What’s New in SQLAlchemy 1.3?

About this Document

This document describes changes between SQLAlchemy version 1.2 and SQLAlchemy version 1.3.


This guide introduces what’s new in SQLAlchemy version 1.3 and also documents changes which affect users migrating their applications from the 1.2 series of SQLAlchemy to 1.3.

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


Deprecation warnings are emitted for all deprecated elements; new deprecations added

Release 1.3 ensures that all behaviors and APIs that are deprecated, including all those that have been long listed as “legacy” for years, are emitting DeprecationWarning warnings. This includes when making use of parameters such as Session.weak_identity_map and classes such as MapperExtension. While all deprecations have been noted in the documentation, often they did not use a proper restructured text directive, or include in what version they were deprecated. Whether or not a particular API feature actually emitted a deprecation warning was not consistent. The general attitude was that most or all of these deprecated features were treated as long-term legacy features with no plans to remove them.

The change includes that all documented deprecations now use a proper restructured text directive in the documentation with a version number, the verbiage that the feature or use case will be removed in a future release is made explicit (e.g., no more legacy forever use cases), and that use of any such feature or use case will definitely emit a DeprecationWarning, which in Python 3 as well as when using modern testing tools like Pytest are now made more explicit in the standard error stream. The goal is that these long deprecated features, going back as far as version 0.7 or 0.6, should start being removed entirely, rather than keeping them around as “legacy” features. Additionally, some major new deprecations are being added as of version 1.3. As SQLAlchemy has 14 years of real world use by thousands of developers, it’s possible to point to a single stream of use cases that blend together well, and to trim away features and patterns that work against this single way of working.

The larger context is that SQLAlchemy seeks to adjust to the coming Python 3-only world, as well as a type-annotated world, and towards this goal there are tentative plans for a major rework of SQLAlchemy which would hopefully greatly reduce the cognitive load of the API as well as perform a major pass over the great many differences in implementation and use between Core and ORM. As these two systems evolved dramatically after SQLAlchemy’s first release, in particular the ORM still retains lots of “bolted on” behaviors that keep the wall of separation between Core and ORM too high. By focusing the API ahead of time on a single pattern for each supported use case, the eventual job of migrating to a significantly altered API becomes simpler.

For the most major deprecations being added in 1.3, see the linked sections below.


New Features and Improvements - ORM

Relationship to AliasedClass replaces the need for non primary mappers

The “non primary mapper” is a mapper() created in the Classical Mappings style, which acts as an additional mapper against an already mapped class against a different kind of selectable. The non primary mapper has its roots in the 0.1, 0.2 series of SQLAlchemy where it was anticipated that the mapper() object was to be the primary query construction interface, before the Query object existed.

With the advent of Query and later the AliasedClass construct, most use cases for the non primary mapper went away. This was a good thing since SQLAlchemy also moved away from “classical” mappings altogether around the 0.5 series in favor of the declarative system.

One use case remained around for non primary mappers when it was realized that some very hard-to-define relationship() configurations could be made possible when a non-primary mapper with an alternative selectable was made as the mapping target, rather than trying to construct a relationship.primaryjoin that encompassed all the complexity of a particular inter-object relationship.

As this use case became more popular, its limitations became apparent, including that the non primary mapper is difficult to configure against a selectable that adds new columns, that the mapper does not inherit the relationships of the original mapping, that relationships which are configured explicitly on the non primary mapper do not function well with loader options, and that the non primary mapper also doesn’t provide a fully functional namespace of column-based attributes which can be used in queries (which again, in the old 0.1 - 0.4 days, one would use Table objects directly with the ORM).

The missing piece was to allow the relationship() to refer directly to the AliasedClass. The AliasedClass already does everything we want the non primary mapper to do; it allows an existing mapped class to be loaded from an alternative selectable, it inherits all the attributes and relationships of the existing mapper, it works extremely well with loader options, and it provides a class-like object that can be mixed into queries just like the class itself. With this change, the recipes that were formerly for non primary mappers at Configuring how Relationship Joins are changed to aliased class.

At Relationship to Aliased Class, the original non primary mapper looked like:

j = join(B, D, D.b_id == B.id).join(C, C.id == D.c_id)

B_viacd = mapper(
    B, j, non_primary=True, primary_key=[j.c.b_id],
        "id": j.c.b_id,  # so that 'id' looks the same as before
        "c_id": j.c.c_id,   # needed for disambiguation
        "d_c_id": j.c.d_c_id,  # needed for disambiguation
        "b_id": [j.c.b_id, j.c.d_b_id],
        "d_id": j.c.d_id,

A.b = relationship(B_viacd, primaryjoin=A.b_id == B_viacd.c.b_id)

The properties were necessary in order to re-map the additional columns so that they did not conflict with the existing columns mapped to B, as well as it was necessary to define a new primary key.

With the new approach, all of this verbosity goes away, and the additional columns are referred towards directly when making the relationship:

j = join(B, D, D.b_id == B.id).join(C, C.id == D.c_id)

B_viacd = aliased(B, j, flat=True)

A.b = relationship(B_viacd, primaryjoin=A.b_id == j.c.b_id)

The non primary mapper is now deprecated with the eventual goal to be that classical mappings as a feature go away entirely. The Declarative API would become the single means of mapping which hopefully will allow internal improvements and simplifications, as well as a clearer documentation story.


selectin loading no longer uses JOIN for simple one-to-many

The “selectin” loading feature added in 1.2 introduced an extremely performant new way to eagerly load collections, in many cases much faster than that of “subquery” eager loading, as it does not rely upon restating the original SELECT query and instead uses a simple IN clause. However, the “selectin” load still relied upon rendering a JOIN between the parent and related tables, since it needs the parent primary key values in the row in order to match rows up. In 1.3, a new optimization is added which will omit this JOIN in the most common case of a simple one-to-many load, where the related row already contains the primary key of the parent row expressed in its foreign key columns. This again provides for a dramatic performance improvement as the ORM now can load large numbers of collections all in one query without using JOIN or subqueries at all.

