What’s New in SQLAlchemy 1.1?

About this Document

This document describes changes between SQLAlchemy version 1.0 and SQLAlchemy version 1.1.


This guide introduces what’s new in SQLAlchemy version 1.1, and also documents changes which affect users migrating their applications from the 1.0 series of SQLAlchemy to 1.1.

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

Platform / Installer Changes

Setuptools is now required for install

SQLAlchemy’s setup.py file has for many years supported operation both with Setuptools installed and without; supporting a “fallback” mode that uses straight Distutils. As a Setuptools-less Python environment is now unheard of, and in order to support the featureset of Setuptools more fully, in particular to support py.test’s integration with it as well as things like “extras”, setup.py now depends on Setuptools fully.


Enabling / Disabling C Extension builds is only via environment variable

The C Extensions build by default during install as long as it is possible. To disable C extension builds, the DISABLE_SQLALCHEMY_CEXT environment variable was made available as of SQLAlchemy 0.8.6 / 0.9.4. The previous approach of using the --without-cextensions argument has been removed, as it relies on deprecated features of setuptools.


New Features and Improvements - ORM

New Session lifecycle events

The Session has long supported events that allow some degree of tracking of state changes to objects, including SessionEvents.before_attach(), SessionEvents.after_attach(), and SessionEvents.before_flush(). The Session documentation also documents major object states at Quickie Intro to Object States. However, there has never been system of tracking objects specifically as they pass through these transitions. Additionally, the status of “deleted” objects has historically been murky as the objects act somewhere between the “persistent” and “detached” states.

To clean up this area and allow the realm of session state transition to be fully transparent, a new series of events have been added that are intended to cover every possible way that an object might transition between states, and additionally the “deleted” status has been given its own official state name within the realm of session object states.

New State Transition Events

Transitions between all states of an object such as persistent, pending and others can now be intercepted in terms of a session-level event intended to cover a specific transition. Transitions as objects move into a Session, move out of a Session, and even all the transitions which occur when the transaction is rolled back using Session.rollback() are explicitly present in the interface of SessionEvents.

In total, there are ten new events. A summary of these events is in a newly written documentation section Object Lifecycle Events.

New Object State “deleted” is added, deleted objects no longer “persistent”

The persistent state of an object in the Session has always been documented as an object that has a valid database identity; however in the case of objects that were deleted within a flush, they have always been in a grey area where they are not really “detached” from the Session yet, because they can still be restored within a rollback, but are not really “persistent” because their database identity has been deleted and they aren’t present in the identity map.

To resolve this grey area given the new events, a new object state deleted is introduced. This state exists between the “persistent” and “detached” states. An object that is marked for deletion via Session.delete() remains in the “persistent” state until a flush proceeds; at that point, it is removed from the identity map, moves to the “deleted” state, and the SessionEvents.persistent_to_deleted() hook is invoked. If the Session object’s transaction is rolled back, the object is restored as persistent; the SessionEvents.deleted_to_persistent() transition is called. Otherwise if the Session object’s transaction is committed, the SessionEvents.deleted_to_detached() transition is invoked.

Additionally, the InstanceState.persistent accessor no longer returns True for an object that is in the new “deleted” state; instead, the InstanceState.deleted accessor has been enhanced to reliably report on this new state. When the object is detached, the InstanceState.deleted returns False and the InstanceState.detached accessor is True instead. To determine if an object was deleted either in the current transaction or in a previous transaction, use the InstanceState.was_deleted accessor.

Strong Identity Map is Deprecated

One of the inspirations for the new series of transition events was to enable leak-proof tracking of objects as they move in and out of the identity map, so that a “strong reference” may be maintained mirroring the object moving in and out of this map. With this new capability, there is no longer any need for the Session.weak_identity_map parameter and the corresponding StrongIdentityMap object. This option has remained in SQLAlchemy for many years as the “strong-referencing” behavior used to be the only behavior available, and many applications were written to assume this behavior. It has long been recommended that strong-reference tracking of objects not be an intrinsic job of the Session and instead be an application-level construct built as needed by the application; the new event model allows even the exact behavior of the strong identity map to be replicated. See Session Referencing Behavior for a new recipe illustrating how to replace the strong identity map.


New init_scalar() event intercepts default values at ORM level

The ORM produces a value of None when an attribute that has not been set is first accessed, for a non-persistent object:

>>> obj = MyObj()
>>> obj.some_value

There’s a use case for this in-Python value to correspond to that of a Core-generated default value, even before the object is persisted. To suit this use case a new event AttributeEvents.init_scalar() is added. The new example active_column_defaults.py at Attribute Instrumentation illustrates a sample use, so the effect can instead be:

>>> obj = MyObj()
>>> obj.some_value
"my default"


Changes regarding “unhashable” types, impacts deduping of ORM rows

The Query object has a well-known behavior of “deduping” returned rows that contain at least one ORM-mapped entity (e.g., a full mapped object, as opposed to individual column values). The primary purpose of this is so that the handling of entities works smoothly in conjunction with the identity map, including to accommodate for the duplicate entities normally represented within joined eager loading, as well as when joins are used for the purposes of filtering on additional columns.

This deduplication relies upon the hashability of the elements within the row. With the introduction of PostgreSQL’s special types like ARRAY, HSTORE and JSON, the experience of types within rows being unhashable and encountering problems here is more prevalent than it was previously.

In fact, SQLAlchemy has since version 0.8 included a flag on datatypes that are noted as “unhashable”, however this flag was not used consistently on built in types. As described in ARRAY and JSON types now correctly specify “unhashable”, this flag is now set consistently for all of PostgreSQL’s “structural” types.

The “unhashable” flag is also set on the NullType type, as NullType is used to refer to any expression of unknown type.

Since NullType is applied to most usages of func, as func doesn’t actually know anything about the function names given in most cases, using func() will often disable row deduping unless explicit typing is applied. The following examples illustrate func.substr() applied to a string expression, and func.date() applied to a datetime expression; both examples will return duplicate rows due to the joined eager load unless explicit typing is applied:

result = (
    session.query(func.substr(A.some_thing, 0, 4), A).options(joinedload(A.bs)).all()

users = (
        func.date(User.date_created, "start of month").label("month"),

The above examples, in order to retain deduping, should be specified as:

result = (
    session.query(func.substr(A.some_thing, 0, 4, type_=String), A)

users = (
        func.date(User.date_created, "start of month", type_=DateTime).label("month"),

Additionally, the treatment of a so-called “unhashable” type is slightly different than its been in previous releases; internally we are using the id() function to get a “hash value” from these structures, just as we would any ordinary mapped object. This replaces the previous approach which applied a counter to the object.


Specific checks added for passing mapped classes, instances as SQL literals

The typing system now has specific checks for passing of SQLAlchemy “inspectable” objects in contexts where they would otherwise be handled as literal values. Any SQLAlchemy built-in object that is legal to pass as a SQL value (which is not already a ClauseElement instance) includes a method __clause_element__() which provides a valid SQL expression for that object. For SQLAlchemy objects that don’t provide this, such as mapped classes, mappers, and mapped instances, a more informative error message is emitted rather than allowing the DBAPI to receive the object and fail later. An example is illustrated below, where a string-based attribute User.name is compared to a full instance of User(), rather than against a string value:

>>> some_user = User()
>>> q = s.query(User).filter(User.name == some_user)
sqlalchemy.exc.ArgumentError: Object <__main__.User object at 0x103167e90> is not legal as a SQL literal value

The exception is now immediate when the comparison is made between User.name == some_user. Previously, a comparison like the above would produce a SQL expression that would only fail once resolved into a DBAPI execution call; the mapped User object would ultimately become a bound parameter that would be rejected by the DBAPI.

Note that in the above example, the expression fails because User.name is a string-based (e.g. column oriented) attribute. The change does not impact the usual case of comparing a many-to-one relationship attribute to an object, which is handled distinctly:

>>> # Address.user refers to the User mapper, so
>>> # this is of course still OK!
>>> q = s.query(Address).filter(Address.user == some_user)


New Indexable ORM extension

The Indexable extension is an extension to the hybrid attribute feature which allows the construction of attributes which refer to specific elements of an “indexable” data type, such as an array or JSON field:

class Person(Base):
    __tablename__ = "person"

    id = Column(Integer, primary_key=True)
    data = Column(JSON)

    name = index_property("data", "name")

Above, the name attribute will read/write the field "name" from the JSON column data, after initializing it to an empty dictionary:

>>> person = Person(name="foobar")
>>> person.name

The extension also triggers a change event when the attribute is modified, so that there’s no need to use MutableDict in order to track this change.

See also


New options allowing explicit persistence of NULL over a default

Related to the new JSON-NULL support added to PostgreSQL as part of JSON “null” is inserted as expected with ORM operations, omitted when not present, the base TypeEngine class now supports a method TypeEngine.evaluates_none() which allows a positive set of the None value on an attribute to be persisted as NULL, rather than omitting the column from the INSERT statement, which has the effect of using the column-level default. This allows a mapper-level configuration of the existing object-level technique of assigning null() to the attribute.


Further Fixes to single-table inheritance querying

Continuing from 1.0’s Change to single-table-inheritance criteria when using from_self(), count(), the Query should no longer inappropriately add the “single inheritance” criteria when the query is against a subquery expression such as an exists:

class Widget(Base):
    __tablename__ = "widget"
    id = Column(Integer, primary_key=True)
    type = Column(String)
    data = Column(String)
    __mapper_args__ = {"polymorphic_on": type}

class FooWidget(Widget):
    __mapper_args__ = {"polymorphic_identity": "foo"}

q = session.query(FooWidget).filter(FooWidget.data == "bar").exists()



FROM widget
WHERE widget.data = :data_1 AND widget.type IN (:type_1)) AS anon_1

The IN clause on the inside is appropriate, in order to limit to FooWidget objects, however previously the IN clause would also be generated a second time on the outside of the subquery.


