[orm] [usecase]

Improve error message when using Query.filter_by() in a query where the first entity is not a mapped class.

References: #5326

[orm] [usecase]

Added a new parameter query_expression.default_expr to the query_expression() construct, which will be appled to queries automatically if the with_expression() option is not used. Pull request courtesy Haoyu Sun.

References: #5198

[engine] [bug]

Further refinements to the fixes to the “reset” agent fixed in #5326, which now emits a warning when it is not being correctly invoked and corrects for the behavior. Additional scenarios have been identified and fixed where this warning was being emitted.

References: #5326

[engine] [bug]

Fixed issue in URL object where stringifying the object would not URL encode special characters, preventing the URL from being re-consumable as a real URL. Pull request courtesy Miguel Grinberg.

References: #5341

[sql] [usecase]

Added a “.schema” parameter to the table() construct, allowing ad-hoc table expressions to also include a schema name. Pull request courtesy Dylan Modesitt.

References: #5309

[sql] [bug]

Correctly apply self_group in type_coerce element.

The type coerce element did not correctly apply grouping rules when using in an expression

References: #5344

[sql] [bug]

Added Select.with_hint() output to the generic SQL string that is produced when calling str() on a statement. Previously, this clause would be omitted under the assumption that it was dialect specific. The hint text is presented within brackets to indicate the rendering of such hints varies among backends.

References: #5353

[sql] [schema]

Introduce IdentityOptions to store common parameters for sequences and identity columns.

References: #5324

[sql] [change] [sybase]

Added .offset support to sybase dialect. Pull request courtesy Alan D. Snow.

References: #5294

[schema] [bug]

Fixed issue where dialect_options were omitted when a database object (e.g., Table) was copied using tometadata().

References: #5276

[mysql] [usecase]

Implemented row-level locking support for mysql. Pull request courtesy Quentin Somerville.

References: #4860

[sqlite] [usecase]

SQLite 3.31 added support for computed column. This change enables their support in SQLAlchemy when targeting SQLite.

References: #5297

[sqlite] [bug]

Added “exists” to the list of reserved words for SQLite so that this word will be quoted when used as a label or column name. Pull request courtesy Thodoris Sotiropoulos.

References: #5395

[mssql] [bug]

Refined the logic used by the SQL Server dialect to interpret multi-part schema names that contain many dots, to not actually lose any dots if the name does not have bracking or quoting used, and additionally to support a “dbname” token that has many parts including that it may have multiple, independently-bracketed sections.

References: #5364, #5366

[mssql] [bug] [pyodbc]

Fixed an issue in the pyodbc connector such that a warning about pyodbc “drivername” would be emitted when using a totally empty URL. Empty URLs are normal when producing a non-connected dialect object or when using the “creator” argument to create_engine(). The warning now only emits if the driver name is missing but other parameters are still present.

References: #5346

[mssql] [bug]

Fixed issue with assembling the ODBC connection string for the pyodbc DBAPI. Tokens containing semicolons and/or braces “{}” were not being correctly escaped, causing the ODBC driver to misinterpret the connection string attributes.

References: #5373

[mssql] [bug]

Fixed issue where datetime.time parameters were being converted to datetime.datetime, making them incompatible with comparisons like >= against an actual TIME column.

References: #5339

[mssql] [bug]

Fixed an issue where the is_disconnect function in the SQL Server pyodbc dialect was incorrectly reporting the disconnect state when the exception messsage had a substring that matched a SQL Server ODBC error code.

References: #5359

[mssql] [change]

Moved the supports_sane_rowcount_returning = False requirement from the PyODBCConnector level to the MSDialect_pyodbc since pyodbc does work properly in some circumstances.

References: #5321

[oracle] [bug] [reflection]

Fixed bug in Oracle dialect where indexes that contain the full set of primary key columns would be mistaken as the primary key index itself, which is omitted, even if there were multiples. The check has been refined to compare the name of the primary key constraint against the index name itself, rather than trying to guess based on the columns present in the index.

References: #5421

[change] [examples]

Added new option --raw to the examples.performance suite which will dump the raw profile test for consumption by any number of profiling visualizer tools. Removed the “runsnake” option as runsnake is very hard to build at this point;

[orm] [usecase]

Added an accessor Comparator.expressions which provides access to the group of columns mapped under a multi-column ColumnProperty attribute.

References: #5262

[orm] [usecase]

Introduce relationship.sync_backref flag in a relationship to control if the synchronization events that mutate the in-Python attributes are added. This supersedes the previous change #5149, which warned that viewonly=True relationship target of a back_populates or backref configuration would be disallowed.

References: #5237

[orm] [bug]

Fixed bug where using with_polymorphic() as the target of a join via RelationshipComparator.of_type() on a mapper that already has a subquery-based with_polymorphic setting that’s equivalent to the one requested would not correctly alias the ON clause in the join.

References: #5288

[orm] [bug]

Fixed issue in the area of where loader options such as selectinload() interact with the baked query system, such that the caching of a query is not supposed to occur if the loader options themselves have elements such as with_polymorphic() objects in them that currently are not cache-compatible. The baked loader could sometimes not fully invalidate itself in these some of these scenarios leading to missed eager loads.

References: #5303

[orm] [bug]

Modified the internal “identity set” implementation, which is a set that hashes objects on their id() rather than their hash values, to not actually call the __hash__() method of the objects, which are typically user-mapped objects. Some methods were calling this method as a side effect of the implementation.

References: #5304

[orm] [bug]

An informative error message is raised when an ORM many-to-one comparison is attempted against an object that is not an actual mapped instance. Comparisons such as those to scalar subqueries aren’t supported; generalized comparison with subqueries is better achieved using Comparator.has().

References: #5269

[engine] [bug]

Fixed fairly critical issue where the DBAPI connection could be returned to the connection pool while still in an un-rolled-back state. The reset agent responsible for rolling back the connection could be corrupted in the case that the transaction was “closed” without being rolled back or committed, which can occur in some scenarios when using ORM sessions and emitting .close() in a certain pattern involving savepoints. The fix ensures that the reset agent is always active.

References: #5326

[schema] [bug]

Fixed issue where an Index that is deferred in being associated with a table, such as as when it contains a Column that is not associated with any Table yet, would fail to attach correctly if it also contained a non table-oriented expression.

References: #5298

[schema] [bug]

A warning is emitted when making use of the MetaData.sorted_tables attribute as well as the sort_tables() function, and the given tables cannot be correctly sorted due to a cyclic dependency between foreign key constraints. In this case, the functions will no longer sort the involved tables by foreign key, and a warning will be emitted. Other tables that are not part of the cycle will still be returned in dependency order. Previously, the sorted_table routines would return a collection that would unconditionally omit all foreign keys when a cycle was detected, and no warning was emitted.

References: #5316

[schema]

Add comment attribute to Column __repr__ method.

References: #4138

[postgresql] [usecase]

Added support for columns or type ARRAY of Enum, JSON or JSONB in PostgreSQL. Previously a workaround was required in these use cases.

References: #5265

[postgresql] [usecase]

Raise an explicit CompileError when adding a table with a column of type ARRAY of Enum configured with Enum.native_enum set to False when Enum.create_constraint is not set to False

References: #5266

[mssql] [bug] [reflection]

Fix a regression introduced by the reflection of computed column in MSSQL when using the legacy TDS version 4.2. The dialect will try to detect the protocol version of first connect and run in compatibility mode if it cannot detect it.

References: #5255

[mssql] [bug] [reflection]

Fix a regression introduced by the reflection of computed column in MSSQL when using SQL server versions before 2012, which does not support the concat function.

References: #5271

[oracle] [bug]

Some modifications to how the cx_oracle dialect sets up per-column outputtype handlers for LOB and numeric datatypes to adjust for potential changes coming in cx_Oracle 8.

References: #5246

[oracle] [bug] [performance]

Changed the implementation of fetching CLOB and BLOB objects to use cx_Oracle’s native implementation which fetches CLOB/BLOB objects inline with other result columns, rather than performing a separate fetch. As always, this can be disabled by setting auto_convert_lobs to False.

As part of this change, the behavior of a CLOB that was given a blank string on INSERT now returns None on SELECT, which is now consistent with that of VARCHAR on Oracle.

References: #5314

[firebird] [change]

Adjusted dialect loading for firebird:// URIs so the external sqlalchemy-firebird dialect will be used if it has been installed, otherwise fall back to the (now deprecated) internal Firebird dialect.

References: #5278

[orm] [bug]

Fixed bug in selectinload() loading option where two or more loaders that represent different relationships with the same string key name as referenced from a single with_polymorphic() construct with multiple subclass mappers would fail to invoke each subqueryload separately, instead making use of a single string-based slot that would prevent the other loaders from being invoked.

References: #5228

[orm] [bug]

Fixed issue where a lazyload that uses session-local “get” against a target many-to-one relationship where an object with the correct primary key is present, however it’s an instance of a sibling class, does not correctly return None as is the case when the lazy loader actually emits a load for that row.

References: #5210

[orm] [performance]

Modified the queries used by subqueryload and selectinload to no longer ORDER BY the primary key of the parent entity; this ordering was there to allow the rows as they come in to be copied into lists directly with a minimal level of Python-side collation. However, these ORDER BY clauses can negatively impact the performance of the query as in many scenarios these columns are derived from a subquery or are otherwise not actual primary key columns such that SQL planners cannot make use of indexes. The Python-side collation uses the native itertools.group_by() to collate the incoming rows, and has been modified to allow multiple row-groups-per-parent to be assembled together using list.extend(), which should still allow for relatively fast Python-side performance. There will still be an ORDER BY present for a relationship that includes an explicit order_by parameter, however this is the only ORDER BY that will be added to the query for both kinds of loading.

References: #5162

[bug] [declarative] [orm]

The string argument accepted as the first positional argument by the relationship() function when using the Declarative API is no longer interpreted using the Python eval() function; instead, the name is dot separated and the names are looked up directly in the name resolution dictionary without treating the value as a Python expression. However, passing a string argument to the other relationship() parameters that necessarily must accept Python expressions will still use eval(); the documentation has been clarified to ensure that there is no ambiguity that this is in use.

References: #5238

[sql] [types]

Add ability to literal compile a DateTime, Date or Time when using the string dialect for debugging purposes. This change does not impact real dialect implementation that retain their current behavior.

References: #5052

[schema] [reflection]

Added support for reflection of “computed” columns, which are now returned as part of the structure returned by Inspector.get_columns(). When reflecting full Table objects, computed columns will be represented using the Computed construct.

References: #5063

[postgresql] [bug]

Fixed issue where a “covering” index, e.g. those which have an INCLUDE clause, would be reflected including all the columns in INCLUDE clause as regular columns. A warning is now emitted if these additional columns are detected indicating that they are currently ignored. Note that full support for “covering” indexes is part of #4458. Pull request courtesy Marat Sharafutdinov.