Given a mapping:

class A(Base):
    __tablename__ = 'a'

    id = Column(Integer, primary_key=True)
    bs = relationship("B", lazy="selectin")

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

In the 1.2 version of “selectin” loading, a load of A to B looks like:

SELECT a.id AS a_id FROM a
SELECT a_1.id AS a_1_id, b.id AS b_id, b.a_id AS b_a_id
FROM a AS a_1 JOIN b ON a_1.id = b.a_id
WHERE a_1.id IN (?, ?, ?, ?, ?, ?, ?, ?, ?, ?) ORDER BY a_1.id
(1, 2, 3, 4, 5, 6, 7, 8, 9, 10)

With the new behavior, the load looks like:

SELECT a.id AS a_id FROM a
SELECT b.a_id AS b_a_id, b.id AS b_id FROM b
WHERE b.a_id IN (?, ?, ?, ?, ?, ?, ?, ?, ?, ?) ORDER BY b.a_id
(1, 2, 3, 4, 5, 6, 7, 8, 9, 10)

The behavior is being released as automatic, using a similar heuristic that lazy loading uses in order to determine if related entities can be fetched directly from the identity map. However, as with most querying features, the feature’s implementation became more complex as a result of advanced scenarios regarding polymorphic loading. If problems are encountered, users should report a bug, however the change also includes a flag relationship.omit_join which can be set to False on the relationship() to disable the optimization.


Improvement to the behavior of many-to-one query expressions

When building a query that compares a many-to-one relationship to an object value, such as:

u1 = session.query(User).get(5)

query = session.query(Address).filter(Address.user == u1)

The above expression Address.user == u1, which ultimately compiles to a SQL expression normally based on the primary key columns of the User object like "address.user_id = 5", uses a deferred callable in order to retrieve the value 5 within the bound expression until as late as possible. This is to suit both the use case where the Address.user == u1 expression may be against a User object that isn’t flushed yet which relies upon a server- generated primary key value, as well as that the expression always returns the correct result even if the primary key value of u1 has been changed since the expression was created.

However, a side effect of this behavior is that if u1 ends up being expired by the time the expression is evaluated, it results in an additional SELECT statement, and in the case that u1 was also detached from the Session, it would raise an error:

u1 = session.query(User).get(5)

query = session.query(Address).filter(Address.user == u1)


query.all()  # <-- would raise DetachedInstanceError

The expiration / expunging of the object can occur implicitly when the Session is committed and the u1 instance falls out of scope, as the Address.user == u1 expression does not strongly reference the object itself, only its InstanceState.

The fix is to allow the Address.user == u1 expression to evaluate the value 5 based on attempting to retrieve or load the value normally at expression compilation time as it does now, but if the object is detached and has been expired, it is retrieved from a new mechanism upon the InstanceState which will memoize the last known value for a particular attribute on that state when that attribute is expired. This mechanism is only enabled for a specific attribute / InstanceState when needed by the expression feature to conserve performance / memory overhead.

Originally, simpler approaches such as evaluating the expression immediately with various arrangements for trying to load the value later if not present were attempted, however the difficult edge case is that of the value of a column attribute (typically a natural primary key) that is being changed. In order to ensure that an expression like Address.user == u1 always returns the correct answer for the current state of u1, it will return the current database-persisted value for a persistent object, unexpiring via SELECT query if necessary, and for a detached object it will return the most recent known value, regardless of when the object was expired using a new feature within the InstanceState that tracks the last known value of a column attribute whenever the attribute is to be expired.

Modern attribute API features are used to indicate specific error messages when the value cannot be evaluated, the two cases of which are when the column attributes have never been set, and when the object was already expired when the first evaluation was made and is now detached. In all cases, DetachedInstanceError is no longer raised.


Many-to-one replacement won’t raise for “raiseload” or detached for “old” object

Given the case where a lazy load would proceed on a many-to-one relationship in order to load the “old” value, if the relationship does not specify the relationship.active_history flag, an assertion will not be raised for a detached object:

a1 = session.query(Address).filter_by(id=5).one()


a1.user = some_user

Above, when the .user attribute is replaced on the detached a1 object, a DetachedInstanceError would be raised as the attribute is attempting to retrieve the previous value of .user from the identity map. The change is that the operation now proceeds without the old value being loaded.

The same change is also made to the lazy="raise" loader strategy:

class Address(Base):
    # ...

    user = relationship("User", ..., lazy="raise")

Previously, the association of a1.user would invoke the “raiseload” exception as a result of the attribute attempting to retrieve the previous value. This assertion is now skipped in the case of loading the “old” value.


“del” implemented for ORM attributes

The Python del operation was not really usable for mapped attributes, either scalar columns or object references. Support has been added for this to work correctly, where the del operation is roughly equivalent to setting the attribute to the None value:

some_object = session.query(SomeObject).get(5)

del some_object.some_attribute   # from a SQL perspective, works like "= None"


info dictionary added to InstanceState

Added the .info dictionary to the InstanceState class, the object that comes from calling inspect() on a mapped object. This allows custom recipes to add additional information about an object that will be carried along with that object’s full lifecycle in memory:

from sqlalchemy import inspect

u1 = User(id=7, name='ed')

inspect(u1).info['user_info'] = '7|ed'


Horizontal Sharding extension supports bulk update and delete methods

The ShardedQuery extension object supports the Query.update() and Query.delete() bulk update/delete methods. The query_chooser callable is consulted when they are called in order to run the update/delete across multiple shards based on given criteria.


Association Proxy Improvements

While not for any particular reason, the Association Proxy extension had many improvements this cycle.

Association proxy has new cascade_scalar_deletes flag

Given a mapping as:

class A(Base):
    __tablename__ = 'test_a'
    id = Column(Integer, primary_key=True)
    ab = relationship(
        'AB', backref='a', uselist=False)
    b = association_proxy(
        'ab', 'b', creator=lambda b: AB(b=b),

class B(Base):
    __tablename__ = 'test_b'
    id = Column(Integer, primary_key=True)
    ab = relationship('AB', backref='b', cascade='all, delete-orphan')

class AB(Base):
    __tablename__ = 'test_ab'
    a_id = Column(Integer, ForeignKey(A.id), primary_key=True)
    b_id = Column(Integer, ForeignKey(B.id), primary_key=True)

An assignment to A.b will generate an AB object:

a.b = B()

The A.b association is scalar, and includes a new flag AssociationProxy.cascade_scalar_deletes. When set, setting A.b to None will remove A.ab as well. The default behavior remains that it leaves a.ab in place:

a.b = None
assert a.ab is None

While it at first seemed intuitive that this logic should just look at the “cascade” attribute of the existing relationship, it’s not clear from that alone if the proxied object should be removed, hence the behavior is made available as an explicit option.

