What’s New in SQLAlchemy 2.0?

Note for Readers

SQLAlchemy 2.0’s transition documents are separated into two documents - one which details major API shifts from the 1.x to 2.x series, and the other which details new features and behaviors relative to SQLAlchemy 1.4:

Readers who have not yet updated their 1.4 application to follow SQLAlchemy 2.0 engine and ORM conventions may navigate to SQLAlchemy 2.0 - Major Migration Guide for a guide to ensuring SQLAlchemy 2.0 compatibility, which is a prerequisite for having working code under version 2.0.

About this Document

This document describes changes between SQLAlchemy version 1.4 and SQLAlchemy version 2.0, independent of the major changes between 1.x style and 2.0 style usage. Readers should start with the SQLAlchemy 2.0 - Major Migration Guide document to get an overall picture of the major compatibility changes between the 1.x and 2.x series.

Aside from the major 1.x->2.x migration path, the next largest paradigm shift in SQLAlchemy 2.0 is deep integration with PEP 484 typing practices and current capabilities, particularly within the ORM. New type-driven ORM declarative styles inspired by Python dataclasses, as well as new integrations with dataclasses themselves, complement an overall approach that no longer requires stubs and also goes very far towards providing a type-aware method chain from SQL statement to result set.

The prominence of Python typing is significant not only so that type checkers like mypy can run without plugins; more significantly it allows IDEs like vscode and pycharm to take a much more active role in assisting with the composition of a SQLAlchemy application.

New Typing Support in Core and ORM - Stubs / Extensions no longer used

The approach to typing for Core and ORM has been completely reworked, compared to the interim approach that was provided in version 1.4 via the sqlalchemy2-stubs package. The new approach begins at the most fundamental element in SQLAlchemy which is the Column, or more accurately the ColumnElement that underlies all SQL expressions that have a type. This expression-level typing then extends into the area of statement construction, statement execution, and result sets, and finally into the ORM where new declarative forms allow for fully typed ORM models that integrate all the way from statement to result set.

Tip

Typing support should be considered beta level software for the 2.0 series. Typing details are subject to change however significant backwards-incompatible changes are not planned.

SQL Expression / Statement / Result Set Typing

This section provides background and examples for SQLAlchemy’s new SQL expression typing approach, which extends from base ColumnElement constructs through SQL statements and result sets and into realm of ORM mapping.

Rationale and Overview

Tip

This section is an architectural discussion. Skip ahead to SQL Expression Typing - Examples to just see what the new typing looks like.

In sqlalchemy2-stubs, SQL expressions were typed as generics that then referred to a TypeEngine object such as Integer, DateTime, or String as their generic argument (such as Column[Integer]). This was itself a departure from what the original Dropbox sqlalchemy-stubs package did, where Column and its foundational constructs were directly generic on Python types, such as int, datetime and str. It was hoped that since Integer / DateTime / String themselves are generic against int / datetime / str, there would be ways to maintain both levels of information and to be able to extract the Python type from a column expression via the TypeEngine as an intermediary construct. However, this is not the case, as PEP 484 doesn’t really have a rich enough feature set for this to be viable, lacking capabilities such as higher kinded TypeVars.

So after a deep assessment of the current capabilities of PEP 484, SQLAlchemy 2.0 has realized the original wisdom of sqlalchemy-stubs in this area and returned to linking column expressions directly to Python types. This does mean that if one has SQL expressions to different subtypes, like Column(VARCHAR) vs. Column(Unicode), the specifics of those two String subtypes is not carried along as the type only carries along str, but in practice this is usually not an issue and it is generally vastly more useful that the Python type is immediately present, as it represents the in-Python data one will be storing and receiving for this column directly.

Concretely, this means that an expression like Column('id', Integer) is typed as Column[int]. This allows for a viable pipeline of SQLAlchemy construct -> Python datatype to be set up, without the need for typing plugins. Crucially, it allows full interoperability with the ORM’s paradigm of using select() and Row constructs that reference ORM mapped class types (e.g. a Row containing instances of user-mapped instances, such as the User and Address examples used in our tutorials). While Python typing currently has very limited support for customization of tuple-types (where PEP 646, the first pep that attempts to deal with tuple-like objects, was intentionally limited in its functionality and by itself is not yet viable for arbitrary tuple manipulation), a fairly decent approach has been devised that allows for basic select() -> Result -> Row typing to function, including for ORM classes, where at the point at which a Row object is to be unpacked into individual column entries, a small typing-oriented accessor is added that allows the individual Python values to maintain the Python type linked to the SQL expression from which they originated (translation: it works).

SQL Expression Typing - Examples

A brief tour of typing behaviors. Comments indicate what one would see hovering over the code in vscode (or roughly what typing tools would display when using the reveal_type() helper):

  • Simple Python Types Assigned to SQL Expressions

    # (variable) str_col: ColumnClause[str]
    str_col = column("a", String)
    
    # (variable) int_col: ColumnClause[int]
    int_col = column("a", Integer)
    
    # (variable) expr1: ColumnElement[str]
    expr1 = str_col + "x"
    
    # (variable) expr2: ColumnElement[int]
    expr2 = int_col + 10
    
    # (variable) expr3: ColumnElement[bool]
    expr3 = int_col == 15
  • Individual SQL expressions assigned to select() constructs, as well as any row-returning construct, including row-returning DML such as Insert with Insert.returning(), are packed into a Tuple[] type which retains the Python type for each element.

    # (variable) stmt: Select[Tuple[str, int]]
    stmt = select(str_col, int_col)
    
    # (variable) stmt: ReturningInsert[Tuple[str, int]]
    ins_stmt = insert(table("t")).returning(str_col, int_col)
  • The Tuple[] type from any row returning construct, when invoked with an .execute() method, carries through to Result and Row. In order to unpack the Row object as a tuple, the Row.tuple() or Row.t accessor essentially casts the Row into the corresponding Tuple[] (though remains the same Row object at runtime).

    with engine.connect() as conn:
        # (variable) stmt: Select[Tuple[str, int]]
        stmt = select(str_col, int_col)
    
        # (variable) result: Result[Tuple[str, int]]
        result = conn.execute(stmt)
    
        # (variable) row: Row[Tuple[str, int]] | None
        row = result.first()
    
        if row is not None:
            # for typed tuple unpacking or indexed access,
            # use row.tuple() or row.t  (this is the small typing-oriented accessor)
            strval, intval = row.t
    
            # (variable) strval: str
            strval
    
            # (variable) intval: int
            intval
  • Scalar values for single-column statements do the right thing with methods like Connection.scalar(), Result.scalars(), etc.

    # (variable) data: Sequence[str]
    data = connection.execute(select(str_col)).scalars().all()
  • The above support for row-returning constructs works the best with ORM mapped classes, as a mapped class can list out specific types for its members. The example below sets up a class using new type-aware syntaxes, described in the following section:

    from sqlalchemy.orm import DeclarativeBase
    from sqlalchemy.orm import Mapped
    from sqlalchemy.orm import mapped_column
    
    
    class Base(DeclarativeBase):
        pass
    
    
    class User(Base):
        __tablename__ = "user_account"
    
        id: Mapped[int] = mapped_column(primary_key=True)
        name: Mapped[str]
        addresses: Mapped[List["Address"]] = relationship()
    
    
    class Address(Base):
        __tablename__ = "address"
    
        id: Mapped[int] = mapped_column(primary_key=True)
        email_address: Mapped[str]
        user_id = mapped_column(ForeignKey("user_account.id"))

    With the above mapping, the attributes are typed and express themselves all the way from statement to result set:

    with Session(engine) as session:
        # (variable) stmt: Select[Tuple[int, str]]
        stmt_1 = select(User.id, User.name)
    
        # (variable) result_1: Result[Tuple[int, str]]
        result_1 = session.execute(stmt_1)
    
        # (variable) intval: int
        # (variable) strval: str
        intval, strval = result_1.one().t

    Mapped classes themselves are also types, and behave the same way, such as a SELECT against two mapped classes:

    with Session(engine) as session:
        # (variable) stmt: Select[Tuple[User, Address]]
        stmt_2 = select(User, Address).join_from(User, Address)
    
        # (variable) result_2: Result[Tuple[User, Address]]
        result_2 = session.execute(stmt_2)
    
        # (variable) user_obj: User
        # (variable) address_obj: Address
        user_obj, address_obj = result_2.one().t

    When selecting mapped classes, constructs like aliased work as well, maintaining the column-level attributes of the original mapped class as well as the return type expected from a statement:

    with Session(engine) as session:
        # this is in fact an Annotated type, but typing tools don't
        # generally display this
    
        # (variable) u1: Type[User]
        u1 = aliased(User)
    
        # (variable) stmt: Select[Tuple[User, User, str]]
        stmt = select(User, u1, User.name).filter(User.id == 5)
    
        # (variable) result: Result[Tuple[User, User, str]]
        result = session.execute(stmt)
  • Core Table does not yet have a decent way to maintain typing of Column objects when accessing them via the Table.c accessor.

