What’s New in SQLAlchemy 0.9?

About this Document

This document describes changes between SQLAlchemy version 0.8, undergoing maintenance releases as of May, 2013, and SQLAlchemy version 0.9, which had its first production release on December 30, 2013.

Document last updated: June 10, 2015

Introduction

This guide introduces what’s new in SQLAlchemy version 0.9, and also documents changes which affect users migrating their applications from the 0.8 series of SQLAlchemy to 0.9.

Please carefully review Behavioral Changes - ORM and Behavioral Changes - Core for potentially backwards-incompatible changes.

Platform Support

Targeting Python 2.6 and Up Now, Python 3 without 2to3

The first achievement of the 0.9 release is to remove the dependency on the 2to3 tool for Python 3 compatibility. To make this more straightforward, the lowest Python release targeted now is 2.6, which features a wide degree of cross-compatibility with Python 3. All SQLAlchemy modules and unit tests are now interpreted equally well with any Python interpreter from 2.6 forward, including the 3.1 and 3.2 interpreters.

#2671

C Extensions Supported on Python 3

The C extensions have been ported to support Python 3 and now build in both Python 2 and Python 3 environments.

#2161

Behavioral Changes - ORM

Composite attributes are now returned as their object form when queried on a per-attribute basis

Using a Query in conjunction with a composite attribute now returns the object type maintained by that composite, rather than being broken out into individual columns. Using the mapping setup at Composite Column Types:

  1. >>> session.query(Vertex.start, Vertex.end).filter(Vertex.start == Point(3, 4)).all()
  2. [(Point(x=3, y=4), Point(x=5, y=6))]

This change is backwards-incompatible with code that expects the individual attribute to be expanded into individual columns. To get that behavior, use the .clauses accessor:

  1. >>> session.query(Vertex.start.clauses, Vertex.end.clauses).filter(
  2. ... Vertex.start == Point(3, 4)
  3. ... ).all()
  4. [(3, 4, 5, 6)]

See also

Column Bundles for ORM queries

#2824

Query.select_from() no longer applies the clause to corresponding entities

The Query.select_from() method has been popularized in recent versions as a means of controlling the first thing that a Query object “selects from”, typically for the purposes of controlling how a JOIN will render.

Consider the following example against the usual User mapping:

  1. select_stmt = select([User]).where(User.id == 7).alias()
  2. q = (
  3. session.query(User)
  4. .join(select_stmt, User.id == select_stmt.c.id)
  5. .filter(User.name == "ed")
  6. )

The above statement predictably renders SQL like the following:

  1. SELECT "user".id AS user_id, "user".name AS user_name
  2. FROM "user" JOIN (SELECT "user".id AS id, "user".name AS name
  3. FROM "user"
  4. WHERE "user".id = :id_1) AS anon_1 ON "user".id = anon_1.id
  5. WHERE "user".name = :name_1

If we wanted to reverse the order of the left and right elements of the JOIN, the documentation would lead us to believe we could use Query.select_from() to do so:

  1. q = (
  2. session.query(User)
  3. .select_from(select_stmt)
  4. .join(User, User.id == select_stmt.c.id)
  5. .filter(User.name == "ed")
  6. )

However, in version 0.8 and earlier, the above use of Query.select_from() would apply the select_stmt to replace the User entity, as it selects from the user table which is compatible with User:

  1. -- SQLAlchemy 0.8 and earlier...
  2. SELECT anon_1.id AS anon_1_id, anon_1.name AS anon_1_name
  3. FROM (SELECT "user".id AS id, "user".name AS name
  4. FROM "user"
  5. WHERE "user".id = :id_1) AS anon_1 JOIN "user" ON anon_1.id = anon_1.id
  6. WHERE anon_1.name = :name_1

The above statement is a mess, the ON clause refers anon_1.id = anon_1.id, our WHERE clause has been replaced with anon_1 as well.

This behavior is quite intentional, but has a different use case from that which has become popular for Query.select_from(). The above behavior is now available by a new method known as Query.select_entity_from(). This is a lesser used behavior that in modern SQLAlchemy is roughly equivalent to selecting from a customized aliased() construct:

  1. select_stmt = select([User]).where(User.id == 7)
  2. user_from_stmt = aliased(User, select_stmt.alias())
  3. q = session.query(user_from_stmt).filter(user_from_stmt.name == "ed")

So with SQLAlchemy 0.9, our query that selects from select_stmt produces the SQL we expect:

  1. -- SQLAlchemy 0.9
  2. SELECT "user".id AS user_id, "user".name AS user_name
  3. FROM (SELECT "user".id AS id, "user".name AS name
  4. FROM "user"
  5. WHERE "user".id = :id_1) AS anon_1 JOIN "user" ON "user".id = id
  6. WHERE "user".name = :name_1

The Query.select_entity_from() method will be available in SQLAlchemy 0.8.2, so applications which rely on the old behavior can transition to this method first, ensure all tests continue to function, then upgrade to 0.9 without issue.

#2736

viewonly=True on relationship() prevents history from taking effect

The viewonly flag on relationship() is applied to prevent changes to the target attribute from having any effect within the flush process. This is achieved by eliminating the attribute from being considered during the flush. However, up until now, changes to the attribute would still register the parent object as “dirty” and trigger a potential flush. The change is that the viewonly flag now prevents history from being set for the target attribute as well. Attribute events like backrefs and user-defined events still continue to function normally.

The change is illustrated as follows:

  1. from sqlalchemy import Column, Integer, ForeignKey, create_engine
  2. from sqlalchemy.orm import backref, relationship, Session
  3. from sqlalchemy.ext.declarative import declarative_base
  4. from sqlalchemy import inspect
  5. Base = declarative_base()
  6. class A(Base):
  7. __tablename__ = "a"
  8. id = Column(Integer, primary_key=True)
  9. class B(Base):
  10. __tablename__ = "b"
  11. id = Column(Integer, primary_key=True)
  12. a_id = Column(Integer, ForeignKey("a.id"))
  13. a = relationship("A", backref=backref("bs", viewonly=True))
  14. e = create_engine("sqlite://")
  15. Base.metadata.create_all(e)
  16. a = A()
  17. b = B()
  18. sess = Session(e)
  19. sess.add_all([a, b])
  20. sess.commit()
  21. b.a = a
  22. assert b in sess.dirty
  23. # before 0.9.0
  24. # assert a in sess.dirty
  25. # assert inspect(a).attrs.bs.history.has_changes()
  26. # after 0.9.0
  27. assert a not in sess.dirty
  28. assert not inspect(a).attrs.bs.history.has_changes()

#2833

Association Proxy SQL Expression Improvements and Fixes

The == and != operators as implemented by an association proxy that refers to a scalar value on a scalar relationship now produces a more complete SQL expression, intended to take into account the “association” row being present or not when the comparison is against None.

