • [orm]

    Fixed bug whereby “passive_deletes=’all’” wasn’t passing the correct symbols to lazy loaders during flush, thereby causing an unwarranted load.

    References: #2013

  • [orm]

    Fixed bug which prevented composite mapped attributes from being used on a mapped select statement.. Note the workings of composite are slated to change significantly in 0.7.

    References: #1997

  • [orm]

    active_history flag also added to composite(). The flag has no effect in 0.6, but is instead a placeholder flag for forwards compatibility, as it applies in 0.7 for composites.

    References: #1976

  • [orm]

    Fixed uow bug whereby expired objects passed to Session.delete() would not have unloaded references or collections taken into account when deleting objects, despite passive_deletes remaining at its default of False.

    References: #2002

  • [orm]

    A warning is emitted when version_id_col is specified on an inheriting mapper when the inherited mapper already has one, if those column expressions are not the same.

    References: #1987

  • [orm]

    ”innerjoin” flag doesn’t take effect along the chain of joinedload() joins if a previous join in that chain is an outer join, thus allowing primary rows without a referenced child row to be correctly returned in results.

    References: #1954

  • [orm]

    Fixed bug regarding “subqueryload” strategy whereby strategy would fail if the entity was an aliased() construct.

    References: #1964

  • [orm]

    Fixed bug regarding “subqueryload” strategy whereby the join would fail if using a multi-level load of the form from A->joined-subclass->C

    References: #2014

  • [orm]

    Fixed indexing of Query objects by -1. It was erroneously transformed to the empty slice -1:0 that resulted in IndexError.

    References: #1968

  • [orm]

    The mapper argument “primary_key” can be passed as a single column as well as a list or tuple. The documentation examples that illustrated it as a scalar value have been changed to lists.

    References: #1971

  • [orm]

    Added active_history flag to relationship() and column_property(), forces attribute events to always load the “old” value, so that it’s available to attributes.get_history().

    References: #1961

  • [orm]

    Query.get() will raise if the number of params in a composite key is too large, as well as too small.

    References: #1977

  • [orm]

    Backport of “optimized get” fix from 0.7, improves the generation of joined-inheritance “load expired row” behavior.

    References: #1992

  • [orm]

    A little more verbiage to the “primaryjoin” error, in an unusual condition that the join condition “works” for viewonly but doesn’t work for non-viewonly, and foreign_keys wasn’t used - adds “foreign_keys” to the suggestion. Also add “foreign_keys” to the suggestion for the generic “direction” error.

examples

  • [examples]

    The versioning example now supports detection of changes in an associated relationship().

engine

  • [engine]

    The “unicode warning” against non-unicode bind data is now raised only when the Unicode type is used explicitly; not when convert_unicode=True is used on the engine or String type.

  • [engine]

    Fixed memory leak in C version of Decimal result processor.

    References: #1978

  • [engine]

    Implemented sequence check capability for the C version of RowProxy, as well as 2.7 style “collections.Sequence” registration for RowProxy.

    References: #1871

  • [engine]

    Threadlocal engine methods rollback(), commit(), prepare() won’t raise if no transaction is in progress; this was a regression introduced in 0.6.

    References: #1998

  • [engine]

    Threadlocal engine returns itself upon begin(), begin_nested(); engine then implements contextmanager methods to allow the “with” statement.

    References: #2004

sql

  • [sql]

    Fixed operator precedence rules for multiple chains of a single non-associative operator. I.e. “x - (y - z)” will compile as “x - (y - z)” and not “x - y - z”. Also works with labels, i.e. “x - (y - z).label(‘foo’)”

    References: #1984

  • [sql]

    The ‘info’ attribute of Column is copied during Column.copy(), i.e. as occurs when using columns in declarative mixins.

    References: #1967

  • [sql]

    Added a bind processor for booleans which coerces to int, for DBAPIs such as pymssql that naively call str() on values.

  • [sql]

    CheckConstraint will copy its ‘initially’, ‘deferrable’, and ‘_create_rule’ attributes within a copy()/tometadata()

    References: #2000

postgresql

  • [postgresql]

    Single element tuple expressions inside an IN clause parenthesize correctly, also from

    References: #1984

  • [postgresql]

    Ensured every numeric, float, int code, scalar + array, are recognized by psycopg2 and pg8000’s “numeric” base type.

    References: #1955

  • [postgresql]

    Added as_uuid=True flag to the UUID type, will receive and return values as Python UUID() objects rather than strings. Currently, the UUID type is only known to work with psycopg2.

    References: #1956

  • [postgresql]

    Fixed bug whereby KeyError would occur with non-ENUM supported PG versions after a pool dispose+recreate would occur.

    References: #1989

mysql

  • [mysql]

    Fixed error handling for Jython + zxjdbc, such that has_table() property works again. Regression from 0.6.3 (we don’t have a Jython buildbot, sorry)

    References: #1960

sqlite

  • [sqlite]

    The REFERENCES clause in a CREATE TABLE that includes a remote schema to another table with the same schema name now renders the remote name without the schema clause, as required by SQLite.

    References: #1851

  • [sqlite]

    On the same theme, the REFERENCES clause in a CREATE TABLE that includes a remote schema to a different schema than that of the parent table doesn’t render at all, as cross-schema references do not appear to be supported.

mssql

  • [mssql]

    The rewrite of index reflection in was unfortunately not tested correctly, and returned incorrect results. This regression is now fixed.

    References: #1770

oracle

  • [oracle]

    The cx_oracle “decimal detection” logic, which takes place for result set columns with ambiguous numeric characteristics, now uses the decimal point character determined by the locale/ NLS_LANG setting, using an on-first-connect detection of this character. cx_oracle 5.0.3 or greater is also required when using a non-period-decimal-point NLS_LANG setting..

    References: #1953

firebird

  • [firebird]

    Firebird numeric type now checks for Decimal explicitly, lets float() pass right through, thereby allowing special values such as float(‘inf’).

    References: #2012

misc

  • [declarative]

    An error is raised if __table_args__ is not in tuple or dict format, and is not None.

    References: #1972

  • [sqlsoup]

    Added “map_to()” method to SqlSoup, which is a “master” method which accepts explicit arguments for each aspect of the selectable and mapping, including a base class per mapping.

    References: #1975

  • [sqlsoup]

    Mapped selectables used with the map(), with_labels(), join() methods no longer put the given argument into the internal “cache” dictionary. Particularly since the join() and select() objects are created in the method itself this was pretty much a pure memory leaking behavior.

0.6.5

Released: Sun Oct 24 2010

orm

  • [orm]

    Added a new “lazyload” option “immediateload”. Issues the usual “lazy” load operation automatically as the object is populated. The use case here is when loading objects to be placed in an offline cache, or otherwise used after the session isn’t available, and straight ‘select’ loading, not ‘joined’ or ‘subquery’, is desired.

    References: #1914

  • [orm]

    New Query methods: query.label(name), query.as_scalar(), return the query’s statement as a scalar subquery with /without label; query.with_entities(*ent), replaces the SELECT list of the query with new entities. Roughly equivalent to a generative form of query.values() which accepts mapped entities as well as column expressions.

    References: #1920

  • [orm]

    Fixed recursion bug which could occur when moving an object from one reference to another, with backrefs involved, where the initiating parent was a subclass (with its own mapper) of the previous parent.

  • [orm]

    Fixed a regression in 0.6.4 which occurred if you passed an empty list to “include_properties” on mapper()

    References: #1918

  • [orm]

    Fixed labeling bug in Query whereby the NamedTuple would mis-apply labels if any of the column expressions were un-labeled.

  • [orm]

    Patched a case where query.join() would adapt the right side to the right side of the left’s join inappropriately

    References: #1925

  • [orm]

    Query.select_from() has been beefed up to help ensure that a subsequent call to query.join() will use the select_from() entity, assuming it’s a mapped entity and not a plain selectable, as the default “left” side, not the first entity in the Query object’s list of entities.

  • [orm]

    The exception raised by Session when it is used subsequent to a subtransaction rollback (which is what happens when a flush fails in autocommit=False mode) has now been reworded (this is the “inactive due to a rollback in a subtransaction” message). In particular, if the rollback was due to an exception during flush(), the message states this is the case, and reiterates the string form of the original exception that occurred during flush. If the session is closed due to explicit usage of subtransactions (not very common), the message just states this is the case.

  • [orm]

    The exception raised by Mapper when repeated requests to its initialization are made after initialization already failed no longer assumes the “hasattr” case, since there’s other scenarios in which this message gets emitted, and the message also does not compound onto itself multiple times - you get the same message for each attempt at usage. The misnomer “compiles” is being traded out for “initialize”.

  • [orm]

    Fixed bug in query.update() where ‘evaluate’ or ‘fetch’ expiration would fail if the column expression key was a class attribute with a different keyname as the actual column name.

    References: #1935

  • [orm]

    Added an assertion during flush which ensures that no NULL-holding identity keys were generated on “newly persistent” objects. This can occur when user defined code inadvertently triggers flushes on not-fully-loaded objects.

  • [orm]

    lazy loads for relationship attributes now use the current state, not the “committed” state, of foreign and primary key attributes when issuing SQL, if a flush is not in process. Previously, only the database-committed state would be used. In particular, this would cause a many-to-one get()-on-lazyload operation to fail, as autoflush is not triggered on these loads when the attributes are determined and the “committed” state may not be available.

    References: #1910

  • [orm]

    A new flag on relationship(), load_on_pending, allows the lazy loader to fire off on pending objects without a flush taking place, as well as a transient object that’s been manually “attached” to the session. Note that this flag blocks attribute events from taking place when an object is loaded, so backrefs aren’t available until after a flush. The flag is only intended for very specific use cases.

  • [orm]

    Another new flag on relationship(), cascade_backrefs, disables the “save-update” cascade when the event was initiated on the “reverse” side of a bidirectional relationship. This is a cleaner behavior so that many-to-ones can be set on a transient object without it getting sucked into the child object’s session, while still allowing the forward collection to cascade. We might default this to False in 0.7.

  • [orm]

    Slight improvement to the behavior of “passive_updates=False” when placed only on the many-to-one side of a relationship; documentation has been clarified that passive_updates=False should really be on the one-to-many side.

  • [orm]

    Placing passive_deletes=True on a many-to-one emits a warning, since you probably intended to put it on the one-to-many side.

  • [orm]

    Fixed bug that would prevent “subqueryload” from working correctly with single table inheritance for a relationship from a subclass - the “where type in (x, y, z)” only gets placed on the inside, instead of repeatedly.

  • [orm]

    When using from_self() with single table inheritance, the “where type in (x, y, z)” is placed on the outside of the query only, instead of repeatedly. May make some more adjustments to this.

  • [orm]

    scoped_session emits a warning when configure() is called if a Session is already present (checks only the current thread)

    References: #1924

  • [orm]

    reworked the internals of mapper.cascade_iterator() to cut down method calls by about 9% in some circumstances.

    References: #1932

engine

  • [engine]

    Fixed a regression in 0.6.4 whereby the change that allowed cursor errors to be raised consistently broke the result.lastrowid accessor. Test coverage has been added for result.lastrowid. Note that lastrowid is only supported by Pysqlite and some MySQL drivers, so isn’t super-useful in the general case.

  • [engine]

    the logging message emitted by the engine when a connection is first used is now “BEGIN (implicit)” to emphasize that DBAPI has no explicit begin().

  • [engine]

    added “views=True” option to metadata.reflect(), will add the list of available views to those being reflected.

    References: #1936

  • [engine]

    engine_from_config() now accepts ‘debug’ for ‘echo’, ‘echo_pool’, ‘force’ for ‘convert_unicode’, boolean values for ‘use_native_unicode’.

    References: #1899

sql

  • [sql]

    Fixed bug in TypeDecorator whereby the dialect-specific type was getting pulled in to generate the DDL for a given type, which didn’t always return the correct result.

  • [sql]

    TypeDecorator can now have a fully constructed type specified as its “impl”, in addition to a type class.

  • [sql]

    TypeDecorator will now place itself as the resulting type for a binary expression where the type coercion rules would normally return its impl type - previously, a copy of the impl type would be returned which would have the TypeDecorator embedded into it as the “dialect” impl, this was probably an unintentional way of achieving the desired effect.

  • [sql]

    TypeDecorator.load_dialect_impl() returns “self.impl” by default, i.e. not the dialect implementation type of “self.impl”. This to support compilation correctly. Behavior can be user-overridden in exactly the same way as before to the same effect.

  • [sql]

    Added type_coerce(expr, type_) expression element. Treats the given expression as the given type when evaluating expressions and processing result rows, but does not affect the generation of SQL, other than an anonymous label.

  • [sql]

    Table.tometadata() now copies Index objects associated with the Table as well.