Improved Session state when a SAVEPOINT is cancelled by the database

A common case with MySQL is that a SAVEPOINT is cancelled when a deadlock occurs within the transaction. The Session has been modified to deal with this failure mode slightly more gracefully, such that the outer, non-savepoint transaction still remains usable:

s = Session()


    # assume the flush fails, flush goes to rollback to the
    # savepoint and that also fails
except Exception as err:
    print("Something broke, and our SAVEPOINT vanished too")

# this is the SAVEPOINT transaction, marked as
# DEACTIVE so the rollback() call succeeds

# this is the outermost transaction, remains ACTIVE
# so rollback() or commit() can succeed

This issue is a continuation of #2696 where we emit a warning so that the original error can be seen when running on Python 2, even though the SAVEPOINT exception takes precedence. On Python 3, exceptions are chained so both failures are reported individually.


Erroneous “new instance X conflicts with persistent instance Y” flush errors fixed

The Session.rollback() method is responsible for removing objects that were INSERTed into the database, e.g. moved from pending to persistent, within that now rolled-back transaction. Objects that make this state change are tracked in a weak-referencing collection, and if an object is garbage collected from that collection, the Session no longer worries about it (it would otherwise not scale for operations that insert many new objects within a transaction). However, an issue arises if the application re-loads that same garbage-collected row within the transaction, before the rollback occurs; if a strong reference to this object remains into the next transaction, the fact that this object was not inserted and should be removed would be lost, and the flush would incorrectly raise an error:

from sqlalchemy import Column, create_engine
from sqlalchemy.orm import Session
from sqlalchemy.ext.declarative import declarative_base

Base = declarative_base()

class A(Base):
    __tablename__ = "a"
    id = Column(Integer, primary_key=True)

e = create_engine("sqlite://", echo=True)

s = Session(e)

# persist an object

# rollback buffer loses reference to A

# load it again, rollback buffer knows nothing
# about it
a1 = s.query(A).first()

# roll back the transaction; all state is expired but the
# "a1" reference remains

# previous "a1" conflicts with the new one because we aren't
# checking that it never got committed

The above program would raise:

FlushError: New instance <User at 0x7f0287eca4d0> with identity key
(<class 'test.orm.test_transaction.User'>, ('u1',)) conflicts
with persistent instance <User at 0x7f02889c70d0>

The bug is that when the above exception is raised, the unit of work is operating upon the original object assuming it’s a live row, when in fact the object is expired and upon testing reveals that it’s gone. The fix tests this condition now, so in the SQL log we see:

BEGIN (implicit)


(1, 0)


BEGIN (implicit)

SELECT a.id AS a_id FROM a WHERE a.id = ?



Above, the unit of work now does a SELECT for the row we’re about to report as a conflict for, sees that it doesn’t exist, and proceeds normally. The expense of this SELECT is only incurred in the case when we would have erroneously raised an exception in any case.


passive_deletes feature for joined-inheritance mappings

A joined-table inheritance mapping may now allow a DELETE to proceed as a result of Session.delete(), which only emits DELETE for the base table, and not the subclass table, allowing configured ON DELETE CASCADE to take place for the configured foreign keys. This is configured using the mapper.passive_deletes option:

from sqlalchemy import Column, Integer, String, ForeignKey, create_engine
from sqlalchemy.orm import Session
from sqlalchemy.ext.declarative import declarative_base

Base = declarative_base()

class A(Base):
    __tablename__ = "a"
    id = Column("id", Integer, primary_key=True)
    type = Column(String)

    __mapper_args__ = {
        "polymorphic_on": type,
        "polymorphic_identity": "a",
        "passive_deletes": True,

class B(A):
    __tablename__ = "b"
    b_table_id = Column("b_table_id", Integer, primary_key=True)
    bid = Column("bid", Integer, ForeignKey("a.id", ondelete="CASCADE"))
    data = Column("data", String)

    __mapper_args__ = {"polymorphic_identity": "b"}

With the above mapping, the mapper.passive_deletes option is configured on the base mapper; it takes effect for all non-base mappers that are descendants of the mapper with the option set. A DELETE for an object of type B no longer needs to retrieve the primary key value of b_table_id if unloaded, nor does it need to emit a DELETE statement for the table itself:


Will emit SQL as:

DELETE FROM a WHERE a.id = %(id)s
-- {'id': 1}

As always, the target database must have foreign key support with ON DELETE CASCADE enabled.


Same-named backrefs will not raise an error when applied to concrete inheritance subclasses

The following mapping has always been possible without issue:

class A(Base):
    __tablename__ = "a"
    id = Column(Integer, primary_key=True)
    b = relationship("B", foreign_keys="B.a_id", backref="a")

class A1(A):
    __tablename__ = "a1"
    id = Column(Integer, primary_key=True)
    b = relationship("B", foreign_keys="B.a1_id", backref="a1")
    __mapper_args__ = {"concrete": True}

class B(Base):
    __tablename__ = "b"
    id = Column(Integer, primary_key=True)

    a_id = Column(ForeignKey("a.id"))
    a1_id = Column(ForeignKey("a1.id"))

Above, even though class A and class A1 have a relationship named b, no conflict warning or error occurs because class A1 is marked as “concrete”.

However, if the relationships were configured the other way, an error would occur:

class A(Base):
    __tablename__ = "a"
    id = Column(Integer, primary_key=True)

class A1(A):
    __tablename__ = "a1"
    id = Column(Integer, primary_key=True)
    __mapper_args__ = {"concrete": True}

class B(Base):
    __tablename__ = "b"
    id = Column(Integer, primary_key=True)

    a_id = Column(ForeignKey("a.id"))
    a1_id = Column(ForeignKey("a1.id"))

    a = relationship("A", backref="b")
    a1 = relationship("A1", backref="b")

The fix enhances the backref feature so that an error is not emitted, as well as an additional check within the mapper logic to bypass warning for an attribute being replaced.


Same-named relationships on inheriting mappers no longer warn

When creating two mappers in an inheritance scenario, placing a relationship on both with the same name would emit the warning “relationship ‘<name>’ on mapper <name> supersedes the same relationship on inherited mapper ‘<name>’; this can cause dependency issues during flush”. An example is as follows:

class A(Base):
    __tablename__ = "a"
    id = Column(Integer, primary_key=True)
    bs = relationship("B")

class ASub(A):
    __tablename__ = "a_sub"
    id = Column(Integer, ForeignKey("a.id"), primary_key=True)
    bs = relationship("B")

class B(Base):
    __tablename__ = "b"
    id = Column(Integer, primary_key=True)
    a_id = Column(ForeignKey("a.id"))

This warning dates back to the 0.4 series in 2007 and is based on a version of the unit of work code that has since been entirely rewritten. Currently, there is no known issue with the same-named relationships being placed on a base class and a descendant class, so the warning is lifted. However, note that this use case is likely not prevalent in real world use due to the warning. While rudimentary test support is added for this use case, it is possible that some new issue with this pattern may be identified.

New in version 1.1.0b3.


Hybrid properties and methods now propagate the docstring as well as .info

A hybrid method or property will now reflect the __doc__ value present in the original docstring:

class A(Base):
    __tablename__ = "a"
    id = Column(Integer, primary_key=True)

    name = Column(String)

    def some_name(self):
        """The name field"""
        return self.name

The above value of A.some_name.__doc__ is now honored:

>>> A.some_name.__doc__
The name field

However, to accomplish this, the mechanics of hybrid properties necessarily becomes more complex. Previously, the class-level accessor for a hybrid would be a simple pass-through, that is, this test would succeed:

>>> assert A.name is A.some_name

With the change, the expression returned by A.some_name is wrapped inside of its own QueryableAttribute wrapper:

>>> A.some_name
<sqlalchemy.orm.attributes.hybrid_propertyProxy object at 0x7fde03888230>

A lot of testing went into making sure this wrapper works correctly, including for elaborate schemes like that of the Custom Value Object recipe, however we’ll be looking to see that no other regressions occur for users.

As part of this change, the hybrid_property.info collection is now also propagated from the hybrid descriptor itself, rather than from the underlying expression. That is, accessing A.some_name.info now returns the same dictionary that you’d get from inspect(A).all_orm_descriptors['some_name'].info:

>>> A.some_name.info["foo"] = "bar"
>>> from sqlalchemy import inspect
>>> inspect(A).all_orm_descriptors["some_name"].info
{'foo': 'bar'}

Note that this .info dictionary is separate from that of a mapped attribute which the hybrid descriptor may be proxying directly; this is a behavioral change from 1.0. The wrapper will still proxy other useful attributes of a mirrored attribute such as QueryableAttribute.property and QueryableAttribute.class_.


Session.merge resolves pending conflicts the same as persistent

The Session.merge() method will now track the identities of objects given within a graph to maintain primary key uniqueness before emitting an INSERT. When duplicate objects of the same identity are encountered, non-primary-key attributes are overwritten as the objects are encountered, which is essentially non-deterministic. This behavior matches that of how persistent objects, that is objects that are already located in the database via primary key, are already treated, so this behavior is more internally consistent.


u1 = User(id=7, name="x")
u1.orders = [
    Order(description="o1", address=Address(id=1, email_address="a")),
    Order(description="o2", address=Address(id=1, email_address="b")),
    Order(description="o3", address=Address(id=1, email_address="c")),

sess = Session()

Above, we merge a User object with three new Order objects, each referring to a distinct Address object, however each is given the same primary key. The current behavior of Session.merge() is to look in the identity map for this Address object, and use that as the target. If the object is present, meaning that the database already has a row for Address with primary key “1”, we can see that the email_address field of the Address will be overwritten three times, in this case with the values a, b and finally c.

However, if the Address row for primary key “1” were not present, Session.merge() would instead create three separate Address instances, and we’d then get a primary key conflict upon INSERT. The new behavior is that the proposed primary key for these Address objects are tracked in a separate dictionary so that we merge the state of the three proposed Address objects onto one Address object to be inserted.

It may have been preferable if the original case emitted some kind of warning that conflicting data were present in a single merge-tree, however the non-deterministic merging of values has been the behavior for many years for the persistent case; it now matches for the pending case. A feature that warns for conflicting values could still be feasible for both cases but would add considerable performance overhead as each column value would have to be compared during the merge.