Consider this mapping:

  1. class A(Base):
  2. __tablename__ = "a"
  3. id = Column(Integer, primary_key=True)
  4. b_id = Column(Integer, ForeignKey("b.id"), primary_key=True)
  5. b = relationship("B")
  6. b_value = association_proxy("b", "value")
  7. class B(Base):
  8. __tablename__ = "b"
  9. id = Column(Integer, primary_key=True)
  10. value = Column(String)

Up through 0.8, a query like the following:

  1. s.query(A).filter(A.b_value == None).all()

would produce:

  1. SELECT a.id AS a_id, a.b_id AS a_b_id
  2. FROM a
  3. WHERE EXISTS (SELECT 1
  4. FROM b
  5. WHERE b.id = a.b_id AND b.value IS NULL)

In 0.9, it now produces:

  1. SELECT a.id AS a_id, a.b_id AS a_b_id
  2. FROM a
  3. WHERE (EXISTS (SELECT 1
  4. FROM b
  5. WHERE b.id = a.b_id AND b.value IS NULL)) OR a.b_id IS NULL

The difference being, it not only checks b.value, it also checks if a refers to no b row at all. This will return different results versus prior versions, for a system that uses this type of comparison where some parent rows have no association row.

More critically, a correct expression is emitted for A.b_value != None. In 0.8, this would return True for A rows that had no b:

  1. SELECT a.id AS a_id, a.b_id AS a_b_id
  2. FROM a
  3. WHERE NOT (EXISTS (SELECT 1
  4. FROM b
  5. WHERE b.id = a.b_id AND b.value IS NULL))

Now in 0.9, the check has been reworked so that it ensures the A.b_id row is present, in addition to B.value being non-NULL:

  1. SELECT a.id AS a_id, a.b_id AS a_b_id
  2. FROM a
  3. WHERE EXISTS (SELECT 1
  4. FROM b
  5. WHERE b.id = a.b_id AND b.value IS NOT NULL)

In addition, the has() operator is enhanced such that you can call it against a scalar column value with no criterion only, and it will produce criteria that checks for the association row being present or not:

  1. s.query(A).filter(A.b_value.has()).all()

output:

  1. SELECT a.id AS a_id, a.b_id AS a_b_id
  2. FROM a
  3. WHERE EXISTS (SELECT 1
  4. FROM b
  5. WHERE b.id = a.b_id)

This is equivalent to A.b.has(), but allows one to query against b_value directly.

#2751

Association Proxy Missing Scalar returns None

An association proxy from a scalar attribute to a scalar will now return None if the proxied object isn’t present. This is consistent with the fact that missing many-to-ones return None in SQLAlchemy, so should the proxied value. E.g.:

  1. from sqlalchemy import *
  2. from sqlalchemy.orm import *
  3. from sqlalchemy.ext.declarative import declarative_base
  4. from sqlalchemy.ext.associationproxy import association_proxy
  5. Base = declarative_base()
  6. class A(Base):
  7. __tablename__ = "a"
  8. id = Column(Integer, primary_key=True)
  9. b = relationship("B", uselist=False)
  10. bname = association_proxy("b", "name")
  11. class B(Base):
  12. __tablename__ = "b"
  13. id = Column(Integer, primary_key=True)
  14. a_id = Column(Integer, ForeignKey("a.id"))
  15. name = Column(String)
  16. a1 = A()
  17. # this is how m2o's always have worked
  18. assert a1.b is None
  19. # but prior to 0.9, this would raise AttributeError,
  20. # now returns None just like the proxied value.
  21. assert a1.bname is None

#2810

attributes.get_history() will query from the DB by default if value not present

A bugfix regarding get_history() allows a column-based attribute to query out to the database for an unloaded value, assuming the passive flag is left at its default of PASSIVE_OFF. Previously, this flag would not be honored. Additionally, a new method AttributeState.load_history() is added to complement the AttributeState.history attribute, which will emit loader callables for an unloaded attribute.

This is a small change demonstrated as follows:

  1. from sqlalchemy import Column, Integer, String, create_engine, inspect
  2. from sqlalchemy.orm import Session, attributes
  3. from sqlalchemy.ext.declarative import declarative_base
  4. Base = declarative_base()
  5. class A(Base):
  6. __tablename__ = "a"
  7. id = Column(Integer, primary_key=True)
  8. data = Column(String)
  9. e = create_engine("sqlite://", echo=True)
  10. Base.metadata.create_all(e)
  11. sess = Session(e)
  12. a1 = A(data="a1")
  13. sess.add(a1)
  14. sess.commit() # a1 is now expired
  15. # history doesn't emit loader callables
  16. assert inspect(a1).attrs.data.history == (None, None, None)
  17. # in 0.8, this would fail to load the unloaded state.
  18. assert attributes.get_history(a1, "data") == (
  19. (),
  20. [
  21. "a1",
  22. ],
  23. (),
  24. )
  25. # load_history() is now equivalent to get_history() with
  26. # passive=PASSIVE_OFF ^ INIT_OK
  27. assert inspect(a1).attrs.data.load_history() == (
  28. (),
  29. [
  30. "a1",
  31. ],
  32. (),
  33. )

#2787

Behavioral Changes - Core

Type objects no longer accept ignored keyword arguments

Up through the 0.8 series, most type objects accepted arbitrary keyword arguments which were silently ignored:

  1. from sqlalchemy import Date, Integer
  2. # storage_format argument here has no effect on any backend;
  3. # it needs to be on the SQLite-specific type
  4. d = Date(storage_format="%(day)02d.%(month)02d.%(year)04d")
  5. # display_width argument here has no effect on any backend;
  6. # it needs to be on the MySQL-specific type
  7. i = Integer(display_width=5)

This was a very old bug for which a deprecation warning was added to the 0.8 series, but because nobody ever runs Python with the “-W” flag, it was mostly never seen:

  1. $ python -W always::DeprecationWarning ~/dev/sqlalchemy/test.py
  2. /Users/classic/dev/sqlalchemy/test.py:5: SADeprecationWarning: Passing arguments to
  3. type object constructor <class 'sqlalchemy.types.Date'> is deprecated
  4. d = Date(storage_format="%(day)02d.%(month)02d.%(year)04d")
  5. /Users/classic/dev/sqlalchemy/test.py:9: SADeprecationWarning: Passing arguments to
  6. type object constructor <class 'sqlalchemy.types.Integer'> is deprecated
  7. i = Integer(display_width=5)

As of the 0.9 series the “catch all” constructor is removed from TypeEngine, and these meaningless arguments are no longer accepted.

The correct way to make use of dialect-specific arguments such as storage_format and display_width is to use the appropriate dialect-specific types:

  1. from sqlalchemy.dialects.sqlite import DATE
  2. from sqlalchemy.dialects.mysql import INTEGER
  3. d = DATE(storage_format="%(day)02d.%(month)02d.%(year)04d")
  4. i = INTEGER(display_width=5)

What about the case where we want the dialect-agnostic type also? We use the TypeEngine.with_variant() method:

  1. from sqlalchemy import Date, Integer
  2. from sqlalchemy.dialects.sqlite import DATE
  3. from sqlalchemy.dialects.mysql import INTEGER
  4. d = Date().with_variant(
  5. DATE(storage_format="%(day)02d.%(month)02d.%(year)04d"), "sqlite"
  6. )
  7. i = Integer().with_variant(INTEGER(display_width=5), "mysql")

TypeEngine.with_variant() isn’t new, it was added in SQLAlchemy 0.7.2. So code that is running on the 0.8 series can be corrected to use this approach and tested before upgrading to 0.9.

None can no longer be used as a “partial AND” constructor

None can no longer be used as the “backstop” to form an AND condition piecemeal. This pattern was not a documented pattern even though some SQLAlchemy internals made use of it:

  1. condition = None
  2. for cond in conditions:
  3. condition = condition & cond
  4. if condition is not None:
  5. stmt = stmt.where(condition)

The above sequence, when conditions is non-empty, will on 0.9 produce SELECT .. WHERE <condition> AND NULL. The None is no longer implicitly ignored, and is instead consistent with when None is interpreted in other contexts besides that of a conjunction.