References: #5205

[mysql] [bug]

Fixed issue in MySQL dialect when connecting to a pseudo-MySQL database such as that provided by ProxySQL, the up front check for isolation level when it returns no row will not prevent the dialect from continuing to connect. A warning is emitted that the isolation level could not be detected.

References: #5239

[sqlite] [usecase]

Implemented AUTOCOMMIT isolation level for SQLite when using pysqlite.

References: #5164

[mssql] [usecase] [mysql] [oracle]

Added support for ColumnOperators.is_distinct_from() and ColumnOperators.isnot_distinct_from() to SQL Server, MySQL, and Oracle.

References: #5137

[oracle] [usecase]

Implemented AUTOCOMMIT isolation level for Oracle when using cx_Oracle. Also added a fixed default isolation level of READ COMMITTED for Oracle.

References: #5200

[oracle] [bug] [reflection]

Fixed regression / incorrect fix caused by fix for #5146 where the Oracle dialect reads from the “all_tab_comments” view to get table comments but fails to accommodate for the current owner of the table being requested, causing it to read the wrong comment if multiple tables of the same name exist in multiple schemas.

References: #5146

[bug] [tests]

Fixed an issue that prevented the test suite from running with the recently released py.test 5.4.0.

References: #5201

[enum] [types]

The Enum type now supports the parameter Enum.length to specify the length of the VARCHAR column to create when using non native enums by setting Enum.native_enum to False

References: #5183

[installer]

Ensured that the “pyproject.toml” file is not included in builds, as the presence of this file indicates to pip that a pep-517 installation process should be used. As this mode of operation appears to be not well supported by current tools / distros, these problems are avoided within the scope of SQLAlchemy installation by omitting the file.

References: #5207

[orm] [bug]

Adjusted the error message emitted by Query.join() when a left hand side can’t be located that the Query.select_from() method is the best way to resolve the issue. Also, within the 1.3 series, used a deterministic ordering when determining the FROM clause from a given column entity passed to Query so that the same expression is determined each time.

References: #5194

[orm] [bug]

Fixed regression in 1.3.14 due to #4849 where a sys.exc_info() call failed to be invoked correctly when a flush error would occur. Test coverage has been added for this exception case.

References: #5196

[general] [bug] [py3k]

Applied an explicit “cause” to most if not all internally raised exceptions that are raised from within an internal exception catch, to avoid misleading stacktraces that suggest an error within the handling of an exception. While it would be preferable to suppress the internally caught exception in the way that the __suppress_context__ attribute would, there does not as yet seem to be a way to do this without suppressing an enclosing user constructed context, so for now it exposes the internally caught exception as the cause so that full information about the context of the error is maintained.

References: #4849

[orm] [usecase]

Added a new flag InstanceEvents.restore_load_context and SessionEvents.restore_load_context which apply to the InstanceEvents.load(), InstanceEvents.refresh(), and SessionEvents.loaded_as_persistent() events, which when set will restore the “load context” of the object after the event hook has been called. This ensures that the object remains within the “loader context” of the load operation that is already ongoing, rather than the object being transferred to a new load context due to refresh operations which may have occurred in the event. A warning is now emitted when this condition occurs, which recommends use of the flag to resolve this case. The flag is “opt-in” so that there is no risk introduced to existing applications.

The change additionally adds support for the raw=True flag to session lifecycle events.

References: #5129

[orm] [bug]

Fixed regression caused in 1.3.13 by #5056 where a refactor of the ORM path registry system made it such that a path could no longer be compared to an empty tuple, which can occur in a particular kind of joined eager loading path. The “empty tuple” use case has been resolved so that the path registry is compared to a path registry in all cases; the PathRegistry object itself now implements __eq__() and __ne__() methods which will take place for all equality comparisons and continue to succeed in the not anticipated case that a non- PathRegistry object is compared, while emitting a warning that this object should not be the subject of the comparison.

References: #5110

[orm] [bug]

Setting a relationship to viewonly=True which is also the target of a back_populates or backref configuration will now emit a warning and eventually be disallowed. back_populates refers specifically to mutation of an attribute or collection, which is disallowed when the attribute is subject to viewonly=True. The viewonly attribute is not subject to persistence behaviors which means it will not reflect correct results when it is locally mutated.

References: #5149

[orm] [bug]

Fixed an additional regression in the same area as that of #5080 introduced in 1.3.0b3 via #4468 where the ability to create a joined option across a with_polymorphic() into a relationship against the base class of that with_polymorphic, and then further into regular mapped relationships would fail as the base class component would not add itself to the load path in a way that could be located by the loader strategy. The changes applied in #5080 have been further refined to also accommodate this scenario.

References: #5121

[engine] [bug]

Expanded the scope of cursor/connection cleanup when a statement is executed to include when the result object fails to be constructed, or an after_cursor_execute() event raises an error, or autocommit / autoclose fails. This allows the DBAPI cursor to be cleaned up on failure and for connectionless execution allows the connection to be closed out and returned to the connection pool, where previously it waiting until garbage collection would trigger a pool return.

References: #5182

[sql] [bug] [postgresql]

Fixed bug where a CTE of an INSERT/UPDATE/DELETE that also uses RETURNING could then not be SELECTed from directly, as the internal state of the compiler would try to treat the outer SELECT as a DELETE statement itself and access nonexistent state.

References: #5181

[postgresql] [bug]

Fixed issue where the “schema_translate_map” feature would not work with a PostgreSQL native enumeration type (i.e. Enum, ENUM) in that while the “CREATE TYPE” statement would be emitted with the correct schema, the schema would not be rendered in the CREATE TABLE statement at the point at which the enumeration was referenced.

References: #5158

[postgresql] [bug] [reflection]

Fixed bug where PostgreSQL reflection of CHECK constraints would fail to parse the constraint if the SQL text contained newline characters. The regular expression has been adjusted to accommodate for this case. Pull request courtesy Eric Borczuk.

References: #5170

[mysql] [bug]

Fixed issue in MySQL Insert.on_duplicate_key_update() construct where using a SQL function or other composed expression for a column argument would not properly render the VALUES keyword surrounding the column itself.

References: #5173

[mssql] [bug]

Fixed issue where the DATETIMEOFFSET type would not accommodate for the None value, introduced as part of the series of fixes for this type first introduced in #4983, #5045. Additionally, added support for passing a backend-specific date formatted string through this type, as is typically allowed for date/time types on most other DBAPIs.

References: #5132

[oracle] [bug]

Fixed a reflection bug where table comments could only be retrieved for tables actually owned by the user but not for tables visible to the user but owned by someone else. Pull request courtesy Dave Hirschfeld.

References: #5146

[usecase] [ext]

Added keyword arguments to the MutableList.sort() function so that a key function as well as the “reverse” keyword argument can be provided.

References: #5114

[bug] [performance]

Revised an internal change to the test system added as a result of #5085 where a testing-related module per dialect would be loaded unconditionally upon making use of that dialect, pulling in SQLAlchemy’s testing framework as well as the ORM into the module import space. This would only impact initial startup time and memory to a modest extent, however it’s best that these additional modules aren’t reverse-dependent on straight Core usage.

References: #5180

[bug] [installation]

Vendored the inspect.formatannotation function inside of sqlalchemy.util.compat, which is needed for the vendored version of inspect.formatargspec. The function is not documented in cPython and is not guaranteed to be available in future Python versions.

References: #5138

[orm] [bug] [engine]

Added test support and repaired a wide variety of unnecessary reference cycles created for short-lived objects, mostly in the area of ORM queries. Thanks much to Carson Ip for the help on this.

References: #5050, #5056, #5071

[orm] [bug]

Fixed regression in loader options introduced in 1.3.0b3 via #4468 where the ability to create a loader option using PropComparator.of_type() targeting an aliased entity that is an inheriting subclass of the entity which the preceding relationship refers to would fail to produce a matching path. See also #5082 fixed in this same release which involves a similar kind of issue.

References: #5107

[orm] [bug]

Fixed regression in joined eager loading introduced in 1.3.0b3 via #4468 where the ability to create a joined option across a with_polymorphic() into a polymorphic subclass using RelationshipProperty.of_type() and then further along regular mapped relationships would fail as the polymorphic subclass would not add itself to the load path in a way that could be located by the loader strategy. A tweak has been made to resolve this scenario.

References: #5082

[orm] [bug]

Repaired a warning in the ORM flush process that was not covered by test coverage when deleting objects that use the “version_id” feature. This warning is generally unreachable unless using a dialect that sets the “supports_sane_rowcount” flag to False, which is not typically the case however is possible for some MySQL configurations as well as older Firebird drivers, and likely some third party dialects.

References: #5068

[orm] [bug]

Fixed bug where usage of joined eager loading would not properly wrap the query inside of a subquery when Query.group_by() were used against the query. When any kind of result-limiting approach is used, such as DISTINCT, LIMIT, OFFSET, joined eager loading embeds the row-limited query inside of a subquery so that the collection results are not impacted. For some reason, the presence of GROUP BY was never included in this criterion, even though it has a similar effect as using DISTINCT. Additionally, the bug would prevent using GROUP BY at all for a joined eager load query for most database platforms which forbid non-aggregated, non-grouped columns from being in the query, as the additional columns for the joined eager load would not be accepted by the database.

References: #5065

[orm] [performance]

Identified a performance issue in the system by which a join is constructed based on a mapped relationship. The clause adaption system would be used for the majority of join expressions including in the common case where no adaptation is needed. The conditions under which this adaptation occur have been refined so that average non-aliased joins along a simple relationship without a “secondary” table use about 70% less function calls.

[engine] [bug]

Fixed issue where the collection of value processors on a Compiled object would be mutated when “expanding IN” parameters were used with a datatype that has bind value processors; in particular, this would mean that when using statement caching and/or baked queries, the same compiled._bind_processors collection would be mutated concurrently. Since these processors are the same function for a given bind parameter namespace every time, there was no actual negative effect of this issue, however, the execution of a Compiled object should never be causing any changes in its state, especially given that they are intended to be thread-safe and reusable once fully constructed.

References: #5048

[sql] [usecase]

A function created using GenericFunction can now specify that the name of the function should be rendered with or without quotes by assigning the quoted_name construct to the .name element of the object. Prior to 1.3.4, quoting was never applied to function names, and some quoting was introduced in #4467 but no means to force quoting for a mixed case name was available. Additionally, the quoted_name construct when used as the name will properly register its lowercase name in the function registry so that the name continues to be available via the func. registry.

References: #5079

[postgresql] [usecase]

Added support for prefixes to the CTE construct, to allow support for Postgresql 12 “MATERIALIZED” and “NOT MATERIALIZED” phrases. Pull request courtesy Marat Sharafutdinov.

References: #5040

[postgresql] [bug]

Fixed issue where the PostgreSQL dialect would fail to parse a reflected CHECK constraint that was a boolean-valued function (as opposed to a boolean-valued expression).

References: #5039

[postgresql] [tests]

Improved detection of two phase transactions requirement for the PostgreSQL database by testing that max_prepared_transactions is set to a value greater than 0. Pull request courtesy Federico Caselli.