Fix involving many-to-one object moves with user-initiated foreign key manipulations

A bug has been fixed involving the mechanics of replacing a many-to-one reference to an object with another object. During the attribute operation, the location of the object that was previously referred to now makes use of the database-committed foreign key value, rather than the current foreign key value. The main effect of the fix is that a backref event towards a collection will fire off more accurately when a many-to-one change is made, even if the foreign key attribute was manually moved to the new value beforehand. Assume a mapping of the classes Parent and SomeClass, where SomeClass.parent refers to Parent and Parent.items refers to the collection of SomeClass objects:

some_object = SomeClass()
some_object.parent_id = some_parent.id
some_object.parent = some_parent

Above, we’ve made a pending object some_object, manipulated its foreign key towards Parent to refer to it, then we actually set up the relationship. Before the bug fix, the backref would not have fired off:

# before the fix
assert some_object not in some_parent.items

The fix now is that when we seek to locate the previous value of some_object.parent, we disregard the parent id that’s been manually set, and we look for the database-committed value. In this case, it’s None because the object is pending, so the event system logs some_object.parent as a net change:

# after the fix, backref fired off for some_object.parent = some_parent
assert some_object in some_parent.items

While it is discouraged to manipulate foreign key attributes that are managed by relationships, there is limited support for this use case. Applications that manipulate foreign keys in order to allow loads to proceed will often make use of the Session.enable_relationship_loading() and RelationshipProperty.load_on_pending features, which cause relationships to emit lazy loads based on in-memory foreign key values that aren’t persisted. Whether or not these features are in use, this behavioral improvement will now be apparent.


Improvements to the Query.correlate method with polymorphic entities

In recent SQLAlchemy versions, the SQL generated by many forms of “polymorphic” queries has a more “flat” form than it used to, where a JOIN of several tables is no longer bundled into a subquery unconditionally. To accommodate this, the Query.correlate() method now extracts the individual tables from such a polymorphic selectable and ensures that all are part of the “correlate” for the subquery. Assuming the Person/Manager/Engineer->Company setup from the mapping documentation, using with_polymorphic:

    .filter(Company.company_id == Person.company_id)
    == "Elbonia, Inc."

The above query now produces:

SELECT people.name AS people_name
FROM people
LEFT OUTER JOIN engineers ON people.person_id = engineers.person_id
LEFT OUTER JOIN managers ON people.person_id = managers.person_id
WHERE (SELECT companies.name
FROM companies
WHERE companies.company_id = people.company_id) = ?

Before the fix, the call to correlate(Person) would inadvertently attempt to correlate to the join of Person, Engineer and Manager as a single unit, so Person wouldn’t be correlated:

-- old, incorrect query
SELECT people.name AS people_name
FROM people
LEFT OUTER JOIN engineers ON people.person_id = engineers.person_id
LEFT OUTER JOIN managers ON people.person_id = managers.person_id
WHERE (SELECT companies.name
FROM companies, people
WHERE companies.company_id = people.company_id) = ?

Using correlated subqueries against polymorphic mappings still has some unpolished edges. If for example Person is polymorphically linked to a so-called “concrete polymorphic union” query, the above subquery may not correctly refer to this subquery. In all cases, a way to refer to the “polymorphic” entity fully is to create an aliased() object from it first:

# works with all SQLAlchemy versions and all types of polymorphic
# aliasing.

paliased = aliased(Person)
    .filter(Company.company_id == paliased.company_id)
    == "Elbonia, Inc."

The aliased() construct guarantees that the “polymorphic selectable” is wrapped in a subquery. By referring to it explicitly in the correlated subquery, the polymorphic form is correctly used.


Stringify of Query will consult the Session for the correct dialect

Calling str() on a Query object will consult the Session for the correct “bind” to use, in order to render the SQL that would be passed to the database. In particular this allows a Query that refers to dialect-specific SQL constructs to be renderable, assuming the Query is associated with an appropriate Session. Previously, this behavior would only take effect if the MetaData to which the mappings were associated were itself bound to the target Engine.

If neither the underlying MetaData nor the Session are associated with any bound Engine, then the fallback to the “default” dialect is used to generate the SQL string.


Joined eager loading where the same entity is present multiple times in one row

A fix has been made to the case has been made whereby an attribute will be loaded via joined eager loading, even if the entity was already loaded from the row on a different “path” that doesn’t include the attribute. This is a deep use case that’s hard to reproduce, but the general idea is as follows:

class A(Base):
    __tablename__ = "a"
    id = Column(Integer, primary_key=True)
    b_id = Column(ForeignKey("b.id"))
    c_id = Column(ForeignKey("c.id"))

    b = relationship("B")
    c = relationship("C")

class B(Base):
    __tablename__ = "b"
    id = Column(Integer, primary_key=True)
    c_id = Column(ForeignKey("c.id"))

    c = relationship("C")

class C(Base):
    __tablename__ = "c"
    id = Column(Integer, primary_key=True)
    d_id = Column(ForeignKey("d.id"))
    d = relationship("D")

class D(Base):
    __tablename__ = "d"
    id = Column(Integer, primary_key=True)

c_alias_1 = aliased(C)
c_alias_2 = aliased(C)

q = s.query(A)
q = q.join(A.b).join(c_alias_1, B.c).join(c_alias_1.d)
q = q.options(
    contains_eager(A.b).contains_eager(B.c, alias=c_alias_1).contains_eager(C.d)
q = q.join(c_alias_2, A.c)
q = q.options(contains_eager(A.c, alias=c_alias_2))

The above query emits SQL like this:

    d.id AS d_id,
    c_1.id AS c_1_id, c_1.d_id AS c_1_d_id,
    b.id AS b_id, b.c_id AS b_c_id,
    c_2.id AS c_2_id, c_2.d_id AS c_2_d_id,
    a.id AS a_id, a.b_id AS a_b_id, a.c_id AS a_c_id
    JOIN b ON b.id = a.b_id
    JOIN c AS c_1 ON c_1.id = b.c_id
    JOIN d ON d.id = c_1.d_id
    JOIN c AS c_2 ON c_2.id = a.c_id

We can see that the c table is selected from twice; once in the context of A.b.c -> c_alias_1 and another in the context of A.c -> c_alias_2. Also, we can see that it is quite possible that the C identity for a single row is the same for both c_alias_1 and c_alias_2, meaning two sets of columns in one row result in only one new object being added to the identity map.

The query options above only call for the attribute C.d to be loaded in the context of c_alias_1, and not c_alias_2. So whether or not the final C object we get in the identity map has the C.d attribute loaded depends on how the mappings are traversed, which while not completely random, is essentially non-deterministic. The fix is that even if the loader for c_alias_1 is processed after that of c_alias_2 for a single row where they both refer to the same identity, the C.d element will still be loaded. Previously, the loader did not seek to modify the load of an entity that was already loaded via a different path. The loader that reaches the entity first has always been non-deterministic, so this fix may be detectable as a behavioral change in some situations and not others.

The fix includes tests for two variants of the “multiple paths to one entity” case, and the fix should hopefully cover all other scenarios of this nature.


New MutableList and MutableSet helpers added to the mutation tracking extension

New helper classes MutableList and MutableSet have been added to the Mutation Tracking extension, to complement the existing MutableDict helper.


New “raise” / “raise_on_sql” loader strategies

To assist with the use case of preventing unwanted lazy loads from occurring after a series of objects are loaded, the new “lazy=’raise’” and “lazy=’raise_on_sql’” strategies and corresponding loader option raiseload() may be applied to a relationship attribute which will cause it to raise InvalidRequestError when a non-eagerly-loaded attribute is accessed for read. The two variants test for either a lazy load of any variety, including those that would only return None or retrieve from the identity map:

>>> from sqlalchemy.orm import raiseload
>>> a1 = s.query(A).options(raiseload(A.some_b)).first()
>>> a1.some_b
Traceback (most recent call last):
sqlalchemy.exc.InvalidRequestError: 'A.some_b' is not available due to lazy='raise'

Or a lazy load only where SQL would be emitted:

>>> from sqlalchemy.orm import raiseload
>>> a1 = s.query(A).options(raiseload(A.some_b, sql_only=True)).first()
>>> a1.some_b
Traceback (most recent call last):
sqlalchemy.exc.InvalidRequestError: 'A.bs' is not available due to lazy='raise_on_sql'


Mapper.order_by is deprecated

This old parameter from the very first versions of SQLAlchemy was part of the original design of the ORM which featured the Mapper object as a public-facing query structure. This role has long since been replaced by the Query object, where we use Query.order_by() to indicate the ordering of results in a way that works consistently for any combination of SELECT statements, entities and SQL expressions. There are many areas in which Mapper.order_by doesn’t work as expected (or what would be expected is not clear), such as when queries are combined into unions; these cases are not supported.


New Features and Improvements - Core

Engines now invalidate connections, run error handlers for BaseException

New in version 1.1: this change is a late add to the 1.1 series just prior to 1.1 final, and is not present in the 1.1 beta releases.

The Python BaseException class is below that of Exception but is the identifiable base for system-level exceptions such as KeyboardInterrupt, SystemExit, and notably the GreenletExit exception that’s used by eventlet and gevent. This exception class is now intercepted by the exception- handling routines of Connection, and includes handling by the ConnectionEvents.handle_error() event. The Connection is now invalidated by default in the case of a system level exception that is not a subclass of Exception, as it is assumed an operation was interrupted and the connection may be in an unusable state. The MySQL drivers are most targeted by this change however the change is across all DBAPIs.

Note that upon invalidation, the immediate DBAPI connection used by Connection is disposed, and the Connection, if still being used subsequent to the exception raise, will use a new DBAPI connection for subsequent operations upon next use; however, the state of any transaction in progress is lost and the appropriate .rollback() method must be called if applicable before this re-use can proceed.

In order to identify this change, it was straightforward to demonstrate a pymysql or mysqlclient / MySQL-Python connection moving into a corrupted state when these exceptions occur in the middle of the connection doing its work; the connection would then be returned to the connection pool where subsequent uses would fail, or even before returning to the pool would cause secondary failures in context managers that call .rollback() upon the exception catch. The behavior here is expected to reduce the incidence of the MySQL error “commands out of sync”, as well as the ResourceClosedError which can occur when the MySQL driver fails to report cursor.description correctly, when running under greenlet conditions where greenlets are killed, or where KeyboardInterrupt exceptions are handled without exiting the program entirely.

