Working with Engines and Connections

This section details direct usage of the Engine, Connection, and related objects. Its important to note that when using the SQLAlchemy ORM, these objects are not generally accessed; instead, the Session object is used as the interface to the database. However, for applications that are built around direct usage of textual SQL statements and/or SQL expression constructs without involvement by the ORM’s higher level management services, the Engine and Connection are king (and queen?) - read on.

Basic Usage

Recall from Engine Configuration that an Engine is created via the create_engine() call:

engine = create_engine("mysql+mysqldb://scott:tiger@localhost/test")

The typical usage of create_engine() is once per particular database URL, held globally for the lifetime of a single application process. A single Engine manages many individual DBAPI connections on behalf of the process and is intended to be called upon in a concurrent fashion. The Engine is not synonymous to the DBAPI connect() function, which represents just one connection resource - the Engine is most efficient when created just once at the module level of an application, not per-object or per-function call.

tip

When using an Engine with multiple Python processes, such as when using os.fork or Python multiprocessing, it’s important that the engine is initialized per process. See Using Connection Pools with Multiprocessing or os.fork() for details.

The most basic function of the Engine is to provide access to a Connection, which can then invoke SQL statements. To emit a textual statement to the database looks like:

from sqlalchemy import text

with engine.connect() as connection:
    result = connection.execute(text("select username from users"))
    for row in result:
        print("username:", row.username)

Above, the Engine.connect() method returns a Connection object, and by using it in a Python context manager (e.g. the with: statement) the Connection.close() method is automatically invoked at the end of the block. The Connection, is a proxy object for an actual DBAPI connection. The DBAPI connection is retrieved from the connection pool at the point at which Connection is created.

The object returned is known as CursorResult, which references a DBAPI cursor and provides methods for fetching rows similar to that of the DBAPI cursor. The DBAPI cursor will be closed by the CursorResult when all of its result rows (if any) are exhausted. A CursorResult that returns no rows, such as that of an UPDATE statement (without any returned rows), releases cursor resources immediately upon construction.

When the Connection is closed at the end of the with: block, the referenced DBAPI connection is released to the connection pool. From the perspective of the database itself, the connection pool will not actually “close” the connection assuming the pool has room to store this connection for the next use. When the connection is returned to the pool for reuse, the pooling mechanism issues a rollback() call on the DBAPI connection so that any transactional state or locks are removed (this is known as Reset On Return), and the connection is ready for its next use.

Our example above illustrated the execution of a textual SQL string, which should be invoked by using the text() construct to indicate that we’d like to use textual SQL. The Connection.execute() method can of course accommodate more than that; see Working with Data in the SQLAlchemy Unified Tutorial for a tutorial.

Using Transactions

Note

This section describes how to use transactions when working directly with Engine and Connection objects. When using the SQLAlchemy ORM, the public API for transaction control is via the Session object, which makes usage of the Transaction object internally. See Managing Transactions for further information.

Commit As You Go

The Connection object always emits SQL statements within the context of a transaction block. The first time the Connection.execute() method is called to execute a SQL statement, this transaction is begun automatically, using a behavior known as autobegin. The transaction remains in place for the scope of the Connection object until the Connection.commit() or Connection.rollback() methods are called. Subsequent to the transaction ending, the Connection waits for the Connection.execute() method to be called again, at which point it autobegins again.

This calling style is known as commit as you go, and is illustrated in the example below:

with engine.connect() as connection:
    connection.execute(some_table.insert(), {"x": 7, "y": "this is some data"})
    connection.execute(
        some_other_table.insert(), {"q": 8, "p": "this is some more data"}
    )

    connection.commit()  # commit the transaction

the Python DBAPI is where autobegin actually happens

The design of “commit as you go” is intended to be complementary to the design of the DBAPI, which is the underlying database interface that SQLAlchemy interacts with. In the DBAPI, the connection object does not assume changes to the database will be automatically committed, instead requiring in the default case that the connection.commit() method is called in order to commit changes to the database. It should be noted that the DBAPI itself does not have a begin() method at all. All Python DBAPIs implement “autobegin” as the primary means of managing transactions, and handle the job of emitting a statement like BEGIN on the connection when SQL statements are first emitted. SQLAlchemy’s API is basically re-stating this behavior in terms of higher level Python objects.

In “commit as you go” style, we can call upon Connection.commit() and Connection.rollback() methods freely within an ongoing sequence of other statements emitted using Connection.execute(); each time the transaction is ended, and a new statement is emitted, a new transaction begins implicitly:

with engine.connect() as connection:
    connection.execute(text("<some statement>"))
    connection.commit()  # commits "some statement"

    # new transaction starts
    connection.execute(text("<some other statement>"))
    connection.rollback()  # rolls back "some other statement"

    # new transaction starts
    connection.execute(text("<a third statement>"))
    connection.commit()  # commits "a third statement"

Added in version 2.0: “commit as you go” style is a new feature of SQLAlchemy 2.0. It is also available in SQLAlchemy 1.4’s “transitional” mode when using a “future” style engine.

