What’s New in SQLAlchemy 1.0?

About this Document

This document describes changes between SQLAlchemy version 0.9, undergoing maintenance releases as of May, 2014, and SQLAlchemy version 1.0, released in April, 2015.

Document last updated: June 9, 2015

Introduction

This guide introduces what’s new in SQLAlchemy version 1.0, and also documents changes which affect users migrating their applications from the 0.9 series of SQLAlchemy to 1.0.

Please carefully review the sections on behavioral changes for potentially backwards-incompatible changes in behavior.

New Features and Improvements - ORM

New Session Bulk INSERT/UPDATE API

A new series of Session methods which provide hooks directly into the unit of work’s facility for emitting INSERT and UPDATE statements has been created. When used correctly, this expert-oriented system can allow ORM-mappings to be used to generate bulk insert and update statements batched into executemany groups, allowing the statements to proceed at speeds that rival direct use of the Core.

See also

Bulk Operations - introduction and full documentation

#3100

New Performance Example Suite

Inspired by the benchmarking done for the Bulk Operations feature as well as for the How can I profile a SQLAlchemy powered application? section of the FAQ, a new example section has been added which features several scripts designed to illustrate the relative performance profile of various Core and ORM techniques. The scripts are organized into use cases, and are packaged under a single console interface such that any combination of demonstrations can be run, dumping out timings, Python profile results and/or RunSnake profile displays.

See also

Performance

“Baked” Queries

The “baked” query feature is an unusual new approach which allows for straightforward construction an invocation of Query objects using caching, which upon successive calls features vastly reduced Python function call overhead (over 75%). By specifying a Query object as a series of lambdas which are only invoked once, a query as a pre-compiled unit begins to be feasible:

  1. from sqlalchemy.ext import baked
  2. from sqlalchemy import bindparam
  3. bakery = baked.bakery()
  4. def search_for_user(session, username, email=None):
  5. baked_query = bakery(lambda session: session.query(User))
  6. baked_query += lambda q: q.filter(User.name == bindparam('username'))
  7. baked_query += lambda q: q.order_by(User.id)
  8. if email:
  9. baked_query += lambda q: q.filter(User.email == bindparam('email'))
  10. result = baked_query(session).params(username=username, email=email).all()
  11. return result

See also

Baked Queries

#3054

Improvements to declarative mixins, @declared_attr and related features

The declarative system in conjunction with declared_attr has been overhauled to support new capabilities.

A function decorated with declared_attr is now called only after any mixin-based column copies are generated. This means the function can call upon mixin-established columns and will receive a reference to the correct Column object:

  1. class HasFooBar(object):
  2. foobar = Column(Integer)
  3. @declared_attr
  4. def foobar_prop(cls):
  5. return column_property('foobar: ' + cls.foobar)
  6. class SomeClass(HasFooBar, Base):
  7. __tablename__ = 'some_table'
  8. id = Column(Integer, primary_key=True)

Above, SomeClass.foobar_prop will be invoked against SomeClass, and SomeClass.foobar will be the final Column object that is to be mapped to SomeClass, as opposed to the non-copied object present directly on HasFooBar, even though the columns aren’t mapped yet.

The declared_attr function now memoizes the value that’s returned on a per-class basis, so that repeated calls to the same attribute will return the same value. We can alter the example to illustrate this:

  1. class HasFooBar(object):
  2. @declared_attr
  3. def foobar(cls):
  4. return Column(Integer)
  5. @declared_attr
  6. def foobar_prop(cls):
  7. return column_property('foobar: ' + cls.foobar)
  8. class SomeClass(HasFooBar, Base):
  9. __tablename__ = 'some_table'
  10. id = Column(Integer, primary_key=True)

Previously, SomeClass would be mapped with one particular copy of the foobar column, but the foobar_prop by calling upon foobar a second time would produce a different column. The value of SomeClass.foobar is now memoized during declarative setup time, so that even before the attribute is mapped by the mapper, the interim column value will remain consistent no matter how many times the declared_attr is called upon.

The two behaviors above should help considerably with declarative definition of many types of mapper properties that derive from other attributes, where the declared_attr function is called upon from other declared_attr functions locally present before the class is actually mapped.

For a pretty slim edge case where one wishes to build a declarative mixin that establishes distinct columns per subclass, a new modifier declared_attr.cascading is added. With this modifier, the decorated function will be invoked individually for each class in the mapped inheritance hierarchy. While this is already the behavior for special attributes such as __table_args__ and __mapper_args__, for columns and other properties the behavior by default assumes that attribute is affixed to the base class only, and just inherited from subclasses. With declared_attr.cascading, individual behaviors can be applied:

  1. class HasIdMixin(object):
  2. @declared_attr.cascading
  3. def id(cls):
  4. if has_inherited_table(cls):
  5. return Column(ForeignKey('myclass.id'), primary_key=True)
  6. else:
  7. return Column(Integer, primary_key=True)
  8. class MyClass(HasIdMixin, Base):
  9. __tablename__ = 'myclass'
  10. # ...
  11. class MySubClass(MyClass):
  12. ""
  13. # ...

See also

Mixing in Columns in Inheritance Scenarios

Finally, the AbstractConcreteBase class has been reworked so that a relationship or other mapper property can be set up inline on the abstract base:

  1. from sqlalchemy import Column, Integer, ForeignKey
  2. from sqlalchemy.orm import relationship
  3. from sqlalchemy.ext.declarative import (declarative_base, declared_attr,
  4. AbstractConcreteBase)
  5. Base = declarative_base()
  6. class Something(Base):
  7. __tablename__ = u'something'
  8. id = Column(Integer, primary_key=True)
  9. class Abstract(AbstractConcreteBase, Base):
  10. id = Column(Integer, primary_key=True)
  11. @declared_attr
  12. def something_id(cls):
  13. return Column(ForeignKey(Something.id))
  14. @declared_attr
  15. def something(cls):
  16. return relationship(Something)
  17. class Concrete(Abstract):
  18. __tablename__ = u'cca'
  19. __mapper_args__ = {'polymorphic_identity': 'cca', 'concrete': True}

The above mapping will set up a table cca with both an id and a something_id column, and Concrete will also have a relationship something. The new feature is that Abstract will also have an independently configured relationship something that builds against the polymorphic union of the base.

#3150 #2670 #3149 #2952 #3050

ORM full object fetches 25% faster

The mechanics of the loading.py module as well as the identity map have undergone several passes of inlining, refactoring, and pruning, so that a raw load of rows now populates ORM-based objects around 25% faster. Assuming a 1M row table, a script like the following illustrates the type of load that’s improved the most:

  1. import time
  2. from sqlalchemy import Integer, Column, create_engine, Table
  3. from sqlalchemy.orm import Session
  4. from sqlalchemy.ext.declarative import declarative_base
  5. Base = declarative_base()
  6. class Foo(Base):
  7. __table__ = Table(
  8. 'foo', Base.metadata,
  9. Column('id', Integer, primary_key=True),
  10. Column('a', Integer(), nullable=False),
  11. Column('b', Integer(), nullable=False),
  12. Column('c', Integer(), nullable=False),
  13. )
  14. engine = create_engine(
  15. 'mysql+mysqldb://scott:tiger@localhost/test', echo=True)
  16. sess = Session(engine)
  17. now = time.time()
  18. # avoid using all() so that we don't have the overhead of building
  19. # a large list of full objects in memory
  20. for obj in sess.query(Foo).yield_per(100).limit(1000000):
  21. pass
  22. print("Total time: %d" % (time.time() - now))

Local MacBookPro results bench from 19 seconds for 0.9 down to 14 seconds for 1.0. The Query.yield_per() call is always a good idea when batching huge numbers of rows, as it prevents the Python interpreter from having to allocate a huge amount of memory for all objects and their instrumentation at once. Without the Query.yield_per(), the above script on the MacBookPro is 31 seconds on 0.9 and 26 seconds on 1.0, the extra time spent setting up very large memory buffers.

New KeyedTuple implementation dramatically faster

We took a look into the KeyedTuple implementation in the hopes of improving queries like this:

  1. rows = sess.query(Foo.a, Foo.b, Foo.c).all()

The KeyedTuple class is used rather than Python’s collections.namedtuple(), because the latter has a very complex type-creation routine that benchmarks much slower than KeyedTuple. However, when fetching hundreds of thousands of rows, collections.namedtuple() quickly overtakes KeyedTuple which becomes dramatically slower as instance invocation goes up. What to do? A new type that hedges between the approaches of both. Benching all three types for “size” (number of rows returned) and “num” (number of distinct queries), the new “lightweight keyed tuple” either outperforms both, or lags very slightly behind the faster object, based on which scenario. In the “sweet spot”, where we are both creating a good number of new types as well as fetching a good number of rows, the lightweight object totally smokes both namedtuple and KeyedTuple:

  1. -----------------
  2. size=10 num=10000 # few rows, lots of queries
  3. namedtuple: 3.60302400589 # namedtuple falls over
  4. keyedtuple: 0.255059957504 # KeyedTuple very fast
  5. lw keyed tuple: 0.582715034485 # lw keyed trails right on KeyedTuple
  6. -----------------
  7. size=100 num=1000 # <--- sweet spot
  8. namedtuple: 0.365247011185
  9. keyedtuple: 0.24896979332
  10. lw keyed tuple: 0.0889317989349 # lw keyed blows both away!
  11. -----------------
  12. size=10000 num=100
  13. namedtuple: 0.572599887848
  14. keyedtuple: 2.54251694679
  15. lw keyed tuple: 0.613876104355
  16. -----------------
  17. size=1000000 num=10 # few queries, lots of rows
  18. namedtuple: 5.79669594765 # namedtuple very fast
  19. keyedtuple: 28.856498003 # KeyedTuple falls over
  20. lw keyed tuple: 6.74346804619 # lw keyed trails right on namedtuple

#3176

Significant Improvements in Structural Memory Use

Structural memory use has been improved via much more significant use of __slots__ for many internal objects. This optimization is particularly geared towards the base memory size of large applications that have lots of tables and columns, and reduces memory size for a variety of high-volume objects including event listening internals, comparator objects and parts of the ORM attribute and loader strategy system.