The behavior is distinct from the usual auto-invalidation feature, in that it does not assume that the backend database itself has been shut down or restarted; it does not recycle the entire connection pool as is the case for usual DBAPI disconnect exceptions.

This change should be a net improvement for all users with the exception of any application that currently intercepts ``KeyboardInterrupt`` or ``GreenletExit`` and wishes to continue working within the same transaction. Such an operation is theoretically possible with other DBAPIs that do not appear to be impacted by KeyboardInterrupt such as psycopg2. For these DBAPIs, the following workaround will disable the connection from being recycled for specific exceptions:

engine = create_engine("postgresql+psycopg2://")

@event.listens_for(engine, "handle_error")
def cancel_disconnect(ctx):
    if isinstance(ctx.original_exception, KeyboardInterrupt):
        ctx.is_disconnect = False



One of the most widely requested features is support for common table expressions (CTE) that work with INSERT, UPDATE, DELETE, and is now implemented. An INSERT/UPDATE/DELETE can both draw from a WITH clause that’s stated at the top of the SQL, as well as can be used as a CTE itself in the context of a larger statement.

As part of this change, an INSERT from SELECT that includes a CTE will now render the CTE at the top of the entire statement, rather than nested in the SELECT statement as was the case in 1.0.

Below is an example that renders UPDATE, INSERT and SELECT all in one statement:

>>> from sqlalchemy import table, column, select, literal, exists
>>> orders = table(
...     "orders",
...     column("region"),
...     column("amount"),
...     column("product"),
...     column("quantity"),
... )
>>> upsert = (
...     orders.update()
...     .where(orders.c.region == "Region1")
...     .values(amount=1.0, product="Product1", quantity=1)
...     .returning(*(orders.c._all_columns))
...     .cte("upsert")
... )
>>> insert = orders.insert().from_select(
...     orders.c.keys(),
...     select([literal("Region1"), literal(1.0), literal("Product1"), literal(1)]).where(
...         ~exists(upsert.select())
...     ),
... )
>>> print(insert)  # Note: formatting added for clarity
WITH upsert AS (UPDATE orders SET amount=:amount, product=:product, quantity=:quantity WHERE orders.region = :region_1 RETURNING orders.region, orders.amount, orders.product, orders.quantity ) INSERT INTO orders (region, amount, product, quantity) SELECT :param_1 AS anon_1, :param_2 AS anon_2, :param_3 AS anon_3, :param_4 AS anon_4 WHERE NOT ( EXISTS ( SELECT upsert.region, upsert.amount, upsert.product, upsert.quantity FROM upsert))


Support for RANGE and ROWS specification within window functions

New over.range_ and over.rows parameters allow RANGE and ROWS expressions for window functions:

>>> from sqlalchemy import func

>>> print(func.row_number().over(order_by="x", range_=(-5, 10)))
>>> print(func.row_number().over(order_by="x", rows=(None, 0)))
>>> print(func.row_number().over(order_by="x", range_=(-2, None)))

over.range_ and over.rows are specified as 2-tuples and indicate negative and positive values for specific ranges, 0 for “CURRENT ROW”, and None for UNBOUNDED.


Support for the SQL LATERAL keyword

The LATERAL keyword is currently known to only be supported by PostgreSQL 9.3 and greater, however as it is part of the SQL standard support for this keyword is added to Core. The implementation of Select.lateral() employs special logic beyond just rendering the LATERAL keyword to allow for correlation of tables that are derived from the same FROM clause as the selectable, e.g. lateral correlation:

>>> from sqlalchemy import table, column, select, true
>>> people = table("people", column("people_id"), column("age"), column("name"))
>>> books = table("books", column("book_id"), column("owner_id"))
>>> subq = (
...     select([books.c.book_id])
...     .where(books.c.owner_id == people.c.people_id)
...     .lateral("book_subq")
... )
>>> print(select([people]).select_from(people.join(subq, true())))
SELECT people.people_id, people.age, people.name FROM people JOIN LATERAL (SELECT books.book_id AS book_id FROM books WHERE books.owner_id = people.people_id) AS book_subq ON true



The SQL standard TABLESAMPLE can be rendered using the FromClause.tablesample() method, which returns a TableSample construct similar to an alias:

from sqlalchemy import func

selectable = people.tablesample(func.bernoulli(1), name="alias", seed=func.random())
stmt = select([selectable.c.people_id])

Assuming people with a column people_id, the above statement would render as:

SELECT alias.people_id FROM
people AS alias TABLESAMPLE bernoulli(:bernoulli_1)
REPEATABLE (random())


The .autoincrement directive is no longer implicitly enabled for a composite primary key column

SQLAlchemy has always had the convenience feature of enabling the backend database’s “autoincrement” feature for a single-column integer primary key; by “autoincrement” we mean that the database column will include whatever DDL directives the database provides in order to indicate an auto-incrementing integer identifier, such as the SERIAL keyword on PostgreSQL or AUTO_INCREMENT on MySQL, and additionally that the dialect will receive these generated values from the execution of a Table.insert() construct using techniques appropriate to that backend.

What’s changed is that this feature no longer turns on automatically for a composite primary key; previously, a table definition such as:

    Column("x", Integer, primary_key=True),
    Column("y", Integer, primary_key=True),

Would have “autoincrement” semantics applied to the 'x' column, only because it’s first in the list of primary key columns. In order to disable this, one would have to turn off autoincrement on all columns:

# old way
    Column("x", Integer, primary_key=True, autoincrement=False),
    Column("y", Integer, primary_key=True, autoincrement=False),

With the new behavior, the composite primary key will not have autoincrement semantics unless a column is marked explicitly with autoincrement=True:

# column 'y' will be SERIAL/AUTO_INCREMENT/ auto-generating
    Column("x", Integer, primary_key=True),
    Column("y", Integer, primary_key=True, autoincrement=True),

In order to anticipate some potential backwards-incompatible scenarios, the Table.insert() construct will perform more thorough checks for missing primary key values on composite primary key columns that don’t have autoincrement set up; given a table such as:

    Column("x", Integer, primary_key=True),
    Column("y", Integer, primary_key=True),

An INSERT emitted with no values for this table will produce this warning:

SAWarning: Column 'b.x' is marked as a member of the primary
key for table 'b', but has no Python-side or server-side default
generator indicated, nor does it indicate 'autoincrement=True',
and no explicit value is passed.  Primary key columns may not
store NULL. Note that as of SQLAlchemy 1.1, 'autoincrement=True'
must be indicated explicitly for composite (e.g. multicolumn)
primary keys if AUTO_INCREMENT/SERIAL/IDENTITY behavior is
expected for one of the columns in the primary key. CREATE TABLE
statements are impacted by this change as well on most backends.

For a column that is receiving primary key values from a server-side default or something less common such as a trigger, the presence of a value generator can be indicated using FetchedValue:

    Column("x", Integer, primary_key=True, server_default=FetchedValue()),
    Column("y", Integer, primary_key=True, server_default=FetchedValue()),

For the very unlikely case where a composite primary key is actually intended to store NULL in one or more of its columns (only supported on SQLite and MySQL), specify the column with nullable=True:

    Column("x", Integer, primary_key=True),
    Column("y", Integer, primary_key=True, nullable=True),

In a related change, the autoincrement flag may be set to True on a column that has a client-side or server-side default. This typically will not have much impact on the behavior of the column during an INSERT.



New operators ColumnOperators.is_distinct_from() and ColumnOperators.isnot_distinct_from() allow the IS DISTINCT FROM and IS NOT DISTINCT FROM sql operation:

>>> print(column("x").is_distinct_from(None))

Handling is provided for NULL, True and False:

>>> print(column("x").isnot_distinct_from(False))

For SQLite, which doesn’t have this operator, “IS” / “IS NOT” is rendered, which on SQLite works for NULL unlike other backends:

>>> from sqlalchemy.dialects import sqlite
>>> print(column("x").is_distinct_from(None).compile(dialect=sqlite.dialect()))

Core and ORM support for FULL OUTER JOIN

The new flag FromClause.outerjoin.full, available at the Core and ORM level, instructs the compiler to render FULL OUTER JOIN where it would normally render LEFT OUTER JOIN:

stmt = select([t1]).select_from(t1.outerjoin(t2, full=True))

The flag also works at the ORM level:

q = session.query(MyClass).outerjoin(MyOtherClass, full=True)


ResultSet column matching enhancements; positional column setup for textual SQL

A series of improvements were made to the ResultProxy system in the 1.0 series as part of #918, which reorganizes the internals to match cursor-bound result columns with table/ORM metadata positionally, rather than by matching names, for compiled SQL constructs that contain full information about the result rows to be returned. This allows a dramatic savings on Python overhead as well as much greater accuracy in linking ORM and Core SQL expressions to result rows. In 1.1, this reorganization has been taken further internally, and also has been made available to pure-text SQL constructs via the use of the recently added TextClause.columns() method.

TextAsFrom.columns() now works positionally

The TextClause.columns() method, added in 0.9, accepts column-based arguments positionally; in 1.1, when all columns are passed positionally, the correlation of these columns to the ultimate result set is also performed positionally. The key advantage here is that textual SQL can now be linked to an ORM- level result set without the need to deal with ambiguous or duplicate column names, or with having to match labeling schemes to ORM-level labeling schemes. All that’s needed now is the same ordering of columns within the textual SQL and the column arguments passed to TextClause.columns():

from sqlalchemy import text

stmt = text(
    "SELECT users.id, addresses.id, users.id, "
    "users.name, addresses.email_address AS email "
    "FROM users JOIN addresses ON users.id=addresses.user_id "
    "WHERE users.id = 1"
).columns(User.id, Address.id, Address.user_id, User.name, Address.email_address)

query = session.query(User).from_statement(stmt).options(contains_eager(User.addresses))
result = query.all()

Above, the textual SQL contains the column “id” three times, which would normally be ambiguous. Using the new feature, we can apply the mapped columns from the User and Address class directly, even linking the Address.user_id column to the users.id column in textual SQL for fun, and the Query object will receive rows that are correctly targetable as needed, including for an eager load.