Begin Once

The Connection object provides a more explicit transaction management style known as begin once. In contrast to “commit as you go”, “begin once” allows the start point of the transaction to be stated explicitly, and allows that the transaction itself may be framed out as a context manager block so that the end of the transaction is instead implicit. To use “begin once”, the Connection.begin() method is used, which returns a Transaction object which represents the DBAPI transaction. This object also supports explicit management via its own Transaction.commit() and Transaction.rollback() methods, but as a preferred practice also supports the context manager interface, where it will commit itself when the block ends normally and emit a rollback if an exception is raised, before propagating the exception outwards. Below illustrates the form of a “begin once” block:

with engine.connect() as connection:
    with connection.begin():
        connection.execute(some_table.insert(), {"x": 7, "y": "this is some data"})
        connection.execute(
            some_other_table.insert(), {"q": 8, "p": "this is some more data"}
        )

    # transaction is committed

Connect and Begin Once from the Engine

A convenient shorthand form for the above “begin once” block is to use the Engine.begin() method at the level of the originating Engine object, rather than performing the two separate steps of Engine.connect() and Connection.begin(); the Engine.begin() method returns a special context manager that internally maintains both the context manager for the Connection as well as the context manager for the Transaction normally returned by the Connection.begin() method:

with engine.begin() as connection:
    connection.execute(some_table.insert(), {"x": 7, "y": "this is some data"})
    connection.execute(
        some_other_table.insert(), {"q": 8, "p": "this is some more data"}
    )

# transaction is committed, and Connection is released to the connection
# pool

Tip

Within the Engine.begin() block, we can call upon the Connection.commit() or Connection.rollback() methods, which will end the transaction normally demarcated by the block ahead of time. However, if we do so, no further SQL operations may be emitted on the Connection until the block ends:

>>> from sqlalchemy import create_engine
>>> e = create_engine("sqlite://", echo=True)
>>> with e.begin() as conn:
...     conn.commit()
...     conn.begin()
2021-11-08 09:49:07,517 INFO sqlalchemy.engine.Engine BEGIN (implicit)
2021-11-08 09:49:07,517 INFO sqlalchemy.engine.Engine COMMIT
Traceback (most recent call last):
...
sqlalchemy.exc.InvalidRequestError: Can't operate on closed transaction inside
context manager.  Please complete the context manager before emitting
further commands.

Mixing Styles

The “commit as you go” and “begin once” styles can be freely mixed within a single Engine.connect() block, provided that the call to Connection.begin() does not conflict with the “autobegin” behavior. To accomplish this, Connection.begin() should only be called either before any SQL statements have been emitted, or directly after a previous call to Connection.commit() or Connection.rollback():

with engine.connect() as connection:
    with connection.begin():
        # run statements in a "begin once" block
        connection.execute(some_table.insert(), {"x": 7, "y": "this is some data"})

    # transaction is committed

    # run a new statement outside of a block. The connection
    # autobegins
    connection.execute(
        some_other_table.insert(), {"q": 8, "p": "this is some more data"}
    )

    # commit explicitly
    connection.commit()

    # can use a "begin once" block here
    with connection.begin():
        # run more statements
        connection.execute(...)

When developing code that uses “begin once”, the library will raise InvalidRequestError if a transaction was already “autobegun”.

Setting Transaction Isolation Levels including DBAPI Autocommit

Most DBAPIs support the concept of configurable transaction isolation levels. These are traditionally the four levels “READ UNCOMMITTED”, “READ COMMITTED”, “REPEATABLE READ” and “SERIALIZABLE”. These are usually applied to a DBAPI connection before it begins a new transaction, noting that most DBAPIs will begin this transaction implicitly when SQL statements are first emitted.

DBAPIs that support isolation levels also usually support the concept of true “autocommit”, which means that the DBAPI connection itself will be placed into a non-transactional autocommit mode. This usually means that the typical DBAPI behavior of emitting “BEGIN” to the database automatically no longer occurs, but it may also include other directives. SQLAlchemy treats the concept of “autocommit” like any other isolation level; in that it is an isolation level that loses not only “read committed” but also loses atomicity.

Tip

It is important to note, as will be discussed further in the section below at Understanding the DBAPI-Level Autocommit Isolation Level, that “autocommit” isolation level like any other isolation level does not affect the “transactional” behavior of the Connection object, which continues to call upon DBAPI .commit() and .rollback() methods (they just have no net effect under autocommit), and for which the .begin() method assumes the DBAPI will start a transaction implicitly (which means that SQLAlchemy’s “begin” does not change autocommit mode).

SQLAlchemy dialects should support these isolation levels as well as autocommit to as great a degree as possible.