  • [sql]

    Table.tometadata() issues a warning if the given Table is already present in the target MetaData - the existing Table object is returned.

  • [sql]

    An informative error message is raised if a Column which has not yet been assigned a name, i.e. as in declarative, is used in a context where it is exported to the columns collection of an enclosing select() construct, or if any construct involving that column is compiled before its name is assigned.

  • [sql]

    as_scalar(), label() can be called on a selectable which contains a Column that is not yet named.

    References: #1862

  • [sql]

    Fixed recursion overflow which could occur when operating with two expressions both of type “NullType”, but not the singleton NULLTYPE instance.

    References: #1907

postgresql

  • [postgresql]

    Added “as_tuple” flag to ARRAY type, returns results as tuples instead of lists to allow hashing.

  • [postgresql]

    Fixed bug which prevented “domain” built from a custom type such as “enum” from being reflected.

    References: #1933

mysql

  • [mysql]

    Fixed bug involving reflection of CURRENT_TIMESTAMP default used with ON UPDATE clause, thanks to Taavi Burns

    References: #1940

mssql

  • [mssql]

    Fixed reflection bug which did not properly handle reflection of unknown types.

    References: #1946

  • [mssql]

    Fixed bug where aliasing of tables with “schema” would fail to compile properly.

    References: #1943

  • [mssql]

    Rewrote the reflection of indexes to use sys. catalogs, so that column names of any configuration (spaces, embedded commas, etc.) can be reflected. Note that reflection of indexes requires SQL Server 2005 or greater.

    References: #1770

  • [mssql]

    mssql+pymssql dialect now honors the “port” portion of the URL instead of discarding it.

    References: #1952

oracle

  • [oracle]

    The implicit_returning argument to create_engine() is now honored regardless of detected version of Oracle. Previously, the flag would be forced to False if server version info was < 10.

    References: #1878

tests

  • [tests]

    the NoseSQLAlchemyPlugin has been moved to a new package “sqlalchemy_nose” which installs along with “sqlalchemy”. This so that the “nosetests” script works as always but also allows the –with-coverage option to turn on coverage before SQLAlchemy modules are imported, allowing coverage to work correctly.

misc

  • [declarative]

    @classproperty (soon/now @declared_attr) takes effect for __mapper_args__, __table_args__, __tablename__ on a base class that is not a mixin, as well as mixins.

    References: #1922

  • [declarative]

    @classproperty ‘s official name/location for usage with declarative is sqlalchemy.ext.declarative.declared_attr. Same thing, but moving there since it is more of a “marker” that’s specific to declarative, not just an attribute technique.

    References: #1915

  • [declarative]

    Fixed bug whereby columns on a mixin wouldn’t propagate correctly to a single-table, or joined-table, inheritance scheme where the attribute name is different than that of the column.,.

    References: #1930, #1931

  • [declarative]

    A mixin can now specify a column that overrides a column of the same name associated with a superclass. Thanks to Oystein Haaland.

  • [informix]

    Major cleanup / modernization of the Informix dialect for 0.6, courtesy Florian Apolloner.

    References: #1906

  • [misc]

    CircularDependencyError now has .cycles and .edges members, which are the set of elements involved in one or more cycles, and the set of edges as 2-tuples.

    References: #1890

0.6.4

Released: Tue Sep 07 2010

orm

  • [orm]

    The name ConcurrentModificationError has been changed to StaleDataError, and descriptive error messages have been revised to reflect exactly what the issue is. Both names will remain available for the foreseeable future for schemes that may be specifying ConcurrentModificationError in an “except:” clause.

  • [orm]

    Added a mutex to the identity map which mutexes remove operations against iteration methods, which now pre-buffer before returning an iterable. This because asynchronous gc can remove items via the gc thread at any time.

    References: #1891

  • [orm]

    The Session class is now present in sqlalchemy.orm.*. We’re moving away from the usage of create_session(), which has non-standard defaults, for those situations where a one-step Session constructor is desired. Most users should stick with sessionmaker() for general use, however.

  • [orm]

    query.with_parent() now accepts transient objects and will use the non-persistent values of their pk/fk attributes in order to formulate the criterion. Docs are also clarified as to the purpose of with_parent().

  • [orm]

    The include_properties and exclude_properties arguments to mapper() now accept Column objects as members in addition to strings. This so that same-named Column objects, such as those within a join(), can be disambiguated.

  • [orm]

    A warning is now emitted if a mapper is created against a join or other single selectable that includes multiple columns with the same name in its .c. collection, and those columns aren’t explicitly named as part of the same or separate attributes (or excluded). In 0.7 this warning will be an exception. Note that this warning is not emitted when the combination occurs as a result of inheritance, so that attributes still allow being overridden naturally.. In 0.7 this will be improved further.

    References: #1896

  • [orm]

    The primary_key argument to mapper() can now specify a series of columns that are only a subset of the calculated “primary key” columns of the mapped selectable, without an error being raised. This helps for situations where a selectable’s effective primary key is simpler than the number of columns in the selectable that are actually marked as “primary_key”, such as a join against two tables on their primary key columns.

    References: #1896

  • [orm]

    An object that’s been deleted now gets a flag ‘deleted’, which prohibits the object from being re-add()ed to the session, as previously the object would live in the identity map silently until its attributes were accessed. The make_transient() function now resets this flag along with the “key” flag.

  • [orm]

    make_transient() can be safely called on an already transient instance.

  • [orm]

    a warning is emitted in mapper() if the polymorphic_on column is not present either in direct or derived form in the mapped selectable or in the with_polymorphic selectable, instead of silently ignoring it. Look for this to become an exception in 0.7.

  • [orm]

    Another pass through the series of error messages emitted when relationship() is configured with ambiguous arguments. The “foreign_keys” setting is no longer mentioned, as it is almost never needed and it is preferable users set up correct ForeignKey metadata, which is now the recommendation. If ‘foreign_keys’ is used and is incorrect, the message suggests the attribute is probably unnecessary. Docs for the attribute are beefed up. This because all confused relationship() users on the ML appear to be attempting to use foreign_keys due to the message, which only confuses them further since Table metadata is much clearer.

  • [orm]

    If the “secondary” table has no ForeignKey metadata and no foreign_keys is set, even though the user is passing screwed up information, it is assumed that primary/secondaryjoin expressions should consider only and all cols in “secondary” to be foreign. It’s not possible with “secondary” for the foreign keys to be elsewhere in any case. A warning is now emitted instead of an error, and the mapping succeeds.

    References: #1877

  • [orm]

    Moving an o2m object from one collection to another, or vice versa changing the referenced object by an m2o, where the foreign key is also a member of the primary key, will now be more carefully checked during flush if the change in value of the foreign key on the “many” side is the result of a change in the primary key of the “one” side, or if the “one” is just a different object. In one case, a cascade-capable DB would have cascaded the value already and we need to look at the “new” PK value to do an UPDATE, in the other we need to continue looking at the “old”. We now look at the “old”, assuming passive_updates=True, unless we know it was a PK switch that triggered the change.

    References: #1856

  • [orm]

    The value of version_id_col can be changed manually, and this will result in an UPDATE of the row. Versioned UPDATEs and DELETEs now use the “committed” value of the version_id_col in the WHERE clause and not the pending changed value. The version generator is also bypassed if manual changes are present on the attribute.

    References: #1857

  • [orm]

    Repaired the usage of merge() when used with concrete inheriting mappers. Such mappers frequently have so-called “concrete” attributes, which are subclass attributes that “disable” propagation from the parent - these needed to allow a merge() operation to pass through without effect.

  • [orm]

    Specifying a non-column based argument for column_mapped_collection, including string, text() etc., will raise an error message that specifically asks for a column element, no longer misleads with incorrect information about text() or literal().

    References: #1863

  • [orm]

    Similarly, for relationship(), foreign_keys, remote_side, order_by - all column-based expressions are enforced - lists of strings are explicitly disallowed since this is a very common error

  • [orm]

    Dynamic attributes don’t support collection population - added an assertion for when set_committed_value() is called, as well as when joinedload() or subqueryload() options are applied to a dynamic attribute, instead of failure / silent failure.

    References: #1864

  • [orm]

    Fixed bug whereby generating a Query derived from one which had the same column repeated with different label names, typically in some UNION situations, would fail to propagate the inner columns completely to the outer query.

    References: #1852

  • [orm]

    object_session() raises the proper UnmappedInstanceError when presented with an unmapped instance.

    References: #1881

  • [orm]

    Applied further memoizations to calculated Mapper properties, with significant (~90%) runtime mapper.py call count reduction in heavily polymorphic mapping configurations.

  • [orm]

    mapper _get_col_to_prop private method used by the versioning example is deprecated; now use mapper.get_property_by_column() which will remain the public method for this.

  • [orm]

    the versioning example works correctly now if versioning on a col that was formerly NULL.

examples

  • [examples]

    The beaker_caching example has been reorganized such that the Session, cache manager, declarative_base are part of environment, and custom cache code is portable and now within “caching_query.py”. This allows the example to be easier to “drop in” to existing projects.

  • [examples]

    the history_meta versioning recipe sets “unique=False” when copying columns, so that the versioning table handles multiple rows with repeating values.

    References: #1887

engine

  • [engine]

    Calling fetchone() or similar on a result that has already been exhausted, has been closed, or is not a result-returning result now raises ResourceClosedError, a subclass of InvalidRequestError, in all cases, regardless of backend. Previously, some DBAPIs would raise ProgrammingError (i.e. pysqlite), others would return None leading to downstream breakages (i.e. MySQL-python).

  • [engine]

    Fixed bug in Connection whereby if a “disconnect” event occurred in the “initialize” phase of the first connection pool connect, an AttributeError would be raised when the Connection would attempt to invalidate the DBAPI connection.

    References: #1894

  • [engine]

    Connection, ResultProxy, as well as Session use ResourceClosedError for all “this connection/transaction/result is closed” types of errors.

  • [engine]

    Connection.invalidate() can be called more than once and subsequent calls do nothing.

sql

  • [sql]

    Calling execute() on an alias() construct is pending deprecation for 0.7, as it is not itself an “executable” construct. It currently “proxies” its inner element and is conditionally “executable” but this is not the kind of ambiguity we like these days.

  • [sql]

    The execute() and scalar() methods of ClauseElement are now moved appropriately to the Executable subclass. ClauseElement.execute()/ scalar() are still present and are pending deprecation in 0.7, but note these would always raise an error anyway if you were not an Executable (unless you were an alias(), see previous note).

  • [sql]

    Added basic math expression coercion for Numeric->Integer, so that resulting type is Numeric regardless of the direction of the expression.

  • [sql]

    Changed the scheme used to generate truncated “auto” index names when using the “index=True” flag on Column. The truncation only takes place with the auto-generated name, not one that is user-defined (an error would be raised instead), and the truncation scheme itself is now based on a fragment of an md5 hash of the identifier name, so that multiple indexes on columns with similar names still have unique names.

    References: #1855

  • [sql]

    The generated index name also is based on a “max index name length” attribute which is separate from the “max identifier length” - this to appease MySQL who has a max length of 64 for index names, separate from their overall max length of 255.

    References: #1412

  • [sql]

    the text() construct, if placed in a column oriented situation, will at least return NULLTYPE for its type instead of None, allowing it to be used a little more freely for ad-hoc column expressions than before. literal_column() is still the better choice, however.

  • [sql]

    Added full description of parent table/column, target table/column in error message raised when ForeignKey can’t resolve target.

  • [sql]

    Fixed bug whereby replacing composite foreign key columns in a reflected table would cause an attempt to remove the reflected constraint from the table a second time, raising a KeyError.

    References: #1865

  • [sql]

    the _Label construct, i.e. the one that is produced whenever you say somecol.label(), now counts itself in its “proxy_set” unioned with that of its contained column’s proxy set, instead of directly returning that of the contained column. This allows column correspondence operations which depend on the identity of the _Labels themselves to return the correct result

  • [sql]

    fixes ORM bug.

    References: #1852

postgresql

  • [postgresql]

    Fixed the psycopg2 dialect to use its set_isolation_level() method instead of relying upon the base “SET SESSION ISOLATION” command, as psycopg2 resets the isolation level on each new transaction otherwise.

mssql

  • [mssql]

    Fixed “default schema” query to work with pymssql backend.

oracle

  • [oracle]

    Added ROWID type to the Oracle dialect, for those cases where an explicit CAST might be needed.

    References: #1879

  • [oracle]

    Oracle reflection of indexes has been tuned so that indexes which include some or all primary key columns, but not the same set of columns as that of the primary key, are reflected. Indexes which contain the identical columns as that of the primary key are skipped within reflection, as the index in that case is assumed to be the auto-generated primary key index. Previously, any index with PK columns present would be skipped. Thanks to Kent Bower for the patch.