A bench that makes use of heapy measure the startup size of Nova illustrates a difference of about 3.7 fewer megs, or 46%, taken up by SQLAlchemy’s objects, associated dictionaries, as well as weakrefs, within a basic import of “nova.db.sqlalchemy.models”:

  1. # reported by heapy, summation of SQLAlchemy objects +
  2. # associated dicts + weakref-related objects with core of Nova imported:
  3. Before: total count 26477 total bytes 7975712
  4. After: total count 18181 total bytes 4236456
  5. # reported for the Python module space overall with the
  6. # core of Nova imported:
  7. Before: Partition of a set of 355558 objects. Total size = 61661760 bytes.
  8. After: Partition of a set of 346034 objects. Total size = 57808016 bytes.

UPDATE statements are now batched with executemany() in a flush

UPDATE statements can now be batched within an ORM flush into more performant executemany() call, similarly to how INSERT statements can be batched; this will be invoked within flush based on the following criteria:

  • two or more UPDATE statements in sequence involve the identical set of columns to be modified.

  • The statement has no embedded SQL expressions in the SET clause.

  • The mapping does not use a mapper.version_id_col, or the backend dialect supports a “sane” rowcount for an executemany() operation; most DBAPIs support this correctly now.

Session.get_bind() handles a wider variety of inheritance scenarios

The Session.get_bind() method is invoked whenever a query or unit of work flush process seeks to locate the database engine that corresponds to a particular class. The method has been improved to handle a variety of inheritance-oriented scenarios, including:

  • Binding to a Mixin or Abstract Class:

    1. class MyClass(SomeMixin, Base):
    2. __tablename__ = 'my_table'
    3. # ...
    4. session = Session(binds={SomeMixin: some_engine})
  • Binding to inherited concrete subclasses individually based on table:

    ``` class BaseClass(Base):

    1. __tablename__ = 'base'
    2. # ...

    class ConcreteSubClass(BaseClass):

    1. __tablename__ = 'concrete'
    2. # ...
    3. __mapper_args__ = {'concrete': True}
  1. session = Session(binds={
  2. base_table: some_engine,
  3. concrete_table: some_other_engine
  4. })
  5. ```

#3035

Session.get_bind() will receive the Mapper in all relevant Query cases

A series of issues were repaired where the Session.get_bind() would not receive the primary Mapper of the Query, even though this mapper was readily available (the primary mapper is the single mapper, or alternatively the first mapper, that is associated with a Query object).

The Mapper object, when passed to Session.get_bind(), is typically used by sessions that make use of the Session.binds parameter to associate mappers with a series of engines (although in this use case, things frequently “worked” in most cases anyway as the bind would be located via the mapped table object), or more specifically implement a user-defined Session.get_bind() method that provies some pattern of selecting engines based on mappers, such as horizontal sharding or a so-called “routing” session that routes queries to different backends.

These scenarios include:

  • Query.count():

    1. session.query(User).count()
  • Query.update() and Query.delete(), both for the UPDATE/DELETE statement as well as for the SELECT used by the “fetch” strategy:

    1. session.query(User).filter(User.id == 15).update(
    2. {"name": "foob"}, synchronize_session='fetch')
    3. session.query(User).filter(User.id == 15).delete(
    4. synchronize_session='fetch')
  • Queries against individual columns:

    1. session.query(User.id, User.name).all()
  • SQL functions and other expressions against indirect mappings such as column_property:

    1. class User(Base):
    2. # ...
    3. score = column_property(func.coalesce(self.tables.users.c.name, None)))
    4. session.query(func.max(User.score)).scalar()

#3227 #3242 #1326

.info dictionary improvements

The InspectionAttr.info collection is now available on every kind of object that one would retrieve from the Mapper.all_orm_descriptors collection. This includes hybrid_property and association_proxy(). However, as these objects are class-bound descriptors, they must be accessed separately from the class to which they are attached in order to get at the attribute. Below this is illustrated using the Mapper.all_orm_descriptors namespace:

  1. class SomeObject(Base):
  2. # ...
  3. @hybrid_property
  4. def some_prop(self):
  5. return self.value + 5
  6. inspect(SomeObject).all_orm_descriptors.some_prop.info['foo'] = 'bar'

It is also available as a constructor argument for all SchemaItem objects (e.g. ForeignKey, UniqueConstraint etc.) as well as remaining ORM constructs such as synonym().

#2971

#2963

ColumnProperty constructs work a lot better with aliases, order_by

A variety of issues regarding column_property() have been fixed, most specifically with regards to the aliased() construct as well as the “order by label” logic introduced in 0.9 (see Label constructs can now render as their name alone in an ORDER BY).

Given a mapping like the following:

  1. class A(Base):
  2. __tablename__ = 'a'
  3. id = Column(Integer, primary_key=True)
  4. class B(Base):
  5. __tablename__ = 'b'
  6. id = Column(Integer, primary_key=True)
  7. a_id = Column(ForeignKey('a.id'))
  8. A.b = column_property(
  9. select([func.max(B.id)]).where(B.a_id == A.id).correlate(A)
  10. )

A simple scenario that included “A.b” twice would fail to render correctly:

  1. print(sess.query(A, a1).order_by(a1.b))

This would order by the wrong column:

  1. SELECT a.id AS a_id, (SELECT max(b.id) AS max_1 FROM b
  2. WHERE b.a_id = a.id) AS anon_1, a_1.id AS a_1_id,
  3. (SELECT max(b.id) AS max_2
  4. FROM b WHERE b.a_id = a_1.id) AS anon_2
  5. FROM a, a AS a_1 ORDER BY anon_1

New output:

  1. SELECT a.id AS a_id, (SELECT max(b.id) AS max_1
  2. FROM b WHERE b.a_id = a.id) AS anon_1, a_1.id AS a_1_id,
  3. (SELECT max(b.id) AS max_2
  4. FROM b WHERE b.a_id = a_1.id) AS anon_2
  5. FROM a, a AS a_1 ORDER BY anon_2

There were also many scenarios where the “order by” logic would fail to order by label, for example if the mapping were “polymorphic”:

  1. class A(Base):
  2. __tablename__ = 'a'
  3. id = Column(Integer, primary_key=True)
  4. type = Column(String)
  5. __mapper_args__ = {'polymorphic_on': type, 'with_polymorphic': '*'}

The order_by would fail to use the label, as it would be anonymized due to the polymorphic loading:

  1. SELECT a.id AS a_id, a.type AS a_type, (SELECT max(b.id) AS max_1
  2. FROM b WHERE b.a_id = a.id) AS anon_1
  3. FROM a ORDER BY (SELECT max(b.id) AS max_2
  4. FROM b WHERE b.a_id = a.id)

Now that the order by label tracks the anonymized label, this now works:

  1. SELECT a.id AS a_id, a.type AS a_type, (SELECT max(b.id) AS max_1
  2. FROM b WHERE b.a_id = a.id) AS anon_1
  3. FROM a ORDER BY anon_1

Included in these fixes are a variety of heisenbugs that could corrupt the state of an aliased() construct such that the labeling logic would again fail; these have also been fixed.

#3148 #3188

New Features and Improvements - Core

Select/Query LIMIT / OFFSET may be specified as an arbitrary SQL expression

The Select.limit() and Select.offset() methods now accept any SQL expression, in addition to integer values, as arguments. The ORM Query object also passes through any expression to the underlying Select object. Typically this is used to allow a bound parameter to be passed, which can be substituted with a value later:

  1. sel = select([table]).limit(bindparam('mylimit')).offset(bindparam('myoffset'))

Dialects which don’t support non-integer LIMIT or OFFSET expressions may continue to not support this behavior; third party dialects may also need modification in order to take advantage of the new behavior. A dialect which currently uses the ._limit or ._offset attributes will continue to function for those cases where the limit/offset was specified as a simple integer value. However, when a SQL expression is specified, these two attributes will instead raise a CompileError on access. A third-party dialect which wishes to support the new feature should now call upon the ._limit_clause and ._offset_clause attributes to receive the full SQL expression, rather than the integer value.

The use_alter flag on ForeignKeyConstraint is (usually) no longer needed

The MetaData.create_all() and MetaData.drop_all() methods will now make use of a system that automatically renders an ALTER statement for foreign key constraints that are involved in mutually-dependent cycles between tables, without the need to specify ForeignKeyConstraint.use_alter. Additionally, the foreign key constraints no longer need to have a name in order to be created via ALTER; only the DROP operation requires a name. In the case of a DROP, the feature will ensure that only constraints which have explicit names are actually included as ALTER statements. In the case of an unresolvable cycle within a DROP, the system emits a succinct and clear error message now if the DROP cannot proceed.

The ForeignKeyConstraint.use_alter and ForeignKey.use_alter flags remain in place, and continue to have the same effect of establishing those constraints for which ALTER is required during a CREATE/DROP scenario.

As of version 1.0.1, special logic takes over in the case of SQLite, which does not support ALTER, in the case that during a DROP, the given tables have an unresolvable cycle; in this case a warning is emitted, and the tables are dropped with no ordering, which is usually fine on SQLite unless constraints are enabled. To resolve the warning and proceed with at least a partial ordering on a SQLite database, particularly one where constraints are enabled, re-apply “use_alter” flags to those ForeignKey and ForeignKeyConstraint objects which should be explicitly omitted from the sort.

See also

Creating/Dropping Foreign Key Constraints via ALTER - full description of the new behavior.