    Since Table is set up as an instance of a class, and the Table.c accessor typically accesses Column objects dynamically by name, there’s not yet an established typing approach for this; some alternative syntax would be needed.

  • ORM classes, scalars, etc. work great.

    The typical use case of selecting ORM classes, as scalars or tuples, all works, both 2.0 and 1.x style queries, getting back the exact type either by itself or contained within the appropriate container such as Sequence[], List[] or Iterator[]:

    # (variable) users1: Sequence[User]
    users1 = session.scalars(select(User)).all()
    
    # (variable) user: User
    user = session.query(User).one()
    
    # (variable) user_iter: Iterator[User]
    user_iter = iter(session.scalars(select(User)))
  • Legacy Query gains tuple typing as well.

    The typing support for Query goes well beyond what sqlalchemy-stubs or sqlalchemy2-stubs offered, where both scalar-object as well as tuple-typed Query objects will retain result level typing for most cases:

    # (variable) q1: RowReturningQuery[Tuple[int, str]]
    q1 = session.query(User.id, User.name)
    
    # (variable) rows: List[Row[Tuple[int, str]]]
    rows = q1.all()
    
    # (variable) q2: Query[User]
    q2 = session.query(User)
    
    # (variable) users: List[User]
    users = q2.all()

the catch - all stubs must be uninstalled

A key caveat with the typing support is that all SQLAlchemy stubs packages must be uninstalled for typing to work. When running mypy against a Python virtualenv, this is only a matter of uninstalling those packages. However, a SQLAlchemy stubs package is also currently part of typeshed, which itself is bundled into some typing tools such as Pylance, so it may be necessary in some cases to locate the files for these packages and delete them, if they are in fact interfering with the new typing working correctly.

Once SQLAlchemy 2.0 is released in final status, typeshed will remove SQLAlchemy from its own stubs source.

ORM Declarative Models

SQLAlchemy 1.4 introduced the first SQLAlchemy-native ORM typing support using a combination of sqlalchemy2-stubs and the Mypy Plugin. In SQLAlchemy 2.0, the Mypy plugin remains available, and has been updated to work with SQLAlchemy 2.0’s typing system. However, it should now be considered deprecated, as applications now have a straightforward path to adopting the new typing support that does not use plugins or stubs.

Overview

The fundamental approach for the new system is that mapped column declarations, when using a fully Declarative model (that is, not hybrid declarative or imperative configurations, which are unchanged), are first derived at runtime by inspecting the type annotation on the left side of each attribute declaration, if present. Left hand type annotations are expected to be contained within the Mapped generic type, otherwise the attribute is not considered to be a mapped attribute. The attribute declaration may then refer to the mapped_column() construct on the right hand side, which is used to provide additional Core-level schema information about the Column to be produced and mapped. This right hand side declaration is optional if a Mapped annotation is present on the left side; if no annotation is present on the left side, then the mapped_column() may be used as an exact replacement for the Column directive where it will provide for more accurate (but not exact) typing behavior of the attribute, even though no annotation is present.

The approach is inspired by the approach of Python dataclasses which starts with an annotation on the left, then allows for an optional dataclasses.field() specification on the right; the key difference from the dataclasses approach is that SQLAlchemy’s approach is strictly opt-in, where existing mappings that use Column without any type annotations continue to work as they always have, and the mapped_column() construct may be used as a direct replacement for Column without any explicit type annotations. Only for exact attribute-level Python types to be present is the use of explicit annotations with Mapped required. These annotations may be used on an as-needed, per-attribute basis for those attributes where specific types are helpful; non-annotated attributes that use mapped_column() will be typed as Any at the instance level.

Migrating an Existing Mapping

Transitioning to the new ORM approach begins as more verbose, but becomes more succinct than was previously possible as the available new features are used fully. The following steps detail a typical transition and then continue on to illustrate some more options.

Step one - declarative_base() is superseded by DeclarativeBase.

One observed limitation in Python typing is that there seems to be no ability to have a class dynamically generated from a function which then is understood by typing tools as a base for new classes. To solve this problem without plugins, the usual call to declarative_base() can be replaced with using the DeclarativeBase class, which produces the same Base object as usual, except that typing tools understand it:

from sqlalchemy.orm import DeclarativeBase


class Base(DeclarativeBase):
    pass
Step two - replace Declarative use of Column with mapped_column()

The mapped_column() is an ORM-typing aware construct that can be swapped directly for the use of Column. Given a 1.x style mapping as:

from sqlalchemy import Column
from sqlalchemy.orm import relationship
from sqlalchemy.orm import DeclarativeBase


class Base(DeclarativeBase):
    pass


class User(Base):
    __tablename__ = "user_account"

    id = Column(Integer, primary_key=True)
    name = Column(String(30), nullable=False)
    fullname = Column(String)
    addresses = relationship("Address", back_populates="user")


class Address(Base):
    __tablename__ = "address"

    id = Column(Integer, primary_key=True)
    email_address = Column(String, nullable=False)
    user_id = Column(ForeignKey("user_account.id"), nullable=False)
    user = relationship("User", back_populates="addresses")

We replace Column with mapped_column(); no arguments need to change:

from sqlalchemy.orm import DeclarativeBase
from sqlalchemy.orm import mapped_column
from sqlalchemy.orm import relationship


class Base(DeclarativeBase):
    pass


class User(Base):
    __tablename__ = "user_account"

    id = mapped_column(Integer, primary_key=True)
    name = mapped_column(String(30), nullable=False)
    fullname = mapped_column(String)
    addresses = relationship("Address", back_populates="user")


class Address(Base):
    __tablename__ = "address"

    id = mapped_column(Integer, primary_key=True)
    email_address = mapped_column(String, nullable=False)
    user_id = mapped_column(ForeignKey("user_account.id"), nullable=False)
    user = relationship("User", back_populates="addresses")

The individual columns above are not yet typed with Python types, and are instead typed as Mapped[Any]; this is because we can declare any column either with Optional or not, and there’s no way to have a “guess” in place that won’t cause typing errors when we type it explicitly.

However, at this step, our above mapping has appropriate descriptor types set up for all attributes and may be used in queries as well as for instance-level manipulation, all of which will pass mypy –strict mode with no plugins.

Step three - apply exact Python types as needed using Mapped.

This can be done for all attributes for which exact typing is desired; attributes that are fine being left as Any may be skipped. For context we also illustrate Mapped being used for a relationship() where we apply an exact type. The mapping within this interim step will be more verbose, however with proficiency, this step can be combined with subsequent steps to update mappings more directly:

from typing import List
from typing import Optional
from sqlalchemy.orm import DeclarativeBase
from sqlalchemy.orm import Mapped
from sqlalchemy.orm import mapped_column
from sqlalchemy.orm import relationship


class Base(DeclarativeBase):
    pass


class User(Base):
    __tablename__ = "user_account"

    id: Mapped[int] = mapped_column(Integer, primary_key=True)
    name: Mapped[str] = mapped_column(String(30), nullable=False)
    fullname: Mapped[Optional[str]] = mapped_column(String)
    addresses: Mapped[List["Address"]] = relationship("Address", back_populates="user")


class Address(Base):
    __tablename__ = "address"

    id: Mapped[int] = mapped_column(Integer, primary_key=True)
    email_address: Mapped[str] = mapped_column(String, nullable=False)
    user_id: Mapped[int] = mapped_column(ForeignKey("user_account.id"), nullable=False)
    user: Mapped["User"] = relationship("User", back_populates="addresses")

At this point, our ORM mapping is fully typed and will produce exact-typed select(), Query and Result constructs. We now can proceed to pare down redundancy in the mapping declaration.