The correct code for both 0.8 and 0.9 should read:

  1. from sqlalchemy.sql import and_
  2. if conditions:
  3. stmt = stmt.where(and_(*conditions))

Another variant that works on all backends on 0.9, but on 0.8 only works on backends that support boolean constants:

  1. from sqlalchemy.sql import true
  2. condition = true()
  3. for cond in conditions:
  4. condition = cond & condition
  5. stmt = stmt.where(condition)

On 0.8, this will produce a SELECT statement that always has AND true in the WHERE clause, which is not accepted by backends that don’t support boolean constants (MySQL, MSSQL). On 0.9, the true constant will be dropped within an and_() conjunction.

See also

Improved rendering of Boolean constants, NULL constants, conjunctions

The “password” portion of a create_engine() no longer considers the + sign as an encoded space

For whatever reason, the Python function unquote_plus() was applied to the “password” field of a URL, which is an incorrect application of the encoding rules described in RFC 1738 in that it escaped spaces as plus signs. The stringification of a URL now only encodes “:”, “@”, or “/” and nothing else, and is now applied to both the username and password fields (previously it only applied to the password). On parsing, encoded characters are converted, but plus signs and spaces are passed through as is:

  1. # password: "pass word + other:words"
  2. dbtype://user:pass word + other%3Awords@host/dbname
  3. # password: "apples/oranges"
  4. dbtype://username:apples%2Foranges@hostspec/database
  5. # password: "apples@oranges@@"
  6. dbtype://username:apples%40oranges%40%40@hostspec/database
  7. # password: '', username is "username@"
  8. dbtype://username%40:@hostspec/database

#2873

The precedence rules for COLLATE have been changed

Previously, an expression like the following:

  1. print((column("x") == "somevalue").collate("en_EN"))

would produce an expression like this:

  1. -- 0.8 behavior
  2. (x = :x_1) COLLATE en_EN

The above is misunderstood by MSSQL and is generally not the syntax suggested for any database. The expression will now produce the syntax illustrated by that of most database documentation:

  1. -- 0.9 behavior
  2. x = :x_1 COLLATE en_EN

The potentially backwards incompatible change arises if the ColumnOperators.collate() operator is being applied to the right-hand column, as follows:

  1. print(column("x") == literal("somevalue").collate("en_EN"))

In 0.8, this produces:

  1. x = :param_1 COLLATE en_EN

However in 0.9, will now produce the more accurate, but probably not what you want, form of:

  1. x = (:param_1 COLLATE en_EN)

The ColumnOperators.collate() operator now works more appropriately within an ORDER BY expression as well, as a specific precedence has been given to the ASC and DESC operators which will again ensure no parentheses are generated:

  1. >>> # 0.8
  2. >>> print(column("x").collate("en_EN").desc())
  3. (x COLLATE en_EN) DESC
  4. >>> # 0.9
  5. >>> print(column("x").collate("en_EN").desc())
  6. x COLLATE en_EN DESC

#2879

PostgreSQL CREATE TYPE <x> AS ENUM now applies quoting to values

The ENUM type will now apply escaping to single quote signs within the enumerated values:

  1. >>> from sqlalchemy.dialects import postgresql
  2. >>> type = postgresql.ENUM("one", "two", "three's", name="myenum")
  3. >>> from sqlalchemy.dialects.postgresql import base
  4. >>> print(base.CreateEnumType(type).compile(dialect=postgresql.dialect()))
  5. CREATE TYPE myenum AS ENUM ('one','two','three''s')

Existing workarounds which already escape single quote signs will need to be modified, else they will now double-escape.

#2878

New Features

Event Removal API

Events established using listen() or listens_for() can now be removed using the new remove() function. The target, identifier and fn arguments sent to remove() need to match exactly those which were sent for listening, and the event will be removed from all locations in which it had been established:

  1. @event.listens_for(MyClass, "before_insert", propagate=True)
  2. def my_before_insert(mapper, connection, target):
  3. """listen for before_insert"""
  4. # ...
  5. event.remove(MyClass, "before_insert", my_before_insert)

In the example above, the propagate=True flag is set. This means my_before_insert() is established as a listener for MyClass as well as all subclasses of MyClass. The system tracks everywhere that the my_before_insert() listener function had been placed as a result of this call and removes it as a result of calling remove().

The removal system uses a registry to associate arguments passed to listen() with collections of event listeners, which are in many cases wrapped versions of the original user-supplied function. This registry makes heavy use of weak references in order to allow all the contained contents, such as listener targets, to be garbage collected when they go out of scope.

#2268

New Query Options API; load_only() option

The system of loader options such as joinedload(), subqueryload(), lazyload(), defer(), etc. all build upon a new system known as Load. Load provides a “method chained” (a.k.a. generative) approach to loader options, so that instead of joining together long paths using dots or multiple attribute names, an explicit loader style is given for each path.

While the new way is slightly more verbose, it is simpler to understand in that there is no ambiguity in what options are being applied to which paths; it simplifies the method signatures of the options and provides greater flexibility particularly for column-based options. The old systems are to remain functional indefinitely as well and all styles can be mixed.

Old Way

To set a certain style of loading along every link in a multi-element path, the _all() option has to be used:

  1. query(User).options(joinedload_all("orders.items.keywords"))

New Way

Loader options are now chainable, so the same joinedload(x) method is applied equally to each link, without the need to keep straight between joinedload() and joinedload_all():

  1. query(User).options(joinedload("orders").joinedload("items").joinedload("keywords"))

Old Way

Setting an option on path that is based on a subclass requires that all links in the path be spelled out as class bound attributes, since the PropComparator.of_type() method needs to be called:

  1. session.query(Company).options(
  2. subqueryload_all(Company.employees.of_type(Engineer), Engineer.machines)
  3. )

New Way

Only those elements in the path that actually need PropComparator.of_type() need to be set as a class-bound attribute, string-based names can be resumed afterwards:

  1. session.query(Company).options(
  2. subqueryload(Company.employees.of_type(Engineer)).subqueryload("machines")
  3. )

Old Way

Setting the loader option on the last link in a long path uses a syntax that looks a lot like it should be setting the option for all links in the path, causing confusion:

  1. query(User).options(subqueryload("orders.items.keywords"))

New Way

A path can now be spelled out using defaultload() for entries in the path where the existing loader style should be unchanged. More verbose but the intent is clearer:

  1. query(User).options(defaultload("orders").defaultload("items").subqueryload("keywords"))

The dotted style can still be taken advantage of, particularly in the case of skipping over several path elements:

  1. query(User).options(defaultload("orders.items").subqueryload("keywords"))

Old Way

The defer() option on a path needed to be spelled out with the full path for each column:

  1. query(User).options(defer("orders.description"), defer("orders.isopen"))

New Way

A single Load object that arrives at the target path can have Load.defer() called upon it repeatedly:

  1. query(User).options(defaultload("orders").defer("description").defer("isopen"))

The Load Class

The Load class can be used directly to provide a “bound” target, especially when multiple parent entities are present:

  1. from sqlalchemy.orm import Load
  2. query(User, Address).options(Load(Address).joinedload("entries"))

Load Only

A new option load_only() achieves a “defer everything but” style of load, loading only the given columns and deferring the rest:

  1. from sqlalchemy.orm import load_only
  2. query(User).options(load_only("name", "fullname"))
  3. # specify explicit parent entity
  4. query(User, Address).options(Load(User).load_only("name", "fullname"))
  5. # specify path
  6. query(User).options(joinedload(User.addresses).load_only("email_address"))

Class-specific Wildcards

Using Load, a wildcard may be used to set the loading for all relationships (or perhaps columns) on a given entity, without affecting any others:

  1. # lazyload all User relationships
  2. query(User).options(Load(User).lazyload("*"))
  3. # undefer all User columns
  4. query(User).options(Load(User).undefer("*"))
  5. # lazyload all Address relationships
  6. query(User).options(defaultload(User.addresses).lazyload("*"))
  7. # undefer all Address columns
  8. query(User).options(defaultload(User.addresses).undefer("*"))

#1418

New text() Capabilities

The text() construct gains new methods:

  • TextClause.bindparams() allows bound parameter types and values to be set flexibly:

    1. # setup values
    2. stmt = text(
    3. "SELECT id, name FROM user WHERE name=:name AND timestamp=:timestamp"
    4. ).bindparams(name="ed", timestamp=datetime(2012, 11, 10, 15, 12, 35))
    5. # setup types and/or values
    6. stmt = (
    7. text("SELECT id, name FROM user WHERE name=:name AND timestamp=:timestamp")
    8. .bindparams(bindparam("name", value="ed"), bindparam("timestamp", type_=DateTime()))
    9. .bindparam(timestamp=datetime(2012, 11, 10, 15, 12, 35))
    10. )
  • TextClause.columns() supersedes the typemap option of text(), returning a new construct TextAsFrom:

    ```

    turn a text() into an alias(), with a .c. collection:

    stmt = text(“SELECT id, name FROM user”).columns(id=Integer, name=String) stmt = stmt.alias()

    stmt = select([addresses]).select_from(

    1. addresses.join(stmt), addresses.c.user_id == stmt.c.id

    )

  1. # or into a cte():
  2. stmt = text("SELECT id, name FROM user").columns(id=Integer, name=String)
  3. stmt = stmt.cte("x")
  4. stmt = select([addresses]).select_from(
  5. addresses.join(stmt), addresses.c.user_id == stmt.c.id
  6. )
  7. ```

#2877

INSERT from SELECT

After literally years of pointless procrastination this relatively minor syntactical feature has been added, and is also backported to 0.8.3, so technically isn’t “new” in 0.9. A select() construct or other compatible construct can be passed to the new method Insert.from_select() where it will be used to render an INSERT .. SELECT construct:

  1. >>> from sqlalchemy.sql import table, column
  2. >>> t1 = table("t1", column("a"), column("b"))
  3. >>> t2 = table("t2", column("x"), column("y"))
  4. >>> print(t1.insert().from_select(["a", "b"], t2.select().where(t2.c.y == 5)))
  5. INSERT INTO t1 (a, b) SELECT t2.x, t2.y
  6. FROM t2
  7. WHERE t2.y = :y_1

The construct is smart enough to also accommodate ORM objects such as classes and Query objects:

  1. s = Session()
  2. q = s.query(User.id, User.name).filter_by(name="ed")
  3. ins = insert(Address).from_select((Address.id, Address.email_address), q)

rendering:

  1. INSERT INTO addresses (id, email_address)
  2. SELECT users.id AS users_id, users.name AS users_name
  3. FROM users WHERE users.name = :name_1

#722

New FOR UPDATE support on select(), Query()

An attempt is made to simplify the specification of the FOR UPDATE clause on SELECT statements made within Core and ORM, and support is added for the FOR UPDATE OF SQL supported by PostgreSQL and Oracle.

Using the core GenerativeSelect.with_for_update(), options like FOR SHARE and NOWAIT can be specified individually, rather than linking to arbitrary string codes:

  1. stmt = select([table]).with_for_update(read=True, nowait=True, of=table)

On Posgtresql the above statement might render like:

  1. SELECT table.a, table.b FROM table FOR SHARE OF table NOWAIT

The Query object gains a similar method Query.with_for_update() which behaves in the same way. This method supersedes the existing Query.with_lockmode() method, which translated FOR UPDATE clauses using a different system. At the moment, the “lockmode” string argument is still accepted by the Session.refresh() method.

Floating Point String-Conversion Precision Configurable for Native Floating Point Types

The conversion which SQLAlchemy does whenever a DBAPI returns a Python floating point type which is to be converted into a Python Decimal() necessarily involves an intermediary step which converts the floating point value to a string. The scale used for this string conversion was previously hardcoded to 10, and is now configurable. The setting is available on both the Numeric as well as the Float type, as well as all SQL- and dialect-specific descendant types, using the parameter decimal_return_scale. If the type supports a .scale parameter, as is the case with Numeric and some float types such as DOUBLE, the value of .scale is used as the default for .decimal_return_scale if it is not otherwise specified. If both .scale and .decimal_return_scale are absent, then the default of 10 takes place. E.g.:

  1. from sqlalchemy.dialects.mysql import DOUBLE
  2. import decimal
  3. data = Table(
  4. "data",
  5. metadata,
  6. Column("double_value", mysql.DOUBLE(decimal_return_scale=12, asdecimal=True)),
  7. )
  8. conn.execute(
  9. data.insert(),
  10. double_value=45.768392065789,
  11. )
  12. result = conn.scalar(select([data.c.double_value]))
  13. # previously, this would typically be Decimal("45.7683920658"),
  14. # e.g. trimmed to 10 decimal places
  15. # now we get 12, as requested, as MySQL can support this
  16. # much precision for DOUBLE
  17. assert result == decimal.Decimal("45.768392065789")

#2867

Column Bundles for ORM queries

The Bundle allows for querying of sets of columns, which are then grouped into one name under the tuple returned by the query. The initial purposes of Bundle are 1. to allow “composite” ORM columns to be returned as a single value in a column-based result set, rather than expanding them out into individual columns and 2. to allow the creation of custom result-set constructs within the ORM, using ad-hoc columns and return types, without involving the more heavyweight mechanics of mapped classes.

See also

Composite attributes are now returned as their object form when queried on a per-attribute basis

Grouping Selected Attributes with Bundles

#2824

Server Side Version Counting

The versioning feature of the ORM (now also documented at Configuring a Version Counter) can now make use of server-side version counting schemes, such as those produced by triggers or database system columns, as well as conditional programmatic schemes outside of the version_id_counter function itself. By providing the value False to the version_id_generator parameter, the ORM will use the already-set version identifier, or alternatively fetch the version identifier from each row at the same time the INSERT or UPDATE is emitted. When using a server-generated version identifier, it is strongly recommended that this feature be used only on a backend with strong RETURNING support (PostgreSQL, SQL Server; Oracle also supports RETURNING but the cx_oracle driver has only limited support), else the additional SELECT statements will add significant performance overhead. The example provided at Server Side Version Counters illustrates the usage of the PostgreSQL xmin system column in order to integrate it with the ORM’s versioning feature.