Setting Isolation Level or DBAPI Autocommit for a Connection

For an individual Connection object that’s acquired from Engine.connect(), the isolation level can be set for the duration of that Connection object using the Connection.execution_options() method. The parameter is known as Connection.execution_options.isolation_level and the values are strings which are typically a subset of the following names:

# possible values for Connection.execution_options(isolation_level="<value>")

"AUTOCOMMIT"
"READ COMMITTED"
"READ UNCOMMITTED"
"REPEATABLE READ"
"SERIALIZABLE"

Not every DBAPI supports every value; if an unsupported value is used for a certain backend, an error is raised.

For example, to force REPEATABLE READ on a specific connection, then begin a transaction:

with engine.connect().execution_options(
    isolation_level="REPEATABLE READ"
) as connection:
    with connection.begin():
        connection.execute(text("<statement>"))

Tip

The return value of the Connection.execution_options() method is the same Connection object upon which the method was called, meaning, it modifies the state of the Connection object in place. This is a new behavior as of SQLAlchemy 2.0. This behavior does not apply to the Engine.execution_options() method; that method still returns a copy of the Engine and as described below may be used to construct multiple Engine objects with different execution options, which nonetheless share the same dialect and connection pool.

Note

The Connection.execution_options.isolation_level parameter necessarily does not apply to statement level options, such as that of Executable.execution_options(), and will be rejected if set at this level. This because the option must be set on a DBAPI connection on a per-transaction basis.

Setting Isolation Level or DBAPI Autocommit for an Engine

The Connection.execution_options.isolation_level option may also be set engine wide, as is often preferable. This may be achieved by passing the create_engine.isolation_level parameter to create_engine():

from sqlalchemy import create_engine

eng = create_engine(
    "postgresql://scott:tiger@localhost/test", isolation_level="REPEATABLE READ"
)

With the above setting, each new DBAPI connection the moment it’s created will be set to use a "REPEATABLE READ" isolation level setting for all subsequent operations.

Tip

Prefer to set frequently used isolation levels engine wide as illustrated above compared to using per-engine or per-connection execution options for maximum performance.

Maintaining Multiple Isolation Levels for a Single Engine

The isolation level may also be set per engine, with a potentially greater level of flexibility but with a small per-connection performance overhead, using either the create_engine.execution_options parameter to create_engine() or the Engine.execution_options() method, the latter of which will create a copy of the Engine that shares the dialect and connection pool of the original engine, but has its own per-connection isolation level setting:

from sqlalchemy import create_engine

eng = create_engine(
    "postgresql+psycopg2://scott:tiger@localhost/test",
    execution_options={"isolation_level": "REPEATABLE READ"},
)

With the above setting, the DBAPI connection will be set to use a "REPEATABLE READ" isolation level setting for each new transaction begun; but the connection as pooled will be reset to the original isolation level that was present when the connection first occurred. At the level of create_engine(), the end effect is not any different from using the create_engine.isolation_level parameter.

However, an application that frequently chooses to run operations within different isolation levels may wish to create multiple “sub-engines” of a lead Engine, each of which will be configured to a different isolation level. One such use case is an application that has operations that break into “transactional” and “read-only” operations, a separate Engine that makes use of "AUTOCOMMIT" may be separated off from the main engine:

from sqlalchemy import create_engine

eng = create_engine("postgresql+psycopg2://scott:tiger@localhost/test")

autocommit_engine = eng.execution_options(isolation_level="AUTOCOMMIT")

Above, the Engine.execution_options() method creates a shallow copy of the original Engine. Both eng and autocommit_engine share the same dialect and connection pool. However, the “AUTOCOMMIT” mode will be set upon connections when they are acquired from the autocommit_engine.

The isolation level setting, regardless of which one it is, is unconditionally reverted when a connection is returned to the connection pool.

Note

The execution options approach, whether used engine wide or per connection, incurs a small performance penalty as isolation level instructions are sent on connection acquire as well as connection release. Consider the engine-wide isolation setting at Setting Isolation Level or DBAPI Autocommit for an Engine so that connections are configured at the target isolation level permanently as they are pooled.

See also

SQLite Transaction Isolation

PostgreSQL Transaction Isolation

MySQL Transaction Isolation

SQL Server Transaction Isolation

Oracle Database Transaction Isolation

Setting Transaction Isolation Levels / DBAPI AUTOCOMMIT - for the ORM

Using DBAPI Autocommit Allows for a Readonly Version of Transparent Reconnect - a recipe that uses DBAPI autocommit to transparently reconnect to the database for read-only operations

Understanding the DBAPI-Level Autocommit Isolation Level

In the parent section, we introduced the concept of the Connection.execution_options.isolation_level parameter and how it can be used to set database isolation levels, including DBAPI-level “autocommit” which is treated by SQLAlchemy as another transaction isolation level. In this section we will attempt to clarify the implications of this approach.