Additionally, del now works for scalars in a similar manner as setting to None:

del a.b
assert a.ab is None


AssociationProxy stores class-specific state on a per-class basis

The AssociationProxy object makes lots of decisions based on the parent mapped class it is associated with. While the AssociationProxy historically began as a relatively simple “getter”, it became apparent early on that it also needed to make decisions about what kind of attribute it is referring towards, e.g. scalar or collection, mapped object or simple value, and similar. To achieve this, it needs to inspect the mapped attribute or other descriptor or attribute that it refers towards, as referenced from its parent class. However in Python descriptor mechanics, a descriptor only learns about its “parent” class when it is accessed in the context of that class, such as calling MyClass.some_descriptor, which calls the __get__() method which passes in the class. The AssociationProxy object would therefore store state that is specific to that class, but only once this method were called; trying to inspect this state ahead of time without first accessing the AssociationProxy as a descriptor would raise an error. Additionally, it would assume that the first class to be seen by __get__() would be the only parent class it needed to know about. This is despite the fact that if a particular class has inheriting subclasses, the association proxy is really working on behalf of more than one parent class even though it was not explicitly re-used. While even with this shortcoming, the association proxy would still get pretty far with its current behavior, it still leaves shortcomings in some cases as well as the complex problem of determining the best “owner” class.

These problems are now solved in that AssociationProxy no longer modifies its own internal state when __get__() is called; instead, a new object is generated per-class known as AssociationProxyInstance which handles all the state specific to a particular mapped parent class (when the parent class is not mapped, no AssociationProxyInstance is generated). The concept of a single “owning class” for the association proxy, which was nonetheless improved in 1.1, has essentially been replaced with an approach where the AP now can treat any number of “owning” classes equally.

To accommodate for applications that want to inspect this state for an AssociationProxy without necessarily calling __get__(), a new method AssociationProxy.for_class() is added that provides direct access to a class-specific AssociationProxyInstance, demonstrated as:

class User(Base):
    # ...

    keywords = association_proxy('kws', 'keyword')

proxy_state = inspect(User).all_orm_descriptors["keywords"].for_class(User)

Once we have the AssociationProxyInstance object, in the above example stored in the proxy_state variable, we can look at attributes specific to the User.keywords proxy, such as target_class:

>>> proxy_state.target_class


AssociationProxy now provides standard column operators for a column-oriented target

Given an AssociationProxy where the target is a database column, and is not an object reference or another association proxy:

class User(Base):
    # ...

    elements = relationship("Element")

    # column-based association proxy
    values = association_proxy("elements", "value")

class Element(Base):
    # ...

    value = Column(String)

The User.values association proxy refers to the Element.value column. Standard column operations are now available, such as like:

>>> print(s.query(User).filter(User.values.like('%foo%')))
SELECT "user".id AS user_id
FROM "user"
FROM element
WHERE "user".id = element.user_id AND element.value LIKE :value_1)


>>> print(s.query(User).filter(User.values == 'foo'))
SELECT "user".id AS user_id
FROM "user"
FROM element
WHERE "user".id = element.user_id AND element.value = :value_1)

When comparing to None, the IS NULL expression is augmented with a test that the related row does not exist at all; this is the same behavior as before:

>>> print(s.query(User).filter(User.values == None))
SELECT "user".id AS user_id
FROM "user"
FROM element
WHERE "user".id = element.user_id AND element.value IS NULL)) OR NOT (EXISTS (SELECT 1
FROM element
WHERE "user".id = element.user_id))

Note that the ColumnOperators.contains() operator is in fact a string comparison operator; this is a change in behavior in that previously, the association proxy used .contains as a list containment operator only. With a column-oriented comparison, it now behaves like a “like”:

>>> print(s.query(User).filter(User.values.contains('foo')))
SELECT "user".id AS user_id
FROM "user"
FROM element
WHERE "user".id = element.user_id AND (element.value LIKE '%' || :value_1 || '%'))

In order to test the User.values collection for simple membership of the value "foo", the equals operator (e.g. User.values == 'foo') should be used; this works in previous versions as well.

When using an object-based association proxy with a collection, the behavior is as before, that of testing for collection membership, e.g. given a mapping:

class User(Base):
    __tablename__ = 'user'

    id = Column(Integer, primary_key=True)
    user_elements = relationship("UserElement")

    # object-based association proxy
    elements = association_proxy("user_elements", "element")

class UserElement(Base):
    __tablename__ = 'user_element'

    id = Column(Integer, primary_key=True)
    user_id = Column(ForeignKey("user.id"))
    element_id = Column(ForeignKey("element.id"))
    element = relationship("Element")

class Element(Base):
    __tablename__ = 'element'

    id = Column(Integer, primary_key=True)
    value = Column(String)

The .contains() method produces the same expression as before, testing the list of User.elements for the presence of an Element object:

>>> print(s.query(User).filter(User.elements.contains(Element(id=1))))
SELECT "user".id AS user_id
FROM "user"
FROM user_element
WHERE "user".id = user_element.user_id AND :param_1 = user_element.element_id)

Overall, the change is enabled based on the architectural change that is part of AssociationProxy stores class-specific state on a per-class basis; as the proxy now spins off additional state when an expression is generated, there is both an object-target and a column-target version of the AssociationProxyInstance class.