Step four - remove mapped_column() directives where no longer needed

All nullable parameters can be implied using Optional[]; in the absence of Optional[], nullable defaults to False. All SQL types without arguments such as Integer and String can be expressed as a Python annotation alone. A mapped_column() directive with no parameters can be removed entirely. relationship() now derives its class from the left hand annotation, supporting forward references as well (as relationship() has supported string-based forward references for ten years already ;) ):

from typing import List
from typing import Optional
from sqlalchemy.orm import DeclarativeBase
from sqlalchemy.orm import Mapped
from sqlalchemy.orm import mapped_column
from sqlalchemy.orm import relationship


class Base(DeclarativeBase):
    pass


class User(Base):
    __tablename__ = "user_account"

    id: Mapped[int] = mapped_column(primary_key=True)
    name: Mapped[str] = mapped_column(String(30))
    fullname: Mapped[Optional[str]]
    addresses: Mapped[List["Address"]] = relationship(back_populates="user")


class Address(Base):
    __tablename__ = "address"

    id: Mapped[int] = mapped_column(primary_key=True)
    email_address: Mapped[str]
    user_id: Mapped[int] = mapped_column(ForeignKey("user_account.id"))
    user: Mapped["User"] = relationship(back_populates="addresses")
Step five - make use of pep-593 Annotated to package common directives into types

This is a radical new capability that presents an alternative, or complementary approach, to declarative mixins as a means to provide type oriented configuration, and also replaces the need for declared_attr decorated functions in most cases.

First, the Declarative mapping allows the mapping of Python type to SQL type, such as str to String, to be customized using registry.type_annotation_map. Using PEP 593 Annotated allows us to create variants of a particular Python type so that the same type, such as str, may be used which each provide variants of String, as below where use of an Annotated str called str50 will indicate String(50):

from typing_extensions import Annotated
from sqlalchemy.orm import DeclarativeBase

str50 = Annotated[str, 50]


# declarative base with a type-level override, using a type that is
# expected to be used in multiple places
class Base(DeclarativeBase):
    type_annotation_map = {
        str50: String(50),
    }

Second, Declarative will extract full mapped_column() definitions from the left hand type if Annotated[] is used, by passing a mapped_column() construct as any argument to the Annotated[] construct (credit to @adriangb01 for illustrating this idea). This capability may be extended in future releases to also include relationship(), composite() and other constructs, but currently is limited to mapped_column(). The example below adds additional Annotated types in addition to our str50 example to illustrate this feature:

from typing_extensions import Annotated
from typing import List
from typing import Optional
from sqlalchemy import ForeignKey
from sqlalchemy import String
from sqlalchemy.orm import DeclarativeBase
from sqlalchemy.orm import Mapped
from sqlalchemy.orm import mapped_column
from sqlalchemy.orm import relationship

# declarative base from previous example
str50 = Annotated[str, 50]


class Base(DeclarativeBase):
    type_annotation_map = {
        str50: String(50),
    }


# set up mapped_column() overrides, using whole column styles that are
# expected to be used in multiple places
intpk = Annotated[int, mapped_column(primary_key=True)]
user_fk = Annotated[int, mapped_column(ForeignKey("user_account.id"))]


class User(Base):
    __tablename__ = "user_account"

    id: Mapped[intpk]
    name: Mapped[str50]
    fullname: Mapped[Optional[str]]
    addresses: Mapped[List["Address"]] = relationship(back_populates="user")


class Address(Base):
    __tablename__ = "address"

    id: Mapped[intpk]
    email_address: Mapped[str50]
    user_id: Mapped[user_fk]
    user: Mapped["User"] = relationship(back_populates="addresses")

Above, columns that are mapped with Mapped[str50], Mapped[intpk], or Mapped[user_fk] draw from both the registry.type_annotation_map as well as the Annotated construct directly in order to re-use pre-established typing and column configurations.

Optional step - turn mapped classes into dataclasses

We can turn mapped classes into dataclasses, where a key advantage is that we can build a strictly-typed __init__() method with explicit positional, keyword only, and default arguments, not to mention we get methods such as __str__() and __repr__() for free. The next section Native Support for Dataclasses Mapped as ORM Models illustrates further transformation of the above model.

Typing is supported from step 3 onwards

With the above examples, any example from “step 3” on forward will include that the attributes of the model are typed and will populate through to select(), Query, and Row objects:

# (variable) stmt: Select[Tuple[int, str]]
stmt = select(User.id, User.name)

with Session(e) as sess:
    for row in sess.execute(stmt):
        # (variable) row: Row[Tuple[int, str]]
        print(row)

    # (variable) users: Sequence[User]
    users = sess.scalars(select(User)).all()

    # (variable) users_legacy: List[User]
    users_legacy = sess.query(User).all()

See also

Declarative Table with mapped_column() - Updated Declarative documentation for Declarative generation and mapping of Table columns.

Using Legacy Mypy-Typed Models

SQLAlchemy applications that use the Mypy plugin with explicit annotations that don’t use Mapped in their annotations are subject to errors under the new system, as such annotations are flagged as errors when using constructs such as relationship().

The section Migration to 2.0 Step Six - Add __allow_unmapped__ to explicitly typed ORM models illustrates how to temporarily disable these errors from being raised for a legacy ORM model that uses explicit annotations.

Native Support for Dataclasses Mapped as ORM Models

The new ORM Declarative features introduced above at ORM Declarative Models introduced the new mapped_column() construct and illustrated type-centric mapping with optional use of PEP 593 Annotated. We can take the mapping one step further by integrating this with Python dataclasses. This new feature is made possible via PEP 681 which allows for type checkers to recognize classes that are dataclass compatible, or are fully dataclasses, but were declared through alternate APIs.

Using the dataclasses feature, mapped classes gain an __init__() method that supports positional arguments as well as customizable default values for optional keyword arguments. As mentioned previously, dataclasses also generate many useful methods such as __str__(), __eq__(). Dataclass serialization methods such as dataclasses.asdict() and dataclasses.astuple() also work, but don’t currently accommodate for self-referential structures, which makes them less viable for mappings that have bidirectional relationships.

SQLAlchemy’s current integration approach converts the user-defined class into a real dataclass to provide runtime functionality; the feature makes use of the existing dataclass feature introduced in SQLAlchemy 1.4 at Python Dataclasses, attrs Supported w/ Declarative, Imperative Mappings to produce an equivalent runtime mapping with a fully integrated configuration style, which is also more correctly typed than was possible with the previous approach.

To support dataclasses in compliance with PEP 681, ORM constructs like mapped_column() and relationship() accept additional PEP 681 arguments init, default, and default_factory which are passed along to the dataclass creation process. These arguments currently must be present in an explicit directive on the right side, just as they would be used with dataclasses.field(); they currently can’t be local to an Annotated construct on the left side. To support the convenient use of Annotated while still supporting dataclass configuration, mapped_column() can merge a minimal set of right-hand arguments with that of an existing mapped_column() construct located on the left side within an Annotated construct, so that most of the succinctness is maintained, as will be seen below.

To enable dataclasses using class inheritance we make use of the MappedAsDataclass mixin, either directly on each class, or on the Base class, as illustrated below where we further modify the example mapping from “Step 5” of ORM Declarative Models:

from typing_extensions import Annotated
from typing import List
from typing import Optional
from sqlalchemy import ForeignKey
from sqlalchemy import String
from sqlalchemy.orm import DeclarativeBase
from sqlalchemy.orm import Mapped
from sqlalchemy.orm import MappedAsDataclass
from sqlalchemy.orm import mapped_column
from sqlalchemy.orm import relationship


class Base(MappedAsDataclass, DeclarativeBase):
    """subclasses will be converted to dataclasses"""


intpk = Annotated[int, mapped_column(primary_key=True)]
str30 = Annotated[str, mapped_column(String(30))]
user_fk = Annotated[int, mapped_column(ForeignKey("user_account.id"))]


class User(Base):
    __tablename__ = "user_account"

    id: Mapped[intpk] = mapped_column(init=False)
    name: Mapped[str30]
    fullname: Mapped[Optional[str]] = mapped_column(default=None)
    addresses: Mapped[List["Address"]] = relationship(
        back_populates="user", default_factory=list
    )


class Address(Base):
    __tablename__ = "address"

    id: Mapped[intpk] = mapped_column(init=False)
    email_address: Mapped[str]
    user_id: Mapped[user_fk] = mapped_column(init=False)
    user: Mapped["User"] = relationship(back_populates="addresses", default=None)

The above mapping has used the @dataclasses.dataclass decorator directly on each mapped class at the same time that the declarative mapping was set up, internally setting up each dataclasses.field() directive as indicated. User / Address structures can be created using positional arguments as configured:

>>> u1 = User("username", fullname="full name", addresses=[Address("email@address")])
>>> u1
User(id=None, name='username', fullname='full name', addresses=[Address(id=None, email_address='email@address', user_id=None, user=...)])

Optimized ORM bulk insert now implemented for all backends other than MySQL

The dramatic performance improvement introduced in the 1.4 series and described at ORM Batch inserts with psycopg2 now batch statements with RETURNING in most cases has now been generalized to all included backends that support RETURNING, which is all backends other than MySQL: SQLite, MariaDB, PostgreSQL (all drivers), and Oracle; SQL Server has support but is temporarily disabled in version 2.0.9 [1]. While the original feature was most critical for the psycopg2 driver which otherwise had major performance issues when using cursor.executemany(), the change is also critical for other PostgreSQL drivers such as asyncpg, as when using RETURNING, single-statement INSERT statements are still unacceptably slow, as well as when using SQL Server that also seems to have very slow executemany speed for INSERT statements regardless of whether or not RETURNING is used.