    References: #1867

  • [oracle]

    Oracle now reflects the names of primary key constraints - also thanks to Kent Bower.

    References: #1868

firebird

  • [firebird]

    Fixed bug whereby a column default would fail to reflect if the “default” keyword were lower case.

misc

  • [declarative]

    if @classproperty is used with a regular class-bound mapper property attribute, it will be called to get the actual attribute value during initialization. Currently, there’s no advantage to using @classproperty on a column or relationship attribute of a declarative class that isn’t a mixin - evaluation is at the same time as if @classproperty weren’t used. But here we at least allow it to function as expected.

  • [declarative]

    Fixed bug where “Can’t add additional column” message would display the wrong name.

  • [informix]

    Applied patches from to get basic Informix functionality up again. We rely upon end-user testing to ensure that Informix is working to some degree.

    References: #1904

  • [documentation]

    The docs have been reorganized such that the “API Reference” section is gone - all the docstrings from there which were public API are moved into the context of the main doc section that talks about it. Main docs divided into “SQLAlchemy Core” and “SQLAlchemy ORM” sections, mapper/relationship docs have been broken out. Lots of sections rewritten and/or reorganized.

0.6.3

Released: Thu Jul 15 2010

orm

  • [orm]

    Removed errant many-to-many load in unitofwork which triggered unnecessarily on expired/unloaded collections. This load now takes place only if passive_updates is False and the parent primary key has changed, or if passive_deletes is False and a delete of the parent has occurred.

    References: #1845

  • [orm]

    Column-entities (i.e. query(Foo.id)) copy their state more fully when queries are derived from themselves + a selectable (i.e. from_self(), union(), etc.), so that join() and such have the correct state to work from.

    References: #1853

  • [orm]

    Fixed bug where Query.join() would fail if querying a non-ORM column then joining without an on clause when a FROM clause is already present, now raises a checked exception the same way it does when the clause is not present.

    References: #1853

  • [orm]

    Improved the check for an “unmapped class”, including the case where the superclass is mapped but the subclass is not. Any attempts to access cls._sa_class_manager.mapper now raise UnmappedClassError().

    References: #1142

  • [orm]

    Added “column_descriptions” accessor to Query, returns a list of dictionaries containing naming/typing information about the entities the Query will return. Can be helpful for building GUIs on top of ORM queries.

mysql

  • [mysql]

    The _extract_error_code() method now works correctly with each MySQL dialect ( MySQL-python, OurSQL, MySQL-Connector-Python, PyODBC). Previously, the reconnect logic would fail for OperationalError conditions, however since MySQLdb and OurSQL have their own reconnect feature, there was no symptom for these drivers here unless one watched the logs.

    References: #1848

oracle

  • [oracle]

    More tweaks to cx_oracle Decimal handling. “Ambiguous” numerics with no decimal place are coerced to int at the connection handler level. The advantage here is that ints come back as ints without SQLA type objects being involved and without needless conversion to Decimal first.

    Unfortunately, some exotic subquery cases can even see different types between individual result rows, so the Numeric handler, when instructed to return Decimal, can’t take full advantage of “native decimal” mode and must run isinstance() on every value to check if its Decimal already. Reopen of

    References: #1840

0.6.2

Released: Tue Jul 06 2010

orm

  • [orm]

    Query.join() will check for a call of the form query.join(target, clause_expression), i.e. missing the tuple, and raise an informative error message that this is the wrong calling form.

  • [orm]

    Fixed bug regarding flushes on self-referential bi-directional many-to-many relationships, where two objects made to mutually reference each other in one flush would fail to insert a row for both sides. Regression from 0.5.

    References: #1824

  • [orm]

    the post_update feature of relationship() has been reworked architecturally to integrate more closely with the new 0.6 unit of work. The motivation for the change is so that multiple “post update” calls, each affecting different foreign key columns of the same row, are executed in a single UPDATE statement, rather than one UPDATE statement per column per row. Multiple row updates are also batched into executemany()s as possible, while maintaining consistent row ordering.

  • [orm]

    Query.statement, Query.subquery(), etc. now transfer the values of bind parameters, i.e. those specified by query.params(), into the resulting SQL expression. Previously the values would not be transferred and bind parameters would come out as None.

  • [orm]

    Subquery-eager-loading now works with Query objects which include params(), as well as get() Queries.

  • [orm]

    Can now call make_transient() on an instance that is referenced by parent objects via many-to-one, without the parent’s foreign key value getting temporarily set to None - this was a function of the “detect primary key switch” flush handler. It now ignores objects that are no longer in the “persistent” state, and the parent’s foreign key identifier is left unaffected.

  • [orm]

    query.order_by() now accepts False, which cancels any existing order_by() state on the Query, allowing subsequent generative methods to be called which do not support ORDER BY. This is not the same as the already existing feature of passing None, which suppresses any existing order_by() settings, including those configured on the mapper. False will make it as though order_by() was never called, while None is an active setting.

  • [orm]

    An instance which is moved to “transient”, has an incomplete or missing set of primary key attributes, and contains expired attributes, will raise an InvalidRequestError if an expired attribute is accessed, instead of getting a recursion overflow.

  • [orm]

    The make_transient() function is now in the generated documentation.

  • [orm]

    make_transient() removes all “loader” callables from the state being made transient, removing any “expired” state - all unloaded attributes reset back to undefined, None/empty on access.

sql

  • [sql]

    The warning emitted by the Unicode and String types with convert_unicode=True no longer embeds the actual value passed. This so that the Python warning registry does not continue to grow in size, the warning is emitted once as per the warning filter settings, and large string values don’t pollute the output.

    References: #1822

  • [sql]

    Fixed bug that would prevent overridden clause compilation from working for “annotated” expression elements, which are often generated by the ORM.

  • [sql]

    The argument to “ESCAPE” of a LIKE operator or similar is passed through render_literal_value(), which may implement escaping of backslashes.

    References: #1400

  • [sql]

    Fixed bug in Enum type which blew away native_enum flag when used with TypeDecorators or other adaption scenarios.

  • [sql]

    Inspector hits bind.connect() when invoked to ensure initialize has been called. the internal name “.conn” is changed to “.bind”, since that’s what it is.

  • [sql]

    Modified the internals of “column annotation” such that a custom Column subclass can safely override _constructor to return Column, for the purposes of making “configurational” column classes that aren’t involved in proxying, etc.

  • [sql]

    Column.copy() takes along the “unique” attribute among others, fixes regarding declarative mixins

    References: #1829

postgresql

  • [postgresql]

    render_literal_value() is overridden which escapes backslashes, currently applies to the ESCAPE clause of LIKE and similar expressions. Ultimately this will have to detect the value of “standard_conforming_strings” for full behavior.

    References: #1400

  • [postgresql]

    Won’t generate “CREATE TYPE” / “DROP TYPE” if using types.Enum on a PG version prior to 8.3 - the supports_native_enum flag is fully honored.

    References: #1836

mysql

  • [mysql]

    MySQL dialect doesn’t emit CAST() for MySQL version detected < 4.0.2. This allows the unicode check on connect to proceed.

    References: #1826

  • [mysql]

    MySQL dialect now detects NO_BACKSLASH_ESCAPES sql mode, in addition to ANSI_QUOTES.

  • [mysql]

    render_literal_value() is overridden which escapes backslashes, currently applies to the ESCAPE clause of LIKE and similar expressions. This behavior is derived from detecting the value of NO_BACKSLASH_ESCAPES.

    References: #1400

mssql

  • [mssql]

    If server_version_info is outside the usual range of (8, ), (9, ), (10, ), a warning is emitted which suggests checking that the FreeTDS version configuration is using 7.0 or 8.0, not 4.2.

    References: #1825

oracle

  • [oracle]

    Fixed ora-8 compatibility flags such that they don’t cache a stale value from before the first database connection actually occurs.

    References: #1819

  • [oracle]

    Oracle’s “native decimal” metadata begins to return ambiguous typing information about numerics when columns are embedded in subqueries as well as when ROWNUM is consulted with subqueries, as we do for limit/offset. We’ve added these ambiguous conditions to the cx_oracle “convert to Decimal()” handler, so that we receive numerics as Decimal in more cases instead of as floats. These are then converted, if requested, into Integer or Float, or otherwise kept as the lossless Decimal.

    References: #1840

firebird

  • [firebird]

    Fixed incorrect signature in do_execute(), error introduced in 0.6.1.

    References: #1823

  • [firebird]

    Firebird dialect adds CHAR, VARCHAR types which accept a “charset” flag, to support Firebird “CHARACTER SET” clause.

    References: #1813

misc

  • [declarative]

    Added support for @classproperty to provide any kind of schema/mapping construct from a declarative mixin, including columns with foreign keys, relationships, column_property, deferred. This solves all such issues on declarative mixins. An error is raised if any MapperProperty subclass is specified on a mixin without using @classproperty.

    References: #1751, #1796, #1805

  • [declarative]

    a mixin class can now define a column that matches one which is present on a __table__ defined on a subclass. It cannot, however, define one that is not present in the __table__, and the error message here now works.

    References: #1821

  • [compiler] [extension]

    The ‘default’ compiler is automatically copied over when overriding the compilation of a built in clause construct, so no KeyError is raised if the user-defined compiler is specific to certain backends and compilation for a different backend is invoked.

    References: #1838

  • [documentation]

    Added documentation for the Inspector.

    References: #1820

  • [documentation]

    Fixed @memoized_property and @memoized_instancemethod decorators so that Sphinx documentation picks up these attributes and methods, such as ResultProxy.inserted_primary_key.

    References: #1830

0.6.1

Released: Mon May 31 2010

orm

  • [orm]

    Fixed regression introduced in 0.6.0 involving improper history accounting on mutable attributes.

    References: #1782

  • [orm]

    Fixed regression introduced in 0.6.0 unit of work refactor that broke updates for bi-directional relationship() with post_update=True.

    References: #1807

  • [orm]

    session.merge() will not expire attributes on the returned instance if that instance is “pending”.

    References: #1789

  • [orm]

    fixed __setstate__ method of CollectionAdapter to not fail during deserialize where parent InstanceState not yet unserialized.

    References: #1802

  • [orm]

    Added internal warning in case an instance without a full PK happened to be expired and then was asked to refresh.

    References: #1797

  • [orm]

    Added more aggressive caching to the mapper’s usage of UPDATE, INSERT, and DELETE expressions. Assuming the statement has no per-object SQL expressions attached, the expression objects are cached by the mapper after the first create, and their compiled form is stored persistently in a cache dictionary for the duration of the related Engine. The cache is an LRUCache for the rare case that a mapper receives an extremely high number of different column patterns as UPDATEs.

sql

  • [sql]

    expr.in_() now accepts a text() construct as the argument. Grouping parenthesis are added automatically, i.e. usage is like col.in_(text(“select id from table”)).

    References: #1793

  • [sql]

    Columns of _Binary type (i.e. LargeBinary, BLOB, etc.) will coerce a “basestring” on the right side into a _Binary as well so that required DBAPI processing takes place.

  • [sql]

    Added table.add_is_dependent_on(othertable), allows manual placement of dependency rules between two Table objects for use within create_all(), drop_all(), sorted_tables.

    References: #1801

  • [sql]

    Fixed bug that prevented implicit RETURNING from functioning properly with composite primary key that contained zeroes.

    References: #1778

  • [sql]

    Fixed errant space character when generating ADD CONSTRAINT for a named UNIQUE constraint.

  • [sql]

    Fixed “table” argument on constructor of ForeignKeyConstraint

    References: #1571

  • [sql]

    Fixed bug in connection pool cursor wrapper whereby if a cursor threw an exception on close(), the logging of the message would fail.

    References: #1786

  • [sql]

    the _make_proxy() method of ColumnClause and Column now use self.__class__ to determine the class of object to be returned instead of hardcoding to ColumnClause/Column, making it slightly easier to produce specific subclasses of these which work in alias/subquery situations.

  • [sql]

    func.XXX() doesn’t inadvertently resolve to non-Function classes (e.g. fixes func.text()).

    References: #1798

mysql

  • [mysql]

    func.sysdate() emits “SYSDATE()”, i.e. with the ending parenthesis, on MySQL.

    References: #1794

sqlite

  • [sqlite]

    Fixed concatenation of constraints when “PRIMARY KEY” constraint gets moved to column level due to SQLite AUTOINCREMENT keyword being rendered.

    References: #1812

oracle

  • [oracle]

    Added a check for cx_oracle versions lower than version 5, in which case the incompatible “output type handler” won’t be used. This will impact decimal accuracy and some unicode handling issues.

    References: #1775

  • [oracle]

    Fixed use_ansi=False mode, which was producing broken WHERE clauses in pretty much all cases.