References: #5057

[mssql] [bug]

Fixed issue where a timezone-aware datetime value being converted to string for use as a parameter value of a DATETIMEOFFSET column was omitting the fractional seconds.

References: #5045

[bug] [ext]

Fixed bug in sqlalchemy.ext.serializer where a unique BindParameter object could conflict with itself if it were present in the mapping itself, as well as the filter condition of the query, as one side would be used against the non-deserialized version and the other side would use the deserialized version. Logic is added to BindParameter similar to its “clone” method which will uniquify the parameter name upon deserialize so that it doesn’t conflict with its original.

References: #5086

[bug] [tests]

Fixed a few test failures which would occur on Windows due to SQLite file locking issues, as well as some timing issues in connection pool related tests; pull request courtesy Federico Caselli.

References: #4946

[orm] [bug]

Fixed issue involving lazy="raise" strategy where an ORM delete of an object would raise for a simple “use-get” style many-to-one relationship that had lazy=”raise” configured. This is inconsistent vs. the change introduced in 1.3 as part of #4353, where it was established that a history operation that does not expect emit SQL should bypass the lazy="raise" check, and instead effectively treat it as lazy="raise_on_sql" for this case. The fix adjusts the lazy loader strategy to not raise for the case where the lazy load was instructed that it should not emit SQL if the object were not present.

References: #4997

[orm] [bug]

Fixed regression introduced in 1.3.0 related to the association proxy refactor in #4351 that prevented composite() attributes from working in terms of an association proxy that references them.

References: #5000

[orm] [bug]

Setting persistence-related flags on relationship() while also setting viewonly=True will now emit a regular warning, as these flags do not make sense for a viewonly=True relationship. In particular, the “cascade” settings have their own warning that is generated based on the individual values, such as “delete, delete-orphan”, that should not apply to a viewonly relationship. Note however that in the case of “cascade”, these settings are still erroneously taking effect even though the relationship is set up as “viewonly”. In 1.4, all persistence-related cascade settings will be disallowed on a viewonly=True relationship in order to resolve this issue.

References: #4993

[orm] [bug] [py3k]

Fixed issue where when assigning a collection to itself as a slice, the mutation operation would fail as it would first erase the assigned collection inadvertently. As an assignment that does not change the contents should not generate events, the operation is now a no-op. Note that the fix only applies to Python 3; in Python 2, the __setitem__ hook isn’t called in this case; __setslice__ is used instead which recreates the list item-by-item in all cases.

References: #4990

[orm] [bug]

Fixed issue where by if the “begin” of a transaction failed at the Core engine/connection level, such as due to network error or database is locked for some transactional recipes, within the context of the Session procuring that connection from the conneciton pool and then immediately returning it, the ORM Session would not close the connection despite this connection not being stored within the state of that Session. This would lead to the connection being cleaned out by the connection pool weakref handler within garbage collection which is an unpreferred codepath that in some special configurations can emit errors in standard error.

References: #5034

[sql] [bug]

Fixed bug where “distinct” keyword passed to select() would not treat a string value as a “label reference” in the same way that the select.distinct() does; it would instead raise unconditionally. This keyword argument and the others passed to select() will ultimately be deprecated for SQLAlchemy 2.0.

References: #5028

[sql] [bug]

Changed the text of the exception for “Can’t resolve label reference” to include other kinds of label coercions, namely that “DISTINCT” is also in this category under the PostgreSQL dialect.

[sqlite] [bug]

Fixed issue to workaround SQLite’s behavior of assigning “numeric” affinity to JSON datatypes, first described at Support for SQLite JSON Added, which returns scalar numeric JSON values as a number and not as a string that can be JSON deserialized. The SQLite-specific JSON deserializer now gracefully degrades for this case as an exception and bypasses deserialization for single numeric values, as from a JSON perspective they are already deserialized.

References: #5014

[mssql] [bug]

Repaired support for the DATETIMEOFFSET datatype on PyODBC, by adding PyODBC-level result handlers as it does not include native support for this datatype. This includes usage of the Python 3 “timezone” tzinfo subclass in order to set up a timezone, which on Python 2 makes use of a minimal backport of “timezone” in sqlalchemy.util.

References: #4983

[orm] [usecase]

Added accessor Query.is_single_entity() to Query, which will indicate if the results returned by this Query will be a list of ORM entities, or a tuple of entities or column expressions. SQLAlchemy hopes to improve upon the behavior of single entity / tuples in future releases such that the behavior would be explicit up front, however this attribute should be helpful with the current behavior. Pull request courtesy Patrick Hayes.

References: #4934

[orm] [bug]

The relationship.omit_join flag was not intended to be manually set to True, and will now emit a warning when this occurs. The omit_join optimization is detected automatically, and the omit_join flag was only intended to disable the optimization in the hypothetical case that the optimization may have interfered with correct results, which has not been observed with the modern version of this feature. Setting the flag to True when it is not automatically detected may cause the selectin load feature to not work correctly when a non-default primary join condition is in use.

References: #4954

[orm] [bug]

A warning is emitted if a primary key value is passed to Query.get() that consists of None for all primary key column positions. Previously, passing a single None outside of a tuple would raise a TypeError and passing a composite None (tuple of None values) would silently pass through. The fix now coerces the single None into a tuple where it is handled consistently with the other None conditions. Thanks to Lev Izraelit for the help with this.

References: #4915

[orm] [bug]

The BakedQuery will not cache a query that was modified by a QueryEvents.before_compile() event, so that compilation hooks that may be applying ad-hoc modifications to queries will take effect on each run. In particular this is helpful for events that modify queries used in lazy loading as well as eager loading such as “select in” loading. In order to re-enable caching for a query modified by this event, a new flag bake_ok is added; see Using the before_compile event for details.

A longer term plan to provide a new form of SQL caching should solve this kind of issue more comprehensively.

References: #4947

[orm] [bug]

Fixed ORM bug where a “secondary” table that referred to a selectable which in some way would refer to the local primary table would apply aliasing to both sides of the join condition when a relationship-related join, either via Query.join() or by joinedload(), were generated. The “local” side is now excluded.

References: #4974

[engine] [bug]

Fixed bug where parameter repr as used in logging and error reporting needs additional context in order to distinguish between a list of parameters for a single statement and a list of parameter lists, as the “list of lists” structure could also indicate a single parameter list where the first parameter itself is a list, such as for an array parameter. The engine/connection now passes in an additional boolean indicating how the parameters should be considered. The only SQLAlchemy backend that expects arrays as parameters is that of psycopg2 which uses pyformat parameters, so this issue has not been too apparent, however as other drivers that use positional gain more features it is important that this be supported. It also eliminates the need for the parameter repr function to guess based on the parameter structure passed.

References: #4902

[engine] [bug] [postgresql]

Fixed bug in Inspector where the cache key generation did not take into account arguments passed in the form of tuples, such as the tuple of view name styles to return for the PostgreSQL dialect. This would lead the inspector to cache too generally for a more specific set of criteria. The logic has been adjusted to include every keyword element in the cache, as every argument is expected to be appropriate for a cache else the caching decorator should be bypassed by the dialect.

References: #4955

[sql] [usecase]

Added new accessors to expressions of type JSON to allow for specific datatype access and comparison, covering strings, integers, numeric, boolean elements. This revises the documented approach of CASTing to string when comparing values, instead adding specific functionality into the PostgreSQL, SQlite, MySQL dialects to reliably deliver these basic types in all cases.

References: #4276

[sql] [usecase]

The text() construct now supports “unique” bound parameters, which will dynamically uniquify themselves on compilation thus allowing multiple text() constructs with the same bound parameter names to be combined together.

References: #4933

[sql] [bug] [py3k]

Changed the repr() of the quoted_name construct to use regular string repr() under Python 3, rather than running it through “backslashreplace” escaping, which can be misleading.

References: #4931

[schema] [usecase]

Added DDL support for “computed columns”; these are DDL column specifications for columns that have a server-computed value, either upon SELECT (known as “virtual”) or at the point of which they are INSERTed or UPDATEd (known as “stored”). Support is established for Postgresql, MySQL, Oracle SQL Server and Firebird. Thanks to Federico Caselli for lots of work on this one.

References: #4894

[schema] [bug]

Fixed bug where a table that would have a column label overlap with a plain column name, such as “foo.id AS foo_id” vs. “foo.foo_id”, would prematurely generate the ._label attribute for a column before this overlap could be detected due to the use of the index=True or unique=True flag on the column in conjunction with the default naming convention of "column_0_label". This would then lead to failures when ._label were used later to generate a bound parameter name, in particular those used by the ORM when generating the WHERE clause for an UPDATE statement. The issue has been fixed by using an alternate ._label accessor for DDL generation that does not affect the state of the Column. The accessor also bypasses the key-deduplication step as it is not necessary for DDL, the naming is now consistently "<tablename>_<columnname>" without any subsequent numeric symbols when used in DDL.

References: #4911

[mysql] [bug]

Added “Connection was killed” message interpreted from the base pymysql.Error class in order to detect closed connection, based on reports that this message is arriving via a pymysql.InternalError() object which indicates pymysql is not handling it correctly.

References: #4945

[mssql] [bug]

Fixed issue in MSSQL dialect where an expression-based OFFSET value in a SELECT would be rejected, even though the dialect can render this expression inside of a ROW NUMBER-oriented LIMIT/OFFSET construct.

References: #4973

[mssql] [bug]

Fixed an issue in the Engine.table_names() method where it would feed the dialect’s default schema name back into the dialect level table function, which in the case of SQL Server would interpret it as a dot-tokenized schema name as viewed by the mssql dialect, which would cause the method to fail in the case where the database username actually had a dot inside of it. In 1.3, this method is still used by the MetaData.reflect() function so is a prominent codepath. In 1.4, which is the current master development branch, this issue doesn’t exist, both because MetaData.reflect() isn’t using this method nor does the method pass the default schema name explicitly. The fix nonetheless guards against the default server name value returned by the dialect from being interpreted as dot-tokenized name under any circumstances by wrapping it in quoted_name().

References: #4923

[oracle] [usecase]

Added dialect-level flag encoding_errors to the cx_Oracle dialect, which can be specified as part of create_engine(). This is passed to SQLAlchemy’s unicode decoding converter under Python 2, and to cx_Oracle’s cursor.var() object as the encodingErrors parameter under Python 3, for the very unusual case that broken encodings are present in the target database which cannot be fetched unless error handling is relaxed. The value is ultimately one of the Python “encoding errors” parameters passed to decode().

References: #4799

[oracle] [bug] [firebird]

Modified the approach of “name normalization” for the Oracle and Firebird dialects, which converts from the UPPERCASE-as-case-insensitive convention of these dialects into lowercase-as-case-insensitive for SQLAlchemy, to not automatically apply the quoted_name construct to a name that matches itself under upper or lower case conversion, as is the case for many non-european characters. All names used within metadata structures are converted to quoted_name objects in any case; the change here would only affect the output of some inspection functions.