If we wanted to check out a Connection object and use it “autocommit” mode, we would proceed as follows:

with engine.connect() as connection:
    connection.execution_options(isolation_level="AUTOCOMMIT")
    connection.execute(text("<statement>"))
    connection.execute(text("<statement>"))

Above illustrates normal usage of “DBAPI autocommit” mode. There is no need to make use of methods such as Connection.begin() or Connection.commit(), as all statements are committed to the database immediately. When the block ends, the Connection object will revert the “autocommit” isolation level, and the DBAPI connection is released to the connection pool where the DBAPI connection.rollback() method will normally be invoked, but as the above statements were already committed, this rollback has no change on the state of the database.

It is important to note that “autocommit” mode persists even when the Connection.begin() method is called; the DBAPI will not emit any BEGIN to the database. When Connection.commit() is called, the DBAPI may still emit the “COMMIT” instruction, but this is a no-op at the database level. This usage is also not an error scenario, as it is expected that the “autocommit” isolation level may be applied to code that otherwise was written assuming a transactional context; the “isolation level” is, after all, a configurational detail of the transaction itself just like any other isolation level.

In the example below, statements remain autocommitting regardless of SQLAlchemy-level transaction blocks:

with engine.connect() as connection:
    connection = connection.execution_options(isolation_level="AUTOCOMMIT")

    # this begin() does not affect the DBAPI connection, isolation stays at AUTOCOMMIT
    with connection.begin() as trans:
        connection.execute(text("<statement>"))
        connection.execute(text("<statement>"))

When we run a block like the above with logging turned on, the logging will attempt to indicate that while a DBAPI level .commit() is called, it probably will have no effect due to autocommit mode:

INFO sqlalchemy.engine.Engine BEGIN (implicit)
...
INFO sqlalchemy.engine.Engine COMMIT using DBAPI connection.commit(), has no effect due to autocommit mode

At the same time, even though we are using “DBAPI autocommit”, SQLAlchemy’s transactional semantics, that is, the in-Python behavior of Connection.begin() as well as the behavior of “autobegin”, remain in place, even though these don’t impact the DBAPI connection itself. To illustrate, the code below will raise an error, as Connection.begin() is being called after autobegin has already occurred:

with engine.connect() as connection:
    connection = connection.execution_options(isolation_level="AUTOCOMMIT")

    # "transaction" is autobegin (but has no effect due to autocommit)
    connection.execute(text("<statement>"))

    # this will raise; "transaction" is already begun
    with connection.begin() as trans:
        connection.execute(text("<statement>"))

The above example also demonstrates the same theme that the “autocommit” isolation level is a configurational detail of the underlying database transaction, and is independent of the begin/commit behavior of the SQLAlchemy Connection object. The “autocommit” mode will not interact with Connection.begin() in any way and the Connection does not consult this status when performing its own state changes with regards to the transaction (with the exception of suggesting within engine logging that these blocks are not actually committing). The rationale for this design is to maintain a completely consistent usage pattern with the Connection where DBAPI-autocommit mode can be changed independently without indicating any code changes elsewhere.

Fully preventing ROLLBACK calls under autocommit

Added in version 2.0.43.

A common use case is to use AUTOCOMMIT isolation mode to improve performance, and this is a particularly common practice on MySQL / MariaDB databases. When seeking this pattern, it should be preferred to set AUTOCOMMIT engine wide using the create_engine.isolation_level so that pooled connections are permanently set in autocommit mode. The SQLAlchemy connection pool as well as the Connection will still seek to invoke the DBAPI .rollback() method upon connection release, as their behavior remains agnostic of the isolation level that’s configured on the connection. As this rollback still incurs a network round trip under most if not all DBAPI drivers, this additional network trip may be disabled using the create_engine.skip_autocommit_rollback parameter, which will apply a rule at the basemost portion of the dialect that invokes DBAPI .rollback() to first check if the connection is configured in autocommit, using a method of detection that does not itself incur network overhead:

autocommit_engine = create_engine(
    "mysql+mysqldb://scott:tiger@mysql80/test",
    skip_autocommit_rollback=True,
    isolation_level="AUTOCOMMIT",
)

When DBAPI connections are returned to the pool by the Connection, whether the Connection or the pool attempts to reset the “transaction”, the underlying DBAPI .rollback() method will be blocked based on a positive test of “autocommit”.

If the dialect in use does not support a no-network means of detecting autocommit, the dialect will raise NotImplementedError when a connection release is attempted.

Changing Between Isolation Levels

TL;DR;

prefer to use individual Connection objects each with just one isolation level, rather than switching isolation on a single Connection. The code will be easier to read and less error prone.

Isolation level settings, including autocommit mode, are reset automatically when the connection is released back to the connection pool. Therefore it is preferable to avoid trying to switch isolation levels on a single Connection object as this leads to excess verbosity.