The performance of the new feature provides an almost across-the-board order of magnitude performance increase for basically every driver when INSERTing ORM objects that don’t have a pre-assigned primary key value, as indicated in the table below, in most cases specific to the use of RETURNING which is not normally supported with executemany().

The psycopg2 “fast execution helper” approach consists of transforming an INSERT..RETURNING statement with a single parameter set into a single statement that INSERTs many parameter sets, using multiple “VALUES…” clauses so that it can accommodate many parameter sets at once. Parameter sets are then typically batched into groups of 1000 or similar, so that no single INSERT statement is excessively large, and the INSERT statement is then invoked for each batch of parameters, rather than for each individual parameter set. Primary key values and server defaults are returned by RETURNING, which continues to work as each statement execution is invoked using cursor.execute(), rather than cursor.executemany().

This allows many rows to be inserted in one statement while also being able to return newly-generated primary key values as well as SQL and server defaults. SQLAlchemy historically has always needed to invoke one statement per parameter set, as it relied upon Python DBAPI Features such as cursor.lastrowid which do not support multiple rows.

With most databases now offering RETURNING (with the conspicuous exception of MySQL, given that MariaDB supports it), the new change generalizes the psycopg2 “fast execution helper” approach to all dialects that support RETURNING, which now includes SQlite and MariaDB, and for which no other approach for “executemany plus RETURNING” is possible, which includes SQLite, MariaDB, and all PG drivers. The cx_Oracle and oracledb drivers used for Oracle support RETURNING with executemany natively, and this has also been implemented to provide equivalent performance improvements. With SQLite and MariaDB now offering RETURNING support, ORM use of cursor.lastrowid is nearly a thing of the past, with only MySQL still relying upon it.

For INSERT statements that don’t use RETURNING, traditional executemany() behavior is used for most backends, with the current exception of psycopg2, which has very slow executemany() performance overall and are still improved by the “insertmanyvalues” approach.

Benchmarks

SQLAlchemy includes a Performance Suite within the examples/ directory, where we can make use of the bulk_insert suite to benchmark INSERTs of many rows using both Core and ORM in different ways.

For the tests below, we are inserting 100,000 objects, and in all cases we actually have 100,000 real Python ORM objects in memory, either created up front or generated on the fly. All databases other than SQLite are run over a local network connection, not localhost; this causes the “slower” results to be extremely slow.

Operations that are improved by this feature include:

To get a sense of the scale of the operation, below are performance measurements using the test_flush_no_pk performance suite, which historically represents SQLAlchemy’s worst-case INSERT performance task, where objects that don’t have primary key values need to be INSERTed, and then the newly generated primary key values must be fetched so that the objects can be used for subsequent flush operations, such as establishment within relationships, flushing joined-inheritance models, etc:

@Profiler.profile
def test_flush_no_pk(n):
    """INSERT statements via the ORM (batched with RETURNING if available),
    fetching generated row id"""
    session = Session(bind=engine)
    for chunk in range(0, n, 1000):
        session.add_all(
            [
                Customer(
                    name="customer name %d" % i,
                    description="customer description %d" % i,
                )
                for i in range(chunk, chunk + 1000)
            ]
        )
        session.flush()
    session.commit()

This test can be run from any SQLAlchemy source tree as follows:

python -m examples.performance.bulk_inserts --test test_flush_no_pk

The table below summarizes performance measurements with the latest 1.4 series of SQLAlchemy compared to 2.0, both running the same test:

Driver

SQLA 1.4 Time (secs)

SQLA 2.0 Time (secs)

sqlite+pysqlite2 (memory)

6.204843

3.554856

postgresql+asyncpg (network)

88.292285

4.561492

postgresql+psycopg (network)

N/A (psycopg3)

4.861368

mssql+pyodbc (network)

158.396667

4.825139

oracle+cx_Oracle (network)

92.603953

4.809520

mariadb+mysqldb (network)

71.705197

4.075377

Note

Two additional drivers have no change in performance; the psycopg2 drivers, for which fast executemany was already implemented in SQLAlchemy 1.4, and MySQL, which continues to not offer RETURNING support:

Driver

SQLA 1.4 Time (secs)

SQLA 2.0 Time (secs)

postgresql+psycopg2 (network)

4.704876

4.699883

mysql+mysqldb (network)

77.281997

76.132995

Summary of Changes

The following bullets list the individual changes made within 2.0 in order to get all drivers to this state:

  • RETURNING implemented for SQLite - #6195

  • RETURNING implemented for MariaDB - #7011

  • Fix multi-row RETURNING for Oracle - #6245

  • make insert() executemany() support RETURNING for as many dialects as possible, usually with VALUES() - #6047

  • Emit a warning when RETURNING w/ executemany is used for non-supporting backend (currently no RETURNING backend has this limitation) - #7907

  • The ORM Mapper.eager_defaults parameter now defaults to a a new setting "auto", which will enable “eager defaults” automatically for INSERT statements, when the backend in use supports RETURNING with “insertmanyvalues”. See Fetching Server-Generated Defaults for documentation.

See also

“Insert Many Values” Behavior for INSERT statements - Documentation and background on the new feature as well as how to configure it

ORM-enabled Insert, Upsert, Update and Delete Statements, with ORM RETURNING

SQLAlchemy 1.4 ported the features of the legacy Query object to 2.0 style execution, which meant that the Select construct could be passed to Session.execute() to deliver ORM results. Support was also added for Update and Delete to be passed to Session.execute(), to the degree that they could provide implementations of Query.update() and Query.delete().

The major missing element has been support for the Insert construct. The 1.4 documentation addressed this with some recipes for “inserts” and “upserts” with use of Select.from_statement() to integrate RETURNING into an ORM context. 2.0 now fully closes the gap by integrating direct support for Insert as an enhanced version of the Session.bulk_insert_mappings() method, along with full ORM RETURNING support for all DML structures.

Bulk Insert with RETURNING

Insert can be passed to Session.execute(), with or without Insert.returning(), which when passed with a separate parameter list will invoke the same process as was previously implemented by Session.bulk_insert_mappings(), with additional enhancements. This will optimize the batching of rows making use of the new fast insertmany feature, while also adding support for heterogeneous parameter sets and multiple-table mappings like joined table inheritance:

>>> users = session.scalars(
...     insert(User).returning(User),
...     [
...         {"name": "spongebob", "fullname": "Spongebob Squarepants"},
...         {"name": "sandy", "fullname": "Sandy Cheeks"},
...         {"name": "patrick", "fullname": "Patrick Star"},
...         {"name": "squidward", "fullname": "Squidward Tentacles"},
...         {"name": "ehkrabs", "fullname": "Eugene H. Krabs"},
...     ],
... )
>>> print(users.all())
[User(name='spongebob', fullname='Spongebob Squarepants'),
 User(name='sandy', fullname='Sandy Cheeks'),
 User(name='patrick', fullname='Patrick Star'),
 User(name='squidward', fullname='Squidward Tentacles'),
 User(name='ehkrabs', fullname='Eugene H. Krabs')]

RETURNING is supported for all of these use cases, where the ORM will construct a full result set from multiple statement invocations.