    References: #1790

  • [oracle]

    Re-established support for Oracle 8 with cx_oracle, including that use_ansi is set to False automatically, NVARCHAR2 and NCLOB are not rendered for Unicode, “native unicode” check doesn’t fail, cx_oracle “native unicode” mode is disabled, VARCHAR() is emitted with bytes count instead of char count.

    References: #1808

  • [oracle]

    oracle_xe 5 doesn’t accept a Python unicode object in its connect string in normal Python 2.x mode - so we coerce to str() directly. non-ascii characters aren’t supported in connect strings here since we don’t know what encoding we could use.

    References: #1670

  • [oracle]

    FOR UPDATE is emitted in the syntactically correct position when limit/offset is used, i.e. the ROWNUM subquery. However, Oracle can’t really handle FOR UPDATE with ORDER BY or with subqueries, so its still not very usable, but at least SQLA gets the SQL past the Oracle parser.

    References: #1815

firebird

  • [firebird]

    Added a label to the query used within has_table() and has_sequence() to work with older versions of Firebird that don’t provide labels for result columns.

    References: #1521

  • [firebird]

    Added integer coercion to the “type_conv” attribute when passed via query string, so that it is properly interpreted by Kinterbasdb.

    References: #1779

  • [firebird]

    Added ‘connection shutdown’ to the list of exception strings which indicate a dropped connection.

    References: #1646

misc

  • [engines]

    Fixed building the C extensions on Python 2.4.

    References: #1781

  • [engines]

    Pool classes will reuse the same “pool_logging_name” setting after a dispose() occurs.

  • [engines]

    Engine gains an “execution_options” argument and update_execution_options() method, which will apply to all connections generated by this engine.

  • [sqlsoup]

    the SqlSoup constructor accepts a base argument which specifies the base class to use for mapped classes, the default being object.

    References: #1783

0.6.0

Released: Sun Apr 18 2010

orm

  • [orm]

    Unit of work internals have been rewritten. Units of work with large numbers of objects interdependent objects can now be flushed without recursion overflows as there is no longer reliance upon recursive calls. The number of internal structures now stays constant for a particular session state, regardless of how many relationships are present on mappings. The flow of events now corresponds to a linear list of steps, generated by the mappers and relationships based on actual work to be done, filtered through a single topological sort for correct ordering. Flush actions are assembled using far fewer steps and less memory.

    References: #1081, #1742

  • [orm]

    Along with the UOW rewrite, this also removes an issue introduced in 0.6beta3 regarding topological cycle detection for units of work with long dependency cycles. We now use an algorithm written by Guido (thanks Guido!).

  • [orm]

    one-to-many relationships now maintain a list of positive parent-child associations within the flush, preventing previous parents marked as deleted from cascading a delete or NULL foreign key set on those child objects, despite the end-user not removing the child from the old association.

    References: #1764

  • [orm]

    A collection lazy load will switch off default eagerloading on the reverse many-to-one side, since that loading is by definition unnecessary.

    References: #1495

  • [orm]

    Session.refresh() now does an equivalent expire() on the given instance first, so that the “refresh-expire” cascade is propagated. Previously, refresh() was not affected in any way by the presence of “refresh-expire” cascade. This is a change in behavior versus that of 0.6beta2, where the “lockmode” flag passed to refresh() would cause a version check to occur. Since the instance is first expired, refresh() always upgrades the object to the most recent version.

  • [orm]

    The ‘refresh-expire’ cascade, when reaching a pending object, will expunge the object if the cascade also includes “delete-orphan”, or will simply detach it otherwise.

    References: #1754

  • [orm]

    id(obj) is no longer used internally within topological.py, as the sorting functions now require hashable objects only.

    References: #1756

  • [orm]

    The ORM will set the docstring of all generated descriptors to None by default. This can be overridden using ‘doc’ (or if using Sphinx, attribute docstrings work too).

  • [orm]

    Added kw argument ‘doc’ to all mapper property callables as well as Column(). Will assemble the string ‘doc’ as the ‘__doc__’ attribute on the descriptor.

  • [orm]

    Usage of version_id_col on a backend that supports cursor.rowcount for execute() but not executemany() now works when a delete is issued (already worked for saves, since those don’t use executemany()). For a backend that doesn’t support cursor.rowcount at all, a warning is emitted the same as with saves.

    References: #1761

  • [orm]

    The ORM now short-term caches the “compiled” form of insert() and update() constructs when flushing lists of objects of all the same class, thereby avoiding redundant compilation per individual INSERT/UPDATE within an individual flush() call.

  • [orm]

    internal getattr(), setattr(), getcommitted() methods on ColumnProperty, CompositeProperty, RelationshipProperty have been underscored (i.e. are private), signature has changed.

examples

  • [examples]

    Updated attribute_shard.py example to use a more robust method of searching a Query for binary expressions which compare columns against literal values.

sql

  • [sql]

    Restored some bind-labeling logic from 0.5 which ensures that tables with column names that overlap another column of the form “<tablename>_<columnname>” won’t produce errors if column._label is used as a bind name during an UPDATE. Test coverage which wasn’t present in 0.5 has been added.

    References: #1755

  • [sql]

    somejoin.select(fold_equivalents=True) is no longer deprecated, and will eventually be rolled into a more comprehensive version of the feature for.

    References: #1729

  • [sql]

    the Numeric type raises an enormous warning when expected to convert floats to Decimal from a DBAPI that returns floats. This includes SQLite, Sybase, MS-SQL.

    References: #1759

  • [sql]

    Fixed an error in expression typing which caused an endless loop for expressions with two NULL types.

  • [sql]

    Fixed bug in execution_options() feature whereby the existing Transaction and other state information from the parent connection would not be propagated to the sub-connection.

  • [sql]

    Added new ‘compiled_cache’ execution option. A dictionary where Compiled objects will be cached when the Connection compiles a clause expression into a dialect- and parameter- specific Compiled object. It is the user’s responsibility to manage the size of this dictionary, which will have keys corresponding to the dialect, clause element, the column names within the VALUES or SET clause of an INSERT or UPDATE, as well as the “batch” mode for an INSERT or UPDATE statement.

  • [sql]

    Added get_pk_constraint() to reflection.Inspector, similar to get_primary_keys() except returns a dict that includes the name of the constraint, for supported backends (PG so far).

    References: #1769

  • [sql]

    Table.create() and Table.drop() no longer apply metadata- level create/drop events.

    References: #1771

postgresql

  • [postgresql]

    PostgreSQL now reflects sequence names associated with SERIAL columns correctly, after the name of the sequence has been changed. Thanks to Kumar McMillan for the patch.

    References: #1071

  • [postgresql]

    Repaired missing import in psycopg2._PGNumeric type when unknown numeric is received.

  • [postgresql]

    psycopg2/pg8000 dialects now aware of REAL[], FLOAT[], DOUBLE_PRECISION[], NUMERIC[] return types without raising an exception.

  • [postgresql]

    PostgreSQL reflects the name of primary key constraints, if one exists.

    References: #1769

oracle

  • [oracle]

    Now using cx_oracle output converters so that the DBAPI returns natively the kinds of values we prefer:

  • [oracle]

    NUMBER values with positive precision + scale convert to cx_oracle.STRING and then to Decimal. This allows perfect precision for the Numeric type when using cx_oracle.

    References: #1759

  • [oracle]

    STRING/FIXED_CHAR now convert to unicode natively. SQLAlchemy’s String types then don’t need to apply any kind of conversions.

firebird

  • [firebird]

    The functionality of result.rowcount can be disabled on a per-engine basis by setting ‘enable_rowcount=False’ on create_engine(). Normally, cursor.rowcount is called after any UPDATE or DELETE statement unconditionally, because the cursor is then closed and Firebird requires an open cursor in order to get a rowcount. This call is slightly expensive however so it can be disabled. To re-enable on a per-execution basis, the ‘enable_rowcount=True’ execution option may be used.

misc

  • [engines]

    The C extension now also works with DBAPIs which use custom sequences as row (and not only tuples).

    References: #1757

  • [ext]

    the compiler extension now allows @compiles decorators on base classes that extend to child classes, @compiles decorators on child classes that aren’t broken by a @compiles decorator on the base class.

  • [ext]

    Declarative will raise an informative error message if a non-mapped class attribute is referenced in the string-based relationship() arguments.

  • [ext]

    Further reworked the “mixin” logic in declarative to additionally allow __mapper_args__ as a @classproperty on a mixin, such as to dynamically assign polymorphic_identity.

0.6beta3

Released: Sun Mar 28 2010

orm

  • [orm]

    Major feature: Added new “subquery” loading capability to relationship(). This is an eager loading option which generates a second SELECT for each collection represented in a query, across all parents at once. The query re-issues the original end-user query wrapped in a subquery, applies joins out to the target collection, and loads all those collections fully in one result, similar to “joined” eager loading but using all inner joins and not re-fetching full parent rows repeatedly (as most DBAPIs seem to do, even if columns are skipped). Subquery loading is available at mapper config level using “lazy=’subquery’” and at the query options level using “subqueryload(props..)”, “subqueryload_all(props…)”.

    References: #1675

  • [orm]

    To accommodate the fact that there are now two kinds of eager loading available, the new names for eagerload() and eagerload_all() are joinedload() and joinedload_all(). The old names will remain as synonyms for the foreseeable future.

  • [orm]

    The “lazy” flag on the relationship() function now accepts a string argument for all kinds of loading: “select”, “joined”, “subquery”, “noload” and “dynamic”, where the default is now “select”. The old values of True/ False/None still retain their usual meanings and will remain as synonyms for the foreseeable future.

  • [orm]

    Added with_hint() method to Query() construct. This calls directly down to select().with_hint() and also accepts entities as well as tables and aliases. See with_hint() in the SQL section below.

    References: #921

  • [orm]

    Fixed bug in Query whereby calling q.join(prop).from_self(…). join(prop) would fail to render the second join outside the subquery, when joining on the same criterion as was on the inside.

  • [orm]

    Fixed bug in Query whereby the usage of aliased() constructs would fail if the underlying table (but not the actual alias) were referenced inside the subquery generated by q.from_self() or q.select_from().

  • [orm]

    Fixed bug which affected all eagerload() and similar options such that “remote” eager loads, i.e. eagerloads off of a lazy load such as query(A).options(eagerload(A.b, B.c)) wouldn’t eagerload anything, but using eagerload(“b.c”) would work fine.

  • [orm]

    Query gains an add_columns(*columns) method which is a multi- version of add_column(col). add_column(col) is future deprecated.

  • [orm]

    Query.join() will detect if the end result will be “FROM A JOIN A”, and will raise an error if so.

  • [orm]

    Query.join(Cls.propname, from_joinpoint=True) will check more carefully that “Cls” is compatible with the current joinpoint, and act the same way as Query.join(“propname”, from_joinpoint=True) in that regard.

sql

  • [sql]

    Added with_hint() method to select() construct. Specify a table/alias, hint text, and optional dialect name, and “hints” will be rendered in the appropriate place in the statement. Works for Oracle, Sybase, MySQL.

    References: #921

  • [sql]

    Fixed bug introduced in 0.6beta2 where column labels would render inside of column expressions already assigned a label.

    References: #1747

postgresql

  • [postgresql]

    The psycopg2 dialect will log NOTICE messages via the “sqlalchemy.dialects.postgresql” logger name.

    References: #877

  • [postgresql]

    the TIME and TIMESTAMP types are now available from the postgresql dialect directly, which add the PG-specific argument ‘precision’ to both. ‘precision’ and ‘timezone’ are correctly reflected for both TIME and TIMEZONE types.

    References: #997

mysql

  • [mysql]

    No longer guessing that TINYINT(1) should be BOOLEAN when reflecting - TINYINT(1) is returned. Use Boolean/ BOOLEAN in table definition to get boolean conversion behavior.

    References: #1752

oracle

  • [oracle]

    The Oracle dialect will issue VARCHAR type definitions using character counts, i.e. VARCHAR2(50 CHAR), so that the column is sized in terms of characters and not bytes. Column reflection of character types will also use ALL_TAB_COLUMNS.CHAR_LENGTH instead of ALL_TAB_COLUMNS.DATA_LENGTH. Both of these behaviors take effect when the server version is 9 or higher - for version 8, the old behaviors are used.

    References: #1744

misc

  • [declarative]

    Using a mixin won’t break if the mixin implements an unpredictable __getattribute__(), i.e. Zope interfaces.

    References: #1746

  • [declarative]

    Using @classdecorator and similar on mixins to define __tablename__, __table_args__, etc. now works if the method references attributes on the ultimate subclass.

    References: #1749

  • [declarative]

    relationships and columns with foreign keys aren’t allowed on declarative mixins, sorry.

    References: #1751

  • [ext]

    The sqlalchemy.orm.shard module now becomes an extension, sqlalchemy.ext.horizontal_shard. The old import works with a deprecation warning.