To illustrate how to use “autocommit” in an ad-hoc mode within the scope of a single Connection checkout, the Connection.execution_options.isolation_level parameter must be re-applied with the previous isolation level. The previous section illustrated an attempt to call Connection.begin() in order to start a transaction while autocommit was taking place; we can rewrite that example to actually do so by first reverting the isolation level before we call upon Connection.begin():

# if we wanted to flip autocommit on and off on a single connection/
# which... we usually don't.

with engine.connect() as connection:
    connection.execution_options(isolation_level="AUTOCOMMIT")

    # run statement(s) in autocommit mode
    connection.execute(text("<statement>"))

    # "commit" the autobegun "transaction"
    connection.commit()

    # switch to default isolation level
    connection.execution_options(isolation_level=connection.default_isolation_level)

    # use a begin block
    with connection.begin() as trans:
        connection.execute(text("<statement>"))

Above, to manually revert the isolation level we made use of Connection.default_isolation_level to restore the default isolation level (assuming that’s what we want here). However, it’s probably a better idea to work with the architecture of of the Connection which already handles resetting of isolation level automatically upon checkin. The preferred way to write the above is to use two blocks

# use an autocommit block
with engine.connect().execution_options(isolation_level="AUTOCOMMIT") as connection:
    # run statement in autocommit mode
    connection.execute(text("<statement>"))

# use a regular block
with engine.begin() as connection:
    connection.execute(text("<statement>"))

To sum up:

1. “DBAPI level autocommit” isolation level is entirely independent of the Connection object’s notion of “begin” and “commit”

2. use individual Connection checkouts per isolation level. Avoid trying to change back and forth between “autocommit” on a single connection checkout; let the engine do the work of restoring default isolation levels

Using Server Side Cursors (a.k.a. stream results)

Some backends feature explicit support for the concept of “server side cursors” versus “client side cursors”. A client side cursor here means that the database driver fully fetches all rows from a result set into memory before returning from a statement execution. Drivers such as those of PostgreSQL and MySQL/MariaDB generally use client side cursors by default. A server side cursor, by contrast, indicates that result rows remain pending within the database server’s state as result rows are consumed by the client. The drivers for Oracle Database generally use a “server side” model, for example, and the SQLite dialect, while not using a real “client / server” architecture, still uses an unbuffered result fetching approach that will leave result rows outside of process memory before they are consumed.

What we really mean is “buffered” vs. “unbuffered” results

Server side cursors also imply a wider set of features with relational databases, such as the ability to “scroll” a cursor forwards and backwards. SQLAlchemy does not include any explicit support for these behaviors; within SQLAlchemy itself, the general term “server side cursors” should be considered to mean “unbuffered results” and “client side cursors” means “result rows are buffered into memory before the first row is returned”. To work with a richer “server side cursor” featureset specific to a certain DBAPI driver, see the section Working with the DBAPI cursor directly.

From this basic architecture it follows that a “server side cursor” is more memory efficient when fetching very large result sets, while at the same time may introduce more complexity in the client/server communication process and be less efficient for small result sets (typically less than 10000 rows).

For those dialects that have conditional support for buffered or unbuffered results, there are usually caveats to the use of the “unbuffered”, or server side cursor mode. When using the psycopg2 dialect for example, an error is raised if a server side cursor is used with any kind of DML or DDL statement. When using MySQL drivers with a server side cursor, the DBAPI connection is in a more fragile state and does not recover as gracefully from error conditions nor will it allow a rollback to proceed until the cursor is fully closed.

For this reason, SQLAlchemy’s dialects will always default to the less error prone version of a cursor, which means for PostgreSQL and MySQL dialects it defaults to a buffered, “client side” cursor where the full set of results is pulled into memory before any fetch methods are called from the cursor. This mode of operation is appropriate in the vast majority of cases; unbuffered cursors are not generally useful except in the uncommon case of an application fetching a very large number of rows in chunks, where the processing of these rows can be complete before more rows are fetched.

For database drivers that provide client and server side cursor options, the Connection.execution_options.stream_results and Connection.execution_options.yield_per execution options provide access to “server side cursors” on a per-Connection or per-statement basis. Similar options exist when using an ORM Session as well.

Streaming with a fixed buffer via yield_per

As individual row-fetch operations with fully unbuffered server side cursors are typically more expensive than fetching batches of rows at once, The Connection.execution_options.yield_per execution option configures a Connection or statement to make use of server-side cursors as are available, while at the same time configuring a fixed-size buffer of rows that will retrieve rows from the server in batches as they are consumed. This parameter may be to a positive integer value using the Connection.execution_options() method on Connection or on a statement using the Executable.execution_options() method.

Added in version 1.4.40: Connection.execution_options.yield_per as a Core-only option is new as of SQLAlchemy 1.4.40; for prior 1.4 versions, use Connection.execution_options.stream_results directly in combination with Result.yield_per().

Using this option is equivalent to manually setting the Connection.execution_options.stream_results option, described in the next section, and then invoking the Result.yield_per() method on the Result object with the given integer value. In both cases, the effect this combination has includes:

• server side cursors mode is selected for the given backend, if available and not already the default behavior for that backend

• as result rows are fetched, they will be buffered in batches, where the size of each batch up until the last batch will be equal to the integer argument passed to the Connection.execution_options.yield_per option or the Result.yield_per() method; the last batch is then sized against the remaining rows fewer than this size

• The default partition size used by the Result.partitions() method, if used, will be made equal to this integer size as well.