References: #4931

[oracle] [bug]

The NCHAR datatype will now bind to the cx_Oracle.FIXED_NCHAR DBAPI data bindings when used in a bound parameter, which supplies proper comparison behavior against a variable-length string. Previously, the NCHAR datatype would bind to cx_oracle.NCHAR which is not fixed length; the CHAR datatype already binds to cx_Oracle.FIXED_CHAR so it is now consistent that NCHAR binds to cx_Oracle.FIXED_NCHAR.

References: #4913

[firebird] [bug]

Added additional “disconnect” message “Error writing data to the connection” to Firebird disconnection detection. Pull request courtesy lukens.

References: #4903

[bug] [tests]

Fixed test failures which would occur with newer SQLite as of version 3.30 or greater, due to their addition of nulls ordering syntax as well as new restrictions on aggregate functions. Pull request courtesy Nils Philippsen.

References: #4920

[bug] [installation] [windows]

Added a workaround for a setuptools-related failure that has been observed as occurring on Windows installations, where setuptools is not correctly reporting a build error when the MSVC build dependencies are not installed and therefore not allowing graceful degradation into non C extensions builds.

References: #4967

[mssql] [bug]

Fixed bug in SQL Server dialect with new “max_identifier_length” feature where the mssql dialect already featured this flag, and the implementation did not accommodate for the new initialization hook correctly.

References: #4857

[oracle] [bug]

Fixed regression in Oracle dialect that was inadvertently using max identifier length of 128 characters on Oracle server 12.2 and greater even though the stated contract for the remainder of the 1.3 series is that this value stays at 30 until version SQLAlchemy 1.4. Also repaired issues with the retrieval of the “compatibility” version, and removed the warning emitted when the “v$parameter” view was not accessible as this was causing user confusion.

References: #4857, #4898

[orm] [bug]

Fixed regression in selectinload loader strategy caused by #4775 (released in version 1.3.6) where a many-to-one attribute of None would no longer be populated by the loader. While this was usually not noticeable due to the lazyloader populating None upon get, it would lead to a detached instance error if the object were detached.

References: #4872

[orm] [bug]

Passing a plain string expression to Session.query() is deprecated, as all string coercions were removed in #4481 and this one should have been included. The literal_column() function may be used to produce a textual column expression.

References: #4873

[orm] [bug]

A warning is emitted for a condition in which the Session may implicitly swap an object out of the identity map for another one with the same primary key, detaching the old one, which can be an observed result of load operations which occur within the SessionEvents.after_flush() hook. The warning is intended to notify the user that some special condition has caused this to happen and that the previous object may not be in the expected state.

References: #4890

[engine] [usecase]

Added new create_engine() parameter create_engine.max_identifier_length. This overrides the dialect-coded “max identifier length” in order to accommodate for databases that have recently changed this length and the SQLAlchemy dialect has not yet been adjusted to detect for that version. This parameter interacts with the existing create_engine.label_length parameter in that it establishes the maximum (and default) value for anonymously generated labels. Additionally, post-connection detection of max identifier lengths has been added to the dialect system. This feature is first being used by the Oracle dialect.

References: #4857

[sql] [usecase]

Added an explicit error message for the case when objects passed to Table are not SchemaItem objects, rather than resolving to an attribute error.

References: #4847

[sql] [bug]

Characters that interfere with “pyformat” or “named” formats in bound parameters, namely %, (, ) and the space character, as well as a few other typically undesirable characters, are stripped early for a bindparam() that is using an anonymized name, which is typically generated automatically from a named column which itself includes these characters in its name and does not use a .key, so that they do not interfere either with the SQLAlchemy compiler’s use of string formatting or with the driver-level parsing of the parameter, both of which could be demonstrated before the fix. The change only applies to anonymized parameter names that are generated and consumed internally, not end-user defined names, so the change should have no impact on any existing code. Applies in particular to the psycopg2 driver which does not otherwise quote special parameter names, but also strips leading underscores to suit Oracle (but not yet leading numbers, as some anon parameters are currently entirely numeric/underscore based); Oracle in any case continues to quote parameter names that include special characters.

References: #4837

[sqlite] [usecase]

Added support for sqlite “URI” connections, which allow for sqlite-specific flags to be passed in the query string such as “read only” for Python sqlite3 drivers that support this.

References: #4863

[mssql] [bug]

Added identifier quoting to the schema name applied to the “use” statement which is invoked when a SQL Server multipart schema name is used within a Table that is being reflected, as well as for Inspector methods such as Inspector.get_table_names(); this accommodates for special characters or spaces in the database name. Additionally, the “use” statement is not emitted if the current database matches the target owner database name being passed.

References: #4883

[oracle] [usecase]

The Oracle dialect now emits a warning if Oracle version 12.2 or greater is used, and the create_engine.max_identifier_length parameter is not set. The version in this specific case defaults to that of the “compatibility” version set in the Oracle server configuration, not the actual server version. In version 1.4, the default max_identifier_length for 12.2 or greater will move to 128 characters. In order to maintain forwards compatibility, applications should set create_engine.max_identifier_length to 30 in order to maintain the same length behavior, or to 128 in order to test the upcoming behavior. This length determines among other things how generated constraint names are truncated for statements like CREATE CONSTRAINT and DROP CONSTRAINT, which means a the new length may produce a name-mismatch against a name that was generated with the old length, impacting database migrations.

References: #4857

[oracle] [bug]

Restored adding cx_Oracle.DATETIME to the setinputsizes() call when a SQLAlchemy Date, DateTime or Time datatype is used, as some complex queries require this to be present. This was removed in the 1.2 series for arbitrary reasons.

References: #4886

[bug] [tests]

Fixed unit test regression released in 1.3.8 that would cause failure for Oracle, SQL Server and other non-native ENUM platforms due to new enumeration tests added as part of #4285 enum sortability in the unit of work; the enumerations created constraints that were duplicated on name.

References: #4285

[orm] [usecase]

Added support for the use of an Enum datatype using Python pep-435 enumeration objects as values for use as a primary key column mapped by the ORM. As these values are not inherently sortable, as required by the ORM for primary keys, a new TypeEngine.sort_key_function attribute is added to the typing system which allows any SQL type to implement a sorting for Python objects of its type which is consulted by the unit of work. The Enum type then defines this using the database value of a given enumeration. The sorting scheme can be also be redefined by passing a callable to the Enum.sort_key_function parameter. Pull request courtesy Nicolas Caniart.

References: #4285

[orm] [bug]

Fixed bug where Load objects were not pickleable due to mapper/relationship state in the internal context dictionary. These objects are now converted to picklable using similar techniques as that of other elements within the loader option system that have long been serializable.

References: #4823

[engine] [feature]

Added new parameter create_engine.hide_parameters which when set to True will cause SQL parameters to no longer be logged, nor rendered in the string representation of a StatementError object.

References: #4815

[engine] [bug]

Fixed an issue whereby if the dialect “initialize” process which occurs on first connect would encounter an unexpected exception, the initialize process would fail to complete and then no longer attempt on subsequent connection attempts, leaving the dialect in an un-initialized, or partially initialized state, within the scope of parameters that need to be established based on inspection of a live connection. The “invoke once” logic in the event system has been reworked to accommodate for this occurrence using new, private API features that establish an “exec once” hook that will continue to allow the initializer to fire off on subsequent connections, until it completes without raising an exception. This does not impact the behavior of the existing once=True flag within the event system.

References: #4807

[postgresql] [usecase]

Added support for reflection of CHECK constraints that include the special PostgreSQL qualifier “NOT VALID”, which can be present for CHECK constraints that were added to an exsiting table with the directive that they not be applied to existing data in the table. The PostgreSQL dictionary for CHECK constraints as returned by Inspector.get_check_constraints() may include an additional entry dialect_options which within will contain an entry "not_valid": True if this symbol is detected. Pull request courtesy Bill Finn.

References: #4824

[postgresql] [bug]

Revised the approach for the just added support for the psycopg2 “execute_values()” feature added in 1.3.7 for #4623. The approach relied upon a regular expression that would fail to match for a more complex INSERT statement such as one which had subqueries involved. The new approach matches exactly the string that was rendered as the VALUES clause.

References: #4623

[postgresql] [bug]

Fixed bug where Postgresql operators such as Comparator.contains() and Comparator.contained_by() would fail to function correctly for non-integer values when used against a array object, due to an erroneous assert statement.

References: #4822

[sqlite] [bug] [reflection]

Fixed bug where a FOREIGN KEY that was set up to refer to the parent table by table name only without the column names would not correctly be reflected as far as setting up the “referred columns”, since SQLite’s PRAGMA does not report on these columns if they weren’t given explicitly. For some reason this was harcoded to assume the name of the local column, which might work for some cases but is not correct. The new approach reflects the primary key of the referred table and uses the constraint columns list as the referred columns list, if the remote column(s) aren’t present in the reflected pragma directly.

References: #4810

[orm] [bug]

Fixed regression caused by new selectinload for many-to-one logic where a primaryjoin condition not based on real foreign keys would cause KeyError if a related object did not exist for a given key value on the parent object.

References: #4777

[orm] [bug]

Fixed bug where using Query.first() or a slice expression in conjunction with a query that has an expression based “offset” applied would raise TypeError, due to an “or” conditional against “offset” that did not expect it to be a SQL expression as opposed to an integer or None.

References: #4803

[sql] [bug]

Fixed issue where Index object which contained a mixture of functional expressions which were not resolvable to a particular column, in combination with string-based column names, would fail to initialize its internal state correctly leading to failures during DDL compilation.

References: #4778

[sql] [bug]

Fixed bug where TypeEngine.column_expression() method would not be applied to subsequent SELECT statements inside of a UNION or other _selectable.CompoundSelect, even though the SELECT statements are rendered at the topmost level of the statement. New logic now differentiates between rendering the column expression, which is needed for all SELECTs in the list, vs. gathering the returned data type for the result row, which is needed only for the first SELECT.

References: #4787

[sql] [bug]

Fixed issue where internal cloning of SELECT constructs could lead to a key error if the copy of the SELECT changed its state such that its list of columns changed. This was observed to be occurring in some ORM scenarios which may be unique to 1.3 and above, so is partially a regression fix.

References: #4780

[postgresql] [usecase]

Added new dialect flag for the psycopg2 dialect, executemany_mode which supersedes the previous experimental use_batch_mode flag. executemany_mode supports both the “execute batch” and “execute values” functions provided by psycopg2, the latter which is used for compiled insert() constructs. Pull request courtesy Yuval Dinari.

References: #4623

[mysql] [usecase]

Added reserved words ARRAY and MEMBER to the MySQL reserved words list, as MySQL 8.0 has now made these reserved.