Bulk UPDATE

In a similar manner as that of Insert, passing the Update construct along with a parameter list that includes primary key values to Session.execute() will invoke the same process as previously supported by the Session.bulk_update_mappings() method. This feature does not however support RETURNING, as it uses a SQL UPDATE statement that is invoked using DBAPI executemany:

>>> from sqlalchemy import update
>>> session.execute(
...     update(User),
...     [
...         {"id": 1, "fullname": "Spongebob Squarepants"},
...         {"id": 3, "fullname": "Patrick Star"},
...     ],
... )

INSERT / upsert … VALUES … RETURNING

When using Insert with Insert.values(), the set of parameters may include SQL expressions. Additionally, upsert variants such as those for SQLite, PostgreSQL and MariaDB are also supported. These statements may now include Insert.returning() clauses with column expressions or full ORM entities:

>>> from sqlalchemy.dialects.sqlite import insert as sqlite_upsert
>>> stmt = sqlite_upsert(User).values(
...     [
...         {"name": "spongebob", "fullname": "Spongebob Squarepants"},
...         {"name": "sandy", "fullname": "Sandy Cheeks"},
...         {"name": "patrick", "fullname": "Patrick Star"},
...         {"name": "squidward", "fullname": "Squidward Tentacles"},
...         {"name": "ehkrabs", "fullname": "Eugene H. Krabs"},
...     ]
... )
>>> stmt = stmt.on_conflict_do_update(
...     index_elements=[User.name], set_=dict(fullname=stmt.excluded.fullname)
... )
>>> result = session.scalars(stmt.returning(User))
>>> print(result.all())
[User(name='spongebob', fullname='Spongebob Squarepants'),
User(name='sandy', fullname='Sandy Cheeks'),
User(name='patrick', fullname='Patrick Star'),
User(name='squidward', fullname='Squidward Tentacles'),
User(name='ehkrabs', fullname='Eugene H. Krabs')]

ORM UPDATE / DELETE with WHERE … RETURNING

SQLAlchemy 1.4 also had some modest support for the RETURNING feature to be used with the update() and delete() constructs, when used with Session.execute(). This support has now been upgraded to be fully native, including that the fetch synchronization strategy may also proceed whether or not explicit use of RETURNING is present:

>>> from sqlalchemy import update
>>> stmt = (
...     update(User)
...     .where(User.name == "squidward")
...     .values(name="spongebob")
...     .returning(User)
... )
>>> result = session.scalars(stmt, execution_options={"synchronize_session": "fetch"})
>>> print(result.all())

Improved synchronize_session behavior for ORM UPDATE / DELETE

The default strategy for synchronize_session is now a new value "auto". This strategy will attempt to use the "evaluate" strategy and then automatically fall back to the "fetch" strategy. For all backends other than MySQL / MariaDB, "fetch" uses RETURNING to fetch UPDATE/DELETEd primary key identifiers within the same statement, so is generally more efficient than previous versions (in 1.4, RETURNING was only available for PostgreSQL, SQL Server).

Summary of Changes

Listed tickets for new ORM DML with RETURNING features:

  • convert insert() at ORM level to interpret values() in an ORM context - #7864

  • evaluate feasibility of dml.returning(Entity) to deliver ORM expressions, automatically apply select().from_statement equiv - #7865

  • given ORM insert, try to carry the bulk methods along, re: inheritance - #8360

New “Write Only” relationship strategy supersedes “dynamic”

The lazy="dynamic" loader strategy becomes legacy, in that it is hardcoded to make use of legacy Query. This loader strategy is both not compatible with asyncio, and additionally has many behaviors that implicitly iterate its contents, which defeat the original purpose of the “dynamic” relationship as being for very large collections that should not be implicitly fully loaded into memory at any time.

The “dynamic” strategy is now superseded by a new strategy lazy="write_only". Configuration of “write only” may be achieved using the relationship.lazy parameter of relationship(), or when using type annotated mappings, indicating the WriteOnlyMapped annotation as the mapping style:

from sqlalchemy.orm import WriteOnlyMapped


class Base(DeclarativeBase):
    pass


class Account(Base):
    __tablename__ = "account"
    id: Mapped[int] = mapped_column(primary_key=True)
    identifier: Mapped[str]
    account_transactions: WriteOnlyMapped["AccountTransaction"] = relationship(
        cascade="all, delete-orphan",
        passive_deletes=True,
        order_by="AccountTransaction.timestamp",
    )


class AccountTransaction(Base):
    __tablename__ = "account_transaction"
    id: Mapped[int] = mapped_column(primary_key=True)
    account_id: Mapped[int] = mapped_column(
        ForeignKey("account.id", ondelete="cascade")
    )
    description: Mapped[str]
    amount: Mapped[Decimal]
    timestamp: Mapped[datetime] = mapped_column(default=func.now())

The write-only-mapped collection resembles lazy="dynamic" in that the collection may be assigned up front, and also has methods such as WriteOnlyCollection.add() and WriteOnlyCollection.remove() to modify the collection on an individual item basis:

new_account = Account(
    identifier="account_01",
    account_transactions=[
        AccountTransaction(description="initial deposit", amount=Decimal("500.00")),
        AccountTransaction(description="transfer", amount=Decimal("1000.00")),
        AccountTransaction(description="withdrawal", amount=Decimal("-29.50")),
    ],
)

new_account.account_transactions.add(
    AccountTransaction(description="transfer", amount=Decimal("2000.00"))
)

The bigger difference is on the database loading side, where the collection has no ability to load objects from the database directly; instead, SQL construction methods such as WriteOnlyCollection.select() are used to produce SQL constructs such as Select which are then executed using 2.0 style to load the desired objects in an explicit way:

account_transactions = session.scalars(
    existing_account.account_transactions.select()
    .where(AccountTransaction.amount < 0)
    .limit(10)
).all()

The WriteOnlyCollection also integrates with the new ORM bulk dml features, including support for bulk INSERT and UPDATE/DELETE with WHERE criteria, all including RETURNING support as well. See the complete documentation at Write Only Relationships.

New pep-484 / type annotated mapping support for Dynamic Relationships

Even though “dynamic” relationships are legacy in 2.0, as these patterns are expected to have a long lifespan, type annotated mapping support is now added for “dynamic” relationships in the same way that its available for the new lazy="write_only" approach, using the DynamicMapped annotation:

from sqlalchemy.orm import DynamicMapped


class Base(DeclarativeBase):
    pass


class Account(Base):
    __tablename__ = "account"
    id: Mapped[int] = mapped_column(primary_key=True)
    identifier: Mapped[str]
    account_transactions: DynamicMapped["AccountTransaction"] = relationship(
        cascade="all, delete-orphan",
        passive_deletes=True,
        order_by="AccountTransaction.timestamp",
    )


class AccountTransaction(Base):
    __tablename__ = "account_transaction"
    id: Mapped[int] = mapped_column(primary_key=True)
    account_id: Mapped[int] = mapped_column(
        ForeignKey("account.id", ondelete="cascade")
    )
    description: Mapped[str]
    amount: Mapped[Decimal]
    timestamp: Mapped[datetime] = mapped_column(default=func.now())

The above mapping will provide an Account.account_transactions collection that is typed as returning the AppenderQuery collection type, including its element type, e.g. AppenderQuery[AccountTransaction]. This then allows iteration and queries to yield objects which are typed as AccountTransaction.

#7123

Installation is now fully pep-517 enabled

The source distribution now includes a pyproject.toml file to allow for complete PEP 517 support. In particular this allows a local source build using pip to automatically install the Cython optional dependency.

#7311

C Extensions now ported to Cython

The SQLAlchemy C extensions have been replaced with all new extensions written in Cython. While Cython was evaluated back in 2010 when the C extensions were first created, the nature and focus of the C extensions in use today has changed quite a bit from that time. At the same time, Cython has apparently evolved significantly, as has the Python build / distribution toolchain which made it feasible for us to revisit it.

The move to Cython provides dramatic new advantages with no apparent downsides:

  • The Cython extensions that replace specific C extensions have all benchmarked as faster, often slightly, but sometimes significantly, than virtually all the C code that SQLAlchemy previously included. While this seems amazing, it appears to be a product of non-obvious optimizations within Cython’s implementation that would not be present in a direct Python to C port of a function, as was particularly the case for many of the custom collection types added to the C extensions.

  • Cython extensions are much easier to write, maintain and debug compared to raw C code, and in most cases are line-per-line equivalent to the Python code. It is expected that many more elements of SQLAlchemy will be ported to Cython in the coming releases which should open many new doors to performance improvements that were previously out of reach.

  • Cython is very mature and widely used, including being the basis of some of the prominent database drivers supported by SQLAlchemy including asyncpg, psycopg3 and asyncmy.

Like the previous C extensions, the Cython extensions are pre-built within SQLAlchemy’s wheel distributions which are automatically available to pip from PyPi. Manual build instructions are also unchanged with the exception of the Cython requirement.

#7256

Major Architectural, Performance and API Enhancements for Database Reflection

The internal system by which Table objects and their components are reflected has been completely rearchitected to allow high performance bulk reflection of thousands of tables at once for participating dialects. Currently, the PostgreSQL and Oracle dialects participate in the new architecture, where the PostgreSQL dialect can now reflect a large series of Table objects nearly three times faster, and the Oracle dialect can now reflect a large series of Table objects ten times faster.

The rearchitecture applies most directly to dialects that make use of SELECT queries against system catalog tables to reflect tables, and the remaining included dialect that can benefit from this approach will be the SQL Server dialect. The MySQL/MariaDB and SQLite dialects by contrast make use of non-relational systems to reflect database tables, and were not subject to a pre-existing performance issue.

The new API is backwards compatible with the previous system, and should require no changes to third party dialects to retain compatibility; third party dialects can also opt into the new system by implementing batched queries for schema reflection.

Along with this change, the API and behavior of the Inspector object has been improved and enhanced with more consistent cross-dialect behaviors as well as new methods and new performance features.