These three behaviors are illustrated in the example below:

with engine.connect() as conn:
    with conn.execution_options(yield_per=100).execute(
        text("select * from table")
    ) as result:
        for partition in result.partitions():
            # partition is an iterable that will be at most 100 items
            for row in partition:
                print(f"{row}")

The above example illustrates the combination of yield_per=100 along with using the Result.partitions() method to run processing on rows in batches that match the size fetched from the server. The use of Result.partitions() is optional, and if the Result is iterated directly, a new batch of rows will be buffered for each 100 rows fetched. Calling a method such as Result.all() should not be used, as this will fully fetch all remaining rows at once and defeat the purpose of using yield_per.

Tip

The Result object may be used as a context manager as illustrated above. When iterating with a server-side cursor, this is the best way to ensure the Result object is closed, even if exceptions are raised within the iteration process.

The Connection.execution_options.yield_per option is portable to the ORM as well, used by a Session to fetch ORM objects, where it also limits the amount of ORM objects generated at once. See the section Fetching Large Result Sets with Yield Per - in the ORM Querying Guide for further background on using Connection.execution_options.yield_per with the ORM.

Added in version 1.4.40: Added Connection.execution_options.yield_per as a Core level execution option to conveniently set streaming results, buffer size, and partition size all at once in a manner that is transferable to that of the ORM’s similar use case.

Streaming with a dynamically growing buffer using stream_results

To enable server side cursors without a specific partition size, the Connection.execution_options.stream_results option may be used, which like Connection.execution_options.yield_per may be called on the Connection object or the statement object.

When a Result object delivered using the Connection.execution_options.stream_results option is iterated directly, rows are fetched internally using a default buffering scheme that buffers first a small set of rows, then a larger and larger buffer on each fetch up to a pre-configured limit of 1000 rows. The maximum size of this buffer can be affected using the Connection.execution_options.max_row_buffer execution option:

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

While the Connection.execution_options.stream_results option may be combined with use of the Result.partitions() method, a specific partition size should be passed to Result.partitions() so that the entire result is not fetched. It is usually more straightforward to use the Connection.execution_options.yield_per option when setting up to use the Result.partitions() method.

See also

Fetching Large Result Sets with Yield Per - in the ORM Querying Guide

Result.partitions()

Result.yield_per()

Translation of Schema Names

To support multi-tenancy applications that distribute common sets of tables into multiple schemas, the Connection.execution_options.schema_translate_map execution option may be used to repurpose a set of Table objects to render under different schema names without any changes.

Given a table:

user_table = Table(
    "user",
    metadata_obj,
    Column("id", Integer, primary_key=True),
    Column("name", String(50)),
)

The “schema” of this Table as defined by the Table.schema attribute is None. The Connection.execution_options.schema_translate_map can specify that all Table objects with a schema of None would instead render the schema as user_schema_one:

connection = engine.connect().execution_options(
    schema_translate_map={None: "user_schema_one"}
)

result = connection.execute(user_table.select())

The above code will invoke SQL on the database of the form:

SELECT user_schema_one.user.id, user_schema_one.user.name FROM
user_schema_one.user

That is, the schema name is substituted with our translated name. The map can specify any number of target->destination schemas:

connection = engine.connect().execution_options(
    schema_translate_map={
        None: "user_schema_one",  # no schema name -> "user_schema_one"
        "special": "special_schema",  # schema="special" becomes "special_schema"
        "public": None,  # Table objects with schema="public" will render with no schema
    }
)

The Connection.execution_options.schema_translate_map parameter affects all DDL and SQL constructs generated from the SQL expression language, as derived from the Table or Sequence objects. It does not impact literal string SQL used via the text() construct nor via plain strings passed to Connection.execute().

The feature takes effect only in those cases where the name of the schema is derived directly from that of a Table or Sequence; it does not impact methods where a string schema name is passed directly. By this pattern, it takes effect within the “can create” / “can drop” checks performed by methods such as MetaData.create_all() or MetaData.drop_all() are called, and it takes effect when using table reflection given a Table object. However it does not affect the operations present on the Inspector object, as the schema name is passed to these methods explicitly.

Tip

To use the schema translation feature with the ORM Session, set this option at the level of the Engine, then pass that engine to the Session. The Session uses a new Connection for each transaction:

schema_engine = engine.execution_options(schema_translate_map={...})

session = Session(schema_engine)

...

Warning

When using the ORM Session without extensions, the schema translate feature is only supported as a single schema translate map per Session. It will not work if different schema translate maps are given on a per-statement basis, as the ORM Session does not take current schema translate values into account for individual objects.

To use a single Session with multiple schema_translate_map configurations, the Horizontal Sharding extension may be used. See the example at Horizontal Sharding.