Association Proxy now Strong References the Parent Object

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

As an example, given a mapping with association proxy:

class A(Base):
    __tablename__ = 'a'

    id = Column(Integer, primary_key=True)
    bs = relationship("B")
    b_data = association_proxy('bs', 'data')

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

a1 = A(bs=[B(data='b1'), B(data='b2')])

b_data = a1.b_data

Previously, if a1 were deleted out of scope:

del a1

Trying to iterate the b_data collection after a1 is deleted from scope would raise the error "stale association proxy, parent object has gone out of scope". This is because the association proxy needs to access the actual a1.bs collection in order to produce a view, and prior to this change it maintained only a weak reference to a1. In particular, users would frequently encounter this error when performing an inline operation such as:

collection = session.query(A).filter_by(id=1).first().b_data

Above, because the A object would be garbage collected before the b_data collection were actually used.

The change is that the b_data collection is now maintaining a strong reference to the a1 object, so that it remains present:

assert b_data == ['b1', 'b2']

This change introduces the side effect that if an application is passing around the collection as above, the parent object won’t be garbage collected until the collection is also discarded. As always, if a1 is persistent inside a particular Session, it will remain part of that session’s state until it is garbage collected.

Note that this change may be revised if it leads to problems.


Implemented bulk replace for sets, dicts with AssociationProxy

Assignment of a set or dictionary to an association proxy collection should now work correctly, whereas before it would re-create association proxy members for existing keys, leading to the issue of potential flush failures due to the delete+insert of the same object it now should only create new association objects where appropriate:

class A(Base):
    __tablename__ = "test_a"

    id = Column(Integer, primary_key=True)
    b_rel = relationship(
        "B", collection_class=set, cascade="all, delete-orphan",
    b = association_proxy("b_rel", "value", creator=lambda x: B(value=x))

class B(Base):
    __tablename__ = "test_b"
    __table_args__ = (UniqueConstraint("a_id", "value"),)

    id = Column(Integer, primary_key=True)
    a_id = Column(Integer, ForeignKey("test_a.id"), nullable=False)
    value = Column(String)

# ...

s = Session(e)
a = A(b={"x", "y", "z"})

# re-assign where one B should be deleted, one B added, two
# B's maintained
a.b = {"x", "z", "q"}

# only 'q' was added, so only one new B object.  previously
# all three would have been re-created leading to flush conflicts
# against the deleted ones.
assert len(s.new) == 1


Many-to-one backref checks for collection duplicates during remove operation

When an ORM-mapped collection that existed as a Python sequence, typically a Python list as is the default for relationship(), contained duplicates, and the object were removed from one of its positions but not the other(s), a many-to-one backref would set its attribute to None even though the one-to-many side still represented the object as present. Even though one-to-many collections cannot have duplicates in the relational model, an ORM-mapped relationship() that uses a sequence collection can have duplicates inside of it in memory, with the restriction that this duplicate state can neither be persisted nor retrieved from the database. In particular, having a duplicate temporarily present in the list is intrinsic to a Python “swap” operation. Given a standard one-to-many/many-to-one setup:

class A(Base):
    __tablename__ = 'a'

    id = Column(Integer, primary_key=True)
    bs = relationship("B", backref="a")

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

If we have an A object with two B members, and perform a swap:

a1 = A(bs=[B(), B()])

a1.bs[0], a1.bs[1] = a1.bs[1], a1.bs[0]

During the above operation, interception of the standard Python __setitem__ __delitem__ methods delivers an interim state where the second B() object is present twice in the collection. When the B() object is removed from one of the positions, the B.a backref would set the reference to None, causing the link between the A and B object to be removed during the flush. The same issue can be demonstrated using plain duplicates:

>>> a1 = A()
>>> b1 = B()
>>> a1.bs.append(b1)
>>> a1.bs.append(b1)  # append the same b1 object twice
>>> del a1.bs[1]
>>> a1.bs  # collection is unaffected so far...
[<__main__.B object at 0x7f047af5fb70>]
>>> b1.a   # however b1.a is None
>>> session.add(a1)
>>> session.commit()  # so upon flush + expire....
>>> a1.bs  # the value is gone

The fix ensures that when the backref fires off, which is before the collection is mutated, the collection is checked for exactly one or zero instances of the target item before unsetting the many-to-one side, using a linear search which at the moment makes use of list.search and list.__contains__.

Originally it was thought that an event-based reference counting scheme would need to be used within the collection internals so that all duplicate instances could be tracked throughout the lifecycle of the collection, which would have added a performance/memory/complexity impact to all collection operations, including the very frequent operations of loading and appending. The approach that is taken instead limits the additional expense to the less common operations of collection removal and bulk replacement, and the observed overhead of the linear scan is negligible; linear scans of relationship-bound collections are already used within the unit of work as well as when a collection is bulk replaced.


Key Behavioral Changes - ORM

Query.join() handles ambiguity in deciding the “left” side more explicitly

Historically, given a query like the following:

u_alias = aliased(User)
session.query(User, u_alias).join(Address)

given the standard tutorial mappings, the query would produce a FROM clause as:

FROM users AS users_1, users JOIN addresses ON users.id = addresses.user_id

That is, the JOIN would implicitly be against the first entity that matches. The new behavior is that an exception requests that this ambiguity be resolved:

sqlalchemy.exc.InvalidRequestError: Can't determine which FROM clause to
join from, there are multiple FROMS which can join to this entity.
Try adding an explicit ON clause to help resolve the ambiguity.

The solution is to provide an ON clause, either as an expression:

# join to User
session.query(User, u_alias).join(Address, Address.user_id == User.id)

# join to u_alias
session.query(User, u_alias).join(Address, Address.user_id == u_alias.id)

Or to use the relationship attribute, if available:

# join to User
session.query(User, u_alias).join(Address, User.addresses)

# join to u_alias
session.query(User, u_alias).join(Address, u_alias.addresses)

The change includes that a join can now correctly link to a FROM clause that is not the first element in the list if the join is otherwise non-ambiguous:

session.query(func.current_timestamp(), User).join(Address)

Prior to this enhancement, the above query would raise:

sqlalchemy.exc.InvalidRequestError: Don't know how to join from
CURRENT_TIMESTAMP; please use select_from() to establish the
left entity/selectable of this join

Now the query works fine:

SELECT CURRENT_TIMESTAMP AS current_timestamp_1, users.id AS users_id,
users.name AS users_name, users.fullname AS users_fullname,
users.password AS users_password
FROM users JOIN addresses ON users.id = addresses.user_id

Overall the change is directly towards Python’s “explicit is better than implicit” philosophy.