This change is backwards incompatible with code that passes the columns to the method with a different ordering than is present in the textual statement. It is hoped that this impact will be low due to the fact that this method has always been documented illustrating the columns being passed in the same order as that of the textual SQL statement, as would seem intuitive, even though the internals weren’t checking for this. The method itself was only added as of 0.9 in any case and may not yet have widespread use. Notes on exactly how to handle this behavioral change for applications using it are at TextClause.columns() will match columns positionally, not by name, when passed positionally.

Positional matching is trusted over name-based matching for Core/ORM SQL constructs

Another aspect of this change is that the rules for matching columns have also been modified to rely upon “positional” matching more fully for compiled SQL constructs as well. Given a statement like the following:

ua = users.alias("ua")
stmt = select([users.c.user_id, ua.c.user_id])

The above statement will compile to:

SELECT users.user_id, ua.user_id FROM users, users AS ua

In 1.0, the above statement when executed would be matched to its original compiled construct using positional matching, however because the statement contains the 'user_id' label duplicated, the “ambiguous column” rule would still get involved and prevent the columns from being fetched from a row. As of 1.1, the “ambiguous column” rule does not affect an exact match from a column construct to the SQL column, which is what the ORM uses to fetch columns:

result = conn.execute(stmt)
row = result.first()

# these both match positionally, so no error
user_id = row[users.c.user_id]
ua_id = row[ua.c.user_id]

# this still raises, however
user_id = row["user_id"]

Much less likely to get an “ambiguous column” error message

As part of this change, the wording of the error message Ambiguous column name '<name>' in result set! try 'use_labels' option on select statement. has been dialed back; as this message should now be extremely rare when using the ORM or Core compiled SQL constructs, it merely states Ambiguous column name '<name>' in result set column descriptions, and only when a result column is retrieved using the string name that is actually ambiguous, e.g. row['user_id'] in the above example. It also now refers to the actual ambiguous name from the rendered SQL statement itself, rather than indicating the key or name that was local to the construct being used for the fetch.


Support for Python’s native enum type and compatible forms

The Enum type can now be constructed using any PEP-435 compliant enumerated type. When using this mode, input values and return values are the actual enumerated objects, not the string/integer/etc values:

import enum
from sqlalchemy import Table, MetaData, Column, Enum, create_engine

class MyEnum(enum.Enum):
    one = 1
    two = 2
    three = 3

t = Table("data", MetaData(), Column("value", Enum(MyEnum)))

e = create_engine("sqlite://")

e.execute(t.insert(), {"value": MyEnum.two})
assert e.scalar(t.select()) is MyEnum.two

The Enum.enums collection is now a list instead of a tuple

As part of the changes to Enum, the Enum.enums collection of elements is now a list instead of a tuple. This because lists are appropriate for variable length sequences of homogeneous items where the position of the element is not semantically significant.


Negative integer indexes accommodated by Core result rows

The RowProxy object now accommodates single negative integer indexes like a regular Python sequence, both in the pure Python and C-extension version. Previously, negative values would only work in slices:

>>> from sqlalchemy import create_engine
>>> e = create_engine("sqlite://")
>>> row = e.execute("select 1, 2, 3").first()
>>> row[-1], row[-2], row[1], row[-2:2]
3 2 2 (2,)

The Enum type now does in-Python validation of values

To accommodate for Python native enumerated objects, as well as for edge cases such as that of where a non-native ENUM type is used within an ARRAY and a CHECK constraint is infeasible, the Enum datatype now adds in-Python validation of input values when the Enum.validate_strings flag is used (1.1.0b2):

>>> from sqlalchemy import Table, MetaData, Column, Enum, create_engine
>>> t = Table(
...     "data",
...     MetaData(),
...     Column("value", Enum("one", "two", "three", validate_strings=True)),
... )
>>> e = create_engine("sqlite://")
>>> t.create(e)
>>> e.execute(t.insert(), {"value": "four"})
Traceback (most recent call last):
sqlalchemy.exc.StatementError: (exceptions.LookupError)
"four" is not among the defined enum values
[SQL: u'INSERT INTO data (value) VALUES (?)']
[parameters: [{'value': 'four'}]]

This validation is turned off by default as there are already use cases identified where users don’t want such validation (such as string comparisons). For non-string types, it necessarily takes place in all cases. The check also occurs unconditionally on the result-handling side as well, when values coming from the database are returned.

This validation is in addition to the existing behavior of creating a CHECK constraint when a non-native enumerated type is used. The creation of this CHECK constraint can now be disabled using the new Enum.create_constraint flag.


Non-native boolean integer values coerced to zero/one/None in all cases

The Boolean datatype coerces Python booleans to integer values for backends that don’t have a native boolean type, such as SQLite and MySQL. On these backends, a CHECK constraint is normally set up which ensures the values in the database are in fact one of these two values. However, MySQL ignores CHECK constraints, the constraint is optional, and an existing database might not have this constraint. The Boolean datatype has been repaired such that an incoming Python-side value that is already an integer value is coerced to zero or one, not just passed as-is; additionally, the C-extension version of the int-to-boolean processor for results now uses the same Python boolean interpretation of the value, rather than asserting an exact one or zero value. This is now consistent with the pure-Python int-to-boolean processor and is more forgiving of existing data already within the database. Values of None/NULL are as before retained as None/NULL.


this change had an unintended side effect that the interpretation of non- integer values, such as strings, also changed in behavior such that the string value "0" would be interpreted as “true”, but only on backends that don’t have a native boolean datatype - on “native boolean” backends like PostgreSQL, the string value "0" is passed directly to the driver and is interpreted as “false”. This is an inconsistency that did not occur with the previous implementation. It should be noted that passing strings or any other value outside of None, True, False, 1, 0 to the Boolean datatype is not supported and version 1.2 will raise an error for this scenario (or possibly just emit a warning, TBD). See also #4102.


Large parameter and row values are now truncated in logging and exception displays

A large value present as a bound parameter for a SQL statement, as well as a large value present in a result row, will now be truncated during display within logging, exception reporting, as well as repr() of the row itself:

>>> from sqlalchemy import create_engine
>>> import random
>>> e = create_engine("sqlite://", echo="debug")
>>> some_value = "".join(chr(random.randint(52, 85)) for i in range(5000))
>>> row = e.execute("select ?", [some_value]).first()
... # (lines are wrapped for clarity) ...
2016-02-17 13:23:03,027 INFO sqlalchemy.engine.base.Engine select ?
2016-02-17 13:23:03,027 INFO sqlalchemy.engine.base.Engine
GJ7HQ6 ... (4702 characters truncated) ... J6IK546AJMB4N6S9L;;9AKI;=RJP
2016-02-17 13:23:03,027 DEBUG sqlalchemy.engine.base.Engine Col ('?',)
2016-02-17 13:23:03,027 DEBUG sqlalchemy.engine.base.Engine
Row (u'E6@?>9HPOJB<<BHR:@=TS:5ILU=;JLM<4?B9<S48PTNG9>:=TSTLA;9K;9FPM4M8M@;
>4=4:PGJ7HQ ... (4703 characters truncated) ... J6IK546AJMB4N6S9L;;9AKI;=
>>> print(row)
=4:PGJ7HQ ... (4703 characters truncated) ... J6IK546AJMB4N6S9L;;9AKI;


JSON support added to Core

As MySQL now has a JSON datatype in addition to the PostgreSQL JSON datatype, the core now gains a sqlalchemy.types.JSON datatype that is the basis for both of these. Using this type allows access to the “getitem” operator as well as the “getpath” operator in a way that is agnostic across PostgreSQL and MySQL.

The new datatype also has a series of improvements to the handling of NULL values as well as expression handling.


JSON “null” is inserted as expected with ORM operations, omitted when not present

The JSON type and its descendant types JSON and JSON have a flag JSON.none_as_null which when set to True indicates that the Python value None should translate into a SQL NULL rather than a JSON NULL value. This flag defaults to False, which means that the Python value None should result in a JSON NULL value.

This logic would fail, and is now corrected, in the following circumstances:

1. When the column also contained a default or server_default value, a positive value of None on the mapped attribute that expects to persist JSON “null” would still result in the column-level default being triggered, replacing the None value:

class MyObject(Base):
    # ...

    json_value = Column(JSON(none_as_null=False), default="some default")

# would insert "some default" instead of "'null'",
# now will insert "'null'"
obj = MyObject(json_value=None)

2. When the column did not contain a default or server_default value, a missing value on a JSON column configured with none_as_null=False would still render JSON NULL rather than falling back to not inserting any value, behaving inconsistently vs. all other datatypes:

class MyObject(Base):
    # ...

    some_other_value = Column(String(50))
    json_value = Column(JSON(none_as_null=False))

# would result in NULL for some_other_value,
# but json "'null'" for json_value.  Now results in NULL for both
# (the json_value is omitted from the INSERT)
obj = MyObject()

This is a behavioral change that is backwards incompatible for an application that was relying upon this to default a missing value as JSON null. This essentially establishes that a missing value is distinguished from a present value of None. See JSON Columns will not insert JSON NULL if no value is supplied and no default is established for further detail.

3. When the Session.bulk_insert_mappings() method were used, None would be ignored in all cases:

# would insert SQL NULL and/or trigger defaults,
# now inserts "'null'"
session.bulk_insert_mappings(MyObject, [{"json_value": None}])

The JSON type now implements the TypeEngine.should_evaluate_none flag, indicating that None should not be ignored here; it is configured automatically based on the value of JSON.none_as_null. Thanks to #3061, we can differentiate when the value None is actively set by the user versus when it was never set at all.

The feature applies as well to the new base JSON type and its descendant types.