SQL Compilation Caching

Added in version 1.4: SQLAlchemy now has a transparent query caching system that substantially lowers the Python computational overhead involved in converting SQL statement constructs into SQL strings across both Core and ORM. See the introduction at Transparent SQL Compilation Caching added to All DQL, DML Statements in Core, ORM.

SQLAlchemy includes a comprehensive caching system for the SQL compiler as well as its ORM variants. This caching system is transparent within the Engine and provides that the SQL compilation process for a given Core or ORM SQL statement, as well as related computations which assemble result-fetching mechanics for that statement, will only occur once for that statement object and all others with the identical structure, for the duration that the particular structure remains within the engine’s “compiled cache”. By “statement objects that have the identical structure”, this generally corresponds to a SQL statement that is constructed within a function and is built each time that function runs:

def run_my_statement(connection, parameter):
    stmt = select(table)
    stmt = stmt.where(table.c.col == parameter)
    stmt = stmt.order_by(table.c.id)
    return connection.execute(stmt)

The above statement will generate SQL resembling SELECT id, col FROM table WHERE col = :col ORDER BY id, noting that while the value of parameter is a plain Python object such as a string or an integer, the string SQL form of the statement does not include this value as it uses bound parameters. Subsequent invocations of the above run_my_statement() function will use a cached compilation construct within the scope of the connection.execute() call for enhanced performance.

Note

it is important to note that the SQL compilation cache is caching the SQL string that is passed to the database only, and not the data returned by a query. It is in no way a data cache and does not impact the results returned for a particular SQL statement nor does it imply any memory use linked to fetching of result rows.

While SQLAlchemy has had a rudimentary statement cache since the early 1.x series, and additionally has featured the “Baked Query” extension for the ORM, both of these systems required a high degree of special API use in order for the cache to be effective. The new cache as of 1.4 is instead completely automatic and requires no change in programming style to be effective.

The cache is automatically used without any configurational changes and no special steps are needed in order to enable it. The following sections detail the configuration and advanced usage patterns for the cache.

Configuration

The cache itself is a dictionary-like object called an LRUCache, which is an internal SQLAlchemy dictionary subclass that tracks the usage of particular keys and features a periodic “pruning” step which removes the least recently used items when the size of the cache reaches a certain threshold. The size of this cache defaults to 500 and may be configured using the create_engine.query_cache_size parameter:

engine = create_engine(
    "postgresql+psycopg2://scott:tiger@localhost/test", query_cache_size=1200
)

The size of the cache can grow to be a factor of 150% of the size given, before it’s pruned back down to the target size. A cache of size 1200 above can therefore grow to be 1800 elements in size at which point it will be pruned to 1200.

The sizing of the cache is based on a single entry per unique SQL statement rendered, per engine. SQL statements generated from both the Core and the ORM are treated equally. DDL statements will usually not be cached. In order to determine what the cache is doing, engine logging will include details about the cache’s behavior, described in the next section.

Estimating Cache Performance Using Logging

The above cache size of 1200 is actually fairly large. For small applications, a size of 100 is likely sufficient. To estimate the optimal size of the cache, assuming enough memory is present on the target host, the size of the cache should be based on the number of unique SQL strings that may be rendered for the target engine in use. The most expedient way to see this is to use SQL echoing, which is most directly enabled by using the create_engine.echo flag, or by using Python logging; see the section Configuring Logging for background on logging configuration.

As an example, we will examine the logging produced by the following program:

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

Base = declarative_base()


class A(Base):
    __tablename__ = "a"

    id = Column(Integer, primary_key=True)
    data = Column(String)
    bs = relationship("B")


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


e = create_engine("sqlite://", echo=True)
Base.metadata.create_all(e)

s = Session(e)

s.add_all([A(bs=[B(), B(), B()]), A(bs=[B(), B(), B()]), A(bs=[B(), B(), B()])])
s.commit()

for a_rec in s.scalars(select(A)):
    print(a_rec.bs)

When run, each SQL statement that’s logged will include a bracketed cache statistics badge to the left of the parameters passed. The four types of message we may see are summarized as follows:

• [raw sql] - the driver or the end-user emitted raw SQL using Connection.exec_driver_sql() - caching does not apply

• [no key] - the statement object is a DDL statement that is not cached, or the statement object contains uncacheable elements such as user-defined constructs or arbitrarily large VALUES clauses.

• [generated in Xs] - the statement was a cache miss and had to be compiled, then stored in the cache. it took X seconds to produce the compiled construct. The number X will be in the small fractional seconds.

• [cached since Xs ago] - the statement was a cache hit and did not have to be recompiled. The statement has been stored in the cache since X seconds ago. The number X will be proportional to how long the application has been running and how long the statement has been cached, so for example would be 86400 for a 24 hour period.

Each badge is described in more detail below.