0.6beta2

Released: Sat Mar 20 2010

orm

  • [orm]

    The official name for the relation() function is now relationship(), to eliminate confusion over the relational algebra term. relation() however will remain available in equal capacity for the foreseeable future.

    References: #1740

  • [orm]

    Added “version_id_generator” argument to Mapper, this is a callable that, given the current value of the “version_id_col”, returns the next version number. Can be used for alternate versioning schemes such as uuid, timestamps.

    References: #1692

  • [orm]

    added “lockmode” kw argument to Session.refresh(), will pass through the string value to Query the same as in with_lockmode(), will also do version check for a version_id_col-enabled mapping.

  • [orm]

    Fixed bug whereby calling query(A).join(A.bs).add_entity(B) in a joined inheritance scenario would double-add B as a target and produce an invalid query.

    References: #1188

  • [orm]

    Fixed bug in session.rollback() which involved not removing formerly “pending” objects from the session before re-integrating “deleted” objects, typically occurred with natural primary keys. If there was a primary key conflict between them, the attach of the deleted would fail internally. The formerly “pending” objects are now expunged first.

    References: #1674

  • [orm]

    Removed a lot of logging that nobody really cares about, logging that remains will respond to live changes in the log level. No significant overhead is added.

    References: #1719

  • [orm]

    Fixed bug in session.merge() which prevented dict-like collections from merging.

  • [orm]

    session.merge() works with relations that specifically don’t include “merge” in their cascade options - the target is ignored completely.

  • [orm]

    session.merge() will not expire existing scalar attributes on an existing target if the target has a value for that attribute, even if the incoming merged doesn’t have a value for the attribute. This prevents unnecessary loads on existing items. Will still mark the attr as expired if the destination doesn’t have the attr, though, which fulfills some contracts of deferred cols.

    References: #1681

  • [orm]

    The “allow_null_pks” flag is now called “allow_partial_pks”, defaults to True, acts like it did in 0.5 again. Except, it also is implemented within merge() such that a SELECT won’t be issued for an incoming instance with partially NULL primary key if the flag is False.

    References: #1680

  • [orm]

    Fixed bug in 0.6-reworked “many-to-one” optimizations such that a many-to-one that is against a non-primary key column on the remote table (i.e. foreign key against a UNIQUE column) will pull the “old” value in from the database during a change, since if it’s in the session we will need it for proper history/backref accounting, and we can’t pull from the local identity map on a non-primary key column.

    References: #1737

  • [orm]

    fixed internal error which would occur if calling has() or similar complex expression on a single-table inheritance relation().

    References: #1731

  • [orm]

    query.one() no longer applies LIMIT to the query, this to ensure that it fully counts all object identities present in the result, even in the case where joins may conceal multiple identities for two or more rows. As a bonus, one() can now also be called with a query that issued from_statement() to start with since it no longer modifies the query.

    References: #1688

  • [orm]

    query.get() now returns None if queried for an identifier that is present in the identity map with a different class than the one requested, i.e. when using polymorphic loading.

    References: #1727

  • [orm]

    A major fix in query.join(), when the “on” clause is an attribute of an aliased() construct, but there is already an existing join made out to a compatible target, query properly joins to the right aliased() construct instead of sticking onto the right side of the existing join.

    References: #1706

  • [orm]

    Slight improvement to the fix for to not issue needless updates of the primary key column during a so-called “row switch” operation, i.e. add + delete of two objects with the same PK.

    References: #1362

  • [orm]

    Now uses sqlalchemy.orm.exc.DetachedInstanceError when an attribute load or refresh action fails due to object being detached from any Session. UnboundExecutionError is specific to engines bound to sessions and statements.

  • [orm]

    Query called in the context of an expression will render disambiguating labels in all cases. Note that this does not apply to the existing .statement and .subquery() accessor/method, which still honors the .with_labels() setting that defaults to False.

  • [orm]

    Query.union() retains disambiguating labels within the returned statement, thus avoiding various SQL composition errors which can result from column name conflicts.

    References: #1676

  • [orm]

    Fixed bug in attribute history that inadvertently invoked __eq__ on mapped instances.

  • [orm]

    Some internal streamlining of object loading grants a small speedup for large results, estimates are around 10-15%. Gave the “state” internals a good solid cleanup with less complexity, datamembers, method calls, blank dictionary creates.

  • [orm]

    Documentation clarification for query.delete()

    References: #1689

  • [orm]

    Fixed cascade bug in many-to-one relation() when attribute was set to None, introduced in r6711 (cascade deleted items into session during add()).

  • [orm]

    Calling query.order_by() or query.distinct() before calling query.select_from(), query.with_polymorphic(), or query.from_statement() raises an exception now instead of silently dropping those criterion.

    References: #1736

  • [orm]

    query.scalar() now raises an exception if more than one row is returned. All other behavior remains the same.

    References: #1735

  • [orm]

    Fixed bug which caused “row switch” logic, that is an INSERT and DELETE replaced by an UPDATE, to fail when version_id_col was in use.

    References: #1692

examples

  • [examples]

    Changed the beaker cache example a bit to have a separate RelationCache option for lazyload caching. This object does a lookup among any number of potential attributes more efficiently by grouping several into a common structure. Both FromCache and RelationCache are simpler individually.

sql

  • [sql]

    join() will now simulate a NATURAL JOIN by default. Meaning, if the left side is a join, it will attempt to join the right side to the rightmost side of the left first, and not raise any exceptions about ambiguous join conditions if successful even if there are further join targets across the rest of the left.

    References: #1714

  • [sql]

    The most common result processors conversion function were moved to the new “processors” module. Dialect authors are encouraged to use those functions whenever they correspond to their needs instead of implementing custom ones.

  • [sql]

    SchemaType and subclasses Boolean, Enum are now serializable, including their ddl listener and other event callables.

    References: #1694, #1698

  • [sql]

    Some platforms will now interpret certain literal values as non-bind parameters, rendered literally into the SQL statement. This to support strict SQL-92 rules that are enforced by some platforms including MS-SQL and Sybase. In this model, bind parameters aren’t allowed in the columns clause of a SELECT, nor are certain ambiguous expressions like “?=?”. When this mode is enabled, the base compiler will render the binds as inline literals, but only across strings and numeric values. Other types such as dates will raise an error, unless the dialect subclass defines a literal rendering function for those. The bind parameter must have an embedded literal value already or an error is raised (i.e. won’t work with straight bindparam(‘x’)). Dialects can also expand upon the areas where binds are not accepted, such as within argument lists of functions (which don’t work on MS-SQL when native SQL binding is used).

  • [sql]

    Added “unicode_errors” parameter to String, Unicode, etc. Behaves like the ‘errors’ keyword argument to the standard library’s string.decode() functions. This flag requires that convert_unicode is set to “force” - otherwise, SQLAlchemy is not guaranteed to handle the task of unicode conversion. Note that this flag adds significant performance overhead to row-fetching operations for backends that already return unicode objects natively (which most DBAPIs do). This flag should only be used as an absolute last resort for reading strings from a column with varied or corrupted encodings, which only applies to databases that accept invalid encodings in the first place (i.e. MySQL. not PG, Sqlite, etc.)

  • [sql]

    Added math negation operator support, -x.

  • [sql]

    FunctionElement subclasses are now directly executable the same way any func.foo() construct is, with automatic SELECT being applied when passed to execute().

  • [sql]

    The “type” and “bind” keyword arguments of a func.foo() construct are now local to “func.” constructs and are not part of the FunctionElement base class, allowing a “type” to be handled in a custom constructor or class-level variable.

  • [sql]

    Restored the keys() method to ResultProxy.

  • [sql]

    The type/expression system now does a more complete job of determining the return type from an expression as well as the adaptation of the Python operator into a SQL operator, based on the full left/right/operator of the given expression. In particular the date/time/interval system created for PostgreSQL EXTRACT in has now been generalized into the type system. The previous behavior which often occurred of an expression “column + literal” forcing the type of “literal” to be the same as that of “column” will now usually not occur - the type of “literal” is first derived from the Python type of the literal, assuming standard native Python types + date types, before falling back to that of the known type on the other side of the expression. If the “fallback” type is compatible (i.e. CHAR from String), the literal side will use that. TypeDecorator types override this by default to coerce the “literal” side unconditionally, which can be changed by implementing the coerce_compared_value() method. Also part of.

    References: #1647, #1683

  • [sql]

    Made sqlalchemy.sql.expressions.Executable part of public API, used for any expression construct that can be sent to execute(). FunctionElement now inherits Executable so that it gains execution_options(), which are also propagated to the select() that’s generated within execute(). Executable in turn subclasses _Generative which marks any ClauseElement that supports the @_generative decorator - these may also become “public” for the benefit of the compiler extension at some point.

  • [sql]

    A change to the solution for - an end-user defined bind parameter name that directly conflicts with a column-named bind generated directly from the SET or VALUES clause of an update/insert generates a compile error. This reduces call counts and eliminates some cases where undesirable name conflicts could still occur.

    References: #1579

  • [sql]

    Column() requires a type if it has no foreign keys (this is not new). An error is now raised if a Column() has no type and no foreign keys.

    References: #1705

  • [sql]

    the “scale” argument of the Numeric() type is honored when coercing a returned floating point value into a string on its way to Decimal - this allows accuracy to function on SQLite, MySQL.

    References: #1717

  • [sql]

    the copy() method of Column now copies over uninitialized “on table attach” events. Helps with the new declarative “mixin” capability.

mysql

  • [mysql]

    Fixed reflection bug whereby when COLLATE was present, nullable flag and server defaults would not be reflected.

    References: #1655

  • [mysql]

    Fixed reflection of TINYINT(1) “boolean” columns defined with integer flags like UNSIGNED.

  • [mysql]

    Further fixes for the mysql-connector dialect.

    References: #1668

  • [mysql]

    Composite PK table on InnoDB where the “autoincrement” column isn’t first will emit an explicit “KEY” phrase within CREATE TABLE thereby avoiding errors.

    References: #1496

  • [mysql]

    Added reflection/create table support for a wide range of MySQL keywords.

    References: #1634

  • [mysql]

    Fixed import error which could occur reflecting tables on a Windows host

    References: #1580

sqlite

  • [sqlite]

    Added “native_datetime=True” flag to create_engine(). This will cause the DATE and TIMESTAMP types to skip all bind parameter and result row processing, under the assumption that PARSE_DECLTYPES has been enabled on the connection. Note that this is not entirely compatible with the “func.current_date()”, which will be returned as a string.

    References: #1685

mssql

  • [mssql]

    Re-established support for the pymssql dialect.

  • [mssql]

    Various fixes for implicit returning, reflection, etc. - the MS-SQL dialects aren’t quite complete in 0.6 yet (but are close)

  • [mssql]

    Added basic support for mxODBC.

    References: #1710

  • [mssql]

    Removed the text_as_varchar option.

oracle

  • [oracle]

    ”out” parameters require a type that is supported by cx_oracle. An error will be raised if no cx_oracle type can be found.

  • [oracle]

    Oracle ‘DATE’ now does not perform any result processing, as the DATE type in Oracle stores full date+time objects, that’s what you’ll get. Note that the generic types.Date type will still call value.date() on incoming values, however. When reflecting a table, the reflected type will be ‘DATE’.

  • [oracle]

    Added preliminary support for Oracle’s WITH_UNICODE mode. At the very least this establishes initial support for cx_Oracle with Python 3. When WITH_UNICODE mode is used in Python 2.xx, a large and scary warning is emitted asking that the user seriously consider the usage of this difficult mode of operation.

    References: #1670

  • [oracle]

    The except_() method now renders as MINUS on Oracle, which is more or less equivalent on that platform.

    References: #1712

  • [oracle]

    Added support for rendering and reflecting TIMESTAMP WITH TIME ZONE, i.e. TIMESTAMP(timezone=True).

    References: #651

  • [oracle]

    Oracle INTERVAL type can now be reflected.

misc

  • [py3k]

    Improved the installation/test setup regarding Python 3, now that Distribute runs on Py3k. distribute_setup.py is now included. See README.py3k for Python 3 installation/ testing instructions.

  • [engines]

    Added an optional C extension to speed up the sql layer by reimplementing RowProxy and the most common result processors. The actual speedups will depend heavily on your DBAPI and the mix of datatypes used in your tables, and can vary from a 30% improvement to more than 200%. It also provides a modest (~15-20%) indirect improvement to ORM speed for large queries. Note that it is not built/installed by default. See README for installation instructions.

  • [engines]

    the execution sequence pulls all rowcount/last inserted ID info from the cursor before commit() is called on the DBAPI connection in an “autocommit” scenario. This helps mxodbc with rowcount and is probably a good idea overall.