New JSON.NULL Constant Added

To ensure that an application can always have full control at the value level of whether a JSON, JSON, JSON, or JSONB column should receive a SQL NULL or JSON "null" value, the constant JSON.NULL has been added, which in conjunction with null() can be used to determine fully between SQL NULL and JSON "null", regardless of what JSON.none_as_null is set to:

from sqlalchemy import null
from sqlalchemy.dialects.postgresql import JSON

obj1 = MyObject(json_value=null())  # will *always* insert SQL NULL
obj2 = MyObject(json_value=JSON.NULL)  # will *always* insert JSON string "null"

session.add_all([obj1, obj2])

The feature applies as well to the new base JSON type and its descendant types.


Array support added to Core; new ANY and ALL operators

Along with the enhancements made to the PostgreSQL ARRAY type described in Correct SQL Types are Established from Indexed Access of ARRAY, JSON, HSTORE, the base class of ARRAY itself has been moved to Core in a new class ARRAY.

Arrays are part of the SQL standard, as are several array-oriented functions such as array_agg() and unnest(). In support of these constructs for not just PostgreSQL but also potentially for other array-capable backends in the future such as DB2, the majority of array logic for SQL expressions is now in Core. The ARRAY type still only works on PostgreSQL, however it can be used directly, supporting special array use cases such as indexed access, as well as support for the ANY and ALL:

mytable = Table("mytable", metadata, Column("data", ARRAY(Integer, dimensions=2)))

expr = mytable.c.data[5][6]

expr = mytable.c.data[5].any(12)

In support of ANY and ALL, the ARRAY type retains the same Comparator.any() and Comparator.all() methods from the PostgreSQL type, but also exports these operations to new standalone operator functions any_() and all_(). These two functions work in more of the traditional SQL way, allowing a right-side expression form such as:

from sqlalchemy import any_, all_

select([mytable]).where(12 == any_(mytable.c.data[5]))

For the PostgreSQL-specific operators “contains”, “contained_by”, and “overlaps”, one should continue to use the ARRAY type directly, which provides all functionality of the ARRAY type as well.

The any_() and all_() operators are open-ended at the Core level, however their interpretation by backend databases is limited. On the PostgreSQL backend, the two operators only accept array values. Whereas on the MySQL backend, they only accept subquery values. On MySQL, one can use an expression such as:

from sqlalchemy import any_, all_

subq = select([mytable.c.value])
select([mytable]).where(12 > any_(subq))


New Function features, “WITHIN GROUP”, array_agg and set aggregate functions

With the new ARRAY type we can also implement a pre-typed function for the array_agg() SQL function that returns an array, which is now available using array_agg:

from sqlalchemy import func

stmt = select([func.array_agg(table.c.value)])

A PostgreSQL element for an aggregate ORDER BY is also added via aggregate_order_by:

from sqlalchemy.dialects.postgresql import aggregate_order_by

expr = func.array_agg(aggregate_order_by(table.c.a, table.c.b.desc()))
stmt = select([expr])


SELECT array_agg(table1.a ORDER BY table1.b DESC) AS array_agg_1 FROM table1

The PG dialect itself also provides an array_agg() wrapper to ensure the ARRAY type:

from sqlalchemy.dialects.postgresql import array_agg

stmt = select([array_agg(table.c.value).contains("foo")])

Additionally, functions like percentile_cont(), percentile_disc(), rank(), dense_rank() and others that require an ordering via WITHIN GROUP (ORDER BY <expr>) are now available via the FunctionElement.within_group() modifier:

from sqlalchemy import func

stmt = select(

The above statement would produce SQL similar to:

SELECT department.id, percentile_cont(0.5)
WITHIN GROUP (ORDER BY department.salary DESC)

Placeholders with correct return types are now provided for these functions, and include percentile_cont, percentile_disc, rank, dense_rank, mode, percent_rank, and cume_dist.

#3132 #1370

TypeDecorator now works with Enum, Boolean, “schema” types automatically

The SchemaType types include types such as Enum and Boolean which, in addition to corresponding to a database type, also generate either a CHECK constraint or in the case of PostgreSQL ENUM a new CREATE TYPE statement, will now work automatically with TypeDecorator recipes. Previously, a TypeDecorator for an ENUM had to look like this:

# old way
class MyEnum(TypeDecorator, SchemaType):
    impl = postgresql.ENUM("one", "two", "three", name="myenum")

    def _set_table(self, table):

The TypeDecorator now propagates those additional events so it can be done like any other type:

# new way
class MyEnum(TypeDecorator):
    impl = postgresql.ENUM("one", "two", "three", name="myenum")


Multi-Tenancy Schema Translation for Table objects

To support the use case of an application that uses the same set of Table objects in many schemas, such as schema-per-user, a new execution option Connection.execution_options.schema_translate_map is added. Using this mapping, a set of Table objects can be made on a per-connection basis to refer to any set of schemas instead of the Table.schema to which they were assigned. The translation works for DDL and SQL generation, as well as with the ORM.

For example, if the User class were assigned the schema “per_user”:

class User(Base):
    __tablename__ = "user"
    id = Column(Integer, primary_key=True)

    __table_args__ = {"schema": "per_user"}

On each request, the Session can be set up to refer to a different schema each time:

session = Session()
    execution_options={"schema_translate_map": {"per_user": "account_one"}}

# will query from the ``account_one.user`` table


“Friendly” stringification of Core SQL constructs without a dialect

Calling str() on a Core SQL construct will now produce a string in more cases than before, supporting various SQL constructs not normally present in default SQL such as RETURNING, array indexes, and non-standard datatypes:

>>> from sqlalchemy import table, column
t>>> t = table('x', column('a'), column('b'))
>>> print(t.insert().returning(t.c.a, t.c.b))
INSERT INTO x (a, b) VALUES (:a, :b) RETURNING x.a, x.b

The str() function now calls upon an entirely separate dialect / compiler intended just for plain string printing without a specific dialect set up, so as more “just show me a string!” cases come up, these can be added to this dialect/compiler without impacting behaviors on real dialects.


The type_coerce function is now a persistent SQL element

The type_coerce() function previously would return an object either of type BindParameter or Label, depending on the input. An effect this would have was that in the case where expression transformations were used, such as the conversion of an element from a Column to a BindParameter that’s critical to ORM-level lazy loading, the type coercion information would not be used since it would have been lost already.

To improve this behavior, the function now returns a persistent TypeCoerce container around the given expression, which itself remains unaffected; this construct is evaluated explicitly by the SQL compiler. This allows for the coercion of the inner expression to be maintained no matter how the statement is modified, including if the contained element is replaced with a different one, as is common within the ORM’s lazy loading feature.

The test case illustrating the effect makes use of a heterogeneous primaryjoin condition in conjunction with custom types and lazy loading. Given a custom type that applies a CAST as a “bind expression”:

class StringAsInt(TypeDecorator):
    impl = String

    def column_expression(self, col):
        return cast(col, Integer)

    def bind_expression(self, value):
        return cast(value, String)

Then, a mapping where we are equating a string “id” column on one table to an integer “id” column on the other:

class Person(Base):
    __tablename__ = "person"
    id = Column(StringAsInt, primary_key=True)

    pets = relationship(
            "foreign(Pets.person_id)==cast(type_coerce(Person.id, Integer), Integer)"

class Pets(Base):
    __tablename__ = "pets"
    id = Column("id", Integer, primary_key=True)
    person_id = Column("person_id", Integer)

Above, in the relationship.primaryjoin expression, we are using type_coerce() to handle bound parameters passed via lazyloading as integers, since we already know these will come from our StringAsInt type which maintains the value as an integer in Python. We are then using cast() so that as a SQL expression, the VARCHAR “id” column will be CAST to an integer for a regular non- converted join as with Query.join() or joinedload(). That is, a joinedload of .pets looks like:

SELECT person.id AS person_id, pets_1.id AS pets_1_id,
       pets_1.person_id AS pets_1_person_id
FROM person
LEFT OUTER JOIN pets AS pets_1
ON pets_1.person_id = CAST(person.id AS INTEGER)

Without the CAST in the ON clause of the join, strongly-typed databases such as PostgreSQL will refuse to implicitly compare the integer and fail.

The lazyload case of .pets relies upon replacing the Person.id column at load time with a bound parameter, which receives a Python-loaded value. This replacement is specifically where the intent of our type_coerce() function would be lost. Prior to the change, this lazy load comes out as:

SELECT pets.id AS pets_id, pets.person_id AS pets_person_id
FROM pets
WHERE pets.person_id = CAST(CAST(%(param_1)s AS VARCHAR) AS INTEGER)
-- {'param_1': 5}

Where above, we see that our in-Python value of 5 is CAST first to a VARCHAR, then back to an INTEGER in SQL; a double CAST which works, but is nevertheless not what we asked for.

With the change, the type_coerce() function maintains a wrapper even after the column is swapped out for a bound parameter, and the query now looks like:

SELECT pets.id AS pets_id, pets.person_id AS pets_person_id
FROM pets
WHERE pets.person_id = CAST(%(param_1)s AS INTEGER)
-- {'param_1': 5}

Where our outer CAST that’s in our primaryjoin still takes effect, but the needless CAST that’s in part of the StringAsInt custom type is removed as intended by the type_coerce() function.


Key Behavioral Changes - ORM

JSON Columns will not insert JSON NULL if no value is supplied and no default is established

As detailed in JSON “null” is inserted as expected with ORM operations, omitted when not present, JSON will not render a JSON “null” value if the value is missing entirely. To prevent SQL NULL, a default should be set up. Given the following mapping:

class MyObject(Base):
    # ...

    json_value = Column(JSON(none_as_null=False), nullable=False)

The following flush operation will fail with an integrity error:

obj = MyObject()  # note no json_value
session.commit()  # will fail with integrity error

If the default for the column should be JSON NULL, set this on the Column:

class MyObject(Base):
    # ...

    json_value = Column(JSON(none_as_null=False), nullable=False, default=JSON.NULL)

Or, ensure the value is present on the object:

obj = MyObject(json_value=None)
session.commit()  # will insert JSON NULL

Note that setting None for the default is the same as omitting it entirely; the JSON.none_as_null flag does not impact the value of None passed to Column.default or Column.server_default:

# default=None is the same as omitting it entirely, does not apply JSON NULL
json_value = Column(JSON(none_as_null=False), nullable=False, default=None)

Columns no longer added redundantly with DISTINCT + ORDER BY

A query such as the following will now augment only those columns that are missing from the SELECT list, without duplicates:

q = (
    session.query(User.id, User.name.label("name"))
    .order_by(User.id, User.name, User.fullname)


SELECT DISTINCT user.id AS a_id, user.name AS name,
 user.fullname AS a_fullname
FROM a ORDER BY user.id, user.name, user.fullname

Previously, it would produce:

SELECT DISTINCT user.id AS a_id, user.name AS name, user.name AS a_name,
  user.fullname AS a_fullname
FROM a ORDER BY user.id, user.name, user.fullname

Where above, the user.name column is added unnecessarily. The results would not be affected, as the additional columns are not included in the result in any case, but the columns are unnecessary.