References: #4783

[mysql] [bug]

The MySQL dialects will emit “SET NAMES” at the start of a connection when charset is given to the MySQL driver, to appease an apparent behavior observed in MySQL 8.0 that raises a collation error when a UNION includes string columns unioned against columns of the form CAST(NULL AS CHAR(..)), which is what SQLAlchemy’s polymorphic_union function does. The issue seems to have affected PyMySQL for at least a year, however has recently appeared as of mysqlclient 1.4.4 based on changes in how this DBAPI creates a connection. As the presence of this directive impacts three separate MySQL charset settings which each have intricate effects based on their presense, SQLAlchemy will now emit the directive on new connections to ensure correct behavior.

References: #4804

[mysql] [bug]

Added another fix for an upstream MySQL 8 issue where a case sensitive table name is reported incorrectly in foreign key constraint reflection, this is an extension of the fix first added for #4344 which affects a case sensitive column name. The new issue occurs through MySQL 8.0.17, so the general logic of the 88718 fix remains in place.

References: #4751

[sqlite] [bug]

The dialects that support json are supposed to take arguments json_serializer and json_deserializer at the create_engine() level, however the SQLite dialect calls them _json_serilizer and _json_deserilalizer. The names have been corrected, the old names are accepted with a change warning, and these parameters are now documented as create_engine.json_serializer and create_engine.json_deserializer.

References: #4798

[sqlite] [bug]

Fixed bug where usage of “PRAGMA table_info” in SQLite dialect meant that reflection features to detect for table existence, list of table columns, and list of foreign keys, would default to any table in any attached database, when no schema name was given and the table did not exist in the base schema. The fix explicitly runs PRAGMA for the ‘main’ schema and then the ‘temp’ schema if the ‘main’ returned no rows, to maintain the behavior of tables + temp tables in the “no schema” namespace, attached tables only in the “schema” namespace.

References: #4793

[mssql] [usecase]

Added new try_cast() construct for SQL Server which emits “TRY_CAST” syntax. Pull request courtesy Leonel Atencio.

References: #4782

[bug] [events]

Fixed issue in event system where using the once=True flag with dynamically generated listener functions would cause event registration of future events to fail if those listener functions were garbage collected after they were used, due to an assumption that a listened function is strongly referenced. The “once” wrapped is now modified to strongly reference the inner function persistently, and documentation is updated that using “once” does not imply automatic de-registration of listener functions.

References: #4794

[orm] [feature]

Added new loader option method Load.options() which allows loader options to be constructed hierarchically, so that many sub-options can be applied to a particular path without needing to call defaultload() many times. Thanks to Alessio Bogon for the idea.

References: #4736

[orm] [bug]

Fixed regression caused by #4365 where a join from an entity to itself without using aliases no longer raises an informative error message, instead failing on an assertion. The informative error condition has been restored.

References: #4773

[orm] [bug]

Fixed an issue where the _ORMJoin.join() method, which is a not-internally-used ORM-level method that exposes what is normally an internal process of Query.join(), did not propagate the full and outerjoin keyword arguments correctly. Pull request courtesy Denis Kataev.

References: #4713

[orm] [bug]

Fixed bug where a many-to-one relationship that specified uselist=True would fail to update correctly during a primary key change where a related column needs to change.

References: #4772

[orm] [bug]

Fixed bug where the detection for many-to-one or one-to-one use with a “dynamic” relationship, which is an invalid configuration, would fail to raise if the relationship were configured with uselist=True. The current fix is that it warns, instead of raises, as this would otherwise be backwards incompatible, however in a future release it will be a raise.

References: #4772

[orm] [bug]

Fixed bug where a synonym created against a mapped attribute that does not exist yet, as is the case when it refers to backref before mappers are configured, would raise recursion errors when trying to test for attributes on it which ultimately don’t exist (as occurs when the classes are run through Sphinx autodoc), as the unconfigured state of the synonym would put it into an attribute not found loop.

References: #4767

[orm] [performance]

The optimization applied to selectin loading in #4340 where a JOIN is not needed to eagerly load related items is now applied to many-to-one relationships as well, so that only the related table is queried for a simple join condition. In this case, the related items are queried based on the value of a foreign key column on the parent; if these columns are deferred or otherwise not loaded on any of the parent objects in the collection, the loader falls back to the JOIN method.

References: #4775

[engine] [bug]

Fixed bug where using reflection function such as MetaData.reflect() with an Engine object that had execution options applied to it would fail, as the resulting OptionEngine proxy object failed to include a .engine attribute used within the reflection routines.

References: #4754

[sql] [bug]

Adjusted the initialization for Enum to minimize how often it invokes the .__members__ attribute of a given PEP-435 enumeration object, to suit the case where this attribute is expensive to invoke, as is the case for some popular third party enumeration libraries.

References: #4758

[sql] [bug] [postgresql]

Fixed issue where the array_agg construct in combination with FunctionElement.filter() would not produce the correct operator precedence in combination with the array index operator.

References: #4760

[sql] [bug]

Fixed an unlikely issue where the “corresponding column” routine for unions and other _selectable.CompoundSelect objects could return the wrong column in some overlapping column situtations, thus potentially impacting some ORM operations when set operations are in use, if the underlying select() constructs were used previously in other similar kinds of routines, due to a cached value not being cleared.

References: #4747

[postgresql] [usecase]

Added support for reflection of indexes on PostgreSQL partitioned tables, which was added to PostgreSQL as of version 11.

References: #4771

[postgresql] [usecase]

Added support for multidimensional Postgresql array literals via nesting the array object within another one. The multidimensional array type is detected automatically.

References: #4756

[mysql] [bug]

Fixed bug where the special logic to render “NULL” for the TIMESTAMP datatype when nullable=True would not work if the column’s datatype were a TypeDecorator or a Variant. The logic now ensures that it unwraps down to the original TIMESTAMP so that this special case NULL keyword is correctly rendered when requested.

References: #4743

[mysql] [bug]

Enhanced MySQL/MariaDB version string parsing to accommodate for exotic MariaDB version strings where the “MariaDB” word is embedded among other alphanumeric characters such as “MariaDBV1”. This detection is critical in order to correctly accommodate for API features that have split between MySQL and MariaDB such as the “transaction_isolation” system variable.

References: #4624

[sqlite] [usecase]

Added support for composite (tuple) IN operators with SQLite, by rendering the VALUES keyword for this backend. As other backends such as DB2 are known to use the same syntax, the syntax is enabled in the base compiler using a dialect-level flag tuple_in_values. The change also includes support for “empty IN tuple” expressions for SQLite when using “in_()” between a tuple value and an empty set.

References: #4766

[mssql] [bug]

Ensured that the queries used to reflect indexes and view definitions will explicitly CAST string parameters into NVARCHAR, as many SQL Server drivers frequently treat string values, particularly those with non-ascii characters or larger string values, as TEXT which often don’t compare correctly against VARCHAR characters in SQL Server’s information schema tables for some reason. These CAST operations already take place for reflection queries against SQL Server information_schema. tables but were missing from three additional queries that are against sys. tables.

References: #4745

[orm] [bug]

Fixed a series of related bugs regarding joined table inheritance more than two levels deep, in conjunction with modification to primary key values, where those primary key columns are also linked together in a foreign key relationship as is typical for joined table inheritance. The intermediary table in a three-level inheritance hierarchy will now get its UPDATE if only the primary key value has changed and passive_updates=False (e.g. foreign key constraints not being enforced), whereas before it would be skipped; similarly, with passive_updates=True (e.g. ON UPDATE CASCADE in effect), the third-level table will not receive an UPDATE statement as was the case earlier which would fail since CASCADE already modified it. In a related issue, a relationship linked to a three-level inheritance hierarchy on the primary key of an intermediary table of a joined-inheritance hierarchy will also correctly have its foreign key column updated when the parent object’s primary key is modified, even if that parent object is a subclass of the linked parent class, whereas before these classes would not be counted.

References: #4723

[orm] [bug]

Fixed bug where the Mapper.all_orm_descriptors accessor would return an entry for the Mapper itself under the declarative __mapper___ key, when this is not a descriptor. The .is_attribute flag that’s present on all InspectionAttr objects is now consulted, which has also been modified to be True for an association proxy, as it was erroneously set to False for this object.

References: #4729

[orm] [bug]

Fixed regression in Query.join() where the aliased=True flag would not properly apply clause adaptation to filter criteria, if a previous join were made to the same entity. This is because the adapters were placed in the wrong order. The order has been reversed so that the adapter for the most recent aliased=True call takes precedence as was the case in 1.2 and earlier. This broke the “elementtree” examples among other things.

References: #4704

[orm] [bug] [py3k]

Replaced the Python compatbility routines for getfullargspec() with a fully vendored version from Python 3.3. Originally, Python was emitting deprecation warnings for this function in Python 3.8 alphas. While this change was reverted, it was observed that Python 3 implementations for getfullargspec() are an order of magnitude slower as of the 3.4 series where it was rewritten against Signature. While Python plans to improve upon this situation, SQLAlchemy projects for now are using a simple replacement to avoid any future issues.

References: #4674

[orm] [bug]

Reworked the attribute mechanics used by AliasedClass to no longer rely upon calling __getattribute__ on the MRO of the wrapped class, and to instead resolve the attribute normally on the wrapped class using getattr(), and then unwrap/adapt that. This allows a greater range of attribute styles on the mapped class including special __getattr__() schemes; but it also makes the code simpler and more resilient in general.

References: #4694

[sql] [bug]

Fixed a series of quoting issues which all stemmed from the concept of the literal_column() construct, which when being “proxied” through a subquery to be referred towards by a label that matches its text, the label would not have quoting rules applied to it, even if the string in the Label were set up as a quoted_name construct. Not applying quoting to the text of the Label is a bug because this text is strictly a SQL identifier name and not a SQL expression, and the string should not have quotes embedded into it already unlike the literal_column() which it may be applied towards. The existing behavior of a non-labeled literal_column() being propagated as is on the outside of a subquery is maintained in order to help with manual quoting schemes, although it’s not clear if valid SQL can be generated for such a construct in any case.

References: #4730

[postgresql] [usecase]

Added support for column sorting flags when reflecting indexes for PostgreSQL, including ASC, DESC, NULLSFIRST, NULLSLAST. Also adds this facility to the reflection system in general which can be applied to other dialects in future releases. Pull request courtesy Eli Collins.

References: #4717

[postgresql] [bug]

Fixed bug where PostgreSQL dialect could not correctly reflect an ENUM datatype that has no members, returning a list with None for the get_enums() call and raising a TypeError when reflecting a column which has such a datatype. The inspection now returns an empty list.

References: #4701

[mysql] [bug]

Fixed bug where MySQL ON DUPLICATE KEY UPDATE would not accommodate setting a column to the value NULL. Pull request courtesy Lukáš Banič.

References: #4715

[orm] [bug]

Fixed issue where the AttributeEvents.active_history flag would not be set for an event listener that propgated to a subclass via the AttributeEvents.propagate flag. This bug has been present for the full span of the AttributeEvents system.

References: #4695

[orm] [bug]

Fixed regression where new association proxy system was still not proxying hybrid attributes when they made use of the @hybrid_property.expression decorator to return an alternate SQL expression, or when the hybrid returned an arbitrary PropComparator, at the expression level. This involved further generalization of the heuristics used to detect the type of object being proxied at the level of QueryableAttribute, to better detect if the descriptor ultimately serves mapped classes or column expressions.