FOR UPDATE clause is rendered within the joined eager load subquery as well as outside

This change applies specifically to the use of the joinedload() loading strategy in conjunction with a row limited query, e.g. using Query.first() or Query.limit(), as well as with use of the Query.with_for_update() method.

Given a query as:


The Query object renders a SELECT of the following form when joined eager loading is combined with LIMIT:

SELECT subq.a_id, subq.a_data, b_alias.id, b_alias.data FROM (
    SELECT a.id AS a_id, a.data AS a_data FROM a LIMIT 5
) AS subq LEFT OUTER JOIN b ON subq.a_id=b.a_id

This is so that the limit of rows takes place for the primary entity without affecting the joined eager load of related items. When the above query is combined with “SELECT..FOR UPDATE”, the behavior has been this:

SELECT subq.a_id, subq.a_data, b_alias.id, b_alias.data FROM (
    SELECT a.id AS a_id, a.data AS a_data FROM a LIMIT 5
) AS subq LEFT OUTER JOIN b ON subq.a_id=b.a_id FOR UPDATE

However, MySQL due to https://bugs.mysql.com/bug.php?id=90693 does not lock the rows inside the subquery, unlike that of PostgreSQL and other databases. So the above query now renders as:

SELECT subq.a_id, subq.a_data, b_alias.id, b_alias.data FROM (
    SELECT a.id AS a_id, a.data AS a_data FROM a LIMIT 5 FOR UPDATE
) AS subq LEFT OUTER JOIN b ON subq.a_id=b.a_id FOR UPDATE

On the Oracle dialect, the inner “FOR UPDATE” is not rendered as Oracle does not support this syntax and the dialect skips any “FOR UPDATE” that is against a subquery; it isn’t necessary in any case since Oracle, like PostgreSQL, correctly locks all elements of the returned row.

When using the Query.with_for_update.of modifier, typically on PostgreSQL, the outer “FOR UPDATE” is omitted, and the OF is now rendered on the inside; previously, the OF target would not be converted to accommodate for the subquery correctly. So given:


The query would now render as:

SELECT subq.a_id, subq.a_data, b_alias.id, b_alias.data FROM (
    SELECT a.id AS a_id, a.data AS a_data FROM a LIMIT 5 FOR UPDATE OF a
) AS subq LEFT OUTER JOIN b ON subq.a_id=b.a_id

The above form should be helpful on PostgreSQL additionally since PostgreSQL will not allow the FOR UPDATE clause to be rendered after the LEFT OUTER JOIN target.

Overall, FOR UPDATE remains highly specific to the target database in use and can’t easily be generalized for more complex queries.


passive_deletes=’all’ will leave FK unchanged for object removed from collection

The relationship.passive_deletes option accepts the value "all" to indicate that no foreign key attributes should be modified when the object is flushed, even if the relationship’s collection / reference has been removed. Previously, this did not take place for one-to-many, or one-to-one relationships, in the following situation:

class User(Base):
    __tablename__ = 'users'

    id = Column(Integer, primary_key=True)
    addresses = relationship(

class Address(Base):
    __tablename__ = 'addresses'
    id = Column(Integer, primary_key=True)
    email = Column(String)

    user_id = Column(Integer, ForeignKey('users.id'))
    user = relationship("User")

u1 = session.query(User).first()
address = u1.addresses[0]

# would fail and be set to None
assert address.user_id == u1.id

The fix now includes that address.user_id is left unchanged as per passive_deletes="all". This kind of thing is useful for building custom “version table” schemes and such where rows are archived instead of deleted.


New Features and Improvements - Core

New multi-column naming convention tokens, long name truncation

To suit the case where a MetaData naming convention needs to disambiguate between multiple-column constraints and wishes to use all the columns within the generated constraint name, a new series of naming convention tokens are added, including column_0N_name, column_0_N_name, column_0N_key, column_0_N_key, referred_column_0N_name, referred_column_0_N_name, etc., which render the column name (or key or label) for all columns in the constraint, joined together either with no separator or with an underscore separator. Below we define a convention that will name UniqueConstraint constraints with a name that joins together the names of all columns:

metadata = MetaData(naming_convention={
    "uq": "uq_%(table_name)s_%(column_0_N_name)s"

table = Table(
    'info', metadata,
    Column('a', Integer),
    Column('b', Integer),
    Column('c', Integer),
    UniqueConstraint('a', 'b', 'c')

The CREATE TABLE for the above table will render as:

    a INTEGER,
    b INTEGER,
    c INTEGER,
    CONSTRAINT uq_info_a_b_c UNIQUE (a, b, c)

In addition, long-name truncation logic is now applied to the names generated by naming conventions, in particular to accommodate for multi-column labels that can produce very long names. This logic, which is the same as that used for truncating long label names in a SELECT statement, replaces excess characters that go over the identifier-length limit for the target database with a deterministically generated 4-character hash. For example, on PostgreSQL where identifiers cannot be longer than 63 characters, a long constraint name would normally be generated from the table definition below:

long_names = Table(
    'long_names', metadata,
    Column('information_channel_code', Integer, key='a'),
    Column('billing_convention_name', Integer, key='b'),
    Column('product_identifier', Integer, key='c'),
    UniqueConstraint('a', 'b', 'c')

The truncation logic will ensure a too-long name isn’t generated for the UNIQUE constraint:

CREATE TABLE long_names (
    information_channel_code INTEGER,
    billing_convention_name INTEGER,
    product_identifier INTEGER,
    CONSTRAINT uq_long_names_information_channel_code_billing_conventi_a79e
    UNIQUE (information_channel_code, billing_convention_name, product_identifier)

The above suffix a79e is based on the md5 hash of the long name and will generate the same value every time to produce consistent names for a given schema.