#3282

ResultProxy “auto close” is now a “soft” close

For many releases, the ResultProxy object has always been automatically closed out at the point at which all result rows have been fetched. This was to allow usage of the object without the need to call upon ResultProxy.close() explicitly; as all DBAPI resources had been freed, the object was safe to discard. However, the object maintained a strict “closed” behavior, which meant that any subsequent calls to ResultProxy.fetchone(), ResultProxy.fetchmany() or ResultProxy.fetchall() would now raise a ResourceClosedError:

  1. >>> result = connection.execute(stmt)
  2. >>> result.fetchone()
  3. (1, 'x')
  4. >>> result.fetchone()
  5. None # indicates no more rows
  6. >>> result.fetchone()
  7. exception: ResourceClosedError

This behavior is inconsistent vs. what pep-249 states, which is that you can call upon the fetch methods repeatedly even after results are exhausted. It also interferes with behavior for some implementations of result proxy, such as the BufferedColumnResultProxy used by the cx_oracle dialect for certain datatypes.

To solve this, the “closed” state of the ResultProxy has been broken into two states; a “soft close” which does the majority of what “close” does, in that it releases the DBAPI cursor and in the case of a “close with result” object will also release the connection, and a “closed” state which is everything included by “soft close” as well as establishing the fetch methods as “closed”. The ResultProxy.close() method is now never called implicitly, only the ResultProxy._soft_close() method which is non-public:

  1. >>> result = connection.execute(stmt)
  2. >>> result.fetchone()
  3. (1, 'x')
  4. >>> result.fetchone()
  5. None # indicates no more rows
  6. >>> result.fetchone()
  7. None # still None
  8. >>> result.fetchall()
  9. []
  10. >>> result.close()
  11. >>> result.fetchone()
  12. exception: ResourceClosedError # *now* it raises

#3330 #3329

CHECK Constraints now support the %(column_0_name)s token in naming conventions

The %(column_0_name)s will derive from the first column found in the expression of a CheckConstraint:

  1. metadata = MetaData(
  2. naming_convention={"ck": "ck_%(table_name)s_%(column_0_name)s"}
  3. )
  4. foo = Table('foo', metadata,
  5. Column('value', Integer),
  6. )
  7. CheckConstraint(foo.c.value > 5)

Will render:

  1. CREATE TABLE foo (
  2. value INTEGER,
  3. CONSTRAINT ck_foo_value CHECK (value > 5)
  4. )

The combination of naming conventions with the constraint produced by a SchemaType such as Boolean or Enum will also now make use of all CHECK constraint conventions.

See also

Naming CHECK Constraints

Configuring Naming for Boolean, Enum, and other schema types

#3299

Constraints referring to unattached Columns can auto-attach to the Table when their referred columns are attached

Since at least version 0.8, a Constraint has had the ability to “auto-attach” itself to a Table based on being passed table-attached columns:

  1. from sqlalchemy import Table, Column, MetaData, Integer, UniqueConstraint
  2. m = MetaData()
  3. t = Table('t', m,
  4. Column('a', Integer),
  5. Column('b', Integer)
  6. )
  7. uq = UniqueConstraint(t.c.a, t.c.b) # will auto-attach to Table
  8. assert uq in t.constraints

In order to assist with some cases that tend to come up with declarative, this same auto-attachment logic can now function even if the Column objects are not yet associated with the Table; additional events are established such that when those Column objects are associated, the Constraint is also added:

  1. from sqlalchemy import Table, Column, MetaData, Integer, UniqueConstraint
  2. m = MetaData()
  3. a = Column('a', Integer)
  4. b = Column('b', Integer)
  5. uq = UniqueConstraint(a, b)
  6. t = Table('t', m, a, b)
  7. assert uq in t.constraints # constraint auto-attached

The above feature was a late add as of version 1.0.0b3. A fix as of version 1.0.4 for #3411 ensures that this logic does not occur if the Constraint refers to a mixture of Column objects and string column names; as we do not yet have tracking for the addition of names to a Table:

  1. from sqlalchemy import Table, Column, MetaData, Integer, UniqueConstraint
  2. m = MetaData()
  3. a = Column('a', Integer)
  4. b = Column('b', Integer)
  5. uq = UniqueConstraint(a, 'b')
  6. t = Table('t', m, a, b)
  7. # constraint *not* auto-attached, as we do not have tracking
  8. # to locate when a name 'b' becomes available on the table
  9. assert uq not in t.constraints

Above, the attachment event for column “a” to table “t” will fire off before column “b” is attached (as “a” is stated in the Table constructor before “b”), and the constraint will fail to locate “b” if it were to attempt an attachment. For consistency, if the constraint refers to any string names, the autoattach-on-column-attach logic is skipped.

The original auto-attach logic of course remains in place, if the Table already contains all the target Column objects at the time the Constraint is constructed:

  1. from sqlalchemy import Table, Column, MetaData, Integer, UniqueConstraint
  2. m = MetaData()
  3. a = Column('a', Integer)
  4. b = Column('b', Integer)
  5. t = Table('t', m, a, b)
  6. uq = UniqueConstraint(a, 'b')
  7. # constraint auto-attached normally as in older versions
  8. assert uq in t.constraints

#3341 #3411

INSERT FROM SELECT now includes Python and SQL-expression defaults

Insert.from_select() now includes Python and SQL-expression defaults if otherwise unspecified; the limitation where non-server column defaults aren’t included in an INSERT FROM SELECT is now lifted and these expressions are rendered as constants into the SELECT statement:

  1. from sqlalchemy import Table, Column, MetaData, Integer, select, func
  2. m = MetaData()
  3. t = Table(
  4. 't', m,
  5. Column('x', Integer),
  6. Column('y', Integer, default=func.somefunction()))
  7. stmt = select([t.c.x])
  8. print(t.insert().from_select(['x'], stmt))

Will render:

  1. INSERT INTO t (x, y) SELECT t.x, somefunction() AS somefunction_1
  2. FROM t

The feature can be disabled using Insert.from_select.include_defaults.

Column server defaults now render literal values

The “literal binds” compiler flag is switched on when a DefaultClause, set up by Column.server_default is present as a SQL expression to be compiled. This allows literals embedded in SQL to render correctly, such as:

  1. from sqlalchemy import Table, Column, MetaData, Text
  2. from sqlalchemy.schema import CreateTable
  3. from sqlalchemy.dialects.postgresql import ARRAY, array
  4. from sqlalchemy.dialects import postgresql
  5. metadata = MetaData()
  6. tbl = Table("derp", metadata,
  7. Column("arr", ARRAY(Text),
  8. server_default=array(["foo", "bar", "baz"])),
  9. )
  10. print(CreateTable(tbl).compile(dialect=postgresql.dialect()))

Now renders:

  1. CREATE TABLE derp (
  2. arr TEXT[] DEFAULT ARRAY['foo', 'bar', 'baz']
  3. )

Previously, the literal values "foo", "bar", "baz" would render as bound parameters, which are useless in DDL.

#3087

UniqueConstraint is now part of the Table reflection process

A Table object populated using autoload=True will now include UniqueConstraint constructs as well as Index constructs. This logic has a few caveats for PostgreSQL and MySQL:

PostgreSQL

PostgreSQL has the behavior such that when a UNIQUE constraint is created, it implicitly creates a UNIQUE INDEX corresponding to that constraint as well. The Inspector.get_indexes() and the Inspector.get_unique_constraints() methods will continue to both return these entries distinctly, where Inspector.get_indexes() now features a token duplicates_constraint within the index entry indicating the corresponding constraint when detected. However, when performing full table reflection using Table(..., autoload=True), the Index construct is detected as being linked to the UniqueConstraint, and is not present within the Table.indexes collection; only the UniqueConstraint will be present in the Table.constraints collection. This deduplication logic works by joining to the pg_constraint table when querying pg_index to see if the two constructs are linked.

MySQL

MySQL does not have separate concepts for a UNIQUE INDEX and a UNIQUE constraint. While it supports both syntaxes when creating tables and indexes, it does not store them any differently. The Inspector.get_indexes() and the Inspector.get_unique_constraints() methods will continue to both return an entry for a UNIQUE index in MySQL, where Inspector.get_unique_constraints() features a new token duplicates_index within the constraint entry indicating that this is a dupe entry corresponding to that index. However, when performing full table reflection using Table(..., autoload=True), the UniqueConstraint construct is not part of the fully reflected Table construct under any circumstances; this construct is always represented by a Index with the unique=True setting present in the Table.indexes collection.

See also

PostgreSQL Index Reflection

MySQL / MariaDB Unique Constraints and Reflection

#3184

New systems to safely emit parameterized warnings

For a long time, there has been a restriction that warning messages could not refer to data elements, such that a particular function might emit an infinite number of unique warnings. The key place this occurs is in the Unicode type received non-unicode bind param value warning. Placing the data value in this message would mean that the Python __warningregistry__ for that module, or in some cases the Python-global warnings.onceregistry, would grow unbounded, as in most warning scenarios, one of these two collections is populated with every distinct warning message.

The change here is that by using a special string type that purposely changes how the string is hashed, we can control that a large number of parameterized messages are hashed only on a small set of possible hash values, such that a warning such as Unicode type received non-unicode bind param value can be tailored to be emitted only a specific number of times; beyond that, the Python warnings registry will begin recording them as duplicates.

To illustrate, the following test script will show only ten warnings being emitted for ten of the parameter sets, out of a total of 1000:

  1. from sqlalchemy import create_engine, Unicode, select, cast
  2. import random
  3. import warnings
  4. e = create_engine("sqlite://")
  5. # Use the "once" filter (which is also the default for Python
  6. # warnings). Exactly ten of these warnings will
  7. # be emitted; beyond that, the Python warnings registry will accumulate
  8. # new values as dupes of one of the ten existing.
  9. warnings.filterwarnings("once")
  10. for i in range(1000):
  11. e.execute(select([cast(
  12. ('foo_%d' % random.randint(0, 1000000)).encode('ascii'), Unicode)]))

The format of the warning here is:

  1. /path/lib/sqlalchemy/sql/sqltypes.py:186: SAWarning: Unicode type received
  2. non-unicode bind param value 'foo_4852'. (this warning may be
  3. suppressed after 10 occurrences)

#3178

Key Behavioral Changes - ORM

query.update() now resolves string names into mapped attribute names

The documentation for Query.update() states that the given values dictionary is “a dictionary with attributes names as keys”, implying that these are mapped attribute names. Unfortunately, the function was designed more in mind to receive attributes and SQL expressions and not as much strings; when strings were passed, these strings would be passed through straight to the core update statement without any resolution as far as how these names are represented on the mapped class, meaning the name would have to match that of a table column exactly, not how an attribute of that name was mapped onto the class.