References: #4690

[orm] [bug]

Applied the mapper “configure mutex” against the declarative class mapping process, to guard against the race which can occur if mappers are used while dynamic module import schemes are still in the process of configuring mappers for related classes. This does not guard against all possible race conditions, such as if the concurrent import has not yet encountered the dependent classes as of yet, however it guards against as much as possible within the SQLAlchemy declarative process.

References: #4686

[orm] [bug]

A warning is now emitted for the case where a transient object is being merged into the session with Session.merge() when that object is already transient in the Session. This warns for the case where the object would normally be double-inserted.

References: #4647

[orm] [bug]

Fixed regression in new relationship m2o comparison logic first introduced at Improvement to the behavior of many-to-one query expressions when comparing to an attribute that is persisted as NULL and is in an un-fetched state in the mapped instance. Since the attribute has no explicit default, it needs to default to NULL when accessed in a persistent setting.

References: #4676

[engine] [bug] [postgresql]

Moved the “rollback” which occurs during dialect initialization so that it occurs after additional dialect-specific initialize steps, in particular those of the psycopg2 dialect which would inadvertently leave transactional state on the first new connection, which could interfere with some psycopg2-specific APIs which require that no transaction is started. Pull request courtesy Matthew Wilkes.

References: #4663

[sql] [bug]

Fixed that the GenericFunction class was inadvertently registering itself as one of the named functions. Pull request courtesy Adrien Berchet.

References: #4653

[sql] [bug]

Fixed issue where double negation of a boolean column wouldn’t reset the “NOT” operator.

References: #4618

[sql] [bug]

The GenericFunction namespace is being migrated so that function names are looked up in a case-insensitive manner, as SQL functions do not collide on case sensitive differences nor is this something which would occur with user-defined functions or stored procedures. Lookups for functions declared with GenericFunction now use a case insensitive scheme, however a deprecation case is supported which allows two or more GenericFunction objects with the same name of different cases to exist, which will cause case sensitive lookups to occur for that particular name, while emitting a warning at function registration time. Thanks to Adrien Berchet for a lot of work on this complicated feature.

References: #4569

[postgresql] [bug] [orm]

Fixed an issue where the “number of rows matched” warning would emit even if the dialect reported “supports_sane_multi_rowcount=False”, as is the case for psycogp2 with use_batch_mode=True and others.

References: #4661

[mysql] [bug]

Added support for DROP CHECK constraint which is required by MySQL 8.0.16 to drop a CHECK constraint; MariaDB supports plain DROP CONSTRAINT. The logic distinguishes between the two syntaxes by checking the server version string for MariaDB presence. Alembic migrations has already worked around this issue by implementing its own DROP for MySQL / MariaDB CHECK constraints, however this change implements it straight in Core so that its available for general use. Pull request courtesy Hannes Hansen.

References: #4650

[mssql] [feature]

Added support for SQL Server filtered indexes, via the mssql_where parameter which works similarly to that of the postgresql_where index function in the PostgreSQL dialect.

References: #4657

[mssql] [bug]

Added error code 20047 to “is_disconnect” for pymssql. Pull request courtesy Jon Schuff.

References: #4680

[misc] [bug]

Removed errant “sqla_nose.py” symbol from MANIFEST.in which created an undesirable warning message.

References: #4625

[orm] [bug]

Fixed 1.3 regression in new “ambiguous FROMs” query logic introduced in Query.join() handles ambiguity in deciding the “left” side more explicitly where a Query that explicitly places an entity in the FROM clause with Query.select_from() and also joins to it using Query.join() would later cause an “ambiguous FROM” error if that entity were used in additional joins, as the entity appears twice in the “from” list of the Query. The fix resolves this ambiguity by folding the standalone entity into the join that it’s already a part of in the same way that ultimately happens when the SELECT statement is rendered.

References: #4584

[orm] [bug]

Adjusted the Query.filter_by() method to not call and() internally against multiple criteria, instead passing it off to Query.filter() as a series of criteria, instead of a single criteria. This allows Query.filter_by() to defer to Query.filter()’s treatment of variable numbers of clauses, including the case where the list is empty. In this case, the Query object will not have a .whereclause, which allows subsequent “no whereclause” methods like Query.select_from() to behave consistently.

References: #4606

[postgresql] [bug]

Fixed regression from release 1.3.2 caused by #4562 where a URL that contained only a query string and no hostname, such as for the purposes of specifying a service file with connection information, would no longer be propagated to psycopg2 properly. The change in #4562 has been adjusted to further suit psycopg2’s exact requirements, which is that if there are any connection parameters whatsoever, the “dsn” parameter is no longer required, so in this case the query string parameters are passed alone.

References: #4601

[mssql] [bug]

Fixed issue in SQL Server dialect where if a bound parameter were present in an ORDER BY expression that would ultimately not be rendered in the SQL Server version of the statement, the parameters would still be part of the execution parameters, leading to DBAPI-level errors. Pull request courtesy Matt Lewellyn.

References: #4587

[bug] [pool]

Fixed behavioral regression as a result of deprecating the “use_threadlocal” flag for Pool, where the SingletonThreadPool no longer makes use of this option which causes the “rollback on return” logic to take place when the same Engine is used multiple times in the context of a transaction to connect or implicitly execute, thereby cancelling the transaction. While this is not the recommended way to work with engines and connections, it is nonetheless a confusing behavioral change as when using SingletonThreadPool, the transaction should stay open regardless of what else is done with the same engine in the same thread. The use_threadlocal flag remains deprecated however the SingletonThreadPool now implements its own version of the same logic.

References: #4585

[bug] [ext]

Fixed bug where using copy.copy() or copy.deepcopy() on MutableList would cause the items within the list to be duplicated, due to an inconsistency in how Python pickle and copy both make use of __getstate__() and __setstate__() regarding lists. In order to resolve, a __reduce_ex__ method had to be added to MutableList. In order to maintain backwards compatibility with existing pickles based on __getstate__(), the __setstate__() method remains as well; the test suite asserts that pickles made against the old version of the class can still be deserialized by the pickle module.

References: #4603

[orm] [bug] [ext]

Restored instance-level support for plain Python descriptors, e.g. @property objects, in conjunction with association proxies, in that if the proxied object is not within ORM scope at all, it gets classified as “ambiguous” but is proxed directly. For class level access, a basic class level``__get__()`` now returns the AmbiguousAssociationProxyInstance directly, rather than raising its exception, which is the closest approximation to the previous behavior that returned the AssociationProxy itself that’s possible. Also improved the stringification of these objects to be more descriptive of current state.

References: #4573, #4574

[orm] [bug]

Fixed bug where use of with_polymorphic() or other aliased construct would not properly adapt when the aliased target were used as the Select.correlate_except() target of a subquery used inside of a column_property(). This required a fix to the clause adaption mechanics to properly handle a selectable that shows up in the “correlate except” list, in a similar manner as which occurs for selectables that show up in the “correlate” list. This is ultimately a fairly fundamental bug that has lasted for a long time but it is hard to come across it.

References: #4537

[orm] [bug]

Fixed regression where a new error message that was supposed to raise when attempting to link a relationship option to an AliasedClass without using PropComparator.of_type() would instead raise an AttributeError. Note that in 1.3, it is no longer valid to create an option path from a plain mapper relationship to an AliasedClass without using PropComparator.of_type().

References: #4566

[sql] [bug] [documentation]

Thanks to TypeEngine methods bind_expression, column_expression work with Variant, type-specific types, we no longer need to rely on recipes that subclass dialect-specific types directly, TypeDecorator can now handle all cases. Additionally, the above change made it slightly less likely that a direct subclass of a base SQLAlchemy type would work as expected, which could be misleading. Documentation has been updated to use TypeDecorator for these examples including the PostgreSQL “ArrayOfEnum” example datatype and direct support for the “subclass a type directly” has been removed.

References: #4580

[postgresql] [feature]

Added support for parameter-less connection URLs for the psycopg2 dialect, meaning, the URL can be passed to create_engine() as "postgresql+psycopg2://" with no additional arguments to indicate an empty DSN passed to libpq, which indicates to connect to “localhost” with no username, password, or database given. Pull request courtesy Julian Mehnle.

References: #4562

[postgresql] [bug]

Modified the Select.with_for_update.of parameter so that if a join or other composed selectable is passed, the individual Table objects will be filtered from it, allowing one to pass a join() object to the parameter, as occurs normally when using joined table inheritance with the ORM. Pull request courtesy Raymond Lu.

References: #4550

[orm] [bug] [ext]

Fixed regression where an association proxy linked to a synonym would no longer work, both at instance level and at class level.

References: #4522

[mssql] [bug]

A commit() is emitted after an isolation level change to SNAPSHOT, as both pyodbc and pymssql open an implicit transaction which blocks subsequent SQL from being emitted in the current transaction.

This change is also backported to: 1.2.19

References: #4536

[mssql] [bug]

Fixed regression in SQL Server reflection due to #4393 where the removal of open-ended **kw from the Float datatype caused reflection of this type to fail due to a “scale” argument being passed.

References: #4525

[orm] [feature]

The Query.get() method can now accept a dictionary of attribute keys and values as a means of indicating the primary key value to load; is particularly useful for composite primary keys. Pull request courtesy Sanjana S.

References: #4316

[orm] [feature]

A SQL expression can now be assigned to a primary key attribute for an ORM flush in the same manner as ordinary attributes as described in Embedding SQL Insert/Update Expressions into a Flush where the expression will be evaulated and then returned to the ORM using RETURNING, or in the case of pysqlite, works using the cursor.lastrowid attribute.Requires either a database that supports RETURNING (e.g. Postgresql, Oracle, SQL Server) or pysqlite.

References: #3133

[engine] [feature]

Revised the formatting for StatementError when stringified. Each error detail is broken up over multiple newlines instead of spaced out on a single line. Additionally, the SQL representation now stringifies the SQL statement rather than using repr(), so that newlines are rendered as is. Pull request courtesy Nate Clark.

References: #4500

[sql] [bug]

The Alias class and related subclasses CTE, Lateral and TableSample have been reworked so that it is not possible for a user to construct the objects directly. These constructs require that the standalone construction function or selectable-bound method be used to instantiate new objects.

References: #4509

[schema] [feature]

Added new parameters Table.resolve_fks and MetaData.reflect.resolve_fks which when set to False will disable the automatic reflection of related tables encountered in ForeignKey objects, which can both reduce SQL overhead for omitted tables as well as avoid tables that can’t be reflected for database-specific reasons. Two Table objects present in the same MetaData collection can still refer to each other even if the reflection of the two tables occurred separately.

References: #4517

[orm] [bug]

Improved the behavior of with_polymorphic() in conjunction with loader options, in particular wildcard operations as well as load_only(). The polymorphic object will be more accurately targeted so that column-level options on the entity will correctly take effect.The issue is a continuation of the same kinds of things fixed in #4468.