Note that the truncation logic also raises IdentifierError when a constraint name is explicitly too large for a given dialect. This has been the behavior for an Index object for a long time, but is now applied to other kinds of constraints as well:

from sqlalchemy import Column
from sqlalchemy import Integer
from sqlalchemy import MetaData
from sqlalchemy import Table
from sqlalchemy import UniqueConstraint
from sqlalchemy.dialects import postgresql
from sqlalchemy.schema import AddConstraint

m = MetaData()
t = Table("t", m, Column("x", Integer))
uq = UniqueConstraint(


will output:

sqlalchemy.exc.IdentifierError: Identifier
exceeds maximum length of 63 characters

The exception raise prevents the production of non-deterministic constraint names truncated by the database backend which are then not compatible with database migrations later on.

To apply SQLAlchemy-side truncation rules to the above identifier, use the conv() construct:

uq = UniqueConstraint(

This will again output deterministically truncated SQL as in:

ALTER TABLE t ADD CONSTRAINT this_is_too_long_of_a_name_for_any_database_backend_eve_ac05 UNIQUE (x)

There is not at the moment an option to have the names pass through to allow database-side truncation. This has already been the case for Index names for some time and issues have not been raised.

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


Binary comparison interpretation for SQL functions

This enhancement is implemented at the Core level, however is applicable primarily to the ORM.

A SQL function that compares two elements can now be used as a “comparison” object, suitable for usage in an ORM relationship(), by first creating the function as usual using the func factory, then when the function is complete calling upon the FunctionElement.as_comparison() modifier to produce a BinaryExpression that has a “left” and a “right” side:

class Venue(Base):
    __tablename__ = 'venue'
    id = Column(Integer, primary_key=True)
    name = Column(String)

    descendants = relationship(
            remote(foreign(name)), name + "/"
        ).as_comparison(1, 2) == 1,

Above, the relationship.primaryjoin of the “descendants” relationship will produce a “left” and a “right” expression based on the first and second arguments passed to instr(). This allows features like the ORM lazyload to produce SQL like:

SELECT venue.id AS venue_id, venue.name AS venue_name
FROM venue
WHERE instr(venue.name, (? || ?)) = ? ORDER BY venue.name
('parent1', '/', 1)

and a joinedload, such as:

v1 = s.query(Venue).filter_by(name="parent1").options(

to work as:

SELECT venue.id AS venue_id, venue.name AS venue_name,
  venue_1.id AS venue_1_id, venue_1.name AS venue_1_name
FROM venue LEFT OUTER JOIN venue AS venue_1
  ON instr(venue_1.name, (venue.name || ?)) = ?
WHERE venue.name = ? ORDER BY venue_1.name
('/', 1, 'parent1')

This feature is expected to help with situations such as making use of geometric functions in relationship join conditions, or any case where the ON clause of the SQL join is expressed in terms of a SQL function.


Expanding IN feature now supports empty lists

The “expanding IN” feature introduced in version 1.2 at Late-expanded IN parameter sets allow IN expressions with cached statements now supports empty lists passed to the ColumnOperators.in_() operator. The implementation for an empty list will produce an “empty set” expression that is specific to a target backend, such as “SELECT CAST(NULL AS INTEGER) WHERE 1!=1” for PostgreSQL, “SELECT 1 FROM (SELECT 1) as _empty_set WHERE 1!=1” for MySQL:

>>> from sqlalchemy import create_engine
>>> from sqlalchemy import select, literal_column, bindparam
>>> e = create_engine("postgresql://scott:tiger@localhost/test", echo=True)
>>> with e.connect() as conn:
...      conn.execute(
...          select([literal_column('1')]).
...          where(literal_column('1').in_(bindparam('q', expanding=True))),
...          q=[]
...      )

The feature also works for tuple-oriented IN statements, where the “empty IN” expression will be expanded to support the elements given inside the tuple, such as on PostgreSQL:

>>> from sqlalchemy import create_engine
>>> from sqlalchemy import select, literal_column, tuple_, bindparam
>>> e = create_engine("postgresql://scott:tiger@localhost/test", echo=True)
>>> with e.connect() as conn:
...      conn.execute(
...          select([literal_column('1')]).
...          where(tuple_(50, "somestring").in_(bindparam('q', expanding=True))),
...          q=[]
...      )
SELECT 1 WHERE (%(param_1)s, %(param_2)s)


TypeEngine methods bind_expression, column_expression work with Variant, type-specific types

The TypeEngine.bind_expression() and TypeEngine.column_expression() methods now work when they are present on the “impl” of a particular datatype, allowing these methods to be used by dialects as well as for TypeDecorator and Variant use cases.

The following example illustrates a TypeDecorator that applies SQL-time conversion functions to a LargeBinary. In order for this type to work in the context of a Variant, the compiler needs to drill into the “impl” of the variant expression in order to locate these methods:

from sqlalchemy import TypeDecorator, LargeBinary, func

class CompressedLargeBinary(TypeDecorator):
    impl = LargeBinary

    def bind_expression(self, bindvalue):
        return func.compress(bindvalue, type_=self)

    def column_expression(self, col):
        return func.uncompress(col, type_=self)

MyLargeBinary = LargeBinary().with_variant(CompressedLargeBinary(), "sqlite")

The above expression will render a function within SQL when used on SQLite only:

from sqlalchemy import select, column
from sqlalchemy.dialects import sqlite
print(select([column('x', CompressedLargeBinary)]).compile(dialect=sqlite.dialect()))

will render:

SELECT uncompress(x) AS x

The change also includes that dialects can implement TypeEngine.bind_expression() and TypeEngine.column_expression() on dialect-level implementation types where they will now be used; in particular this will be used for MySQL’s new “binary prefix” requirement as well as for casting decimal bind values for MySQL.


New last-in-first-out strategy for QueuePool

The connection pool usually used by create_engine() is known as QueuePool. This pool uses an object equivalent to Python’s built-in Queue class in order to store database connections waiting to be used. The Queue features first-in-first-out behavior, which is intended to provide a round-robin use of the database connections that are persistently in the pool. However, a potential downside of this is that when the utilization of the pool is low, the re-use of each connection in series means that a server-side timeout strategy that attempts to reduce unused connections is prevented from shutting down these connections. To suit this use case, a new flag create_engine.pool_use_lifo is added which reverses the .get() method of the Queue to pull the connection from the beginning of the queue instead of the end, essentially turning the “queue” into a “stack” (adding a whole new pool called StackPool was considered, however this was too much verbosity).