The string names are now resolved as attribute names in earnest:

  1. class User(Base):
  2. __tablename__ = 'user'
  3. id = Column(Integer, primary_key=True)
  4. name = Column('user_name', String(50))

Above, the column user_name is mapped as name. Previously, a call to Query.update() that was passed strings would have to have been called as follows:

  1. session.query(User).update({'user_name': 'moonbeam'})

The given string is now resolved against the entity:

  1. session.query(User).update({'name': 'moonbeam'})

It is typically preferable to use the attribute directly, to avoid any ambiguity:

  1. session.query(User).update({User.name: 'moonbeam'})

The change also indicates that synonyms and hybrid attributes can be referred to by string name as well:

  1. class User(Base):
  2. __tablename__ = 'user'
  3. id = Column(Integer, primary_key=True)
  4. name = Column('user_name', String(50))
  5. @hybrid_property
  6. def fullname(self):
  7. return self.name
  8. session.query(User).update({'fullname': 'moonbeam'})

#3228

Warnings emitted when comparing objects with None values to relationships

This change is new as of 1.0.1. Some users are performing queries that are essentially of this form:

  1. session.query(Address).filter(Address.user == User(id=None))

This pattern is not currently supported in SQLAlchemy. For all versions, it emits SQL resembling:

  1. SELECT address.id AS address_id, address.user_id AS address_user_id,
  2. address.email_address AS address_email_address
  3. FROM address WHERE ? = address.user_id
  4. (None,)

Note above, there is a comparison WHERE ? = address.user_id where the bound value ? is receiving None, or NULL in SQL. This will always return False in SQL. The comparison here would in theory generate SQL as follows:

  1. SELECT address.id AS address_id, address.user_id AS address_user_id,
  2. address.email_address AS address_email_address
  3. FROM address WHERE address.user_id IS NULL

But right now, it does not. Applications which are relying upon the fact that “NULL = NULL” produces False in all cases run the risk that someday, SQLAlchemy might fix this issue to generate “IS NULL”, and the queries will then produce different results. Therefore with this kind of operation, you will see a warning:

  1. SAWarning: Got None for value of column user.id; this is unsupported
  2. for a relationship comparison and will not currently produce an
  3. IS comparison (but may in a future release)

Note that this pattern was broken in most cases for release 1.0.0 including all of the betas; a value like SYMBOL('NEVER_SET') would be generated. This issue has been fixed, but as a result of identifying this pattern, the warning is now there so that we can more safely repair this broken behavior (now captured in #3373) in a future release.

#3371

A “negated contains or equals” relationship comparison will use the current value of attributes, not the database value

This change is new as of 1.0.1; while we would have preferred for this to be in 1.0.0, it only became apparent as a result of #3371.

Given a mapping:

  1. class A(Base):
  2. __tablename__ = 'a'
  3. id = Column(Integer, primary_key=True)
  4. class B(Base):
  5. __tablename__ = 'b'
  6. id = Column(Integer, primary_key=True)
  7. a_id = Column(ForeignKey('a.id'))
  8. a = relationship("A")

Given A, with primary key of 7, but which we changed to be 10 without flushing:

  1. s = Session(autoflush=False)
  2. a1 = A(id=7)
  3. s.add(a1)
  4. s.commit()
  5. a1.id = 10

A query against a many-to-one relationship with this object as the target will use the value 10 in the bound parameters:

  1. s.query(B).filter(B.a == a1)

Produces:

  1. SELECT b.id AS b_id, b.a_id AS b_a_id
  2. FROM b
  3. WHERE ? = b.a_id
  4. (10,)

However, before this change, the negation of this criteria would not use 10, it would use 7, unless the object were flushed first:

  1. s.query(B).filter(B.a != a1)

Produces (in 0.9 and all versions prior to 1.0.1):

  1. SELECT b.id AS b_id, b.a_id AS b_a_id
  2. FROM b
  3. WHERE b.a_id != ? OR b.a_id IS NULL
  4. (7,)

For a transient object, it would produce a broken query:

  1. SELECT b.id, b.a_id
  2. FROM b
  3. WHERE b.a_id != :a_id_1 OR b.a_id IS NULL
  4. {u'a_id_1': symbol('NEVER_SET')}

This inconsistency has been repaired, and in all queries the current attribute value, in this example 10, will now be used.

#3374

Changes to attribute events and other operations regarding attributes that have no pre-existing value

In this change, the default return value of None when accessing an object is now returned dynamically on each access, rather than implicitly setting the attribute’s state with a special “set” operation when it is first accessed. The visible result of this change is that obj.__dict__ is not implicitly modified on get, and there are also some minor behavioral changes for get_history() and related functions.

Given an object with no state:

  1. >>> obj = Foo()

It has always been SQLAlchemy’s behavior such that if we access a scalar or many-to-one attribute that was never set, it is returned as None:

  1. >>> obj.someattr
  2. None

This value of None is in fact now part of the state of obj, and is not unlike as though we had set the attribute explicitly, e.g. obj.someattr = None. However, the “set on get” here would behave differently as far as history and events. It would not emit any attribute event, and additionally if we view history, we see this:

  1. >>> inspect(obj).attrs.someattr.history
  2. History(added=(), unchanged=[None], deleted=()) # 0.9 and below

That is, it’s as though the attribute were always None and were never changed. This is explicitly different from if we had set the attribute first instead:

  1. >>> obj = Foo()
  2. >>> obj.someattr = None
  3. >>> inspect(obj).attrs.someattr.history
  4. History(added=[None], unchanged=(), deleted=()) # all versions

The above means that the behavior of our “set” operation can be corrupted by the fact that the value was accessed via “get” earlier. In 1.0, this inconsistency has been resolved, by no longer actually setting anything when the default “getter” is used.

  1. >>> obj = Foo()
  2. >>> obj.someattr
  3. None
  4. >>> inspect(obj).attrs.someattr.history
  5. History(added=(), unchanged=(), deleted=()) # 1.0
  6. >>> obj.someattr = None
  7. >>> inspect(obj).attrs.someattr.history
  8. History(added=[None], unchanged=(), deleted=())

The reason the above behavior hasn’t had much impact is because the INSERT statement in relational databases considers a missing value to be the same as NULL in most cases. Whether SQLAlchemy received a history event for a particular attribute set to None or not would usually not matter; as the difference between sending None/NULL or not wouldn’t have an impact. However, as #3060 (described here in Priority of attribute changes on relationship-bound attributes vs. FK-bound may appear to change) illustrates, there are some seldom edge cases where we do in fact want to positively have None set. Also, allowing the attribute event here means it’s now possible to create “default value” functions for ORM mapped attributes.

As part of this change, the generation of the implicit “None” is now disabled for other situations where this used to occur; this includes when an attribute set operation on a many-to-one is received; previously, the “old” value would be “None” if it had been not set otherwise; it now will send the value NEVER_SET, which is a value that may be sent to an attribute listener now. This symbol may also be received when calling on mapper utility functions such as Mapper.primary_key_from_instance(); if the primary key attributes have no setting at all, whereas the value would be None before, it will now be the NEVER_SET symbol, and no change to the object’s state occurs.

#3061

Priority of attribute changes on relationship-bound attributes vs. FK-bound may appear to change

As a side effect of #3060, setting a relationship-bound attribute to None is now a tracked history event which refers to the intention of persisting None to that attribute. As it has always been the case that setting a relationship-bound attribute will trump direct assignment to the foreign key attributes, a change in behavior can be seen here when assigning None. Given a mapping:

  1. class A(Base):
  2. __tablename__ = 'table_a'
  3. id = Column(Integer, primary_key=True)
  4. class B(Base):
  5. __tablename__ = 'table_b'
  6. id = Column(Integer, primary_key=True)
  7. a_id = Column(ForeignKey('table_a.id'))
  8. a = relationship(A)

In 1.0, the relationship-bound attribute takes precedence over the FK-bound attribute in all cases, whether or not the value we assign is a reference to an A object or is None. In 0.9, the behavior is inconsistent and only takes effect if a value is assigned; the None is not considered:

  1. a1 = A(id=1)
  2. a2 = A(id=2)
  3. session.add_all([a1, a2])
  4. session.flush()
  5. b1 = B()
  6. b1.a = a1 # we expect a_id to be '1'; takes precedence in 0.9 and 1.0
  7. b2 = B()
  8. b2.a = None # we expect a_id to be None; takes precedence only in 1.0
  9. b1.a_id = 2
  10. b2.a_id = 2
  11. session.add_all([b1, b2])
  12. session.commit()
  13. assert b1.a is a1 # passes in both 0.9 and 1.0
  14. assert b2.a is None # passes in 1.0, in 0.9 it's a2

#3060

session.expunge() will fully detach an object that’s been deleted

The behavior of Session.expunge() had a bug that caused an inconsistency in behavior regarding deleted objects. The object_session() function as well as the InstanceState.session attribute would still report object as belonging to the Session subsequent to the expunge:

  1. u1 = sess.query(User).first()
  2. sess.delete(u1)
  3. sess.flush()
  4. assert u1 not in sess
  5. assert inspect(u1).session is sess # this is normal before commit
  6. sess.expunge(u1)
  7. assert u1 not in sess
  8. assert inspect(u1).session is None # would fail

Note that it is normal for u1 not in sess to be True while inspect(u1).session still refers to the session, while the transaction is ongoing subsequent to the delete operation and Session.expunge() has not been called; the full detachment normally completes once the transaction is committed. This issue would also impact functions that rely on Session.expunge() such as make_transient().