References: #4469

[bug] [declarative] [orm]

Added some helper exceptions that invoke when a mapping based on AbstractConcreteBase, DeferredReflection, or AutoMap is used before the mapping is ready to be used, which contain descriptive information on the class, rather than falling through into other failure modes that are less informative.

References: #4470

[sql] [bug]

Fully removed the behavior of strings passed directly as components of a select() or Query object being coerced to text() constructs automatically; the warning that has been emitted is now an ArgumentError or in the case of order_by() / group_by() a CompileError. This has emitted a warning since version 1.0 however its presence continues to create concerns for the potential of mis-use of this behavior.

Note that public CVEs have been posted for order_by() / group_by() which are resolved by this commit: CVE-2019-7164 CVE-2019-7548

References: #4481

[sql] [bug]

Quoting is applied to Function names, those which are usually but not necessarily generated from the sqlalchemy.sql.expression.func construct, at compile time if they contain illegal characters, such as spaces or punctuation. The names are as before treated as case insensitive however, meaning if the names contain uppercase or mixed case characters, that alone does not trigger quoting. The case insensitivity is currently maintained for backwards compatibility.

References: #4467

[sql] [bug]

Added “SQL phrase validation” to key DDL phrases that are accepted as plain strings, including ForeignKeyConstraint.on_delete, ForeignKeyConstraint.on_update, ExcludeConstraint.using, ForeignKeyConstraint.initially, for areas where a series of SQL keywords only are expected.Any non-space characters that suggest the phrase would need to be quoted will raise a CompileError. This change is related to the series of changes committed as part of #4481.

References: #4481

[postgresql] [bug]

Fixed issue where using an uppercase name for an index type (e.g. GIST, BTREE, etc. ) or an EXCLUDE constraint would treat it as an identifier to be quoted, rather than rendering it as is. The new behavior converts these types to lowercase and ensures they contain only valid SQL characters.

References: #4473

[bug] [ext]

Implemented a more comprehensive assignment operation (e.g. “bulk replace”) when using association proxy with sets or dictionaries. Fixes the problem of redundant proxy objects being created to replace the old ones, which leads to excessive events and SQL and in the case of unique constraints will cause the flush to fail.

References: #2642

[change] [tests]

The test system has removed support for Nose, which is unmaintained for several years and is producing warnings under Python 3. The test suite is currently standardized on Pytest. Pull request courtesy Parth Shandilya.

References: #4460

[general] [change]

A large change throughout the library has ensured that all objects, parameters, and behaviors which have been noted as deprecated or legacy now emit DeprecationWarning warnings when invoked.As the Python 3 interpreter now defaults to displaying deprecation warnings, as well as that modern test suites based on tools like tox and pytest tend to display deprecation warnings, this change should make it easier to note what API features are obsolete. A major rationale for this change is so that long- deprecated features that nonetheless still see continue to see real world use can finally be removed in the near future; the biggest example of this are the SessionExtension and MapperExtension classes as well as a handful of other pre-event extension hooks, which have been deprecated since version 0.7 but still remain in the library. Another is that several major longstanding behaviors are to be deprecated as well, including the threadlocal engine strategy, the convert_unicode flag, and non primary mappers.

References: #4393

[orm] [feature]

Implemented a new feature whereby the AliasedClass construct can now be used as the target of a relationship(). This allows the concept of “non primary mappers” to no longer be necessary, as the AliasedClass is much easier to configure and automatically inherits all the relationships of the mapped class, as well as preserves the ability for loader options to work normally.

References: #4423

[orm] [feature]

Added new MapperEvents.before_mapper_configured() event. This event complements the other “configure” stage mapper events with a per mapper event that receives each Mapper right before its configure step, and additionally may be used to prevent or delay the configuration of specific Mapper objects using a new return value interfaces.EXT_SKIP. See the documentation link for an example.

References: #4397

[orm] [bug]

Fixed long-standing issue where duplicate collection members would cause a backref to delete the association between the member and its parent object when one of the duplicates were removed, as occurs as a side effect of swapping two objects in one statement.

References: #1103

[orm] [bug]

Extended the fix first made as part of #3287, where a loader option made against a subclass using a wildcard would extend itself to include application of the wildcard to attributes on the super classes as well, to a “bound” loader option as well, e.g. in an expression like Load(SomeSubClass).load_only('foo'). Columns that are part of the parent class of SomeSubClass will also be excluded in the same way as if the unbound option load_only('foo') were used.

References: #4373

[orm] [bug]

Improved error messages emitted by the ORM in the area of loader option traversal. This includes early detection of mis-matched loader strategies along with a clearer explanation why these strategies don’t match.

References: #4433

[orm] [bug]

The “remove” event for collections is now called before the item is removed in the case of the collection.remove() method, as is consistent with the behavior for most other forms of collection item removal (such as __delitem__, replacement under __setitem__). For pop() methods, the remove event still fires after the operation.

[orm] [bug] [engine]

Added accessors for execution options to Core and ORM, via Query.get_execution_options(), Connection.get_execution_options(), Engine.get_execution_options(), and Executable.get_execution_options(). PR courtesy Daniel Lister.

References: #4464

[orm] [bug]

Fixed issue in association proxy due to #3423 which caused the use of custom PropComparator objects with hybrid attributes, such as the one demonstrated in the dictlike-polymorphic example to not function within an association proxy. The strictness that was added in #3423 has been relaxed, and additional logic to accommodate for an association proxy that links to a custom hybrid have been added.

References: #4446

[orm] [bug]

Implemented the .get_history() method, which also implies availability of AttributeState.history, for synonym() attributes. Previously, trying to access attribute history via a synonym would raise an AttributeError.

References: #3777

[orm] [change]

Added a new function close_all_sessions() which takes over the task of the Session.close_all() method, which is now deprecated as this is confusing as a classmethod. Pull request courtesy Augustin Trancart.

References: #4412

[orm declarative] [bug]

Added a __clause_element__() method to ColumnProperty which can allow the usage of a not-fully-declared column or deferred attribute in a declarative mapped class slightly more friendly when it’s used in a constraint or other column-oriented scenario within the class declaration, though this still can’t work in open-ended expressions; prefer to call the ColumnProperty.expression attribute if receiving TypeError.

References: #4372

[engine] [feature]

Added public accessor QueuePool.timeout() that returns the configured timeout for a QueuePool object. Pull request courtesy Irina Delamare.

References: #3689

[engine] [change]

The “threadlocal” engine strategy which has been a legacy feature of SQLAlchemy since around version 0.2 is now deprecated, along with the Pool.threadlocal parameter of Pool which has no effect in most modern use cases.

References: #4393

[sql] [feature]

Amended the AnsiFunction class, the base of common SQL functions like CURRENT_TIMESTAMP, to accept positional arguments like a regular ad-hoc function. This to suit the case that many of these functions on specific backends accept arguments such as “fractional seconds” precision and such. If the function is created with arguments, it renders the parenthesis and the arguments. If no arguments are present, the compiler generates the non-parenthesized form.

References: #4386

[sql] [change]

The create_engine.convert_unicode and String.convert_unicode parameters have been deprecated. These parameters were built back when most Python DBAPIs had little to no support for Python Unicode objects, and SQLAlchemy needed to take on the very complex task of marshalling data and SQL strings between Unicode and bytestrings throughout the system in a performant way. Thanks to Python 3, DBAPIs were compelled to adapt to Unicode-aware APIs and today all DBAPIs supported by SQLAlchemy support Unicode natively, including on Python 2, allowing this long-lived and very complicated feature to finally be (mostly) removed. There are still of course a few Python 2 edge cases where SQLAlchemy has to deal with Unicode however these are handled automatically; in modern use, there should be no need for end-user interaction with these flags.

References: #4393

[mssql] [bug]

The literal_processor for the Unicode and UnicodeText datatypes now render an N character in front of the literal string expression as required by SQL Server for Unicode string values rendered in SQL expressions.

References: #4442

[bug] [ext]

Fixed a regression in 1.3.0b1 caused by #3423 where association proxy objects that access an attribute that’s only present on a polymorphic subclass would raise an AttributeError even though the actual instance being accessed was an instance of that subclass.

References: #4401

[orm] [feature]

Added new feature Query.only_return_tuples(). Causes the Query object to return keyed tuple objects unconditionally even if the query is against a single entity. Pull request courtesy Eric Atkin.

This change is also backported to: 1.2.5

[orm] [feature]

Added new flag Session.bulk_save_objects.preserve_order to the Session.bulk_save_objects() method, which defaults to True. When set to False, the given mappings will be grouped into inserts and updates per each object type, to allow for greater opportunities to batch common operations together. Pull request courtesy Alessandro Cucci.

[orm] [feature]

The “selectin” loader strategy now omits the JOIN in the case of a simple one-to-many load, where it instead relies loads only from the related table, relying upon the foreign key columns of the related table in order to match up to primary keys in the parent table. This optimization can be disabled by setting the relationship.omit_join flag to False. Many thanks to Jayson Reis for the efforts on this.

References: #4340

[orm] [feature]

Added .info dictionary to the InstanceState class, the object that comes from calling inspect() on a mapped object.

References: #4257

[orm] [bug]

Fixed bug where use of Lateral construct in conjunction with Query.join() as well as Query.select_entity_from() would not apply clause adaption to the right side of the join. “lateral” introduces the use case of the right side of a join being correlatable. Previously, adaptation of this clause wasn’t considered. Note that in 1.2 only, a selectable introduced by Query.subquery() is still not adapted due to #4304; the selectable needs to be produced by the select() function to be the right side of the “lateral” join.

This change is also backported to: 1.2.12

References: #4334

[orm] [bug]

Fixed issue regarding passive_deletes=”all”, where the foreign key attribute of an object is maintained with its value even after the object is removed from its parent collection. Previously, the unit of work would set this to NULL even though passive_deletes indicated it should not be modified.

References: #3844

[orm] [bug]

Improved the behavior of a relationship-bound many-to-one object expression such that the retrieval of column values on the related object are now resilient against the object being detached from its parent Session, even if the attribute has been expired. New features within the InstanceState are used to memoize the last known value of a particular column attribute before its expired, so that the expression can still evaluate when the object is detached and expired at the same time. Error conditions are also improved using modern attribute state features to produce more specific messages as needed.

References: #4359

[orm] [bug] [mysql] [postgresql]

The ORM now doubles the “FOR UPDATE” clause within the subquery that renders in conjunction with joined eager loading in some cases, as it has been observed that MySQL does not lock the rows from a subquery. This means the query renders with two FOR UPDATE clauses; note that on some backends such as Oracle, FOR UPDATE clauses on subqueries are silently ignored since they are unnecessary. Additionally, in the case of the “OF” clause used primarily with PostgreSQL, the FOR UPDATE is rendered only on the inner subquery when this is used so that the selectable can be targeted to the table within the SELECT statement.