Additionally, when the PostgreSQL DISTINCT ON format is used by passing expressions to Query.distinct(), the above “column adding” logic is disabled entirely.

When the query is being bundled into a subquery for the purposes of joined eager loading, the “augment column list” rules are necessarily more aggressive so that the ORDER BY can still be satisfied, so this case remains unchanged.


Same-named @validates decorators will now raise an exception

The validates() decorator is only intended to be created once per class for a particular attribute name. Creating more than one now raises an error, whereas previously it would silently pick only the last defined validator:

class A(Base):
    __tablename__ = "a"
    id = Column(Integer, primary_key=True)

    data = Column(String)

    def _validate_data_one(self):
        assert "x" in data

    def _validate_data_two(self):
        assert "y" in data


Will raise:

sqlalchemy.exc.InvalidRequestError: A validation function for mapped attribute 'data'
on mapper Mapper|A|a already exists.


Key Behavioral Changes - Core

TextClause.columns() will match columns positionally, not by name, when passed positionally

The new behavior of the TextClause.columns() method, which itself was recently added as of the 0.9 series, is that when columns are passed positionally without any additional keyword arguments, they are linked to the ultimate result set columns positionally, and no longer on name. It is hoped that the impact of this change will be low due to the fact that the method has always been documented illustrating the columns being passed in the same order as that of the textual SQL statement, as would seem intuitive, even though the internals weren’t checking for this.

An application that is using this method by passing Column objects to it positionally must ensure that the position of those Column objects matches the position in which these columns are stated in the textual SQL.

E.g., code like the following:

stmt = text("SELECT id, name, description FROM table")

# no longer matches by name
stmt = stmt.columns(my_table.c.name, my_table.c.description, my_table.c.id)

Would no longer work as expected; the order of the columns given is now significant:

# correct version
stmt = stmt.columns(my_table.c.id, my_table.c.name, my_table.c.description)

Possibly more likely, a statement that worked like this:

stmt = text("SELECT * FROM table")
stmt = stmt.columns(my_table.c.id, my_table.c.name, my_table.c.description)

is now slightly risky, as the “*” specification will generally deliver columns in the order in which they are present in the table itself. If the structure of the table changes due to schema changes, this ordering may no longer be the same. Therefore when using TextClause.columns(), it’s advised to list out the desired columns explicitly in the textual SQL, though it’s no longer necessary to worry about the names themselves in the textual SQL.

String server_default now literal quoted

A server default passed to Column.server_default as a plain Python string that has quotes embedded is now passed through the literal quoting system:

>>> from sqlalchemy.schema import MetaData, Table, Column, CreateTable
>>> from sqlalchemy.types import String
>>> t = Table("t", MetaData(), Column("x", String(), server_default="hi ' there"))
>>> print(CreateTable(t))
CREATE TABLE t ( x VARCHAR DEFAULT 'hi '' there' )

Previously the quote would render directly. This change may be backwards incompatible for applications with such a use case who were working around the issue.


A UNION or similar of SELECTs with LIMIT/OFFSET/ORDER BY now parenthesizes the embedded selects

An issue that, like others, was long driven by SQLite’s lack of capabilities has now been enhanced to work on all supporting backends. We refer to a query that is a UNION of SELECT statements that themselves contain row-limiting or ordering features which include LIMIT, OFFSET, and/or ORDER BY:


The above query requires parenthesis within each sub-select in order to group the sub-results correctly. Production of the above statement in SQLAlchemy Core looks like:

stmt1 = select([table1.c.x]).order_by(table1.c.y).limit(1)
stmt2 = select([table1.c.x]).order_by(table2.c.y).limit(2)

stmt = union(stmt1, stmt2)

Previously, the above construct would not produce parenthesization for the inner SELECT statements, producing a query that fails on all backends.

The above formats will continue to fail on SQLite; additionally, the format that includes ORDER BY but no LIMIT/SELECT will continue to fail on Oracle. This is not a backwards-incompatible change, because the queries fail without the parentheses as well; with the fix, the queries at least work on all other databases.

In all cases, in order to produce a UNION of limited SELECT statements that also works on SQLite and in all cases on Oracle, the subqueries must be a SELECT of an ALIAS:

stmt1 = select([table1.c.x]).order_by(table1.c.y).limit(1).alias().select()
stmt2 = select([table2.c.x]).order_by(table2.c.y).limit(2).alias().select()

stmt = union(stmt1, stmt2)

This workaround works on all SQLAlchemy versions. In the ORM, it looks like:

stmt1 = session.query(Model1).order_by(Model1.y).limit(1).subquery().select()
stmt2 = session.query(Model2).order_by(Model2.y).limit(1).subquery().select()

stmt = session.query(Model1).from_statement(stmt1.union(stmt2))

The behavior here has many parallels to the “join rewriting” behavior introduced in SQLAlchemy 0.9 in Many JOIN and LEFT OUTER JOIN expressions will no longer be wrapped in (SELECT * FROM ..) AS ANON_1; however in this case we have opted not to add new rewriting behavior to accommodate this case for SQLite. The existing rewriting behavior is very complicated already, and the case of UNIONs with parenthesized SELECT statements is much less common than the “right-nested-join” use case of that feature.


Dialect Improvements and Changes - PostgreSQL


The ON CONFLICT clause of INSERT added to PostgreSQL as of version 9.5 is now supported using a PostgreSQL-specific version of the Insert object, via sqlalchemy.dialects.postgresql.dml.insert(). This Insert subclass adds two new methods Insert.on_conflict_do_update() and Insert.on_conflict_do_nothing() which implement the full syntax supported by PostgreSQL 9.5 in this area:

from sqlalchemy.dialects.postgresql import insert

insert_stmt = insert(my_table).values(id="some_id", data="some data to insert")

do_update_stmt = insert_stmt.on_conflict_do_update(
    index_elements=[my_table.c.id], set_=dict(data="some data to update")


The above will render:

INSERT INTO my_table (id, data)
VALUES (:id, :data)


ARRAY and JSON types now correctly specify “unhashable”

As described in Changes regarding “unhashable” types, impacts deduping of ORM rows, the ORM relies upon being able to produce a hash function for column values when a query’s selected entities mixes full ORM entities with column expressions. The hashable=False flag is now correctly set on all of PG’s “data structure” types, including ARRAY and JSON. The JSONB and HSTORE types already included this flag. For ARRAY, this is conditional based on the ARRAY.as_tuple flag, however it should no longer be necessary to set this flag in order to have an array value present in a composed ORM row.


Correct SQL Types are Established from Indexed Access of ARRAY, JSON, HSTORE

For all three of ARRAY, JSON and HSTORE, the SQL type assigned to the expression returned by indexed access, e.g. col[someindex], should be correct in all cases.

This includes:

  • The SQL type assigned to indexed access of an ARRAY takes into account the number of dimensions configured. An ARRAY with three dimensions will return a SQL expression with a type of ARRAY of one less dimension. Given a column with type ARRAY(Integer, dimensions=3), we can now perform this expression:

    int_expr = col[5][6][7]  # returns an Integer expression object

    Previously, the indexed access to col[5] would return an expression of type Integer where we could no longer perform indexed access for the remaining dimensions, unless we used cast() or type_coerce().

  • The JSON and JSONB types now mirror what PostgreSQL itself does for indexed access. This means that all indexed access for a JSON or JSONB type returns an expression that itself is always JSON or JSONB itself, unless the Comparator.astext modifier is used. This means that whether the indexed access of the JSON structure ultimately refers to a string, list, number, or other JSON structure, PostgreSQL always considers it to be JSON itself unless it is explicitly cast differently. Like the ARRAY type, this means that it is now straightforward to produce JSON expressions with multiple levels of indexed access:

    json_expr = json_col["key1"]["attr1"][5]
  • The “textual” type that is returned by indexed access of HSTORE as well as the “textual” type that is returned by indexed access of JSON and JSONB in conjunction with the Comparator.astext modifier is now configurable; it defaults to TextClause in both cases but can be set to a user-defined type using the JSON.astext_type or HSTORE.text_type parameters.

#3499 #3487

The JSON cast() operation now requires .astext is called explicitly

As part of the changes in Correct SQL Types are Established from Indexed Access of ARRAY, JSON, HSTORE, the workings of the ColumnElement.cast() operator on JSON and JSONB no longer implicitly invoke the Comparator.astext modifier; PostgreSQL’s JSON/JSONB types support CAST operations to each other without the “astext” aspect.

This means that in most cases, an application that was doing this:

expr = json_col["somekey"].cast(Integer)

Will now need to change to this:

expr = json_col["somekey"].astext.cast(Integer)

ARRAY with ENUM will now emit CREATE TYPE for the ENUM

A table definition like the following will now emit CREATE TYPE as expected:

enum = Enum(

class WorkPlacement(Base):
    __tablename__ = "work_placement"
    id = Column(Integer, primary_key=True)
    roles = Column(ARRAY(enum))

e = create_engine("postgresql://scott:tiger@localhost/test", echo=True)


CREATE TYPE work_place_roles AS ENUM (
    'manager', 'place_admin', 'carwash_admin', 'parking_admin',
    'service_admin', 'tire_admin', 'mechanic', 'carwasher',

CREATE TABLE work_placement (
    roles work_place_roles[],
    PRIMARY KEY (id)


Check constraints now reflect

The PostgreSQL dialect now supports reflection of CHECK constraints both within the method Inspector.get_check_constraints() as well as within Table reflection within the Table.constraints collection.