#3139

Joined/Subquery eager loading explicitly disallowed with yield_per

In order to make the Query.yield_per() method easier to use, an exception is raised if any subquery eager loaders, or joined eager loaders that would use collections, are to take effect when yield_per is used, as these are currently not compatible with yield-per (subquery loading could be in theory, however). When this error is raised, the lazyload() option can be sent with an asterisk:

  1. q = sess.query(Object).options(lazyload('*')).yield_per(100)

or use Query.enable_eagerloads():

  1. q = sess.query(Object).enable_eagerloads(False).yield_per(100)

The lazyload() option has the advantage that additional many-to-one joined loader options can still be used:

  1. q = sess.query(Object).options(
  2. lazyload('*'), joinedload("some_manytoone")).yield_per(100)

Changes and fixes in handling of duplicate join targets

Changes here encompass bugs where an unexpected and inconsistent behavior would occur in some scenarios when joining to an entity twice, or to multiple single-table entities against the same table, without using a relationship-based ON clause, as well as when joining multiple times to the same target relationship.

Starting with a mapping as:

  1. from sqlalchemy import Integer, Column, String, ForeignKey
  2. from sqlalchemy.orm import Session, relationship
  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")
  9. class B(Base):
  10. __tablename__ = 'b'
  11. id = Column(Integer, primary_key=True)
  12. a_id = Column(ForeignKey('a.id'))

A query that joins to A.bs twice:

  1. print(s.query(A).join(A.bs).join(A.bs))

Will render:

  1. SELECT a.id AS a_id
  2. FROM a JOIN b ON a.id = b.a_id

The query deduplicates the redundant A.bs because it is attempting to support a case like the following:

  1. s.query(A).join(A.bs).\
  2. filter(B.foo == 'bar').\
  3. reset_joinpoint().join(A.bs, B.cs).filter(C.bar == 'bat')

That is, the A.bs is part of a “path”. As part of #3367, arriving at the same endpoint twice without it being part of a larger path will now emit a warning:

  1. SAWarning: Pathed join target A.bs has already been joined to; skipping

The bigger change involves when joining to an entity without using a relationship-bound path. If we join to B twice:

  1. print(s.query(A).join(B, B.a_id == A.id).join(B, B.a_id == A.id))

In 0.9, this would render as follows:

  1. SELECT a.id AS a_id
  2. FROM a JOIN b ON b.a_id = a.id JOIN b AS b_1 ON b_1.a_id = a.id

This is problematic since the aliasing is implicit and in the case of different ON clauses can lead to unpredictable results.

In 1.0, no automatic aliasing is applied and we get:

  1. SELECT a.id AS a_id
  2. FROM a JOIN b ON b.a_id = a.id JOIN b ON b.a_id = a.id

This will raise an error from the database. While it might be nice if the “duplicate join target” acted identically if we joined both from redundant relationships vs. redundant non-relationship based targets, for now we are only changing the behavior in the more serious case where implicit aliasing would have occurred previously, and only emitting a warning in the relationship case. Ultimately, joining to the same thing twice without any aliasing to disambiguate should raise an error in all cases.

The change also has an impact on single-table inheritance targets. Using a mapping as follows:

  1. from sqlalchemy import Integer, Column, String, ForeignKey
  2. from sqlalchemy.orm import Session, relationship
  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. type = Column(String)
  9. __mapper_args__ = {'polymorphic_on': type, 'polymorphic_identity': 'a'}
  10. class ASub1(A):
  11. __mapper_args__ = {'polymorphic_identity': 'asub1'}
  12. class ASub2(A):
  13. __mapper_args__ = {'polymorphic_identity': 'asub2'}
  14. class B(Base):
  15. __tablename__ = 'b'
  16. id = Column(Integer, primary_key=True)
  17. a_id = Column(Integer, ForeignKey("a.id"))
  18. a = relationship("A", primaryjoin="B.a_id == A.id", backref='b')
  19. s = Session()
  20. print(s.query(ASub1).join(B, ASub1.b).join(ASub2, B.a))
  21. print(s.query(ASub1).join(B, ASub1.b).join(ASub2, ASub2.id == B.a_id))

The two queries at the bottom are equivalent, and should both render the identical SQL:

  1. SELECT a.id AS a_id, a.type AS a_type
  2. FROM a JOIN b ON b.a_id = a.id JOIN a ON b.a_id = a.id AND a.type IN (:type_1)
  3. WHERE a.type IN (:type_2)

The above SQL is invalid, as it renders “a” within the FROM list twice. However, the implicit aliasing bug would occur with the second query only and render this instead:

  1. SELECT a.id AS a_id, a.type AS a_type
  2. FROM a JOIN b ON b.a_id = a.id JOIN a AS a_1
  3. ON a_1.id = b.a_id AND a_1.type IN (:type_1)
  4. WHERE a_1.type IN (:type_2)

Where above, the second join to “a” is aliased. While this seems convenient, it’s not how single-inheritance queries work in general and is misleading and inconsistent.

The net effect is that applications which were relying on this bug will now have an error raised by the database. The solution is to use the expected form. When referring to multiple subclasses of a single-inheritance entity in a query, you must manually use aliases to disambiguate the table, as all the subclasses normally refer to the same table:

  1. asub2_alias = aliased(ASub2)
  2. print(s.query(ASub1).join(B, ASub1.b).join(asub2_alias, B.a.of_type(asub2_alias)))

#3233 #3367

Deferred Columns No Longer Implicitly Undefer

Mapped attributes marked as deferred without explicit undeferral will now remain “deferred” even if their column is otherwise present in the result set in some way. This is a performance enhancement in that an ORM load no longer spends time searching for each deferred column when the result set is obtained. However, for an application that has been relying upon this, an explicit undefer() or similar option should now be used, in order to prevent a SELECT from being emitted when the attribute is accessed.

Deprecated ORM Event Hooks Removed

The following ORM event hooks, some of which have been deprecated since 0.5, have been removed: translate_row, populate_instance, append_result, create_instance. The use cases for these hooks originated in the very early 0.1 / 0.2 series of SQLAlchemy and have long since been unnecessary. In particular, the hooks were largely unusable as the behavioral contracts within these events was strongly linked to the surrounding internals, such as how an instance needs to be created and initialized as well as how columns are located within an ORM-generated row. The removal of these hooks greatly simplifies the mechanics of ORM object loading.

API Change for new Bundle feature when custom row loaders are used

The new Bundle object of 0.9 has a small change in API, when the create_row_processor() method is overridden on a custom class. Previously, the sample code looked like:

  1. from sqlalchemy.orm import Bundle
  2. class DictBundle(Bundle):
  3. def create_row_processor(self, query, procs, labels):
  4. """Override create_row_processor to return values as dictionaries"""
  5. def proc(row, result):
  6. return dict(
  7. zip(labels, (proc(row, result) for proc in procs))
  8. )
  9. return proc

The unused result member is now removed:

  1. from sqlalchemy.orm import Bundle
  2. class DictBundle(Bundle):
  3. def create_row_processor(self, query, procs, labels):
  4. """Override create_row_processor to return values as dictionaries"""
  5. def proc(row):
  6. return dict(
  7. zip(labels, (proc(row) for proc in procs))
  8. )
  9. return proc

See also

Column Bundles

Right inner join nesting now the default for joinedload with innerjoin=True

The behavior of joinedload.innerjoin as well as relationship.innerjoin is now to use “nested” inner joins, that is, right-nested, as the default behavior when an inner join joined eager load is chained to an outer join eager load. In order to get the old behavior of chaining all joined eager loads as outer join when an outer join is present, use innerjoin="unnested".

As introduced in Right-nested inner joins available in joined eager loads from version 0.9, the behavior of innerjoin="nested" is that an inner join eager load chained to an outer join eager load will use a right-nested join. "nested" is now implied when using innerjoin=True:

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

With the new default, this will render the FROM clause in the form:

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

That is, using a right-nested join for the INNER join so that the full result of users can be returned. The use of an INNER join is more efficient than using an OUTER join, and allows the joinedload.innerjoin optimization parameter to take effect in all cases.

To get the older behavior, use innerjoin="unnested":

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

This will avoid right-nested joins and chain the joins together using all OUTER joins despite the innerjoin directive:

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

As noted in the 0.9 notes, the only database backend that has difficulty with right-nested joins is SQLite; SQLAlchemy as of 0.9 converts a right-nested join into a subquery as a join target on SQLite.

See also

Right-nested inner joins available in joined eager loads - description of the feature as introduced in 0.9.4.

#3008

Subqueries no longer applied to uselist=False joined eager loads

Given a joined eager load like the following:

  1. class A(Base):
  2. __tablename__ = 'a'
  3. id = Column(Integer, primary_key=True)
  4. b = relationship("B", uselist=False)
  5. class B(Base):
  6. __tablename__ = 'b'
  7. id = Column(Integer, primary_key=True)
  8. a_id = Column(ForeignKey('a.id'))
  9. s = Session()
  10. print(s.query(A).options(joinedload(A.b)).limit(5))

SQLAlchemy considers the relationship A.b to be a “one to many, loaded as a single value”, which is essentially a “one to one” relationship. However, joined eager loading has always treated the above as a situation where the main query needs to be inside a subquery, as would normally be needed for a collection of B objects where the main query has a LIMIT applied:

  1. SELECT anon_1.a_id AS anon_1_a_id, b_1.id AS b_1_id, b_1.a_id AS b_1_a_id
  2. FROM (SELECT a.id AS a_id
  3. FROM a LIMIT :param_1) AS anon_1
  4. LEFT OUTER JOIN b AS b_1 ON anon_1.a_id = b_1.a_id

However, since the relationship of the inner query to the outer one is that at most only one row is shared in the case of uselist=False (in the same way as a many-to-one), the “subquery” used with LIMIT + joined eager loading is now dropped in this case:

  1. SELECT a.id AS a_id, b_1.id AS b_1_id, b_1.a_id AS b_1_a_id
  2. FROM a LEFT OUTER JOIN b AS b_1 ON a.id = b_1.a_id
  3. LIMIT :param_1

In the case that the LEFT OUTER JOIN returns more than one row, the ORM has always emitted a warning here and ignored additional results for uselist=False, so the results in that error situation should not change.