See also

Server Side Version Counters

#2793

include_backrefs=False option for @validates

The validates() function now accepts an option include_backrefs=True, which will bypass firing the validator for the case where the event initiated from a backref:

  1. from sqlalchemy import Column, Integer, ForeignKey
  2. from sqlalchemy.orm import relationship, validates
  3. from sqlalchemy.ext.declarative import declarative_base
  4. Base = declarative_base()
  5. class A(Base):
  6. __tablename__ = "a"
  7. id = Column(Integer, primary_key=True)
  8. bs = relationship("B", backref="a")
  9. @validates("bs")
  10. def validate_bs(self, key, item):
  11. print("A.bs validator")
  12. return item
  13. class B(Base):
  14. __tablename__ = "b"
  15. id = Column(Integer, primary_key=True)
  16. a_id = Column(Integer, ForeignKey("a.id"))
  17. @validates("a", include_backrefs=False)
  18. def validate_a(self, key, item):
  19. print("B.a validator")
  20. return item
  21. a1 = A()
  22. a1.bs.append(B()) # prints only "A.bs validator"

#1535

PostgreSQL JSON Type

The PostgreSQL dialect now features a JSON type to complement the HSTORE type.

See also

JSON

#2581

Automap Extension

A new extension is added in 0.9.1 known as sqlalchemy.ext.automap. This is an experimental extension which expands upon the functionality of Declarative as well as the DeferredReflection class. Essentially, the extension provides a base class AutomapBase which automatically generates mapped classes and relationships between them based on given table metadata.

The MetaData in use normally might be produced via reflection, but there is no requirement that reflection is used. The most basic usage illustrates how sqlalchemy.ext.automap is able to deliver mapped classes, including relationships, based on a reflected schema:

  1. from sqlalchemy.ext.automap import automap_base
  2. from sqlalchemy.orm import Session
  3. from sqlalchemy import create_engine
  4. Base = automap_base()
  5. # engine, suppose it has two tables 'user' and 'address' set up
  6. engine = create_engine("sqlite:///mydatabase.db")
  7. # reflect the tables
  8. Base.prepare(engine, reflect=True)
  9. # mapped classes are now created with names matching that of the table
  10. # name.
  11. User = Base.classes.user
  12. Address = Base.classes.address
  13. session = Session(engine)
  14. # rudimentary relationships are produced
  15. session.add(Address(email_address="foo@bar.com", user=User(name="foo")))
  16. session.commit()
  17. # collection-based relationships are by default named "<classname>_collection"
  18. print(u1.address_collection)

Beyond that, the AutomapBase class is a declarative base, and supports all the features that declarative does. The “automapping” feature can be used with an existing, explicitly declared schema to generate relationships and missing classes only. Naming schemes and relationship-production routines can be dropped in using callable functions.

It is hoped that the AutomapBase system provides a quick and modernized solution to the problem that the very famous SQLSoup also tries to solve, that of generating a quick and rudimentary object model from an existing database on the fly. By addressing the issue strictly at the mapper configuration level, and integrating fully with existing Declarative class techniques, AutomapBase seeks to provide a well-integrated approach to the issue of expediently auto-generating ad-hoc mappings.

See also

Automap

Behavioral Improvements

Improvements that should produce no compatibility issues except in exceedingly rare and unusual hypothetical cases, but are good to be aware of in case there are unexpected issues.

Many JOIN and LEFT OUTER JOIN expressions will no longer be wrapped in (SELECT * FROM ..) AS ANON_1

For many years, the SQLAlchemy ORM has been held back from being able to nest a JOIN inside the right side of an existing JOIN (typically a LEFT OUTER JOIN, as INNER JOINs could always be flattened):

  1. SELECT a.*, b.*, c.* FROM a LEFT OUTER JOIN (b JOIN c ON b.id = c.id) ON a.id

This was due to the fact that SQLite up until version 3.7.16 cannot parse a statement of the above format:

  1. SQLite version 3.7.15.2 2013-01-09 11:53:05
  2. Enter ".help" for instructions
  3. Enter SQL statements terminated with a ";"
  4. sqlite> create table a(id integer);
  5. sqlite> create table b(id integer);
  6. sqlite> create table c(id integer);
  7. sqlite> select a.id, b.id, c.id from a left outer join (b join c on b.id=c.id) on b.id=a.id;
  8. Error: no such column: b.id

Right-outer-joins are of course another way to work around right-side parenthesization; this would be significantly complicated and visually unpleasant to implement, but fortunately SQLite doesn’t support RIGHT OUTER JOIN either :):

  1. sqlite> select a.id, b.id, c.id from b join c on b.id=c.id
  2. ...> right outer join a on b.id=a.id;
  3. Error: RIGHT and FULL OUTER JOINs are not currently supported

Back in 2005, it wasn’t clear if other databases had trouble with this form, but today it seems clear every database tested except SQLite now supports it (Oracle 8, a very old database, doesn’t support the JOIN keyword at all, but SQLAlchemy has always had a simple rewriting scheme in place for Oracle’s syntax). To make matters worse, SQLAlchemy’s usual workaround of applying a SELECT often degrades performance on platforms like PostgreSQL and MySQL:

  1. SELECT a.*, anon_1.* FROM a LEFT OUTER JOIN (
  2. SELECT b.id AS b_id, c.id AS c_id
  3. FROM b JOIN c ON b.id = c.id
  4. ) AS anon_1 ON a.id=anon_1.b_id

A JOIN like the above form is commonplace when working with joined-table inheritance structures; any time Query.join() is used to join from some parent to a joined-table subclass, or when joinedload() is used similarly, SQLAlchemy’s ORM would always make sure a nested JOIN was never rendered, lest the query wouldn’t be able to run on SQLite. Even though the Core has always supported a JOIN of the more compact form, the ORM had to avoid it.

An additional issue would arise when producing joins across many-to-many relationships where special criteria is present in the ON clause. Consider an eager load join like the following:

  1. session.query(Order).outerjoin(Order.items)

Assuming a many-to-many from Order to Item which actually refers to a subclass like Subitem, the SQL for the above would look like:

  1. SELECT order.id, order.name
  2. FROM order LEFT OUTER JOIN order_item ON order.id = order_item.order_id
  3. LEFT OUTER JOIN item ON order_item.item_id = item.id AND item.type = 'subitem'

What’s wrong with the above query? Basically, that it will load many order / order_item rows where the criteria of item.type == 'subitem' is not true.

As of SQLAlchemy 0.9, an entirely new approach has been taken. The ORM no longer worries about nesting JOINs in the right side of an enclosing JOIN, and it now will render these as often as possible while still returning the correct results. When the SQL statement is passed to be compiled, the dialect compiler will rewrite the join to suit the target backend, if that backend is known to not support a right-nested JOIN (which currently is only SQLite - if other backends have this issue please let us know!).