Key Changes - Core

Coercion of string SQL fragments to text() fully removed

The warnings that were first added in version 1.0, described at Warnings emitted when coercing full SQL fragments into text(), have now been converted into exceptions. Continued concerns have been raised regarding the automatic coercion of string fragments passed to methods like Query.filter() and Select.order_by() being converted to text() constructs, even though this has emitted a warning. In the case of Select.order_by(), Query.order_by(), Select.group_by(), and Query.group_by(), a string label or column name is still resolved into the corresponding expression construct, however if the resolution fails, a CompileError is raised, thus preventing raw SQL text from being rendered directly.


“threadlocal” engine strategy deprecated

The “threadlocal engine strategy” was added around SQLAlchemy 0.2, as a solution to the problem that the standard way of operating in SQLAlchemy 0.1, which can be summed up as “threadlocal everything”, was found to be lacking. In retrospect, it seems fairly absurd that by SQLAlchemy’s first releases which were in every regard “alpha”, that there was concern that too many users had already settled on the existing API to simply change it.

The original usage model for SQLAlchemy looked like this:


result = table.select().execute()



After a few months of real world use, it was clear that trying to pretend a “connection” or a “transaction” was a hidden implementation detail was a bad idea, particularly the moment someone needed to deal with more than one database connection at a time. So the usage paradigm we see today was introduced, minus the context managers since they didn’t yet exist in Python:

conn = engine.connect()
    trans = conn.begin()

    conn.execute(table.insert(), <params>)
    result = conn.execute(table.select())

    conn.execute(table.update(), <params>)


The above paradigm was what people needed, but since it was still kind of verbose (because no context managers), the old way of working was kept around as well and it became the threadlocal engine strategy.

Today, working with Core is much more succinct, and even more succinct than the original pattern, thanks to context managers:

with engine.begin() as conn:
    conn.execute(table.insert(), <params>)
    result = conn.execute(table.select())

    conn.execute(table.update(), <params>)

At this point, any remaining code that is still relying upon the “threadlocal” style will be encouraged via this deprecation to modernize - the feature should be removed totally by the next major series of SQLAlchemy, e.g. 1.4. The connection pool parameter Pool.use_threadlocal is also deprecated as it does not actually have any effect in most cases, as is the Engine.contextual_connect() method, which is normally synonymous with the Engine.connect() method except in the case where the threadlocal engine is in use.


convert_unicode parameters deprecated

The parameters String.convert_unicode and create_engine.convert_unicode are deprecated. The purpose of these parameters was to instruct SQLAlchemy to ensure that incoming Python Unicode objects under Python 2 were encoded to bytestrings before passing to the database, and to expect bytestrings from the database to be converted back to Python Unicode objects. In the pre-Python 3 era, this was an enormous ordeal to get right, as virtually all Python DBAPIs had no Unicode support enabled by default, and most had major issues with the Unicode extensions that they did provide. Eventually, SQLAlchemy added C extensions, one of the primary purposes of these extensions was to speed up the Unicode decode process within result sets.

Once Python 3 was introduced, DBAPIs began to start supporting Unicode more fully, and more importantly, by default. However, the conditions under which a particular DBAPI would or would not return Unicode data from a result, as well as accept Python Unicode values as parameters, remained extremely complicated. This was the beginning of the obsolescence of the “convert_unicode” flags, because they were no longer sufficient as a means of ensuring that encode/decode was occurring only where needed and not where it wasn’t needed. Instead, “convert_unicode” started to be automatically detected by dialects. Part of this can be seen in the “SELECT ‘test plain returns’” and “SELECT ‘test_unicode_returns’” SQL emitted by an engine the first time it connects; the dialect is testing that the current DBAPI with its current settings and backend database connection is returning Unicode by default or not.

The end result is that end-user use of the “convert_unicode” flags should no longer be needed in any circumstances, and if they are, the SQLAlchemy project needs to know what those cases are and why. Currently, hundreds of Unicode round trip tests pass across all major databases without the use of this flag so there is a fairly high level of confidence that they are no longer needed except in arguable non use cases such as accessing mis-encoded data from a legacy database, which would be better suited using custom types.


Dialect Improvements and Changes - PostgreSQL

Added basic reflection support for PostgreSQL partitioned tables

SQLAlchemy can render the “PARTITION BY” sequence within a PostgreSQL CREATE TABLE statement using the flag postgresql_partition_by, added in version 1.2.6. However, the 'p' type was not part of the reflection queries used until now.

Given a schema such as:

dv = Table(
    'data_values', metadata,
    Column('modulus', Integer, nullable=False),
    Column('data', String(30)),

        "CREATE TABLE data_values_4_10 PARTITION OF data_values "
        "FOR VALUES FROM (4) TO (10)")

The two table names 'data_values' and 'data_values_4_10' will come back from Inspector.get_table_names() and additionally the columns will come back from Inspector.get_columns('data_values') as well as Inspector.get_columns('data_values_4_10'). This also extends to the use of Table(..., autoload=True) with these tables.


Dialect Improvements and Changes - MySQL

Protocol-level ping now used for pre-ping

The MySQL dialects including mysqlclient, python-mysql, PyMySQL and mysql-connector-python now use the connection.ping() method for the pool pre-ping feature, described at Disconnect Handling - Pessimistic. This is a much more lightweight ping than the previous method of emitting “SELECT 1” on the connection.

Control of parameter ordering within ON DUPLICATE KEY UPDATE

The order of UPDATE parameters in the ON DUPLICATE KEY UPDATE clause can now be explicitly ordered by passing a list of 2-tuples:

from sqlalchemy.dialects.mysql import insert

insert_stmt = insert(my_table).values(
    data='inserted value')

on_duplicate_key_stmt = insert_stmt.on_duplicate_key_update(
        ("data", "some data"),
        ("updated_at", func.current_timestamp()),

Dialect Improvements and Changes - SQLite

Support for SQLite JSON Added

A new datatype JSON is added which implements SQLite’s json member access functions on behalf of the JSON base datatype. The SQLite JSON_EXTRACT and JSON_QUOTE functions are used by the implementation to provide basic JSON support.