  • [engines]

    Opened up logging a bit such that isEnabledFor() is called more often, so that changes to the log level for engine/pool will be reflected on next connect. This adds a small amount of method call overhead. It’s negligible and will make life a lot easier for all those situations when logging just happens to be configured after create_engine() is called.

    References: #1719

  • [engines]

    The assert_unicode flag is deprecated. SQLAlchemy will raise a warning in all cases where it is asked to encode a non-unicode Python string, as well as when a Unicode or UnicodeType type is explicitly passed a bytestring. The String type will do nothing for DBAPIs that already accept Python unicode objects.

  • [engines]

    Bind parameters are sent as a tuple instead of a list. Some backend drivers will not accept bind parameters as a list.

  • [engines]

    threadlocal engine wasn’t properly closing the connection upon close() - fixed that.

  • [engines]

    Transaction object doesn’t rollback or commit if it isn’t “active”, allows more accurate nesting of begin/rollback/commit.

  • [engines]

    Python unicode objects as binds result in the Unicode type, not string, thus eliminating a certain class of unicode errors on drivers that don’t support unicode binds.

  • [engines]

    Added “logging_name” argument to create_engine(), Pool() constructor as well as “pool_logging_name” argument to create_engine() which filters down to that of Pool. Issues the given string name within the “name” field of logging messages instead of the default hex identifier string.

    References: #1555

  • [engines]

    The visit_pool() method of Dialect is removed, and replaced with on_connect(). This method returns a callable which receives the raw DBAPI connection after each one is created. The callable is assembled into a first_connect/connect pool listener by the connection strategy if non-None. Provides a simpler interface for dialects.

  • [engines]

    StaticPool now initializes, disposes and recreates without opening a new connection - the connection is only opened when first requested. dispose() also works on AssertionPool now.

    References: #1728

  • [metadata] [ticket: 1673]

    Added the ability to strip schema information when using “tometadata” by passing “schema=None” as an argument. If schema is not specified then the table’s schema is retained.

  • [declarative]

    DeclarativeMeta exclusively uses cls.__dict__ (not dict_) as the source of class information; _as_declarative exclusively uses the dict_ passed to it as the source of class information (which when using DeclarativeMeta is cls.__dict__). This should in theory make it easier for custom metaclasses to modify the state passed into _as_declarative.

  • [declarative]

    declarative now accepts mixin classes directly, as a means to provide common functional and column-based elements on all subclasses, as well as a means to propagate a fixed set of __table_args__ or __mapper_args__ to subclasses. For custom combinations of __table_args__/__mapper_args__ from an inherited mixin to local, descriptors can now be used. New details are all up in the Declarative documentation. Thanks to Chris Withers for putting up with my strife on this.

    References: #1707

  • [declarative]

    the __mapper_args__ dict is copied when propagating to a subclass, and is taken straight off the class __dict__ to avoid any propagation from the parent. mapper inheritance already propagates the things you want from the parent mapper.

    References: #1393

  • [declarative]

    An exception is raised when a single-table subclass specifies a column that is already present on the base class.

    References: #1732

  • [sybase]

    Implemented a preliminary working dialect for Sybase, with sub-implementations for Python-Sybase as well as Pyodbc. Handles table creates/drops and basic round trip functionality. Does not yet include reflection or comprehensive support of unicode/special expressions/etc.

  • [documentation]

    Major cleanup work in the docs to link class, function, and method names into the API docs.

    References: #1700

0.6beta1

Released: Wed Feb 03 2010

orm

  • [orm]

    Changes to query.update() and query.delete():
    • the ‘expire’ option on query.update() has been renamed to ‘fetch’, thus matching that of query.delete(). ‘expire’ is deprecated and issues a warning.

    • query.update() and query.delete() both default to ‘evaluate’ for the synchronize strategy.

    • the ‘synchronize’ strategy for update() and delete() raises an error on failure. There is no implicit fallback onto “fetch”. Failure of evaluation is based on the structure of criteria, so success/failure is deterministic based on code structure.

  • [orm]

    Enhancements on many-to-one relations:
    • 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.

    • specifying a foreign key with a declarative column, i.e. ForeignKey(MyRelatedClass.id) doesn’t break the “use_get” condition from taking place

    • relation(), eagerload(), and eagerload_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 relation(). Should generally be set for many-to-one, not nullable foreign key relations to allow improved join performance.

    • the behavior of eagerloading 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.

    References: #1186, #1492, #1544

  • [orm]

    Enhancements / Changes on Session.merge():

  • [orm]

    the “dont_load=True” flag on Session.merge() is deprecated and is now “load=False”.

  • [orm]

    Session.merge() is performance optimized, using half the call counts for “load=False” mode compared to 0.5 and significantly fewer SQL queries in the case of collections for “load=True” mode.

  • [orm]

    merge() will not issue a needless merge of attributes if the given instance is the same instance which is already present.

  • [orm]

    merge() now also merges the “options” associated with a given state, i.e. those passed through query.options() which follow along with an instance, such as options to eagerly- or lazyily- load various attributes. This is essential for the construction of highly integrated caching schemes. This is a subtle behavioral change vs. 0.5.

  • [orm]

    A bug was fixed regarding the serialization of the “loader path” present on an instance’s state, which is also necessary when combining the usage of merge() with serialized state and associated options that should be preserved.

  • [orm]

    The all new merge() is showcased in a new comprehensive example of how to integrate Beaker with SQLAlchemy. See the notes in the “examples” note below.

  • [orm]

    Primary key values can now be changed on a joined-table inheritance object, and ON UPDATE CASCADE will be taken into account when the flush happens. Set the new “passive_updates” flag to False on mapper() when using SQLite or MySQL/MyISAM.

    References: #1362

  • [orm]

    flush() now detects when a primary key column was updated by an ON UPDATE CASCADE operation from another primary key, and can then locate the row for a subsequent UPDATE on the new PK value. This occurs when a relation() is there to establish the relationship as well as passive_updates=True.

    References: #1671

  • [orm]

    the “save-update” cascade will now cascade the pending removed values from a scalar or collection attribute into the new session during an add() operation. This so that the flush() operation will also delete or modify rows of those disconnected items.

  • [orm]

    Using a “dynamic” loader with a “secondary” table now produces a query where the “secondary” table is not aliased. This allows the secondary Table object to be used in the “order_by” attribute of the relation(), and also allows it to be used in filter criterion against the dynamic relation.

    References: #1531

  • [orm]

    relation() with uselist=False will emit a warning when an eager or lazy load locates more than one valid value for the row. This may be due to primaryjoin/secondaryjoin conditions which aren’t appropriate for an eager LEFT OUTER JOIN or for other conditions.

    References: #1643

  • [orm]

    an explicit check occurs when a synonym() is used with map_column=True, when a ColumnProperty (deferred or otherwise) exists separately in the properties dictionary sent to mapper with the same keyname. Instead of silently replacing the existing property (and possible options on that property), an error is raised.

    References: #1633

  • [orm]

    a “dynamic” loader sets up its query criterion at construction time so that the actual query is returned from non-cloning accessors like “statement”.

  • [orm]

    the “named tuple” objects returned when iterating a Query() are now pickleable.

  • [orm]

    mapping to a select() construct now requires that you make an alias() out of it distinctly. This to eliminate confusion over such issues as

    References: #1542

  • [orm]

    query.join() has been reworked to provide more consistent behavior and more flexibility (includes)

    References: #1537

  • [orm]

    query.select_from() accepts multiple clauses to produce multiple comma separated entries within the FROM clause. Useful when selecting from multiple-homed join() clauses.

  • [orm]

    query.select_from() also accepts mapped classes, aliased() constructs, and mappers as arguments. In particular this helps when querying from multiple joined-table classes to ensure the full join gets rendered.

  • [orm]

    query.get() can be used with a mapping to an outer join where one or more of the primary key values are None.

    References: #1135

  • [orm]

    query.from_self(), query.union(), others which do a “SELECT * from (SELECT…)” type of nesting will do a better job translating column expressions within the subquery to the columns clause of the outer query. This is potentially backwards incompatible with 0.5, in that this may break queries with literal expressions that do not have labels applied (i.e. literal(‘foo’), etc.)

    References: #1568

  • [orm]

    relation primaryjoin and secondaryjoin now check that they are column-expressions, not just clause elements. this prohibits things like FROM expressions being placed there directly.

    References: #1622

  • [orm]

    expression.null() is fully understood the same way None is when comparing an object/collection-referencing attribute within query.filter(), filter_by(), etc.

    References: #1415

  • [orm]

    added “make_transient()” helper function which transforms a persistent/ detached instance into a transient one (i.e. deletes the instance_key and removes from any session.)

    References: #1052

  • [orm]

    the allow_null_pks flag on mapper() is deprecated, and the feature is turned “on” by default. This means that a row which has a non-null value for any of its primary key columns will be considered an identity. The need for this scenario typically only occurs when mapping to an outer join.

    References: #1339

  • [orm]

    the mechanics of “backref” have been fully merged into the finer grained “back_populates” system, and take place entirely within the _generate_backref() method of RelationProperty. This makes the initialization procedure of RelationProperty simpler and allows easier propagation of settings (such as from subclasses of RelationProperty) into the reverse reference. The internal BackRef() is gone and backref() returns a plain tuple that is understood by RelationProperty.

  • [orm]

    The version_id_col feature on mapper() will raise a warning when used with dialects that don’t support “rowcount” adequately.

    References: #1569

  • [orm]

    added “execution_options()” to Query, to so options can be passed to the resulting statement. Currently only Select-statements have these options, and the only option used is “stream_results”, and the only dialect which knows “stream_results” is psycopg2.

  • [orm]

    Query.yield_per() will set the “stream_results” statement option automatically.

  • [orm]

    Deprecated or removed:
    • ’allow_null_pks’ flag on mapper() is deprecated. It does nothing now and the setting is “on” in all cases.

    • ’transactional’ flag on sessionmaker() and others is removed. Use ‘autocommit=True’ to indicate ‘transactional=False’.

    • ’polymorphic_fetch’ argument on mapper() is removed. Loading can be controlled using the ‘with_polymorphic’ option.

    • ’select_table’ argument on mapper() is removed. Use ‘with_polymorphic=(“*”, <some selectable>)’ for this functionality.

    • ’proxy’ argument on synonym() is removed. This flag did nothing throughout 0.5, as the “proxy generation” behavior is now automatic.

    • Passing a single list of elements to eagerload(), eagerload_all(), contains_eager(), lazyload(), defer(), and undefer() instead of multiple positional *args is deprecated.

    • Passing a single list of elements to query.order_by(), query.group_by(), query.join(), or query.outerjoin() instead of multiple positional *args is deprecated.

    • query.iterate_instances() is removed. Use query.instances().

    • Query.query_from_parent() is removed. Use the sqlalchemy.orm.with_parent() function to produce a “parent” clause, or alternatively query.with_parent().

    • query._from_self() is removed, use query.from_self() instead.

    • the “comparator” argument to composite() is removed. Use “comparator_factory”.

    • RelationProperty._get_join() is removed.

    • the ‘echo_uow’ flag on Session is removed. Use logging on the “sqlalchemy.orm.unitofwork” name.

    • session.clear() is removed. use session.expunge_all().

    • session.save(), session.update(), session.save_or_update() are removed. Use session.add() and session.add_all().

    • the “objects” flag on session.flush() remains deprecated.

    • the “dont_load=True” flag on session.merge() is deprecated in favor of “load=False”.

    • ScopedSession.mapper remains deprecated. See the usage recipe at https://www.sqlalchemy.org/trac/wiki/UsageRecipes/SessionAwareMapper

    • passing an InstanceState (internal SQLAlchemy state object) to attributes.init_collection() or attributes.get_history() is deprecated. These functions are public API and normally expect a regular mapped object instance.

    • the ‘engine’ parameter to declarative_base() is removed. Use the ‘bind’ keyword argument.

sql

  • [sql]

    the “autocommit” flag on select() and text() as well as select().autocommit() are deprecated - now call .execution_options(autocommit=True) on either of those constructs, also available directly on Connection and orm.Query.

  • [sql]

    the autoincrement flag on column now indicates the column which should be linked to cursor.lastrowid, if that method is used. See the API docs for details.

  • [sql]

    an executemany() now requires that all bound parameter sets require that all keys are present which are present in the first bound parameter set. The structure and behavior of an insert/update statement is very much determined by the first parameter set, including which defaults are going to fire off, and a minimum of guesswork is performed with all the rest so that performance is not impacted. For this reason defaults would otherwise silently “fail” for missing parameters, so this is now guarded against.

    References: #1566

  • [sql]

    returning() support is native to insert(), update(), delete(). Implementations of varying levels of functionality exist for PostgreSQL, Firebird, MSSQL and Oracle. returning() can be called explicitly with column expressions which are then returned in the resultset, usually via fetchone() or first().

    insert() constructs will also use RETURNING implicitly to get newly generated primary key values, if the database version in use supports it (a version number check is performed). This occurs if no end-user returning() was specified.