So a regular query(Parent).join(Subclass) will now usually produce a simpler expression:

  1. SELECT parent.id AS parent_id
  2. FROM parent JOIN (
  3. base_table JOIN subclass_table
  4. ON base_table.id = subclass_table.id) ON parent.id = base_table.parent_id

Joined eager loads like query(Parent).options(joinedload(Parent.subclasses)) will alias the individual tables instead of wrapping in an ANON_1:

  1. SELECT parent.*, base_table_1.*, subclass_table_1.* FROM parent
  2. LEFT OUTER JOIN (
  3. base_table AS base_table_1 JOIN subclass_table AS subclass_table_1
  4. ON base_table_1.id = subclass_table_1.id)
  5. ON parent.id = base_table_1.parent_id

Many-to-many joins and eagerloads will right nest the “secondary” and “right” tables:

  1. SELECT order.id, order.name
  2. FROM order LEFT OUTER JOIN
  3. (order_item JOIN item ON order_item.item_id = item.id AND item.type = 'subitem')
  4. ON order_item.order_id = order.id

All of these joins, when rendered with a Select statement that specifically specifies use_labels=True, which is true for all the queries the ORM emits, are candidates for “join rewriting”, which is the process of rewriting all those right-nested joins into nested SELECT statements, while maintaining the identical labeling used by the Select. So SQLite, the one database that won’t support this very common SQL syntax even in 2013, shoulders the extra complexity itself, with the above queries rewritten as:

  1. -- sqlite only!
  2. SELECT parent.id AS parent_id
  3. FROM parent JOIN (
  4. SELECT base_table.id AS base_table_id,
  5. base_table.parent_id AS base_table_parent_id,
  6. subclass_table.id AS subclass_table_id
  7. FROM base_table JOIN subclass_table ON base_table.id = subclass_table.id
  8. ) AS anon_1 ON parent.id = anon_1.base_table_parent_id
  9. -- sqlite only!
  10. SELECT parent.id AS parent_id, anon_1.subclass_table_1_id AS subclass_table_1_id,
  11. anon_1.base_table_1_id AS base_table_1_id,
  12. anon_1.base_table_1_parent_id AS base_table_1_parent_id
  13. FROM parent LEFT OUTER JOIN (
  14. SELECT base_table_1.id AS base_table_1_id,
  15. base_table_1.parent_id AS base_table_1_parent_id,
  16. subclass_table_1.id AS subclass_table_1_id
  17. FROM base_table AS base_table_1
  18. JOIN subclass_table AS subclass_table_1 ON base_table_1.id = subclass_table_1.id
  19. ) AS anon_1 ON parent.id = anon_1.base_table_1_parent_id
  20. -- sqlite only!
  21. SELECT "order".id AS order_id
  22. FROM "order" LEFT OUTER JOIN (
  23. SELECT order_item_1.order_id AS order_item_1_order_id,
  24. order_item_1.item_id AS order_item_1_item_id,
  25. item.id AS item_id, item.type AS item_type
  26. FROM order_item AS order_item_1
  27. JOIN item ON item.id = order_item_1.item_id AND item.type IN (?)
  28. ) AS anon_1 ON "order".id = anon_1.order_item_1_order_id

Note

As of SQLAlchemy 1.1, the workarounds present in this feature for SQLite will automatically disable themselves when SQLite version 3.7.16 or greater is detected, as SQLite has repaired support for right-nested joins.

The Join.alias(), aliased() and with_polymorphic() functions now support a new argument, flat=True, which is used to construct aliases of joined-table entities without embedding into a SELECT. This flag is not on by default, to help with backwards compatibility - but now a “polymorphic” selectable can be joined as a target without any subqueries generated:

  1. employee_alias = with_polymorphic(Person, [Engineer, Manager], flat=True)
  2. session.query(Company).join(Company.employees.of_type(employee_alias)).filter(
  3. or_(Engineer.primary_language == "python", Manager.manager_name == "dilbert")
  4. )

Generates (everywhere except SQLite):

  1. SELECT companies.company_id AS companies_company_id, companies.name AS companies_name
  2. FROM companies JOIN (
  3. people AS people_1
  4. LEFT OUTER JOIN engineers AS engineers_1 ON people_1.person_id = engineers_1.person_id
  5. LEFT OUTER JOIN managers AS managers_1 ON people_1.person_id = managers_1.person_id
  6. ) ON companies.company_id = people_1.company_id
  7. WHERE engineers.primary_language = %(primary_language_1)s
  8. OR managers.manager_name = %(manager_name_1)s

#2369 #2587

Right-nested inner joins available in joined eager loads

As of version 0.9.4, the above mentioned right-nested joining can be enabled in the case of a joined eager load where an “outer” join is linked to an “inner” on the right side.

Normally, a joined eager load chain like the following:

  1. query(User).options(
  2. joinedload("orders", innerjoin=False).joinedload("items", innerjoin=True)
  3. )

Would not produce an inner join; because of the LEFT OUTER JOIN from user->order, joined eager loading could not use an INNER join from order->items without changing the user rows that are returned, and would instead ignore the “chained” innerjoin=True directive. How 0.9.0 should have delivered this would be that instead of:

  1. FROM users LEFT OUTER JOIN orders ON <onclause> LEFT OUTER JOIN items ON <onclause>

the new “right-nested joins are OK” logic would kick in, and we’d get:

  1. FROM users LEFT OUTER JOIN (orders JOIN items ON <onclause>) ON <onclause>

Since we missed the boat on that, to avoid further regressions we’ve added the above functionality by specifying the string "nested" to joinedload.innerjoin:

  1. query(User).options(
  2. joinedload("orders", innerjoin=False).joinedload("items", innerjoin="nested")
  3. )

This feature is new in 0.9.4.

#2976

ORM can efficiently fetch just-generated INSERT/UPDATE defaults using RETURNING

The Mapper has long supported an undocumented flag known as eager_defaults=True. The effect of this flag is that when an INSERT or UPDATE proceeds, and the row is known to have server-generated default values, a SELECT would immediately follow it in order to “eagerly” load those new values. Normally, the server-generated columns are marked as “expired” on the object, so that no overhead is incurred unless the application actually accesses these columns soon after the flush. The eager_defaults flag was therefore not of much use as it could only decrease performance, and was present only to support exotic event schemes where users needed default values to be available immediately within the flush process.

In 0.9, as a result of the version id enhancements, eager_defaults can now emit a RETURNING clause for these values, so on a backend with strong RETURNING support in particular PostgreSQL, the ORM can fetch newly generated default and SQL expression values inline with the INSERT or UPDATE. eager_defaults, when enabled, makes use of RETURNING automatically when the target backend and Table supports “implicit returning”.

Subquery Eager Loading will apply DISTINCT to the innermost SELECT for some queries

In an effort to reduce the number of duplicate rows that can be generated by subquery eager loading when a many-to-one relationship is involved, a DISTINCT keyword will be applied to the innermost SELECT when the join is targeting columns that do not comprise the primary key, as in when loading along a many to one.

That is, when subquery loading on a many-to-one from A->B:

  1. SELECT b.id AS b_id, b.name AS b_name, anon_1.b_id AS a_b_id
  2. FROM (SELECT DISTINCT a_b_id FROM a) AS anon_1
  3. JOIN b ON b.id = anon_1.a_b_id

Since a.b_id is a non-distinct foreign key, DISTINCT is applied so that redundant a.b_id are eliminated. The behavior can be turned on or off unconditionally for a particular relationship() using the flag distinct_target_key, setting the value to True for unconditionally on, False for unconditionally off, and None for the feature to take effect when the target SELECT is against columns that do not comprise a full primary key. In 0.9, None is the default.

The option is also backported to 0.8 where the distinct_target_key option defaults to False.

While the feature here is designed to help performance by eliminating duplicate rows, the DISTINCT keyword in SQL itself can have a negative performance impact. If columns in the SELECT are not indexed, DISTINCT will likely perform an ORDER BY on the rowset which can be expensive. By keeping the feature limited just to foreign keys which are hopefully indexed in any case, it’s expected that the new defaults are reasonable.

The feature also does not eliminate every possible dupe-row scenario; if a many-to-one is present elsewhere in the chain of joins, dupe rows may still be present.