“Plain” and “Materialized” views can be inspected separately

The new argument PGInspector.get_view_names.include allows specification of which sub-types of views should be returned:

from sqlalchemy import inspect

insp = inspect(engine)

plain_views = insp.get_view_names(include="plain")
all_views = insp.get_view_names(include=("plain", "materialized"))


Added tablespace option to Index

The Index object now accepts the argument postgresql_tablespace in order to specify TABLESPACE, the same way as accepted by the Table object.


Support for PyGreSQL

The PyGreSQL DBAPI is now supported.

The “postgres” module is removed

The sqlalchemy.dialects.postgres module, long deprecated, is removed; this has emitted a warning for many years and projects should be calling upon sqlalchemy.dialects.postgresql. Engine URLs of the form postgres:// will still continue to function, however.


The new parameters GenerativeSelect.with_for_update.skip_locked and GenerativeSelect.with_for_update.key_share in both Core and ORM apply a modification to a “SELECT…FOR UPDATE” or “SELECT…FOR SHARE” query on the PostgreSQL backend:


    stmt = select([table]).with_for_update(key_share=True)

    stmt = select([table]).with_for_update(skip_locked=True)

    stmt = select([table]).with_for_update(read=True, key_share=True)

Dialect Improvements and Changes - MySQL

MySQL JSON Support

A new type JSON is added to the MySQL dialect supporting the JSON type newly added to MySQL 5.7. This type provides both persistence of JSON as well as rudimentary indexed-access using the JSON_EXTRACT function internally. An indexable JSON column that works across MySQL and PostgreSQL can be achieved by using the JSON datatype common to both MySQL and PostgreSQL.


Added support for AUTOCOMMIT “isolation level”

The MySQL dialect now accepts the value “AUTOCOMMIT” for the create_engine.isolation_level and Connection.execution_options.isolation_level parameters:

connection = engine.connect()
connection = connection.execution_options(isolation_level="AUTOCOMMIT")

The isolation level makes use of the various “autocommit” attributes provided by most MySQL DBAPIs.


No more generation of an implicit KEY for composite primary key w/ AUTO_INCREMENT

The MySQL dialect had the behavior such that if a composite primary key on an InnoDB table featured AUTO_INCREMENT on one of its columns which was not the first column, e.g.:

t = Table(
    Column("x", Integer, primary_key=True, autoincrement=False),
    Column("y", Integer, primary_key=True, autoincrement=True),

DDL such as the following would be generated:

CREATE TABLE some_table (
    PRIMARY KEY (x, y),
    KEY idx_autoinc_y (y)

Note the above “KEY” with an auto-generated name; this is a change that found its way into the dialect many years ago in response to the issue that the AUTO_INCREMENT would otherwise fail on InnoDB without this additional KEY.

This workaround has been removed and replaced with the much better system of just stating the AUTO_INCREMENT column first within the primary key:

CREATE TABLE some_table (
    PRIMARY KEY (y, x)

To maintain explicit control of the ordering of primary key columns, use the PrimaryKeyConstraint construct explicitly (1.1.0b2) (along with a KEY for the autoincrement column as required by MySQL), e.g.:

t = Table(
    Column("x", Integer, primary_key=True),
    Column("y", Integer, primary_key=True, autoincrement=True),
    PrimaryKeyConstraint("x", "y"),

Along with the change The .autoincrement directive is no longer implicitly enabled for a composite primary key column, composite primary keys with or without auto increment are now easier to specify; Column.autoincrement now defaults to the value "auto" and the autoincrement=False directives are no longer needed:

t = Table(
    Column("x", Integer, primary_key=True),
    Column("y", Integer, primary_key=True, autoincrement=True),

Dialect Improvements and Changes - SQLite

Right-nested join workaround lifted for SQLite version 3.7.16

In version 0.9, the feature introduced by Many JOIN and LEFT OUTER JOIN expressions will no longer be wrapped in (SELECT * FROM ..) AS ANON_1 went through lots of effort to support rewriting of joins on SQLite to always use subqueries in order to achieve a “right-nested-join” effect, as SQLite has not supported this syntax for many years. Ironically, the version of SQLite noted in that migration note,, was the last version of SQLite to actually have this limitation! The next release was 3.7.16 and support for right nested joins was quietly added. In 1.1, the work to identify the specific SQLite version and source commit where this change was made was done (SQLite’s changelog refers to it with the cryptic phrase “Enhance the query optimizer to exploit transitive join constraints” without linking to any issue number, change number, or further explanation), and the workarounds present in this change are now lifted for SQLite when the DBAPI reports that version 3.7.16 or greater is in effect.


Dotted column names workaround lifted for SQLite version 3.10.0

The SQLite dialect has long had a workaround for an issue where the database driver does not report the correct column names for some SQL result sets, in particular when UNION is used. The workaround is detailed at Dotted Column Names, and requires that SQLAlchemy assume that any column name with a dot in it is actually a tablename.columnname combination delivered via this buggy behavior, with an option to turn it off via the sqlite_raw_colnames execution option.

As of SQLite version 3.10.0, the bug in UNION and other queries has been fixed; like the change described in Right-nested join workaround lifted for SQLite version 3.7.16, SQLite’s changelog only identifies it cryptically as “Added the colUsed field to sqlite3_index_info for use by the sqlite3_module.xBestIndex method”, however SQLAlchemy’s translation of these dotted column names is no longer required with this version, so is turned off when version 3.10.0 or greater is detected.

Overall, the SQLAlchemy ResultProxy as of the 1.0 series relies much less on column names in result sets when delivering results for Core and ORM SQL constructs, so the importance of this issue was already lessened in any case.


Improved Support for Remote Schemas

The SQLite dialect now implements Inspector.get_schema_names() and additionally has improved support for tables and indexes that are created and reflected from a remote schema, which in SQLite is a database that is assigned a name via the ATTACH statement; previously, the``CREATE INDEX`` DDL didn’t work correctly for a schema-bound table and the Inspector.get_foreign_keys() method will now indicate the given schema in the results. Cross-schema foreign keys aren’t supported.

Reflection of the name of PRIMARY KEY constraints

The SQLite backend now takes advantage of the “sqlite_master” view of SQLite in order to extract the name of the primary key constraint of a table from the original DDL, in the same way that is achieved for foreign key constraints in recent SQLAlchemy versions.


Check constraints now reflect

The SQLite dialect now supports reflection of CHECK constraints both within the method Inspector.get_check_constraints() as well as within Table reflection within the Table.constraints collection.

ON DELETE and ON UPDATE foreign key phrases now reflect

The Inspector will now include ON DELETE and ON UPDATE phrases from foreign key constraints on the SQLite dialect, and the ForeignKeyConstraint object as reflected as part of a Table will also indicate these phrases.

Dialect Improvements and Changes - SQL Server

Added transaction isolation level support for SQL Server

All SQL Server dialects support transaction isolation level settings via the create_engine.isolation_level and Connection.execution_options.isolation_level parameters. The four standard levels are supported as well as SNAPSHOT:

engine = create_engine(
    "mssql+pyodbc://scott:tiger@ms_2008", isolation_level="REPEATABLE READ"


String / varlength types no longer represent “max” explicitly on reflection

When reflecting a type such as String, TextClause, etc. which includes a length, an “un-lengthed” type under SQL Server would copy the “length” parameter as the value "max":

>>> from sqlalchemy import create_engine, inspect
>>> engine = create_engine("mssql+pyodbc://scott:tiger@ms_2008", echo=True)
>>> engine.execute("create table s (x varchar(max), y varbinary(max))")
>>> insp = inspect(engine)
>>> for col in insp.get_columns("s"):
...     print(col["type"].__class__, col["type"].length)
<class 'sqlalchemy.sql.sqltypes.VARCHAR'> max
<class 'sqlalchemy.dialects.mssql.base.VARBINARY'> max

The “length” parameter in the base types is expected to be an integer value or None only; None indicates unbounded length which the SQL Server dialect interprets as “max”. The fix then is so that these lengths come out as None, so that the type objects work in non-SQL Server contexts:

>>> for col in insp.get_columns("s"):
...     print(col["type"].__class__, col["type"].length)
<class 'sqlalchemy.sql.sqltypes.VARCHAR'> None
<class 'sqlalchemy.dialects.mssql.base.VARBINARY'> None

Applications which may have been relying on a direct comparison of the “length” value to the string “max” should consider the value of None to mean the same thing.


Support for “non clustered” on primary key to allow clustered elsewhere

The mssql_clustered flag available on UniqueConstraint, PrimaryKeyConstraint, Index now defaults to None, and can be set to False which will render the NONCLUSTERED keyword in particular for a primary key, allowing a different index to be used as “clustered”.

The legacy_schema_aliasing flag is now set to False

SQLAlchemy 1.0.5 introduced the legacy_schema_aliasing flag to the MSSQL dialect, allowing so-called “legacy mode” aliasing to be turned off. This aliasing attempts to turn schema-qualified tables into aliases; given a table such as:

account_table = Table(
    Column("id", Integer, primary_key=True),
    Column("info", String(100)),

The legacy mode of behavior will attempt to turn a schema-qualified table name into an alias:

>>> eng = create_engine("mssql+pymssql://mydsn", legacy_schema_aliasing=True)
>>> print(account_table.select().compile(eng))
SELECT account_1.id, account_1.info FROM customer_schema.account AS account_1

However, this aliasing has been shown to be unnecessary and in many cases produces incorrect SQL.

In SQLAlchemy 1.1, the legacy_schema_aliasing flag now defaults to False, disabling this mode of behavior and allowing the MSSQL dialect to behave normally with schema-qualified tables. For applications which may rely on this behavior, set the flag back to True.


Dialect Improvements and Changes - Oracle

Support for SKIP LOCKED

The new parameter GenerativeSelect.with_for_update.skip_locked in both Core and ORM will generate the “SKIP LOCKED” suffix for a “SELECT…FOR UPDATE” or “SELECT.. FOR SHARE” query.