  • [sql]

    union(), intersect(), except() and other “compound” types of statements have more consistent behavior w.r.t. parenthesizing. Each compound element embedded within another will now be grouped with parenthesis - previously, the first compound element in the list would not be grouped, as SQLite doesn’t like a statement to start with parenthesis. However, PostgreSQL in particular has precedence rules regarding INTERSECT, and it is more consistent for parenthesis to be applied equally to all sub-elements. So now, the workaround for SQLite is also what the workaround for PG was previously - when nesting compound elements, the first one usually needs “.alias().select()” called on it to wrap it inside of a subquery.

    References: #1665

  • [sql]

    insert() and update() constructs can now embed bindparam() objects using names that match the keys of columns. These bind parameters will circumvent the usual route to those keys showing up in the VALUES or SET clause of the generated SQL.

    References: #1579

  • [sql]

    the Binary type now returns data as a Python string (or a “bytes” type in Python 3), instead of the built- in “buffer” type. This allows symmetric round trips of binary data.

    References: #1524

  • [sql]

    Added a tuple_() construct, allows sets of expressions to be compared to another set, typically with IN against composite primary keys or similar. Also accepts an IN with multiple columns. The “scalar select can have only one column” error message is removed - will rely upon the database to report problems with col mismatch.

  • [sql]

    User-defined “default” and “onupdate” callables which accept a context should now call upon “context.current_parameters” to get at the dictionary of bind parameters currently being processed. This dict is available in the same way regardless of single-execute or executemany-style statement execution.

  • [sql]

    multi-part schema names, i.e. with dots such as “dbo.master”, are now rendered in select() labels with underscores for dots, i.e. “dbo_master_table_column”. This is a “friendly” label that behaves better in result sets.

    References: #1428

  • [sql]

    removed needless “counter” behavior with select() labelnames that match a column name in the table, i.e. generates “tablename_id” for “id”, instead of “tablename_id_1” in an attempt to avoid naming conflicts, when the table has a column actually named “tablename_id” - this is because the labeling logic is always applied to all columns so a naming conflict will never occur.

  • [sql]

    calling expr.in_([]), i.e. with an empty list, emits a warning before issuing the usual “expr != expr” clause. The “expr != expr” can be very expensive, and it’s preferred that the user not issue in_() if the list is empty, instead simply not querying, or modifying the criterion as appropriate for more complex situations.

    References: #1628

  • [sql]

    Added “execution_options()” to select()/text(), which set the default options for the Connection. See the note in “engines”.

  • [sql]

    Deprecated or removed:
    • ”scalar” flag on select() is removed, use select.as_scalar().

    • ”shortname” attribute on bindparam() is removed.

    • postgres_returning, firebird_returning flags on insert(), update(), delete() are deprecated, use the new returning() method.

    • fold_equivalents flag on join is deprecated (will remain until is implemented)

    References: #1131

schema

  • [schema]

    the __contains__() method of MetaData now accepts strings or Table objects as arguments. If given a Table, the argument is converted to table.key first, i.e. “[schemaname.]<tablename>”

    References: #1541

  • [schema]

    deprecated MetaData.connect() and ThreadLocalMetaData.connect() have been removed - send the “bind” attribute to bind a metadata.

  • [schema]

    deprecated metadata.table_iterator() method removed (use sorted_tables)

  • [schema]

    deprecated PassiveDefault - use DefaultClause.

  • [schema]

    the “metadata” argument is removed from DefaultGenerator and subclasses, but remains locally present on Sequence, which is a standalone construct in DDL.

  • [schema]

    Removed public mutability from Index and Constraint objects:

    • ForeignKeyConstraint.append_element()

    • Index.append_column()

    • UniqueConstraint.append_column()

    • PrimaryKeyConstraint.add()

    • PrimaryKeyConstraint.remove()

    These should be constructed declaratively (i.e. in one construction).

  • [schema]

    The “start” and “increment” attributes on Sequence now generate “START WITH” and “INCREMENT BY” by default, on Oracle and PostgreSQL. Firebird doesn’t support these keywords right now.

    References: #1545

  • [schema]

    UniqueConstraint, Index, PrimaryKeyConstraint all accept lists of column names or column objects as arguments.

  • [schema]

    Other removed things:
    • Table.key (no idea what this was for)

    • Table.primary_key is not assignable - use table.append_constraint(PrimaryKeyConstraint(…))

    • Column.bind (get via column.table.bind)

    • Column.metadata (get via column.table.metadata)

    • Column.sequence (use column.default)

    • ForeignKey(constraint=some_parent) (is now private _constraint)

  • [schema]

    The use_alter flag on ForeignKey is now a shortcut option for operations that can be hand-constructed using the DDL() event system. A side effect of this refactor is that ForeignKeyConstraint objects with use_alter=True will not be emitted on SQLite, which does not support ALTER for foreign keys.

  • [schema]

    ForeignKey and ForeignKeyConstraint objects now correctly copy() all their public keyword arguments.

    References: #1605

postgresql

  • [postgresql]

    New dialects: pg8000, zxjdbc, and pypostgresql on py3k.

  • [postgresql]

    The “postgres” dialect is now named “postgresql” ! Connection strings look like:

    postgresql://scott:tiger@localhost/test postgresql+pg8000://scott:tiger@localhost/test

    The “postgres” name remains for backwards compatibility in the following ways:

    • There is a “postgres.py” dummy dialect which allows old URLs to work, i.e. postgres://scott:tiger@localhost/test

    • The “postgres” name can be imported from the old “databases” module, i.e. “from sqlalchemy.databases import postgres” as well as “dialects”, “from sqlalchemy.dialects.postgres import base as pg”, will send a deprecation warning.

    • Special expression arguments are now named “postgresql_returning” and “postgresql_where”, but the older “postgres_returning” and “postgres_where” names still work with a deprecation warning.

  • [postgresql]

    ”postgresql_where” now accepts SQL expressions which can also include literals, which will be quoted as needed.

  • [postgresql]

    The psycopg2 dialect now uses psycopg2’s “unicode extension” on all new connections, which allows all String/Text/etc. types to skip the need to post-process bytestrings into unicode (an expensive step due to its volume). Other dialects which return unicode natively (pg8000, zxjdbc) also skip unicode post-processing.

  • [postgresql]

    Added new ENUM type, which exists as a schema-level construct and extends the generic Enum type. Automatically associates itself with tables and their parent metadata to issue the appropriate CREATE TYPE/DROP TYPE commands as needed, supports unicode labels, supports reflection.

    References: #1511

  • [postgresql]

    INTERVAL supports an optional “precision” argument corresponding to the argument that PG accepts.

  • [postgresql]

    using new dialect.initialize() feature to set up version-dependent behavior.

  • [postgresql]

    somewhat better support for % signs in table/column names; psycopg2 can’t handle a bind parameter name of %(foobar)s however and SQLA doesn’t want to add overhead just to treat that one non-existent use case.

    References: #1279

  • [postgresql]

    Inserting NULL into a primary key + foreign key column will allow the “not null constraint” error to raise, not an attempt to execute a nonexistent “col_id_seq” sequence.

    References: #1516

  • [postgresql]

    autoincrement SELECT statements, i.e. those which select from a procedure that modifies rows, now work with server-side cursor mode (the named cursor isn’t used for such statements.)

  • [postgresql]

    postgresql dialect can properly detect pg “devel” version strings, i.e. “8.5devel”

    References: #1636

  • [postgresql]

    The psycopg2 now respects the statement option “stream_results”. This option overrides the connection setting “server_side_cursors”. If true, server side cursors will be used for the statement. If false, they will not be used, even if “server_side_cursors” is true on the connection.

    References: #1619

mysql

  • [mysql]

    New dialects: oursql, a new native dialect, MySQL Connector/Python, a native Python port of MySQLdb, and of course zxjdbc on Jython.

  • [mysql]

    VARCHAR/NVARCHAR will not render without a length, raises an error before passing to MySQL. Doesn’t impact CAST since VARCHAR is not allowed in MySQL CAST anyway, the dialect renders CHAR/NCHAR in those cases.

  • [mysql]

    all the _detect_XXX() functions now run once underneath dialect.initialize()

  • [mysql]

    somewhat better support for % signs in table/column names; MySQLdb can’t handle % signs in SQL when executemany() is used, and SQLA doesn’t want to add overhead just to treat that one non-existent use case.

    References: #1279

  • [mysql]

    the BINARY and MSBinary types now generate “BINARY” in all cases. Omitting the “length” parameter will generate “BINARY” with no length. Use BLOB to generate an unlengthed binary column.

  • [mysql]

    the “quoting=’quoted’” argument to MSEnum/ENUM is deprecated. It’s best to rely upon the automatic quoting.

  • [mysql]

    ENUM now subclasses the new generic Enum type, and also handles unicode values implicitly, if the given labelnames are unicode objects.

  • [mysql]

    a column of type TIMESTAMP now defaults to NULL if “nullable=False” is not passed to Column(), and no default is present. This is now consistent with all other types, and in the case of TIMESTAMP explicitly renders “NULL” due to MySQL’s “switching” of default nullability for TIMESTAMP columns.

    References: #1539

sqlite

  • [sqlite]

    DATE, TIME and DATETIME types can now take optional storage_format and regexp argument. storage_format can be used to store those types using a custom string format. regexp allows to use a custom regular expression to match string values from the database.

  • [sqlite]

    Time and DateTime types now use by a default a stricter regular expression to match strings from the database. Use the regexp argument if you are using data stored in a legacy format.

  • [sqlite]

    __legacy_microseconds__ on SQLite Time and DateTime types is not supported anymore. You should use the storage_format argument instead.

  • [sqlite]

    Date, Time and DateTime types are now stricter in what they accept as bind parameters: Date type only accepts date objects (and datetime ones, because they inherit from date), Time only accepts time objects, and DateTime only accepts date and datetime objects.

  • [sqlite]

    Table() supports a keyword argument “sqlite_autoincrement”, which applies the SQLite keyword “AUTOINCREMENT” to the single integer primary key column when generating DDL. Will prevent generation of a separate PRIMARY KEY constraint.

    References: #1016

mssql

  • [mssql]

    MSSQL + Pyodbc + FreeTDS now works for the most part, with possible exceptions regarding binary data as well as unicode schema identifiers.

  • [mssql]

    the “has_window_funcs” flag is removed. LIMIT/OFFSET usage will use ROW NUMBER as always, and if on an older version of SQL Server, the operation fails. The behavior is exactly the same except the error is raised by SQL server instead of the dialect, and no flag setting is required to enable it.

  • [mssql]

    the “auto_identity_insert” flag is removed. This feature always takes effect when an INSERT statement overrides a column that is known to have a sequence on it. As with “has_window_funcs”, if the underlying driver doesn’t support this, then you can’t do this operation in any case, so there’s no point in having a flag.

  • [mssql]

    using new dialect.initialize() feature to set up version-dependent behavior.

  • [mssql]

    removed references to sequence which is no longer used. implicit identities in mssql work the same as implicit sequences on any other dialects. Explicit sequences are enabled through the use of “default=Sequence()”. See the MSSQL dialect documentation for more information.

oracle

  • [oracle]

    unit tests pass 100% with cx_oracle !

  • [oracle]

    support for cx_Oracle’s “native unicode” mode which does not require NLS_LANG to be set. Use the latest 5.0.2 or later of cx_oracle.

  • [oracle]

    an NCLOB type is added to the base types.

  • [oracle]

    use_ansi=False won’t leak into the FROM/WHERE clause of a statement that’s selecting from a subquery that also uses JOIN/OUTERJOIN.

  • [oracle]

    added native INTERVAL type to the dialect. This supports only the DAY TO SECOND interval type so far due to lack of support in cx_oracle for YEAR TO MONTH.

    References: #1467

  • [oracle]

    usage of the CHAR type results in cx_oracle’s FIXED_CHAR dbapi type being bound to statements.

  • [oracle]

    the Oracle dialect now features NUMBER which intends to act justlike Oracle’s NUMBER type. It is the primary numeric type returned by table reflection and attempts to return Decimal()/float/int based on the precision/scale parameters.

    References: #885

  • [oracle]

    func.char_length is a generic function for LENGTH

  • [oracle]

    ForeignKey() which includes onupdate=<value> will emit a warning, not emit ON UPDATE CASCADE which is unsupported by oracle

  • [oracle]

    the keys() method of RowProxy() now returns the result column names normalized to be SQLAlchemy case insensitive names. This means they will be lower case for case insensitive names, whereas the DBAPI would normally return them as UPPERCASE names. This allows row keys() to be compatible with further SQLAlchemy operations.

  • [oracle]

    using new dialect.initialize() feature to set up version-dependent behavior.