#2836

Backref handlers can now propagate more than one level deep

The mechanism by which attribute events pass along their “initiator”, that is the object associated with the start of the event, has been changed; instead of a AttributeImpl being passed, a new object Event is passed instead; this object refers to the AttributeImpl as well as to an “operation token”, representing if the operation is an append, remove, or replace operation.

The attribute event system no longer looks at this “initiator” object in order to halt a recursive series of attribute events. Instead, the system of preventing endless recursion due to mutually-dependent backref handlers has been moved to the ORM backref event handlers specifically, which now take over the role of ensuring that a chain of mutually-dependent events (such as append to collection A.bs, set many-to-one attribute B.a in response) doesn’t go into an endless recursion stream. The rationale here is that the backref system, given more detail and control over event propagation, can finally allow operations more than one level deep to occur; the typical scenario is when a collection append results in a many-to-one replacement operation, which in turn should cause the item to be removed from a previous collection:

  1. class Parent(Base):
  2. __tablename__ = "parent"
  3. id = Column(Integer, primary_key=True)
  4. children = relationship("Child", backref="parent")
  5. class Child(Base):
  6. __tablename__ = "child"
  7. id = Column(Integer, primary_key=True)
  8. parent_id = Column(ForeignKey("parent.id"))
  9. p1 = Parent()
  10. p2 = Parent()
  11. c1 = Child()
  12. p1.children.append(c1)
  13. assert c1.parent is p1 # backref event establishes c1.parent as p1
  14. p2.children.append(c1)
  15. assert c1.parent is p2 # backref event establishes c1.parent as p2
  16. assert c1 not in p1.children # second backref event removes c1 from p1.children

Above, prior to this change, the c1 object would still have been present in p1.children, even though it is also present in p2.children at the same time; the backref handlers would have stopped at replacing c1.parent with p2 instead of p1. In 0.9, using the more detailed Event object as well as letting the backref handlers make more detailed decisions about these objects, the propagation can continue onto removing c1 from p1.children while maintaining a check against the propagation from going into an endless recursive loop.

End-user code which a. makes use of the AttributeEvents.set(), AttributeEvents.append(), or AttributeEvents.remove() events, and b. initiates further attribute modification operations as a result of these events may need to be modified to prevent recursive loops, as the attribute system no longer stops a chain of events from propagating endlessly in the absence of the backref event handlers. Additionally, code which depends upon the value of the initiator will need to be adjusted to the new API, and furthermore must be ready for the value of initiator to change from its original value within a string of backref-initiated events, as the backref handlers may now swap in a new initiator value for some operations.

#2789

The typing system now handles the task of rendering “literal bind” values

A new method is added to TypeEngine TypeEngine.literal_processor() as well as TypeDecorator.process_literal_param() for TypeDecorator which take on the task of rendering so-called “inline literal parameters” - parameters that normally render as “bound” values, but are instead being rendered inline into the SQL statement due to the compiler configuration. This feature is used when generating DDL for constructs such as CheckConstraint, as well as by Alembic when using constructs such as op.inline_literal(). Previously, a simple “isinstance” check checked for a few basic types, and the “bind processor” was used unconditionally, leading to such issues as strings being encoded into utf-8 prematurely.

Custom types written with TypeDecorator should continue to work in “inline literal” scenarios, as the TypeDecorator.process_literal_param() falls back to TypeDecorator.process_bind_param() by default, as these methods usually handle a data manipulation, not as much how the data is presented to the database. TypeDecorator.process_literal_param() can be specified to specifically produce a string representing how a value should be rendered into an inline DDL statement.

#2838

Schema identifiers now carry along their own quoting information

This change simplifies the Core’s usage of so-called “quote” flags, such as the quote flag passed to Table and Column. The flag is now internalized within the string name itself, which is now represented as an instance of quoted_name, a string subclass. The IdentifierPreparer now relies solely on the quoting preferences reported by the quoted_name object rather than checking for any explicit quote flags in most cases. The issue resolved here includes that various case-sensitive methods such as Engine.has_table() as well as similar methods within dialects now function with explicitly quoted names, without the need to complicate or introduce backwards-incompatible changes to those APIs (many of which are 3rd party) with the details of quoting flags - in particular, a wider range of identifiers now function correctly with the so-called “uppercase” backends like Oracle, Firebird, and DB2 (backends that store and report upon table and column names using all uppercase for case insensitive names).

The quoted_name object is used internally as needed; however if other keywords require fixed quoting preferences, the class is available publicly.

#2812

Improved rendering of Boolean constants, NULL constants, conjunctions

New capabilities have been added to the true() and false() constants, in particular in conjunction with and_() and or_() functions as well as the behavior of the WHERE/HAVING clauses in conjunction with these types, boolean types overall, and the null() constant.

Starting with a table such as this:

  1. from sqlalchemy import Table, Boolean, Integer, Column, MetaData
  2. t1 = Table("t", MetaData(), Column("x", Boolean()), Column("y", Integer))

A select construct will now render the boolean column as a binary expression on backends that don’t feature true/false constant behavior:

  1. >>> from sqlalchemy import select, and_, false, true
  2. >>> from sqlalchemy.dialects import mysql, postgresql
  3. >>> print(select([t1]).where(t1.c.x).compile(dialect=mysql.dialect()))
  4. SELECT t.x, t.y FROM t WHERE t.x = 1

The and_() and or_() constructs will now exhibit quasi “short circuit” behavior, that is truncating a rendered expression, when a true() or false() constant is present:

  1. >>> print(
  2. ... select([t1]).where(and_(t1.c.y > 5, false())).compile(dialect=postgresql.dialect())
  3. ... )
  4. SELECT t.x, t.y FROM t WHERE false

true() can be used as the base to build up an expression:

  1. >>> expr = true()
  2. >>> expr = expr & (t1.c.y > 5)
  3. >>> print(select([t1]).where(expr))
  4. SELECT t.x, t.y FROM t WHERE t.y > :y_1

The boolean constants true() and false() themselves render as 0 = 1 and 1 = 1 for a backend with no boolean constants:

  1. >>> print(select([t1]).where(and_(t1.c.y > 5, false())).compile(dialect=mysql.dialect()))
  2. SELECT t.x, t.y FROM t WHERE 0 = 1

Interpretation of None, while not particularly valid SQL, is at least now consistent:

  1. >>> print(select([t1.c.x]).where(None))
  2. SELECT t.x FROM t WHERE NULL
  3. >>> print(select([t1.c.x]).where(None).where(None))
  4. SELECT t.x FROM t WHERE NULL AND NULL
  5. >>> print(select([t1.c.x]).where(and_(None, None)))
  6. SELECT t.x FROM t WHERE NULL AND NULL

#2804

Label constructs can now render as their name alone in an ORDER BY

For the case where a Label is used in both the columns clause as well as the ORDER BY clause of a SELECT, the label will render as just its name in the ORDER BY clause, assuming the underlying dialect reports support of this feature.

E.g. an example like:

  1. from sqlalchemy.sql import table, column, select, func
  2. t = table("t", column("c1"), column("c2"))
  3. expr = (func.foo(t.c.c1) + t.c.c2).label("expr")
  4. stmt = select([expr]).order_by(expr)
  5. print(stmt)

Prior to 0.9 would render as:

  1. SELECT foo(t.c1) + t.c2 AS expr
  2. FROM t ORDER BY foo(t.c1) + t.c2

And now renders as:

  1. SELECT foo(t.c1) + t.c2 AS expr
  2. FROM t ORDER BY expr

The ORDER BY only renders the label if the label isn’t further embedded into an expression within the ORDER BY, other than a simple ASC or DESC.