#3249

query.update() / query.delete() raises if used with join(), select_from(), from_self()

A warning is emitted in SQLAlchemy 0.9.10 (not yet released as of June 9, 2015) when the Query.update() or Query.delete() methods are invoked against a query which has also called upon Query.join(), Query.outerjoin(), Query.select_from() or Query.from_self(). These are unsupported use cases which silently fail in the 0.9 series up until 0.9.10 where it emits a warning. In 1.0, these cases raise an exception.

#3349

query.update() with synchronize_session='evaluate' raises on multi-table update

The “evaluator” for Query.update() won’t work with multi-table updates, and needs to be set to synchronize_session=False or synchronize_session='fetch' when multiple tables are present. The new behavior is that an explicit exception is now raised, with a message to change the synchronize setting. This is upgraded from a warning emitted as of 0.9.7.

#3117

Resurrect Event has been Removed

The “resurrect” ORM event has been removed entirely. This event ceased to have any function since version 0.8 removed the older “mutable” system from the unit of work.

Change to single-table-inheritance criteria when using from_self(), count()

Given a single-table inheritance mapping, such as:

  1. class Widget(Base):
  2. __table__ = 'widget_table'
  3. class FooWidget(Widget):
  4. pass

Using Query.from_self() or Query.count() against a subclass would produce a subquery, but then add the “WHERE” criteria for subtypes to the outside:

  1. sess.query(FooWidget).from_self().all()

rendering:

  1. SELECT
  2. anon_1.widgets_id AS anon_1_widgets_id,
  3. anon_1.widgets_type AS anon_1_widgets_type
  4. FROM (SELECT widgets.id AS widgets_id, widgets.type AS widgets_type,
  5. FROM widgets) AS anon_1
  6. WHERE anon_1.widgets_type IN (?)

The issue with this is that if the inner query does not specify all columns, then we can’t add the WHERE clause on the outside (it actually tries, and produces a bad query). This decision apparently goes way back to 0.6.5 with the note “may need to make more adjustments to this”. Well, those adjustments have arrived! So now the above query will render:

  1. SELECT
  2. anon_1.widgets_id AS anon_1_widgets_id,
  3. anon_1.widgets_type AS anon_1_widgets_type
  4. FROM (SELECT widgets.id AS widgets_id, widgets.type AS widgets_type,
  5. FROM widgets
  6. WHERE widgets.type IN (?)) AS anon_1

So that queries that don’t include “type” will still work!:

  1. sess.query(FooWidget.id).count()

Renders:

  1. SELECT count(*) AS count_1
  2. FROM (SELECT widgets.id AS widgets_id
  3. FROM widgets
  4. WHERE widgets.type IN (?)) AS anon_1

#3177

single-table-inheritance criteria added to all ON clauses unconditionally

When joining to a single-table inheritance subclass target, the ORM always adds the “single table criteria” when joining on a relationship. Given a mapping as:

  1. class Widget(Base):
  2. __tablename__ = 'widget'
  3. id = Column(Integer, primary_key=True)
  4. type = Column(String)
  5. related_id = Column(ForeignKey('related.id'))
  6. related = relationship("Related", backref="widget")
  7. __mapper_args__ = {'polymorphic_on': type}
  8. class FooWidget(Widget):
  9. __mapper_args__ = {'polymorphic_identity': 'foo'}
  10. class Related(Base):
  11. __tablename__ = 'related'
  12. id = Column(Integer, primary_key=True)

It’s been the behavior for quite some time that a JOIN on the relationship will render a “single inheritance” clause for the type:

  1. s.query(Related).join(FooWidget, Related.widget).all()

SQL output:

  1. SELECT related.id AS related_id
  2. FROM related JOIN widget ON related.id = widget.related_id AND widget.type IN (:type_1)

Above, because we joined to a subclass FooWidget, Query.join() knew to add the AND widget.type IN ('foo') criteria to the ON clause.

The change here is that the AND widget.type IN() criteria is now appended to any ON clause, not just those generated from a relationship, including one that is explicitly stated:

  1. # ON clause will now render as
  2. # related.id = widget.related_id AND widget.type IN (:type_1)
  3. s.query(Related).join(FooWidget, FooWidget.related_id == Related.id).all()

As well as the “implicit” join when no ON clause of any kind is stated:

  1. # ON clause will now render as
  2. # related.id = widget.related_id AND widget.type IN (:type_1)
  3. s.query(Related).join(FooWidget).all()

Previously, the ON clause for these would not include the single-inheritance criteria. Applications that are already adding this criteria to work around this will want to remove its explicit use, though it should continue to work fine if the criteria happens to be rendered twice in the meantime.

See also

Changes and fixes in handling of duplicate join targets

#3222

Key Behavioral Changes - Core

Warnings emitted when coercing full SQL fragments into text()

Since SQLAlchemy’s inception, there has always been an emphasis on not getting in the way of the usage of plain text. The Core and ORM expression systems were intended to allow any number of points at which the user can just use plain text SQL expressions, not just in the sense that you can send a full SQL string to Connection.execute(), but that you can send strings with SQL expressions into many functions, such as Select.where(), Query.filter(), and Select.order_by().

Note that by “SQL expressions” we mean a full fragment of a SQL string, such as:

  1. # the argument sent to where() is a full SQL expression
  2. stmt = select([sometable]).where("somecolumn = 'value'")

and we are not talking about string arguments, that is, the normal behavior of passing string values that become parameterized:

  1. # This is a normal Core expression with a string argument -
  2. # we aren't talking about this!!
  3. stmt = select([sometable]).where(sometable.c.somecolumn == 'value')

The Core tutorial has long featured an example of the use of this technique, using a select() construct where virtually all components of it are specified as straight strings. However, despite this long-standing behavior and example, users are apparently surprised that this behavior exists, and when asking around the community, I was unable to find any user that was in fact not surprised that you can send a full string into a method like Query.filter().

So the change here is to encourage the user to qualify textual strings when composing SQL that is partially or fully composed from textual fragments. When composing a select as below:

  1. stmt = select(["a", "b"]).where("a = b").select_from("sometable")

The statement is built up normally, with all the same coercions as before. However, one will see the following warnings emitted:

  1. SAWarning: Textual column expression 'a' should be explicitly declared
  2. with text('a'), or use column('a') for more specificity
  3. (this warning may be suppressed after 10 occurrences)
  4. SAWarning: Textual column expression 'b' should be explicitly declared
  5. with text('b'), or use column('b') for more specificity
  6. (this warning may be suppressed after 10 occurrences)
  7. SAWarning: Textual SQL expression 'a = b' should be explicitly declared
  8. as text('a = b') (this warning may be suppressed after 10 occurrences)
  9. SAWarning: Textual SQL FROM expression 'sometable' should be explicitly
  10. declared as text('sometable'), or use table('sometable') for more
  11. specificity (this warning may be suppressed after 10 occurrences)

These warnings attempt to show exactly where the issue is by displaying the parameters as well as where the string was received. The warnings make use of the Session.get_bind() handles a wider variety of inheritance scenarios so that parameterized warnings can be emitted safely without running out of memory, and as always, if one wishes the warnings to be exceptions, the Python Warnings Filter should be used:

  1. import warnings
  2. warnings.simplefilter("error") # all warnings raise an exception

Given the above warnings, our statement works just fine, but to get rid of the warnings we would rewrite our statement as follows:

  1. from sqlalchemy import select, text
  2. stmt = select([
  3. text("a"),
  4. text("b")
  5. ]).where(text("a = b")).select_from(text("sometable"))

and as the warnings suggest, we can give our statement more specificity about the text if we use column() and table():

  1. from sqlalchemy import select, text, column, table
  2. stmt = select([column("a"), column("b")]).\
  3. where(text("a = b")).select_from(table("sometable"))

Where note also that table() and column() can now be imported from “sqlalchemy” without the “sql” part.

The behavior here applies to select() as well as to key methods on Query, including Query.filter(), Query.from_statement() and Query.having().

ORDER BY and GROUP BY are special cases

There is one case where usage of a string has special meaning, and as part of this change we have enhanced its functionality. When we have a select() or Query that refers to some column name or named label, we might want to GROUP BY and/or ORDER BY known columns or labels:

  1. stmt = select([
  2. user.c.name,
  3. func.count(user.c.id).label("id_count")
  4. ]).group_by("name").order_by("id_count")

In the above statement we expect to see “ORDER BY id_count”, as opposed to a re-statement of the function. The string argument given is actively matched to an entry in the columns clause during compilation, so the above statement would produce as we expect, without warnings (though note that the "name" expression has been resolved to users.name!):

  1. SELECT users.name, count(users.id) AS id_count
  2. FROM users GROUP BY users.name ORDER BY id_count

However, if we refer to a name that cannot be located, then we get the warning again, as below:

  1. stmt = select([
  2. user.c.name,
  3. func.count(user.c.id).label("id_count")
  4. ]).order_by("some_label")

The output does what we say, but again it warns us:

  1. SAWarning: Can't resolve label reference 'some_label'; converting to
  2. text() (this warning may be suppressed after 10 occurrences)
  3. SELECT users.name, count(users.id) AS id_count
  4. FROM users ORDER BY some_label

The above behavior applies to all those places where we might want to refer to a so-called “label reference”; ORDER BY and GROUP BY, but also within an OVER clause as well as a DISTINCT ON clause that refers to columns (e.g. the PostgreSQL syntax).

We can still specify any arbitrary expression for ORDER BY or others using text():

  1. stmt = select([users]).order_by(text("some special expression"))

The upshot of the whole change is that SQLAlchemy now would like us to tell it when a string is sent that this string is explicitly a text() construct, or a column, table, etc., and if we use it as a label name in an order by, group by, or other expression, SQLAlchemy expects that the string resolves to something known, else it should again be qualified with text() or similar.