Performance Overview

The source distribution includes a script test/perf/many_table_reflection.py which benches both existing reflection features as well as new ones. A limited set of its tests may be run on older versions of SQLAlchemy, where here we use it to illustrate differences in performance to invoke metadata.reflect() to reflect 250 Table objects at once over a local network connection:

Dialect

Operation

SQLA 1.4 Time (secs)

SQLA 2.0 Time (secs)

postgresql+psycopg2

metadata.reflect(), 250 tables

8.2

3.3

oracle+cx_oracle

metadata.reflect(), 250 tables

60.4

6.8

Behavioral Changes for Inspector()

For SQLAlchemy-included dialects for SQLite, PostgreSQL, MySQL/MariaDB, Oracle, and SQL Server, the Inspector.has_table(), Inspector.has_sequence(), Inspector.has_index(), Inspector.get_table_names() and Inspector.get_sequence_names() now all behave consistently in terms of caching: they all fully cache their result after being called the first time for a particular Inspector object. Programs that create or drop tables/sequences while calling upon the same Inspector object will not receive updated status after the state of the database has changed. A call to Inspector.clear_cache() or a new Inspector should be used when DDL changes are to be executed. Previously, the Inspector.has_table(), Inspector.has_sequence() methods did not implement caching nor did the Inspector support caching for these methods, while the Inspector.get_table_names() and Inspector.get_sequence_names() methods were, leading to inconsistent results between the two types of method.

Behavior for third party dialects is dependent on whether or not they implement the “reflection cache” decorator for the dialect-level implementation of these methods.

New Methods and Improvements for Inspector()

  • added a method Inspector.has_schema() that returns if a schema is present in the target database

  • added a method Inspector.has_index() that returns if a table has a particular index.

  • Inspection methods such as Inspector.get_columns() that work on a single table at a time should now all consistently raise NoSuchTableError if a table or view is not found; this change is specific to individual dialects, so may not be the case for existing third-party dialects.

  • Separated the handling of “views” and “materialized views”, as in real world use cases, these two constructs make use of different DDL for CREATE and DROP; this includes that there are now separate Inspector.get_view_names() and Inspector.get_materialized_view_names() methods.

#4379

Dialect support for psycopg 3 (a.k.a. “psycopg”)

Added dialect support for the psycopg 3 DBAPI, which despite the number “3” now goes by the package name psycopg, superseding the previous psycopg2 package that for the time being remains SQLAlchemy’s “default” driver for the postgresql dialects. psycopg is a completely reworked and modernized database adapter for PostgreSQL which supports concepts such as prepared statements as well as Python asyncio.

psycopg is the first DBAPI supported by SQLAlchemy which provides both a pep-249 synchronous API as well as an asyncio driver. The same psycopg database URL may be used with the create_engine() and create_async_engine() engine-creation functions, and the corresponding sync or asyncio version of the dialect will be selected automatically.

See also

psycopg

Dialect support for oracledb

Added dialect support for the oracledb DBAPI, which is the renamed, new major release of the popular cx_Oracle driver.

See also

python-oracledb

New Conditional DDL for Constraints and Indexes

A new method Constraint.ddl_if() and Index.ddl_if() allows constructs such as CheckConstraint, UniqueConstraint and Index to be rendered conditionally for a given Table, based on the same kinds of criteria that are accepted by the DDLElement.execute_if() method. In the example below, the CHECK constraint and index will only be produced against a PostgreSQL backend:

meta = MetaData()


my_table = Table(
    "my_table",
    meta,
    Column("id", Integer, primary_key=True),
    Column("num", Integer),
    Column("data", String),
    Index("my_pg_index", "data").ddl_if(dialect="postgresql"),
    CheckConstraint("num > 5").ddl_if(dialect="postgresql"),
)

e1 = create_engine("sqlite://", echo=True)
meta.create_all(e1)  # will not generate CHECK and INDEX


e2 = create_engine("postgresql://scott:tiger@localhost/test", echo=True)
meta.create_all(e2)  # will generate CHECK and INDEX

#7631

DATE, TIME, DATETIME datatypes now support literal rendering on all backends

Literal rendering is now implemented for date and time types for backend specific compilation, including PostgreSQL and Oracle:

>>> import datetime

>>> from sqlalchemy import DATETIME
>>> from sqlalchemy import literal
>>> from sqlalchemy.dialects import oracle
>>> from sqlalchemy.dialects import postgresql

>>> date_literal = literal(datetime.datetime.now(), DATETIME)

>>> print(
...     date_literal.compile(
...         dialect=postgresql.dialect(), compile_kwargs={"literal_binds": True}
...     )
... )
'2022-12-17 11:02:13.575789'
>>> print( ... date_literal.compile( ... dialect=oracle.dialect(), compile_kwargs={"literal_binds": True} ... ) ... )
TO_TIMESTAMP('2022-12-17 11:02:13.575789', 'YYYY-MM-DD HH24:MI:SS.FF')

Previously, such literal rendering only worked when stringifying statements without any dialect given; when attempting to render with a dialect-specific type, a NotImplementedError would be raised, up until SQLAlchemy 1.4.45 where this became a CompileError (part of #8800).

The default rendering is modified ISO-8601 rendering (i.e. ISO-8601 with the T converted to a space) when using literal_binds with the SQL compilers provided by the PostgreSQL, MySQL, MariaDB, MSSQL, Oracle dialects. For Oracle, the ISO format is wrapped inside of an appropriate TO_DATE() function call. The rendering for SQLite is unchanged as this dialect always included string rendering for date values.

#5052

Context Manager Support for Result, AsyncResult

The Result object now supports context manager use, which will ensure the object and its underlying cursor is closed at the end of the block. This is useful in particular with server side cursors, where it’s important that the open cursor object is closed at the end of an operation, even if user-defined exceptions have occurred:

with engine.connect() as conn:
    with conn.execution_options(yield_per=100).execute(
        text("select * from table")
    ) as result:
        for row in result:
            print(f"{row}")

With asyncio use, the AsyncResult and AsyncConnection have been altered to provide for optional async context manager use, as in:

async with async_engine.connect() as conn:
    async with conn.execution_options(yield_per=100).execute(
        text("select * from table")
    ) as result:
        for row in result:
            print(f"{row}")

#8710

Behavioral Changes

This section covers behavioral changes made in SQLAlchemy 2.0 which are not otherwise part of the major 1.4->2.0 migration path; changes here are not expected to have significant effects on backwards compatibility.

New transaction join modes for Session

The behavior of “joining an external transaction into a Session” has been revised and improved, allowing explicit control over how the Session will accommodate an incoming Connection that already has a transaction and possibly a savepoint already established. The new parameter Session.join_transaction_mode includes a series of option values which can accommodate the existing transaction in several ways, most importantly allowing a Session to operate in a fully transactional style using savepoints exclusively, while leaving the externally initiated transaction non-committed and active under all circumstances, allowing test suites to rollback all changes that take place within tests.

The primary improvement this allows is that the recipe documented at Joining a Session into an External Transaction (such as for test suites), which also changed from SQLAlchemy 1.3 to 1.4, is now simplified to no longer require explicit use of an event handler or any mention of an explicit savepoint; by using join_transaction_mode="create_savepoint", the Session will never affect the state of an incoming transaction, and will instead create a savepoint (i.e. “nested transaction”) as its root transaction.

The following illustrates part of the example given at Joining a Session into an External Transaction (such as for test suites); see that section for a full example:

class SomeTest(TestCase):
    def setUp(self):
        # connect to the database
        self.connection = engine.connect()

        # begin a non-ORM transaction
        self.trans = self.connection.begin()

        # bind an individual Session to the connection, selecting
        # "create_savepoint" join_transaction_mode
        self.session = Session(
            bind=self.connection, join_transaction_mode="create_savepoint"
        )

    def tearDown(self):
        self.session.close()

        # rollback non-ORM transaction
        self.trans.rollback()

        # return connection to the Engine
        self.connection.close()

The default mode selected for Session.join_transaction_mode is "conditional_savepoint", which uses "create_savepoint" behavior if the given Connection is itself already on a savepoint. If the given Connection is in a transaction but not a savepoint, the Session will propagate “rollback” calls but not “commit” calls, but will not begin a new savepoint on its own. This behavior is chosen by default for its maximum compatibility with older SQLAlchemy versions as well as that it does not start a new SAVEPOINT unless the given driver is already making use of SAVEPOINT, as support for SAVEPOINT varies not only with specific backend and driver but also configurationally.

The following illustrates a case that worked in SQLAlchemy 1.3, stopped working in SQLAlchemy 1.4, and is now restored in SQLAlchemy 2.0:

engine = create_engine("...")