The above format works on all databases tested, but might have compatibility issues with older database versions (MySQL 4? Oracle 8? etc.). Based on user reports we can add rules that will disable the feature based on database version detection.

#1068

RowProxy now has tuple-sorting behavior

The RowProxy object acts much like a tuple, but up until now would not sort as a tuple if a list of them were sorted using sorted(). The __eq__() method now compares both sides as a tuple and also an __lt__() method has been added:

  1. users.insert().execute(
  2. dict(user_id=1, user_name="foo"),
  3. dict(user_id=2, user_name="bar"),
  4. dict(user_id=3, user_name="def"),
  5. )
  6. rows = users.select().order_by(users.c.user_name).execute().fetchall()
  7. eq_(rows, [(2, "bar"), (3, "def"), (1, "foo")])
  8. eq_(sorted(rows), [(1, "foo"), (2, "bar"), (3, "def")])

#2848

A bindparam() construct with no type gets upgraded via copy when a type is available

The logic which “upgrades” a bindparam() construct to take on the type of the enclosing expression has been improved in two ways. First, the bindparam() object is copied before the new type is assigned, so that the given bindparam() is not mutated in place. Secondly, this same operation occurs when an Insert or Update construct is compiled, regarding the “values” that were set in the statement via the ValuesBase.values() method.

If given an untyped bindparam():

  1. bp = bindparam("some_col")

If we use this parameter as follows:

  1. expr = mytable.c.col == bp

The type for bp remains as NullType, however if mytable.c.col is of type String, then expr.right, that is the right side of the binary expression, will take on the String type. Previously, bp itself would have been changed in place to have String as its type.

Similarly, this operation occurs in an Insert or Update:

  1. stmt = mytable.update().values(col=bp)

Above, bp remains unchanged, but the String type will be used when the statement is executed, which we can see by examining the binds dictionary:

  1. >>> compiled = stmt.compile()
  2. >>> compiled.binds["some_col"].type
  3. String

The feature allows custom types to take their expected effect within INSERT/UPDATE statements without needing to explicitly specify those types within every bindparam() expression.

The potentially backwards-compatible changes involve two unlikely scenarios. Since the bound parameter is cloned, users should not be relying upon making in-place changes to a bindparam() construct once created. Additionally, code which uses bindparam() within an Insert or Update statement which is relying on the fact that the bindparam() is not typed according to the column being assigned towards will no longer function in that way.

#2850

Columns can reliably get their type from a column referred to via ForeignKey

There’s a long standing behavior which says that a Column can be declared without a type, as long as that Column is referred to by a ForeignKeyConstraint, and the type from the referenced column will be copied into this one. The problem has been that this feature never worked very well and wasn’t maintained. The core issue was that the ForeignKey object doesn’t know what target Column it refers to until it is asked, typically the first time the foreign key is used to construct a Join. So until that time, the parent Column would not have a type, or more specifically, it would have a default type of NullType.

While it’s taken a long time, the work to reorganize the initialization of ForeignKey objects has been completed such that this feature can finally work acceptably. At the core of the change is that the ForeignKey.column attribute no longer lazily initializes the location of the target Column; the issue with this system was that the owning Column would be stuck with NullType as its type until the ForeignKey happened to be used.

In the new version, the ForeignKey coordinates with the eventual Column it will refer to using internal attachment events, so that the moment the referencing Column is associated with the MetaData, all ForeignKey objects that refer to it will be sent a message that they need to initialize their parent column. This system is more complicated but works more solidly; as a bonus, there are now tests in place for a wide variety of Column / ForeignKey configuration scenarios and error messages have been improved to be very specific to no less than seven different error conditions.

Scenarios which now work correctly include:

  1. The type on a Column is immediately present as soon as the target Column becomes associated with the same MetaData; this works no matter which side is configured first:

    1. >>> from sqlalchemy import Table, MetaData, Column, Integer, ForeignKey
    2. >>> metadata = MetaData()
    3. >>> t2 = Table("t2", metadata, Column("t1id", ForeignKey("t1.id")))
    4. >>> t2.c.t1id.type
    5. NullType()
    6. >>> t1 = Table("t1", metadata, Column("id", Integer, primary_key=True))
    7. >>> t2.c.t1id.type
    8. Integer()
  2. The system now works with ForeignKeyConstraint as well:

    1. >>> from sqlalchemy import Table, MetaData, Column, Integer, ForeignKeyConstraint
    2. >>> metadata = MetaData()
    3. >>> t2 = Table(
    4. ... "t2",
    5. ... metadata,
    6. ... Column("t1a"),
    7. ... Column("t1b"),
    8. ... ForeignKeyConstraint(["t1a", "t1b"], ["t1.a", "t1.b"]),
    9. ... )
    10. >>> t2.c.t1a.type
    11. NullType()
    12. >>> t2.c.t1b.type
    13. NullType()
    14. >>> t1 = Table(
    15. ... "t1",
    16. ... metadata,
    17. ... Column("a", Integer, primary_key=True),
    18. ... Column("b", Integer, primary_key=True),
    19. ... )
    20. >>> t2.c.t1a.type
    21. Integer()
    22. >>> t2.c.t1b.type
    23. Integer()
  3. It even works for “multiple hops” - that is, a ForeignKey that refers to a Column that refers to another Column:

    1. >>> from sqlalchemy import Table, MetaData, Column, Integer, ForeignKey
    2. >>> metadata = MetaData()
    3. >>> t2 = Table("t2", metadata, Column("t1id", ForeignKey("t1.id")))
    4. >>> t3 = Table("t3", metadata, Column("t2t1id", ForeignKey("t2.t1id")))
    5. >>> t2.c.t1id.type
    6. NullType()
    7. >>> t3.c.t2t1id.type
    8. NullType()
    9. >>> t1 = Table("t1", metadata, Column("id", Integer, primary_key=True))
    10. >>> t2.c.t1id.type
    11. Integer()
    12. >>> t3.c.t2t1id.type
    13. Integer()

#1765

Dialect Changes

Firebird fdb is now the default Firebird dialect.

The fdb dialect is now used if an engine is created without a dialect specifier, i.e. firebird://. fdb is a kinterbasdb compatible DBAPI which per the Firebird project is now their official Python driver.

#2504

Firebird fdb and kinterbasdb set retaining=False by default

Both the fdb and kinterbasdb DBAPIs support a flag retaining=True which can be passed to the commit() and rollback() methods of its connection. The documented rationale for this flag is so that the DBAPI can re-use internal transaction state for subsequent transactions, for the purposes of improving performance. However, newer documentation refers to analyses of Firebird’s “garbage collection” which expresses that this flag can have a negative effect on the database’s ability to process cleanup tasks, and has been reported as lowering performance as a result.

It’s not clear how this flag is actually usable given this information, and as it appears to be only a performance enhancing feature, it now defaults to False. The value can be controlled by passing the flag retaining=True to the create_engine() call. This is a new flag which is added as of 0.8.2, so applications on 0.8.2 can begin setting this to True or False as desired.

See also

sqlalchemy.dialects.firebird.fdb

sqlalchemy.dialects.firebird.kinterbasdb

https://pythonhosted.org/fdb/usage-guide.html#retaining-transactions - information on the “retaining” flag.

#2763