#2992

Python-side defaults invoked for each row individually when using a multivalued insert

Support for Python-side column defaults when using the multi-valued version of Insert.values() were essentially not implemented, and would only work “by accident” in specific situations, when the dialect in use was using a non-positional (e.g. named) style of bound parameter, and when it was not necessary that a Python-side callable be invoked for each row.

The feature has been overhauled so that it works more similarly to that of an “executemany” style of invocation:

  1. import itertools
  2. counter = itertools.count(1)
  3. t = Table(
  4. 'my_table', metadata,
  5. Column('id', Integer, default=lambda: next(counter)),
  6. Column('data', String)
  7. )
  8. conn.execute(t.insert().values([
  9. {"data": "d1"},
  10. {"data": "d2"},
  11. {"data": "d3"},
  12. ]))

The above example will invoke next(counter) for each row individually as would be expected:

  1. INSERT INTO my_table (id, data) VALUES (?, ?), (?, ?), (?, ?)
  2. (1, 'd1', 2, 'd2', 3, 'd3')

Previously, a positional dialect would fail as a bind would not be generated for additional positions:

  1. Incorrect number of bindings supplied. The current statement uses 6,
  2. and there are 4 supplied.
  3. [SQL: u'INSERT INTO my_table (id, data) VALUES (?, ?), (?, ?), (?, ?)']
  4. [parameters: (1, 'd1', 'd2', 'd3')]

And with a “named” dialect, the same value for “id” would be re-used in each row (hence this change is backwards-incompatible with a system that relied on this):

  1. INSERT INTO my_table (id, data) VALUES (:id, :data_0), (:id, :data_1), (:id, :data_2)
  2. {u'data_2': 'd3', u'data_1': 'd2', u'data_0': 'd1', 'id': 1}

The system will also refuse to invoke a “server side” default as inline-rendered SQL, since it cannot be guaranteed that a server side default is compatible with this. If the VALUES clause renders for a specific column, then a Python-side value is required; if an omitted value only refers to a server-side default, an exception is raised:

  1. t = Table(
  2. 'my_table', metadata,
  3. Column('id', Integer, primary_key=True),
  4. Column('data', String, server_default='some default')
  5. )
  6. conn.execute(t.insert().values([
  7. {"data": "d1"},
  8. {"data": "d2"},
  9. {},
  10. ]))

will raise:

  1. sqlalchemy.exc.CompileError: INSERT value for column my_table.data is
  2. explicitly rendered as a boundparameter in the VALUES clause; a
  3. Python-side value or SQL expression is required

Previously, the value “d1” would be copied into that of the third row (but again, only with named format!):

  1. INSERT INTO my_table (data) VALUES (:data_0), (:data_1), (:data_0)
  2. {u'data_1': 'd2', u'data_0': 'd1'}

#3288

Event listeners can not be added or removed from within that event’s runner

Removal of an event listener from inside that same event itself would modify the elements of a list during iteration, which would cause still-attached event listeners to silently fail to fire. To prevent this while still maintaining performance, the lists have been replaced with collections.deque(), which does not allow any additions or removals during iteration, and instead raises RuntimeError.

#3163

The INSERT…FROM SELECT construct now implies inline=True

Using Insert.from_select() now implies inline=True on insert(). This helps to fix a bug where an INSERT…FROM SELECT construct would inadvertently be compiled as “implicit returning” on supporting backends, which would cause breakage in the case of an INSERT that inserts zero rows (as implicit returning expects a row), as well as arbitrary return data in the case of an INSERT that inserts multiple rows (e.g. only the first row of many). A similar change is also applied to an INSERT..VALUES with multiple parameter sets; implicit RETURNING will no longer emit for this statement either. As both of these constructs deal with variable numbers of rows, the ResultProxy.inserted_primary_key accessor does not apply. Previously, there was a documentation note that one may prefer inline=True with INSERT..FROM SELECT as some databases don’t support returning and therefore can’t do “implicit” returning, but there’s no reason an INSERT…FROM SELECT needs implicit returning in any case. Regular explicit Insert.returning() should be used to return variable numbers of result rows if inserted data is needed.

#3169

autoload_with now implies autoload=True

A Table can be set up for reflection by passing Table.autoload_with alone:

  1. my_table = Table('my_table', metadata, autoload_with=some_engine)

#3027

DBAPI exception wrapping and handle_error() event improvements

SQLAlchemy’s wrapping of DBAPI exceptions was not taking place in the case where a Connection object was invalidated, and then tried to reconnect and encountered an error; this has been resolved.

Additionally, the recently added ConnectionEvents.handle_error() event is now invoked for errors that occur upon initial connect, upon reconnect, and when create_engine() is used given a custom connection function via create_engine.creator.

The ExceptionContext object has a new datamember ExceptionContext.engine that will always refer to the Engine in use, in those cases when the Connection object is not available (e.g. on initial connect).

#3266

ForeignKeyConstraint.columns is now a ColumnCollection

ForeignKeyConstraint.columns was previously a plain list containing either strings or Column objects, depending on how the ForeignKeyConstraint was constructed and whether it was associated with a table. The collection is now a ColumnCollection, and is only initialized after the ForeignKeyConstraint is associated with a Table. A new accessor ForeignKeyConstraint.column_keys is added to unconditionally return string keys for the local set of columns regardless of how the object was constructed or its current state.

MetaData.sorted_tables accessor is “deterministic”

The sorting of tables resulting from the MetaData.sorted_tables accessor is “deterministic”; the ordering should be the same in all cases regardless of Python hashing. This is done by first sorting the tables by name before passing them to the topological algorithm, which maintains that ordering as it iterates.

Note that this change does not yet apply to the ordering applied when emitting MetaData.create_all() or MetaData.drop_all().

#3084

null(), false() and true() constants are no longer singletons

These three constants were changed to return a “singleton” value in 0.9; unfortunately, that would lead to a query like the following to not render as expected:

  1. select([null(), null()])

rendering only SELECT NULL AS anon_1, because the two null() constructs would come out as the same NULL object, and SQLAlchemy’s Core model is based on object identity in order to determine lexical significance. The change in 0.9 had no importance other than the desire to save on object overhead; in general, an unnamed construct needs to stay lexically unique so that it gets labeled uniquely.

#3170

SQLite/Oracle have distinct methods for temporary table/view name reporting

The Inspector.get_table_names() and Inspector.get_view_names() methods in the case of SQLite/Oracle would also return the names of temporary tables and views, which is not provided by any other dialect (in the case of MySQL at least it is not even possible). This logic has been moved out to two new methods Inspector.get_temp_table_names() and Inspector.get_temp_view_names().

Note that reflection of a specific named temporary table or temporary view, either by Table('name', autoload=True) or via methods like Inspector.get_columns() continues to function for most if not all dialects. For SQLite specifically, there is a bug fix for UNIQUE constraint reflection from temp tables as well, which is #3203.

#3204

Dialect Improvements and Changes - PostgreSQL

Overhaul of ENUM type create/drop rules

The rules for PostgreSQL ENUM have been made more strict with regards to creating and dropping of the TYPE.

An ENUM that is created without being explicitly associated with a MetaData object will be created and dropped corresponding to Table.create() and Table.drop():

  1. table = Table('sometable', metadata,
  2. Column('some_enum', ENUM('a', 'b', 'c', name='myenum'))
  3. )
  4. table.create(engine) # will emit CREATE TYPE and CREATE TABLE
  5. table.drop(engine) # will emit DROP TABLE and DROP TYPE - new for 1.0

This means that if a second table also has an enum named ‘myenum’, the above DROP operation will now fail. In order to accommodate the use case of a common shared enumerated type, the behavior of a metadata-associated enumeration has been enhanced.

An ENUM that is created with being explicitly associated with a MetaData object will not be created or dropped corresponding to Table.create() and Table.drop(), with the exception of Table.create() called with the checkfirst=True flag:

  1. my_enum = ENUM('a', 'b', 'c', name='myenum', metadata=metadata)
  2. table = Table('sometable', metadata,
  3. Column('some_enum', my_enum)
  4. )
  5. # will fail: ENUM 'my_enum' does not exist
  6. table.create(engine)
  7. # will check for enum and emit CREATE TYPE
  8. table.create(engine, checkfirst=True)
  9. table.drop(engine) # will emit DROP TABLE, *not* DROP TYPE
  10. metadata.drop_all(engine) # will emit DROP TYPE
  11. metadata.create_all(engine) # will emit CREATE TYPE

#3319

New PostgreSQL Table options

Added support for PG table options TABLESPACE, ON COMMIT, WITH(OUT) OIDS, and INHERITS, when rendering DDL via the Table construct.

See also

PostgreSQL Table Options

#2051

New get_enums() method with PostgreSQL Dialect

The inspect() method returns a PGInspector object in the case of PostgreSQL, which includes a new PGInspector.get_enums() method that returns information on all available ENUM types:

  1. from sqlalchemy import inspect, create_engine
  2. engine = create_engine("postgresql+psycopg2://host/dbname")
  3. insp = inspect(engine)
  4. print(insp.get_enums())

See also

PGInspector.get_enums()

PostgreSQL Dialect reflects Materialized Views, Foreign Tables

Changes are as follows:

  • the Table construct with autoload=True will now match a name that exists in the database as a materialized view or foreign table.

  • Inspector.get_view_names() will return plain and materialized view names.

  • Inspector.get_table_names() does not change for PostgreSQL, it continues to return only the names of plain tables.

  • A new method PGInspector.get_foreign_table_names() is added which will return the names of tables that are specifically marked as “foreign” in the PostgreSQL schema tables.

The change to reflection involves adding 'm' and 'f' to the list of qualifiers we use when querying pg_class.relkind, but this change is new in 1.0.0 to avoid any backwards-incompatible surprises for those running 0.9 in production.