Note that the name of the datatype itself as rendered in the database is the name “JSON”. This will create a SQLite datatype with “numeric” affinity, which normally should not be an issue except in the case of a JSON value that consists of single integer value. Nevertheless, following an example in SQLite’s own documentation at https://www.sqlite.org/json1.html the name JSON is being used for its familiarity.


Support for SQLite ON CONFLICT in constraints added

SQLite supports a non-standard ON CONFLICT clause that may be specified for standalone constraints as well as some column-inline constraints such as NOT NULL. Support has been added for these clauses via the sqlite_on_conflict keyword added to objects like UniqueConstraint as well as several Column -specific variants:

some_table = Table(
    'some_table', metadata,
    Column('id', Integer, primary_key=True, sqlite_on_conflict_primary_key='FAIL'),
    Column('data', Integer),
    UniqueConstraint('id', 'data', sqlite_on_conflict='IGNORE')

The above table would render in a CREATE TABLE statement as:

CREATE TABLE some_table (
    data INTEGER,


Dialect Improvements and Changes - Oracle

National char datatypes de-emphasized for generic unicode, re-enabled with option

The Unicode and UnicodeText datatypes by default now correspond to the VARCHAR2 and CLOB datatypes on Oracle, rather than NVARCHAR2 and NCLOB (otherwise known as “national” character set types). This will be seen in behaviors such as that of how they render in CREATE TABLE statements, as well as that no type object will be passed to setinputsizes() when bound parameters using Unicode or UnicodeText are used; cx_Oracle handles the string value natively. This change is based on advice from cx_Oracle’s maintainer that the “national” datatypes in Oracle are largely obsolete and are not performant. They also interfere in some situations such as when applied to the format specifier for functions like trunc().

The one case where NVARCHAR2 and related types may be needed is for a database that is not using a Unicode-compliant character set. In this case, the flag use_nchar_for_unicode can be passed to create_engine() to re-enable the old behavior.

As always, using the NVARCHAR2 and NCLOB datatypes explicitly will continue to make use of NVARCHAR2 and NCLOB, including within DDL as well as when handling bound parameters with cx_Oracle’s setinputsizes().

On the read side, automatic Unicode conversion under Python 2 has been added to CHAR/VARCHAR/CLOB result rows, to match the behavior of cx_Oracle under Python 3. In order to mitigate the performance hit that the cx_Oracle dialect had previously with this behavior under Python 2, SQLAlchemy’s very performant (when C extensions are built) native Unicode handlers are used under Python 2. The automatic unicode coercion can be disabled by setting the coerce_to_unicode flag to False. This flag now defaults to True and applies to all string data returned in a result set that isn’t explicitly under Unicode or Oracle’s NVARCHAR2/NCHAR/NCLOB datatypes.


cx_Oracle connect arguments modernized, deprecated parameters removed

A series of modernizations to the parameters accepted by the cx_oracle dialect as well as the URL string:

  • The deprecated parameters auto_setinputsizes, allow_twophase, exclude_setinputsizes are removed.

  • The value of the threaded parameter, which has always been defaulted to True for the SQLAlchemy dialect, is no longer generated by default. The SQLAlchemy Connection object is not considered to be thread-safe itself so there’s no need for this flag to be passed.

  • It’s deprecated to pass threaded to create_engine() itself. To set the value of threaded to True, pass it to either the create_engine.connect_args dictionary or use the query string e.g. oracle+cx_oracle://...?threaded=true.

  • All parameters passed on the URL query string that are not otherwise specially consumed are now passed to the cx_Oracle.connect() function. A selection of these are also coerced either into cx_Oracle constants or booleans including mode, purity, events, and threaded.

  • As was the case earlier, all cx_Oracle .connect() arguments are accepted via the create_engine.connect_args dictionary, the documentation was inaccurate regarding this.


Dialect Improvements and Changes - SQL Server

Support for pyodbc fast_executemany

Pyodbc’s recently added “fast_executemany” mode, available when using the Microsoft ODBC driver, is now an option for the pyodbc / mssql dialect. Pass it via create_engine():

engine = create_engine(


New parameters to affect IDENTITY start and increment, use of Sequence deprecated

SQL Server as of SQL Server 2012 now supports sequences with real CREATE SEQUENCE syntax. In #4235, SQLAchemy will add support for these using Sequence in the same way as for any other dialect. However, the current situation is that Sequence has been repurposed on SQL Server specifically in order to affect the “start” and “increment” parameters for the IDENTITY specification on a primary key column. In order to make the transition towards normal sequences being available as well, using Sequence will emit a deprecation warning throughout the 1.3 series. In order to affect “start” and “increment”, use the new mssql_identity_start and mssql_identity_increment parameters on Column:

test = Table(
    'test', metadata,
        'id', Integer, primary_key=True, mssql_identity_start=100,
    Column('name', String(20))

In order to emit IDENTITY on a non-primary key column, which is a little-used but valid SQL Server use case, use the Column.autoincrement flag, setting it to True on the target column, False on any integer primary key column:

test = Table(
    'test', metadata,
    Column('id', Integer, primary_key=True, autoincrement=False),
    Column('number', Integer, autoincrement=True)



Changed StatementError formatting (newlines and %s)

Two changes are introduced to the string representation for StatementError. The “detail” and “SQL” portions of the string representation are now separated by newlines, and newlines that are present in the original SQL statement are maintained. The goal is to improve readability while still keeping the original error message on one line for logging purposes.

This means that an error message that previously looked like this:

sqlalchemy.exc.StatementError: (sqlalchemy.exc.InvalidRequestError) A value is required for bind parameter 'id' [SQL: 'select * from reviews\nwhere id = ?'] (Background on this error at: http://sqlalche.me/e/cd3x)

Will now look like this:

sqlalchemy.exc.StatementError: (sqlalchemy.exc.InvalidRequestError) A value is required for bind parameter 'id'
[SQL: select * from reviews
where id = ?]
(Background on this error at: http://sqlalche.me/e/cd3x)

The primary impact of this change is that consumers can no longer assume that a complete exception message is on a single line, however the original “error” portion that is generated from the DBAPI driver or SQLAlchemy internals will still be on the first line.