References: #4246

[orm] [bug]

Refactored Query.join() to further clarify the individual components of structuring the join. This refactor adds the ability for Query.join() to determine the most appropriate “left” side of the join when there is more than one element in the FROM list or the query is against multiple entities. If more than one FROM/entity matches, an error is raised that asks for an ON clause to be specified to resolve the ambiguity. In particular this targets the regression we saw in #4363 but is also of general use. The codepaths within Query.join() are now easier to follow and the error cases are decided more specifically at an earlier point in the operation.

References: #4365

[orm] [bug]

Fixed long-standing issue in Query where a scalar subquery such as produced by Query.exists(), Query.as_scalar() and other derivations from Query.statement would not correctly be adapted when used in a new Query that required entity adaptation, such as when the query were turned into a union, or a from_self(), etc. The change removes the “no adaptation” annotation from the select() object produced by the Query.statement accessor.

References: #4304

[orm] [bug]

An informative exception is re-raised when a primary key value is not sortable in Python during an ORM flush under Python 3, such as an Enum that has no __lt__() method; normally Python 3 raises a TypeError in this case. The flush process sorts persistent objects by primary key in Python so the values must be sortable.

References: #4232

[orm] [bug]

Removed the collection converter used by the MappedCollection class. This converter was used only to assert that the incoming dictionary keys matched that of their corresponding objects, and only during a bulk set operation. The converter can interfere with a custom validator or AttributeEvents.bulk_replace() listener that wants to convert incoming values further. The TypeError which would be raised by this converter when an incoming key didn’t match the value is removed; incoming values during a bulk assignment will be keyed to their value-generated key, and not the key that’s explicitly present in the dictionary.

Overall, @converter is superseded by the AttributeEvents.bulk_replace() event handler added as part of #3896.

References: #3604

[orm] [bug]

Added new behavior to the lazy load that takes place when the “old” value of a many-to-one is retrieved, such that exceptions which would be raised due to either lazy="raise" or a detached session error are skipped.

References: #4353

[orm] [bug]

A long-standing oversight in the ORM, the __delete__ method for a many- to-one relationship was non-functional, e.g. for an operation such as del a.b. This is now implemented and is equivalent to setting the attribute to None.

References: #4354

[bug] [declarative] [orm]

Fixed bug where declarative would not update the state of the Mapper as far as what attributes were present, when additional attributes were added or removed after the mapper attribute collections had already been called and memoized. Additionally, a NotImplementedError is now raised if a fully mapped attribute (e.g. column, relationship, etc.) is deleted from a class that is currently mapped, since the mapper will not function correctly if the attribute has been removed.

References: #4133

[engine] [feature]

Added new “lifo” mode to QueuePool, typically enabled by setting the flag create_engine.pool_use_lifo to True. “lifo” mode means the same connection just checked in will be the first to be checked out again, allowing excess connections to be cleaned up from the server side during periods of the pool being only partially utilized. Pull request courtesy Taem Park.

[sql] [feature]

Refactored SQLCompiler to expose a SQLCompiler.group_by_clause() method similar to the SQLCompiler.order_by_clause() and SQLCompiler.limit_clause() methods, which can be overridden by dialects to customize how GROUP BY renders. Pull request courtesy Samuel Chou.

This change is also backported to: 1.2.13

[sql] [feature]

Added Sequence to the “string SQL” system that will render a meaningful string expression ("<next sequence value: my_sequence>") when stringifying without a dialect a statement that includes a “sequence nextvalue” expression, rather than raising a compilation error.

References: #4144

[sql] [feature]

Added new naming convention tokens column_0N_name, column_0_N_name, etc., which will render the names / keys / labels for all columns referenced by a particular constraint in a sequence. In order to accommodate for the length of such a naming convention, the SQL compiler’s auto-truncation feature now applies itself to constraint names as well, which creates a shortened, deterministically generated name for the constraint that will apply to a target backend without going over the character limit of that backend.

The change also repairs two other issues. One is that the column_0_key token wasn’t available even though this token was documented, the other was that the referred_column_0_name token would inadvertently render the .key and not the .name of the column if these two values were different.

References: #3989

[sql] [feature]

Added new logic to the “expanding IN” bound parameter feature whereby if the given list is empty, a special “empty set” expression that is specific to different backends is generated, thus allowing IN expressions to be fully dynamic including empty IN expressions.

References: #4271

[sql] [feature]

The Python builtin dir() is now supported for a SQLAlchemy “properties” object, such as that of a Core columns collection (e.g. .c), mapper.attrs, etc. Allows iPython autocompletion to work as well. Pull request courtesy Uwe Korn.

[sql] [feature]

Added new feature FunctionElement.as_comparison() which allows a SQL function to act as a binary comparison operation that can work within the ORM.

References: #3831

[sql] [bug]

Added “like” based operators as “comparison” operators, including ColumnOperators.startswith() ColumnOperators.endswith() ColumnOperators.ilike() ColumnOperators.notilike() among many others, so that all of these operators can be the basis for an ORM “primaryjoin” condition.

References: #4302

[sql] [bug]

Fixed issue with TypeEngine.bind_expression() and TypeEngine.column_expression() methods where these methods would not work if the target type were part of a Variant, or other target type of a TypeDecorator. Additionally, the SQL compiler now calls upon the dialect-level implementation when it renders these methods so that dialects can now provide for SQL-level processing for built-in types.

References: #3981

[postgresql] [feature]

Added new PG type REGCLASS which assists in casting table names to OID values. Pull request courtesy Sebastian Bank.

This change is also backported to: 1.2.7

References: #4160

[postgresql] [feature]

Added rudimental support for reflection of PostgreSQL partitioned tables, e.g. that relkind=’p’ is added to reflection queries that return table information.

References: #4237

[mysql] [feature]

Support added for the “WITH PARSER” syntax of CREATE FULLTEXT INDEX in MySQL, using the mysql_with_parser keyword argument. Reflection is also supported, which accommodates MySQL’s special comment format for reporting on this option as well. Additionally, the “FULLTEXT” and “SPATIAL” index prefixes are now reflected back into the mysql_prefix index option.

References: #4219

[mysql] [feature]

Added support for the parameters in an ON DUPLICATE KEY UPDATE statement on MySQL to be ordered, since parameter order in a MySQL UPDATE clause is significant, in a similar manner as that described at Parameter-Ordered Updates. Pull request courtesy Maxim Bublis.

[mysql] [feature]

The “pre-ping” feature of the connection pool now uses the ping() method of the DBAPI connection in the case of mysqlclient, PyMySQL and mysql-connector-python. Pull request courtesy Maxim Bublis.

[sqlite] [feature]

Added support for SQLite’s json functionality via the new SQLite implementation for JSON, JSON. The name used for the type is JSON, following an example found at SQLite’s own documentation. Pull request courtesy Ilja Everilä.

References: #3850

[sqlite] [feature]

Implemented the SQLite ON CONFLICT clause as understood at the DDL level, e.g. for primary key, unique, and CHECK constraints as well as specified on a Column to satisfy inline primary key and NOT NULL. Pull request courtesy Denis Kataev.

References: #4360

[mssql] [feature]

Added fast_executemany=True parameter to the SQL Server pyodbc dialect, which enables use of pyodbc’s new performance feature of the same name when using Microsoft ODBC drivers.

References: #4158

[mssql] [bug]

Deprecated the use of Sequence with SQL Server in order to affect the “start” and “increment” of the IDENTITY value, in favor of new parameters mssql_identity_start and mssql_identity_increment which set these parameters directly. Sequence will be used to generate real CREATE SEQUENCE DDL with SQL Server in a future release.

References: #4362

[oracle] [feature]

Added a new event currently used only by the cx_Oracle dialect, DialectEvents.setiputsizes(). The event passes a dictionary of BindParameter objects to DBAPI-specific type objects that will be passed, after conversion to parameter names, to the cx_Oracle cursor.setinputsizes() method. This allows both visibility into the setinputsizes process as well as the ability to alter the behavior of what datatypes are passed to this method.

This change is also backported to: 1.2.9

References: #4290

[oracle] [bug]

Updated the parameters that can be sent to the cx_Oracle DBAPI to both allow for all current parameters as well as for future parameters not added yet. In addition, removed unused parameters that were deprecated in version 1.2, and additionally we are now defaulting “threaded” to False.

References: #4369

[oracle] [bug]

The Oracle dialect will no longer use the NCHAR/NCLOB datatypes represent generic unicode strings or clob fields in conjunction with Unicode and UnicodeText unless the flag use_nchar_for_unicode=True is passed to create_engine() - this includes CREATE TABLE behavior as well as setinputsizes() for bound parameters. On the read side, automatic Unicode conversion under Python 2 has been added to CHAR/VARCHAR/CLOB result rows, to match the behavior of cx_Oracle under Python 3. In order to mitigate the performance hit under Python 2, SQLAlchemy’s very performant (when C extensions are built) native Unicode handlers are used under Python 2.

References: #4242

[feature] [ext]

Added new attribute Query.lazy_loaded_from which is populated with an InstanceState that is using this Query in order to lazy load a relationship. The rationale for this is that it serves as a hint for the horizontal sharding feature to use, such that the identity token of the state can be used as the default identity token to use for the query within id_chooser().

This change is also backported to: 1.2.9

References: #4243

[feature] [ext]

Added new feature BakedQuery.to_query(), which allows for a clean way of using one BakedQuery as a subquery inside of another BakedQuery without needing to refer explicitly to a Session.

References: #4318

[feature] [ext]

The AssociationProxy now has standard column comparison operations such as ColumnOperators.like() and ColumnOperators.startswith() available when the target attribute is a plain column - the EXISTS expression that joins to the target table is rendered as usual, but the column expression is then use within the WHERE criteria of the EXISTS. Note that this alters the behavior of the .contains() method on the association proxy to make use of ColumnOperators.contains() when used on a column-based attribute.

References: #4351

[feature] [ext]

Added support for bulk Query.update() and Query.delete() to the ShardedQuery class within the horizontal sharding extension. This also adds an additional expansion hook to the bulk update/delete methods Query._execute_crud().

References: #4196

[bug] [ext]

Reworked AssociationProxy to store state that’s specific to a parent class in a separate object, so that a single AssociationProxy can serve for multiple parent classes, as is intrinsic to inheritance, without any ambiguity in the state returned by it. A new method AssociationProxy.for_class() is added to allow inspection of class-specific state.

References: #3423

[bug] [ext]

The long-standing behavior of the association proxy collection maintaining only a weak reference to the parent object is reverted; the proxy will now maintain a strong reference to the parent for as long as the proxy collection itself is also in memory, eliminating the “stale association proxy” error. This change is being made on an experimental basis to see if any use cases arise where it causes side effects.

References: #4268

[bug] [ext]

Fixed multiple issues regarding de-association of scalar objects with the association proxy. del now works, and additionally a new flag AssociationProxy.cascade_scalar_deletes is added, which when set to True indicates that setting a scalar attribute to None or deleting via del will also set the source association to None.

References: #4308