#2891

PostgreSQL has_table() now works for temporary tables

This is a simple fix such that “has table” for temporary tables now works, so that code like the following may proceed:

  1. from sqlalchemy import *
  2. metadata = MetaData()
  3. user_tmp = Table(
  4. "user_tmp", metadata,
  5. Column("id", INT, primary_key=True),
  6. Column('name', VARCHAR(50)),
  7. prefixes=['TEMPORARY']
  8. )
  9. e = create_engine("postgresql://scott:tiger@localhost/test", echo='debug')
  10. with e.begin() as conn:
  11. user_tmp.create(conn, checkfirst=True)
  12. # checkfirst will succeed
  13. user_tmp.create(conn, checkfirst=True)

The very unlikely case that this behavior will cause a non-failing application to behave differently, is because PostgreSQL allows a non-temporary table to silently overwrite a temporary table. So code like the following will now act completely differently, no longer creating the real table following the temporary table:

  1. from sqlalchemy import *
  2. metadata = MetaData()
  3. user_tmp = Table(
  4. "user_tmp", metadata,
  5. Column("id", INT, primary_key=True),
  6. Column('name', VARCHAR(50)),
  7. prefixes=['TEMPORARY']
  8. )
  9. e = create_engine("postgresql://scott:tiger@localhost/test", echo='debug')
  10. with e.begin() as conn:
  11. user_tmp.create(conn, checkfirst=True)
  12. m2 = MetaData()
  13. user = Table(
  14. "user_tmp", m2,
  15. Column("id", INT, primary_key=True),
  16. Column('name', VARCHAR(50)),
  17. )
  18. # in 0.9, *will create* the new table, overwriting the old one.
  19. # in 1.0, *will not create* the new table
  20. user.create(conn, checkfirst=True)

#3264

PostgreSQL FILTER keyword

The SQL standard FILTER keyword for aggregate functions is now supported by PostgreSQL as of 9.4. SQLAlchemy allows this using FunctionElement.filter():

  1. func.count(1).filter(True)

See also

FunctionElement.filter()

FunctionFilter

PG8000 dialect supports client side encoding

The create_engine.encoding parameter is now honored by the pg8000 dialect, using on connect handler which emits SET CLIENT_ENCODING matching the selected encoding.

PG8000 native JSONB support

Support for PG8000 versions greater than 1.10.1 has been added, where JSONB is supported natively.

Support for psycopg2cffi Dialect on PyPy

Support for the pypy psycopg2cffi dialect is added.

See also

sqlalchemy.dialects.postgresql.psycopg2cffi

Dialect Improvements and Changes - MySQL

MySQL TIMESTAMP Type now renders NULL / NOT NULL in all cases

The MySQL dialect has always worked around MySQL’s implicit NOT NULL default associated with TIMESTAMP columns by emitting NULL for such a type, if the column is set up with nullable=True. However, MySQL 5.6.6 and above features a new flag explicit_defaults_for_timestamp which repairs MySQL’s non-standard behavior to make it behave like any other type; to accommodate this, SQLAlchemy now emits NULL/NOT NULL unconditionally for all TIMESTAMP columns.

See also

TIMESTAMP Columns and NULL

#3155

MySQL SET Type Overhauled to support empty sets, unicode, blank value handling

The SET type historically not included a system of handling blank sets and empty values separately; as different drivers had different behaviors for treatment of empty strings and empty-string-set representations, the SET type tried only to hedge between these behaviors, opting to treat the empty set as set(['']) as is still the current behavior for the MySQL-Connector-Python DBAPI. Part of the rationale here was that it was otherwise impossible to actually store a blank string within a MySQL SET, as the driver gives us back strings with no way to discern between set(['']) and set(). It was left to the user to determine if set(['']) actually meant “empty set” or not.

The new behavior moves the use case for the blank string, which is an unusual case that isn’t even documented in MySQL’s documentation, into a special case, and the default behavior of SET is now:

  • to treat the empty string '' as returned by MySQL-python into the empty set set();

  • to convert the single-blank value set set(['']) returned by MySQL-Connector-Python into the empty set set();

  • To handle the case of a set type that actually wishes includes the blank value '' in its list of possible values, a new feature (required in this use case) is implemented whereby the set value is persisted and loaded as a bitwise integer value; the flag SET.retrieve_as_bitwise is added in order to enable this.

Using the SET.retrieve_as_bitwise flag allows the set to be persisted and retrieved with no ambiguity of values. Theoretically this flag can be turned on in all cases, as long as the given list of values to the type matches the ordering exactly as declared in the database; it only makes the SQL echo output a bit more unusual.

The default behavior of SET otherwise remains the same, roundtripping values using strings. The string-based behavior now supports unicode fully including MySQL-python with use_unicode=0.

#3283

MySQL internal “no such table” exceptions not passed to event handlers

The MySQL dialect will now disable ConnectionEvents.handle_error() events from firing for those statements which it uses internally to detect if a table exists or not. This is achieved using an execution option skip_user_error_events that disables the handle error event for the scope of that execution. In this way, user code that rewrites exceptions doesn’t need to worry about the MySQL dialect or other dialects that occasionally need to catch SQLAlchemy specific exceptions.

Changed the default value of raise_on_warnings for MySQL-Connector

Changed the default value of “raise_on_warnings” to False for MySQL-Connector. This was set at True for some reason. The “buffered” flag unfortunately must stay at True as MySQLconnector does not allow a cursor to be closed unless all results are fully fetched.

#2515

MySQL boolean symbols “true”, “false” work again

0.9’s overhaul of the IS/IS NOT operators as well as boolean types in #2682 disallowed the MySQL dialect from making use of the “true” and “false” symbols in the context of “IS” / “IS NOT”. Apparently, even though MySQL has no “boolean” type, it supports IS / IS NOT when the special “true” and “false” symbols are used, even though these are otherwise synonymous with “1” and “0” (and IS/IS NOT don’t work with the numerics).

So the change here is that the MySQL dialect remains “non native boolean”, but the true() and false() symbols again produce the keywords “true” and “false”, so that an expression like column.is_(true()) again works on MySQL.

#3186

The match() operator now returns an agnostic MatchType compatible with MySQL’s floating point return value

The return type of a ColumnOperators.match() expression is now a new type called MatchType. This is a subclass of Boolean, that can be intercepted by the dialect in order to produce a different result type at SQL execution time.

Code like the following will now function correctly and return floating points on MySQL:

  1. >>> connection.execute(
  2. ... select([
  3. ... matchtable.c.title.match('Agile Ruby Programming').label('ruby'),
  4. ... matchtable.c.title.match('Dive Python').label('python'),
  5. ... matchtable.c.title
  6. ... ]).order_by(matchtable.c.id)
  7. ... )
  8. [
  9. (2.0, 0.0, 'Agile Web Development with Ruby On Rails'),
  10. (0.0, 2.0, 'Dive Into Python'),
  11. (2.0, 0.0, "Programming Matz's Ruby"),
  12. (0.0, 0.0, 'The Definitive Guide to Django'),
  13. (0.0, 1.0, 'Python in a Nutshell')
  14. ]

#3263

Drizzle Dialect is now an External Dialect

The dialect for Drizzle is now an external dialect, available at https://bitbucket.org/zzzeek/sqlalchemy-drizzle. This dialect was added to SQLAlchemy right before SQLAlchemy was able to accommodate third party dialects well; going forward, all databases that aren’t within the “ubiquitous use” category are third party dialects. The dialect’s implementation hasn’t changed and is still based on the MySQL + MySQLdb dialects within SQLAlchemy. The dialect is as of yet unreleased and in “attic” status; however it passes the majority of tests and is generally in decent working order, if someone wants to pick up on polishing it.

Dialect Improvements and Changes - SQLite

SQLite named and unnamed UNIQUE and FOREIGN KEY constraints will inspect and reflect

UNIQUE and FOREIGN KEY constraints are now fully reflected on SQLite both with and without names. Previously, foreign key names were ignored and unnamed unique constraints were skipped. In particular this will help with Alembic’s new SQLite migration features.

To achieve this, for both foreign keys and unique constraints, the result of PRAGMA foreign_keys, index_list, and index_info is combined with regular expression parsing of the CREATE TABLE statement overall to form a complete picture of the names of constraints, as well as differentiating UNIQUE constraints that were created as UNIQUE vs. unnamed INDEXes.

#3244

#3261

Dialect Improvements and Changes - SQL Server

PyODBC driver name is required with hostname-based SQL Server connections

Connecting to SQL Server with PyODBC using a DSN-less connection, e.g. with an explicit hostname, now requires a driver name - SQLAlchemy will no longer attempt to guess a default:

  1. engine = create_engine("mssql+pyodbc://scott:tiger@myhost:port/databasename?driver=SQL+Server+Native+Client+10.0")

SQLAlchemy’s previously hardcoded default of “SQL Server” is obsolete on Windows, and SQLAlchemy cannot be tasked with guessing the best driver based on operation system/driver detection. Using a DSN is always preferred when using ODBC to avoid this issue entirely.

#3182

SQL Server 2012 large text / binary types render as VARCHAR, NVARCHAR, VARBINARY

The rendering of the TextClause, UnicodeText, and LargeBinary types has been changed for SQL Server 2012 and greater, with options to control the behavior completely, based on deprecation guidelines from Microsoft. See Large Text/Binary Type Deprecation for details.

Dialect Improvements and Changes - Oracle

Improved support for CTEs in Oracle

CTE support has been fixed up for Oracle, and there is also a new feature CTE.with_suffixes() that can assist with Oracle’s special directives:

  1. included_parts = select([
  2. part.c.sub_part, part.c.part, part.c.quantity
  3. ]).where(part.c.part == "p1").\
  4. cte(name="included_parts", recursive=True).\
  5. suffix_with(
  6. "search depth first by part set ord1",
  7. "cycle part set y_cycle to 1 default 0", dialect='oracle')

#3220

New Oracle Keywords for DDL

Keywords such as COMPRESS, ON COMMIT, BITMAP:

Oracle Table Options

Oracle Specific Index Options