# setup outer connection with a transaction and a SAVEPOINT
conn = engine.connect()
trans = conn.begin()
nested = conn.begin_nested()

# bind a Session to that connection and operate upon it, including
# a commit
session = Session(conn)
session.connection()
session.commit()
session.close()

# assert both SAVEPOINT and transaction remain active
assert nested.is_active
nested.rollback()
trans.rollback()

Where above, a Session is joined to a Connection that has a savepoint started on it; the state of these two units remains unchanged after the Session has worked with the transaction. In SQLAlchemy 1.3, the above case worked because the Session would begin a “subtransaction” upon the Connection, which would allow the outer savepoint / transaction to remain unaffected for simple cases as above. Since subtransactions were deprecated in 1.4 and are now removed in 2.0, this behavior was no longer available. The new default behavior improves upon the behavior of “subtransactions” by using a real, second SAVEPOINT instead, so that even calls to Session.rollback() prevent the Session from “breaking out” into the externally initiated SAVEPOINT or transaction.

New code that is joining a transaction-started Connection into a Session should however select a Session.join_transaction_mode explicitly, so that the desired behavior is explicitly defined.

#9015

str(engine.url) will obfuscate the password by default

To avoid leakage of database passwords, calling str() on a URL will now enable the password obfuscation feature by default. Previously, this obfuscation would be in place for __repr__() calls but not __str__(). This change will impact applications and test suites that attempt to invoke create_engine() given the stringified URL from another engine, such as:

>>> e1 = create_engine("postgresql+psycopg2://scott:tiger@localhost/test")
>>> e2 = create_engine(str(e1.url))

The above engine e2 will not have the correct password; it will have the obfuscated string "***".

The preferred approach for the above pattern is to pass the URL object directly, there’s no need to stringify:

>>> e1 = create_engine("postgresql+psycopg2://scott:tiger@localhost/test")
>>> e2 = create_engine(e1.url)

Otherwise, for a stringified URL with cleartext password, use the URL.render_as_string() method, passing the URL.render_as_string.hide_password parameter as False:

>>> e1 = create_engine("postgresql+psycopg2://scott:tiger@localhost/test")
>>> url_string = e1.url.render_as_string(hide_password=False)
>>> e2 = create_engine(url_string)

#8567

Stricter rules for replacement of Columns in Table objects with same-names, keys

Stricter rules are in place for appending of Column objects to Table objects, both moving some previous deprecation warnings to exceptions, and preventing some previous scenarios that would cause duplicate columns to appear in tables, when Table.extend_existing were set to True, for both programmatic Table construction as well as during reflection operations.

  • Under no circumstances should a Table object ever have two or more Column objects with the same name, regardless of what .key they have. An edge case where this was still possible was identified and fixed.

  • Adding a Column to a Table that has the same name or key as an existing Column will always raise DuplicateColumnError (a new subclass of ArgumentError in 2.0.0b4) unless additional parameters are present; Table.append_column.replace_existing for Table.append_column(), and Table.extend_existing for construction of a same-named Table as an existing one, with or without reflection being used. Previously, there was a deprecation warning in place for this scenario.

  • A warning is now emitted if a Table is created, that does include Table.extend_existing, where an incoming Column that has no separate Column.key would fully replace an existing Column that does have a key, which suggests the operation is not what the user intended. This can happen particularly during a secondary reflection step, such as metadata.reflect(extend_existing=True). The warning suggests that the Table.autoload_replace parameter be set to False to prevent this. Previously, in 1.4 and earlier, the incoming column would be added in addition to the existing column. This was a bug and is a behavioral change in 2.0 (as of 2.0.0b4), as the previous key will no longer be present in the column collection when this occurs.

#8925

ORM Declarative Applies Column Orders Differently; Control behavior using sort_order

Declarative has changed the system by which mapped columns that originate from mixin or abstract base classes are sorted along with the columns that are on the declared class itself to place columns from the declared class first, followed by mixin columns. The following mapping:

class Foo:
    col1 = mapped_column(Integer)
    col3 = mapped_column(Integer)


class Bar:
    col2 = mapped_column(Integer)
    col4 = mapped_column(Integer)


class Model(Base, Foo, Bar):
    id = mapped_column(Integer, primary_key=True)
    __tablename__ = "model"

Produces a CREATE TABLE as follows on 1.4:

CREATE TABLE model (
  col1 INTEGER,
  col3 INTEGER,
  col2 INTEGER,
  col4 INTEGER,
  id INTEGER NOT NULL,
  PRIMARY KEY (id)
)

Whereas on 2.0 it produces:

CREATE TABLE model (
  id INTEGER NOT NULL,
  col1 INTEGER,
  col3 INTEGER,
  col2 INTEGER,
  col4 INTEGER,
  PRIMARY KEY (id)
)

For the specific case above, this can be seen as an improvement, as the primary key columns on the Model are now where one would typically prefer. However, this is no comfort for the application that defined models the other way around, as:

class Foo:
    id = mapped_column(Integer, primary_key=True)
    col1 = mapped_column(Integer)
    col3 = mapped_column(Integer)


class Model(Foo, Base):
    col2 = mapped_column(Integer)
    col4 = mapped_column(Integer)
    __tablename__ = "model"

This now produces CREATE TABLE output as:

CREATE TABLE model (
  col2 INTEGER,
  col4 INTEGER,
  id INTEGER NOT NULL,
  col1 INTEGER,
  col3 INTEGER,
  PRIMARY KEY (id)
)

To solve this issue, SQLAlchemy 2.0.4 introduces a new parameter on mapped_column() called mapped_column.sort_order, which is an integer value, defaulting to 0, that can be set to a positive or negative value so that columns are placed before or after other columns, as in the example below:

class Foo:
    id = mapped_column(Integer, primary_key=True, sort_order=-10)
    col1 = mapped_column(Integer, sort_order=-1)
    col3 = mapped_column(Integer)


class Model(Foo, Base):
    col2 = mapped_column(Integer)
    col4 = mapped_column(Integer)
    __tablename__ = "model"

The above model places “id” before all others and “col1” after “id”:

CREATE TABLE model (
  id INTEGER NOT NULL,
  col1 INTEGER,
  col2 INTEGER,
  col4 INTEGER,
  col3 INTEGER,
  PRIMARY KEY (id)
)

Future SQLAlchemy releases may opt to provide an explicit ordering hint for the mapped_column construct, as this ordering is ORM specific.

The Sequence construct reverts to not having any explicit default “start” value; impacts MS SQL Server

Prior to SQLAlchemy 1.4, the Sequence construct would emit only simple CREATE SEQUENCE DDL, if no additional arguments were specified:

>>> # SQLAlchemy 1.3 (and 2.0)
>>> from sqlalchemy import Sequence
>>> from sqlalchemy.schema import CreateSequence
>>> print(CreateSequence(Sequence("my_seq")))
CREATE SEQUENCE my_seq

However, as Sequence support was added for MS SQL Server, where the default start value is inconveniently set to -2**63, version 1.4 decided to default the DDL to emit a start value of 1, if Sequence.start were not otherwise provided:

>>> # SQLAlchemy 1.4 (only)
>>> from sqlalchemy import Sequence
>>> from sqlalchemy.schema import CreateSequence
>>> print(CreateSequence(Sequence("my_seq")))
CREATE SEQUENCE my_seq START WITH 1

This change has introduced other complexities, including that when the Sequence.min_value parameter is included, this default of 1 should in fact default to what Sequence.min_value states, else a min_value that’s below the start_value may be seen as contradictory. As looking at this issue started to become a bit of a rabbit hole of other various edge cases, we decided to instead revert this change and restore the original behavior of Sequence which is to have no opinion, and just emit CREATE SEQUENCE, allowing the database itself to make its decisions on how the various parameters of SEQUENCE should interact with each other.

Therefore, to ensure that the start value is 1 on all backends, the start value of 1 may be indicated explicitly, as below:

>>> # All SQLAlchemy versions
>>> from sqlalchemy import Sequence
>>> from sqlalchemy.schema import CreateSequence
>>> print(CreateSequence(Sequence("my_seq", start=1)))
CREATE SEQUENCE my_seq START WITH 1

Beyond all of that, for autogeneration of integer primary keys on modern backends including PostgreSQL, Oracle, SQL Server, the Identity construct should be preferred, which also works the same way in 1.4 and 2.0 with no changes in behavior.

#7211

“with_variant()” clones the original TypeEngine rather than changing the type

The TypeEngine.with_variant() method, which is used to apply alternate per-database behaviors to a particular type, now returns a copy of the original TypeEngine object with the variant information stored internally, rather than wrapping it inside the Variant class.