  • [oracle]

    using types.BigInteger with Oracle will generate NUMBER(19)

    References: #1125

  • [oracle]

    ”case sensitivity” feature will detect an all-lowercase case-sensitive column name during reflect and add “quote=True” to the generated Column, so that proper quoting is maintained.

firebird

  • [firebird]

    the keys() method of RowProxy() now returns the result column names normalized to be SQLAlchemy case insensitive names. This means they will be lower case for case insensitive names, whereas the DBAPI would normally return them as UPPERCASE names. This allows row keys() to be compatible with further SQLAlchemy operations.

  • [firebird]

    using new dialect.initialize() feature to set up version-dependent behavior.

  • [firebird]

    ”case sensitivity” feature will detect an all-lowercase case-sensitive column name during reflect and add “quote=True” to the generated Column, so that proper quoting is maintained.

misc

  • [major] [release]

    For the full set of feature descriptions, see https://docs.sqlalchemy.org/en/latest/changelog/migration_06.html . This document is a work in progress.

  • [major] [release]

    All bug fixes and feature enhancements from the most recent 0.5 version and below are also included within 0.6.

  • [major] [release]

    Platforms targeted now include Python 2.4/2.5/2.6, Python 3.1, Jython2.5.

  • [engines]

    transaction isolation level may be specified with create_engine(… isolation_level=”…”); available on postgresql and sqlite.

    References: #443

  • [engines]

    Connection has execution_options(), generative method which accepts keywords that affect how the statement is executed w.r.t. the DBAPI. Currently supports “stream_results”, causes psycopg2 to use a server side cursor for that statement, as well as “autocommit”, which is the new location for the “autocommit” option from select() and text(). select() and text() also have .execution_options() as well as ORM Query().

  • [engines]

    fixed the import for entrypoint-driven dialects to not rely upon silly tb_info trick to determine import error status.

    References: #1630

  • [engines]

    added first() method to ResultProxy, returns first row and closes result set immediately.

  • [engines]

    RowProxy objects are now pickleable, i.e. the object returned by result.fetchone(), result.fetchall() etc.

  • [engines]

    RowProxy no longer has a close() method, as the row no longer maintains a reference to the parent. Call close() on the parent ResultProxy instead, or use autoclose.

  • [engines]

    ResultProxy internals have been overhauled to greatly reduce method call counts when fetching columns. Can provide a large speed improvement (up to more than 100%) when fetching large result sets. The improvement is larger when fetching columns that have no type-level processing applied and when using results as tuples (instead of as dictionaries). Many thanks to Elixir’s Gaëtan de Menten for this dramatic improvement !

    References: #1586

  • [engines]

    Databases which rely upon postfetch of “last inserted id” to get at a generated sequence value (i.e. MySQL, MS-SQL) now work correctly when there is a composite primary key where the “autoincrement” column is not the first primary key column in the table.

  • [engines]

    the last_inserted_ids() method has been renamed to the descriptor “inserted_primary_key”.

  • [engines]

    setting echo=False on create_engine() now sets the loglevel to WARN instead of NOTSET. This so that logging can be disabled for a particular engine even if logging for “sqlalchemy.engine” is enabled overall. Note that the default setting of “echo” is None.

    References: #1554

  • [engines]

    ConnectionProxy now has wrapper methods for all transaction lifecycle events, including begin(), rollback(), commit() begin_nested(), begin_prepared(), prepare(), release_savepoint(), etc.

  • [engines]

    Connection pool logging now uses both INFO and DEBUG log levels for logging. INFO is for major events such as invalidated connections, DEBUG for all the acquire/return logging. echo_pool can be False, None, True or “debug” the same way as echo works.

  • [engines]

    All pyodbc-dialects now support extra pyodbc-specific kw arguments ‘ansi’, ‘unicode_results’, ‘autocommit’.

    References: #1621

  • [engines]

    the “threadlocal” engine has been rewritten and simplified and now supports SAVEPOINT operations.

  • [engines]

    deprecated or removed
    • result.last_inserted_ids() is deprecated. Use result.inserted_primary_key

    • dialect.get_default_schema_name(connection) is now public via dialect.default_schema_name.

    • the “connection” argument from engine.transaction() and engine.run_callable() is removed - Connection itself now has those methods. All four methods accept *args and **kwargs which are passed to the given callable, as well as the operating connection.

  • [reflection/inspection]

    Table reflection has been expanded and generalized into a new API called “sqlalchemy.engine.reflection.Inspector”. The Inspector object provides fine-grained information about a wide variety of schema information, with room for expansion, including table names, column names, view definitions, sequences, indexes, etc.

  • [reflection/inspection]

    Views are now reflectable as ordinary Table objects. The same Table constructor is used, with the caveat that “effective” primary and foreign key constraints aren’t part of the reflection results; these have to be specified explicitly if desired.

  • [reflection/inspection]

    The existing autoload=True system now uses Inspector underneath so that each dialect need only return “raw” data about tables and other objects - Inspector is the single place that information is compiled into Table objects so that consistency is at a maximum.

  • [ddl]

    the DDL system has been greatly expanded. the DDL() class now extends the more generic DDLElement(), which forms the basis of many new constructs:

    • CreateTable()

    • DropTable()

    • AddConstraint()

    • DropConstraint()

    • CreateIndex()

    • DropIndex()

    • CreateSequence()

    • DropSequence()

    These support “on” and “execute-at()” just like plain DDL() does. User-defined DDLElement subclasses can be created and linked to a compiler using the sqlalchemy.ext.compiler extension.

  • [ddl]

    The signature of the “on” callable passed to DDL() and DDLElement() is revised as follows:

    ddl

    the DDLElement object itself

    event

    the string event name.

    target

    previously “schema_item”, the Table or MetaData object triggering the event.

    connection

    the Connection object in use for the operation.

    **kw

    keyword arguments. In the case of MetaData before/after create/drop, the list of Table objects for which CREATE/DROP DDL is to be issued is passed as the kw argument “tables”. This is necessary for metadata-level DDL that is dependent on the presence of specific tables.

    The “schema_item” attribute of DDL has been renamed to

    ”target”.

  • [dialect] [refactor]

    Dialect modules are now broken into database dialects plus DBAPI implementations. Connect URLs are now preferred to be specified using dialect+driver://…, i.e. “mysql+mysqldb://scott:tiger@localhost/test”. See the 0.6 documentation for examples.

  • [dialect] [refactor]

    the setuptools entrypoint for external dialects is now called “sqlalchemy.dialects”.

  • [dialect] [refactor]

    the “owner” keyword argument is removed from Table. Use “schema” to represent any namespaces to be prepended to the table name.

  • [dialect] [refactor]

    server_version_info becomes a static attribute.

  • [dialect] [refactor]

    dialects receive an initialize() event on initial connection to determine connection properties.

  • [dialect] [refactor]

    dialects receive a visit_pool event have an opportunity to establish pool listeners.

  • [dialect] [refactor]

    cached TypeEngine classes are cached per-dialect class instead of per-dialect.

  • [dialect] [refactor]

    new UserDefinedType should be used as a base class for new types, which preserves the 0.5 behavior of get_col_spec().

  • [dialect] [refactor]

    The result_processor() method of all type classes now accepts a second argument “coltype”, which is the DBAPI type argument from cursor.description. This argument can help some types decide on the most efficient processing of result values.

  • [dialect] [refactor]

    Deprecated Dialect.get_params() removed.

  • [dialect] [refactor]

    Dialect.get_rowcount() has been renamed to a descriptor “rowcount”, and calls cursor.rowcount directly. Dialects which need to hardwire a rowcount in for certain calls should override the method to provide different behavior.

  • [dialect] [refactor]

    DefaultRunner and subclasses have been removed. The job of this object has been simplified and moved into ExecutionContext. Dialects which support sequences should add a fire_sequence() method to their execution context implementation.

    References: #1566

  • [dialect] [refactor]

    Functions and operators generated by the compiler now use (almost) regular dispatch functions of the form “visit_<opname>” and “visit_<funcname>_fn” to provide customed processing. This replaces the need to copy the “functions” and “operators” dictionaries in compiler subclasses with straightforward visitor methods, and also allows compiler subclasses complete control over rendering, as the full _Function or _BinaryExpression object is passed in.

  • [types]

    The construction of types within dialects has been totally overhauled. Dialects now define publicly available types as UPPERCASE names exclusively, and internal implementation types using underscore identifiers (i.e. are private). The system by which types are expressed in SQL and DDL has been moved to the compiler system. This has the effect that there are much fewer type objects within most dialects. A detailed document on this architecture for dialect authors is in lib/sqlalchemy/dialects/type_migration_guidelines.txt .

  • [types]

    Types no longer make any guesses as to default parameters. In particular, Numeric, Float, NUMERIC, FLOAT, DECIMAL don’t generate any length or scale unless specified.

  • [types]

    types.Binary is renamed to types.LargeBinary, it only produces BLOB, BYTEA, or a similar “long binary” type. New base BINARY and VARBINARY types have been added to access these MySQL/MS-SQL specific types in an agnostic way.

    References: #1664

  • [types]

    String/Text/Unicode types now skip the unicode() check on each result column value if the dialect has detected the DBAPI as returning Python unicode objects natively. This check is issued on first connect using “SELECT CAST ‘some text’ AS VARCHAR(10)” or equivalent, then checking if the returned object is a Python unicode. This allows vast performance increases for native-unicode DBAPIs, including pysqlite/sqlite3, psycopg2, and pg8000.

  • [types]

    Most types result processors have been checked for possible speed improvements. Specifically, the following generic types have been optimized, resulting in varying speed improvements: Unicode, PickleType, Interval, TypeDecorator, Binary. Also the following dbapi-specific implementations have been improved: Time, Date and DateTime on Sqlite, ARRAY on PostgreSQL, Time on MySQL, Numeric(as_decimal=False) on MySQL, oursql and pypostgresql, DateTime on cx_oracle and LOB-based types on cx_oracle.

  • [types]

    Reflection of types now returns the exact UPPERCASE type within types.py, or the UPPERCASE type within the dialect itself if the type is not a standard SQL type. This means reflection now returns more accurate information about reflected types.

  • [types]

    Added a new Enum generic type. Enum is a schema-aware object to support databases which require specific DDL in order to use enum or equivalent; in the case of PG it handles the details of CREATE TYPE, and on other databases without native enum support will by generate VARCHAR + an inline CHECK constraint to enforce the enum.

    References: #1109, #1511

  • [types]

    The Interval type includes a “native” flag which controls if native INTERVAL types (postgresql + oracle) are selected if available, or not. “day_precision” and “second_precision” arguments are also added which propagate as appropriately to these native types. Related to.

    References: #1467

  • [types]

    The Boolean type, when used on a backend that doesn’t have native boolean support, will generate a CHECK constraint “col IN (0, 1)” along with the int/smallint- based column type. This can be switched off if desired with create_constraint=False. Note that MySQL has no native boolean or CHECK constraint support so this feature isn’t available on that platform.

    References: #1589

  • [types]

    PickleType now uses == for comparison of values when mutable=True, unless the “comparator” argument with a comparison function is specified to the type. Objects being pickled will be compared based on identity (which defeats the purpose of mutable=True) if __eq__() is not overridden or a comparison function is not provided.

  • [types]

    The default “precision” and “scale” arguments of Numeric and Float have been removed and now default to None. NUMERIC and FLOAT will be rendered with no numeric arguments by default unless these values are provided.

  • [types]

    AbstractType.get_search_list() is removed - the games that was used for are no longer necessary.

  • [types]

    Added a generic BigInteger type, compiles to BIGINT or NUMBER(19).

    References: #1125

  • [types]

    sqlsoup has been overhauled to explicitly support an 0.5 style session, using autocommit=False, autoflush=True. Default behavior of SQLSoup now requires the usual usage of commit() and rollback(), which have been added to its interface. An explicit Session or scoped_session can be passed to the constructor, allowing these arguments to be overridden.

  • [types]

    sqlsoup db.<sometable>.update() and delete() now call query(cls).update() and delete(), respectively.

  • [types]

    sqlsoup now has execute() and connection(), which call upon the Session methods of those names, ensuring that the bind is in terms of the SqlSoup object’s bind.

  • [types]

    sqlsoup objects no longer have the ‘query’ attribute - it’s not needed for sqlsoup’s usage paradigm and it gets in the way of a column that is actually named ‘query’.

  • [types]

    The signature of the proxy_factory callable passed to association_proxy is now (lazy_collection, creator, value_attr, association_proxy), adding a fourth argument that is the parent AssociationProxy argument. Allows serializability and subclassing of the built in collections.

    References: #1259

  • [types]

    association_proxy now has basic comparator methods .any(), .has(), .contains(), ==, !=, thanks to Scott Torborg.

    References: #1372