While the previous Variant approach was able to maintain all the in-Python behaviors of the original type using dynamic attribute getters, the improvement here is that when calling upon a variant, the returned type remains an instance of the original type, which works more smoothly with type checkers such as mypy and pylance. Given a program as below:

import typing

from sqlalchemy import String
from sqlalchemy.dialects.mysql import VARCHAR

type_ = String(255).with_variant(VARCHAR(255, charset="utf8mb4"), "mysql", "mariadb")

if typing.TYPE_CHECKING:
    reveal_type(type_)

A type checker like pyright will now report the type as:

info: Type of "type_" is "String"

In addition, as illustrated above, multiple dialect names may be passed for single type, in particular this is helpful for the pair of "mysql" and "mariadb" dialects which are considered separately as of SQLAlchemy 1.4.

#6980

Python division operator performs true division for all backends; added floor division

The Core expression language now supports both “true division” (i.e. the / Python operator) and “floor division” (i.e. the // Python operator) including backend-specific behaviors to normalize different databases in this regard.

Given a “true division” operation against two integer values:

expr = literal(5, Integer) / literal(10, Integer)

The SQL division operator on PostgreSQL for example normally acts as “floor division” when used against integers, meaning the above result would return the integer “0”. For this and similar backends, SQLAlchemy now renders the SQL using a form which is equivalent towards:

%(param_1)s / CAST(%(param_2)s AS NUMERIC)

With param_1=5, param_2=10, so that the return expression will be of type NUMERIC, typically as the Python value decimal.Decimal("0.5").

Given a “floor division” operation against two integer values:

expr = literal(5, Integer) // literal(10, Integer)

The SQL division operator on MySQL and Oracle for example normally acts as “true division” when used against integers, meaning the above result would return the floating point value “0.5”. For these and similar backends, SQLAlchemy now renders the SQL using a form which is equivalent towards:

FLOOR(%(param_1)s / %(param_2)s)

With param_1=5, param_2=10, so that the return expression will be of type INTEGER, as the Python value 0.

The backwards-incompatible change here would be if an application using PostgreSQL, SQL Server, or SQLite which relied on the Python “truediv” operator to return an integer value in all cases. Applications which rely upon this behavior should instead use the Python “floor division” operator // for these operations, or for forwards compatibility when using a previous SQLAlchemy version, the floor function:

expr = func.floor(literal(5, Integer) / literal(10, Integer))

The above form would be needed on any SQLAlchemy version prior to 2.0 in order to provide backend-agnostic floor division.

#4926

Session raises proactively when illegal concurrent or reentrant access is detected

The Session can now trap more errors related to illegal concurrent state changes within multithreaded or other concurrent scenarios as well as for event hooks which perform unexpected state changes.

One error that’s been known to occur when a Session is used in multiple threads simultaneously is AttributeError: 'NoneType' object has no attribute 'twophase', which is completely cryptic. This error occurs when a thread calls Session.commit() which internally invokes the SessionTransaction.close() method to end the transactional context, at the same time that another thread is in progress running a query as from Session.execute(). Within Session.execute(), the internal method that acquires a database connection for the current transaction first begins by asserting that the session is “active”, but after this assertion passes, the concurrent call to Session.close() interferes with this state which leads to the undefined condition above.

The change applies guards to all state-changing methods surrounding the SessionTransaction object so that in the above case, the Session.commit() method will instead fail as it will seek to change the state to one that is disallowed for the duration of the already-in-progress method that wants to get the current connection to run a database query.

Using the test script illustrated at #7433, the previous error case looks like:

Traceback (most recent call last):
File "/home/classic/dev/sqlalchemy/test3.py", line 30, in worker
    sess.execute(select(A)).all()
File "/home/classic/tmp/sqlalchemy/lib/sqlalchemy/orm/session.py", line 1691, in execute
    conn = self._connection_for_bind(bind)
File "/home/classic/tmp/sqlalchemy/lib/sqlalchemy/orm/session.py", line 1532, in _connection_for_bind
    return self._transaction._connection_for_bind(
File "/home/classic/tmp/sqlalchemy/lib/sqlalchemy/orm/session.py", line 754, in _connection_for_bind
    if self.session.twophase and self._parent is None:
AttributeError: 'NoneType' object has no attribute 'twophase'

Where the _connection_for_bind() method isn’t able to continue since concurrent access placed it into an invalid state. Using the new approach, the originator of the state change throws the error instead:

File "/home/classic/dev/sqlalchemy/lib/sqlalchemy/orm/session.py", line 1785, in close
   self._close_impl(invalidate=False)
File "/home/classic/dev/sqlalchemy/lib/sqlalchemy/orm/session.py", line 1827, in _close_impl
   transaction.close(invalidate)
File "<string>", line 2, in close
File "/home/classic/dev/sqlalchemy/lib/sqlalchemy/orm/session.py", line 506, in _go
   raise sa_exc.InvalidRequestError(
sqlalchemy.exc.InvalidRequestError: Method 'close()' can't be called here;
method '_connection_for_bind()' is already in progress and this would cause
an unexpected state change to symbol('CLOSED')

The state transition checks intentionally don’t use explicit locks to detect concurrent thread activity, instead relying upon simple attribute set / value test operations that inherently fail when unexpected concurrent changes occur. The rationale is that the approach can detect illegal state changes that occur entirely within a single thread, such as an event handler that runs on session transaction events calls a state-changing method that’s not expected, or under asyncio if a particular Session were shared among multiple asyncio tasks, as well as when using patching-style concurrency approaches such as gevent.

#7433

The SQLite dialect uses QueuePool for file-based databases

The SQLite dialect now defaults to QueuePool when a file based database is used. This is set along with setting the check_same_thread parameter to False. It has been observed that the previous approach of defaulting to NullPool, which does not hold onto database connections after they are released, did in fact have a measurable negative performance impact. As always, the pool class is customizable via the create_engine.poolclass parameter.

#7490

New Oracle FLOAT type with binary precision; decimal precision not accepted directly

A new datatype FLOAT has been added to the Oracle dialect, to accompany the addition of Double and database-specific DOUBLE, DOUBLE_PRECISION and REAL datatypes. Oracle’s FLOAT accepts a so-called “binary precision” parameter that per Oracle documentation is roughly a standard “precision” value divided by 0.3103:

from sqlalchemy.dialects import oracle

Table("some_table", metadata, Column("value", oracle.FLOAT(126)))

A binary precision value of 126 is synonymous with using the DOUBLE_PRECISION datatype, and a value of 63 is equivalent to using the REAL datatype. Other precision values are specific to the FLOAT type itself.

The SQLAlchemy Float datatype also accepts a “precision” parameter, but this is decimal precision which is not accepted by Oracle. Rather than attempting to guess the conversion, the Oracle dialect will now raise an informative error if Float is used with a precision value against the Oracle backend. To specify a Float datatype with an explicit precision value for supporting backends, while also supporting other backends, use the TypeEngine.with_variant() method as follows:

from sqlalchemy.types import Float
from sqlalchemy.dialects import oracle

Table(
    "some_table",
    metadata,
    Column("value", Float(5).with_variant(oracle.FLOAT(16), "oracle")),
)

New RANGE / MULTIRANGE support and changes for PostgreSQL backends

RANGE / MULTIRANGE support has been fully implemented for psycopg2, psycopg3, and asyncpg dialects. The new support uses a new SQLAlchemy-specific Range object that is agnostic of the different backends and does not require the use of backend-specific imports or extension steps. For multirange support, lists of Range objects are used.

Code that used the previous psycopg2-specific types should be modified to use Range, which presents a compatible interface.

The Range object also features comparison support which mirrors that of PostgreSQL. Implemented so far are Range.contains() and Range.contained_by() methods which work in the same way as the PostgreSQL @> and <@. Additional operator support may be added in future releases.

See the documentation at Range and Multirange Types for background on using the new feature.

#7156 #8706

match() operator on PostgreSQL uses plainto_tsquery() rather than to_tsquery()

The Operators.match() function now renders col @@ plainto_tsquery(expr) on the PostgreSQL backend, rather than col @@ to_tsquery(). plainto_tsquery() accepts plain text whereas to_tsquery() accepts specialized query symbols, and is therefore less cross-compatible with other backends.

All PostgreSQL search functions and operators are available through use of func to generate PostgreSQL-specific functions and Operators.bool_op() (a boolean-typed version of Operators.op()) to generate arbitrary operators, in the same manner as they are available in previous versions. See the examples at Full Text Search.

Existing SQLAlchemy projects that make use of PG-specific directives within Operators.match() should make use of func.to_tsquery() directly. To render SQL in exactly the same form as would be present in 1.4, see the version note at Simple plain text matching with match().

#7086