The first statements we see for the above program will be the SQLite dialect checking for the existence of the “a” and “b” tables:

INFO sqlalchemy.engine.Engine PRAGMA temp.table_info("a")
INFO sqlalchemy.engine.Engine [raw sql] ()
INFO sqlalchemy.engine.Engine PRAGMA main.table_info("b")
INFO sqlalchemy.engine.Engine [raw sql] ()

For the above two SQLite PRAGMA statements, the badge reads [raw sql], which indicates the driver is sending a Python string directly to the database using Connection.exec_driver_sql(). Caching does not apply to such statements because they already exist in string form, and there is nothing known about what kinds of result rows will be returned since SQLAlchemy does not parse SQL strings ahead of time.

The next statements we see are the CREATE TABLE statements:

INFO sqlalchemy.engine.Engine
CREATE TABLE a (
  id INTEGER NOT NULL,
  data VARCHAR,
  PRIMARY KEY (id)
)

INFO sqlalchemy.engine.Engine [no key 0.00007s] ()
INFO sqlalchemy.engine.Engine
CREATE TABLE b (
  id INTEGER NOT NULL,
  a_id INTEGER,
  data VARCHAR,
  PRIMARY KEY (id),
  FOREIGN KEY(a_id) REFERENCES a (id)
)

INFO sqlalchemy.engine.Engine [no key 0.00006s] ()

For each of these statements, the badge reads [no key 0.00006s]. This indicates that these two particular statements, caching did not occur because the DDL-oriented CreateTable construct did not produce a cache key. DDL constructs generally do not participate in caching because they are not typically subject to being repeated a second time and DDL is also a database configurational step where performance is not as critical.

The [no key] badge is important for one other reason, as it can be produced for SQL statements that are cacheable except for some particular sub-construct that is not currently cacheable. Examples of this include custom user-defined SQL elements that don’t define caching parameters, as well as some constructs that generate arbitrarily long and non-reproducible SQL strings, the main examples being the Values construct as well as when using “multivalued inserts” with the Insert.values() method.

So far our cache is still empty. The next statements will be cached however, a segment looks like:

INFO sqlalchemy.engine.Engine INSERT INTO a (data) VALUES (?)
INFO sqlalchemy.engine.Engine [generated in 0.00011s] (None,)
INFO sqlalchemy.engine.Engine INSERT INTO a (data) VALUES (?)
INFO sqlalchemy.engine.Engine [cached since 0.0003533s ago] (None,)
INFO sqlalchemy.engine.Engine INSERT INTO a (data) VALUES (?)
INFO sqlalchemy.engine.Engine [cached since 0.0005326s ago] (None,)
INFO sqlalchemy.engine.Engine INSERT INTO b (a_id, data) VALUES (?, ?)
INFO sqlalchemy.engine.Engine [generated in 0.00010s] (1, None)
INFO sqlalchemy.engine.Engine INSERT INTO b (a_id, data) VALUES (?, ?)
INFO sqlalchemy.engine.Engine [cached since 0.0003232s ago] (1, None)
INFO sqlalchemy.engine.Engine INSERT INTO b (a_id, data) VALUES (?, ?)
INFO sqlalchemy.engine.Engine [cached since 0.0004887s ago] (1, None)

Above, we see essentially two unique SQL strings; "INSERT INTO a (data) VALUES (?)" and "INSERT INTO b (a_id, data) VALUES (?, ?)". Since SQLAlchemy uses bound parameters for all literal values, even though these statements are repeated many times for different objects, because the parameters are separate, the actual SQL string stays the same.

Note

the above two statements are generated by the ORM unit of work process, and in fact will be caching these in a separate cache that is local to each mapper. However the mechanics and terminology are the same. The section Disabling or using an alternate dictionary to cache some (or all) statements below will describe how user-facing code can also use an alternate caching container on a per-statement basis.

The caching badge we see for the first occurrence of each of these two statements is [generated in 0.00011s]. This indicates that the statement was not in the cache, was compiled into a String in .00011s and was then cached. When we see the [generated] badge, we know that this means there was a cache miss. This is to be expected for the first occurrence of a particular statement. However, if lots of new [generated] badges are observed for a long-running application that is generally using the same series of SQL statements over and over, this may be a sign that the create_engine.query_cache_size parameter is too small. When a statement that was cached is then evicted from the cache due to the LRU cache pruning lesser used items, it will display the [generated] badge when it is next used.

The caching badge that we then see for the subsequent occurrences of each of these two statements looks like [cached since 0.0003533s ago]. This indicates that the statement was found in the cache, and was originally placed into the cache .0003533 seconds ago. It is important to note that while the [generated] and [cached since] badges refer to a number of seconds, they mean different things; in the case of [generated], the number is a rough timing of how long it took to compile the statement, and will be an extremely small amount of time. In the case of [cached since], this is the total time that a statement has been present in the cache. For an application that’s been running for six hours, this number may read [cached since 21600 seconds ago], and that’s a good thing. Seeing high numbers for “cached since” is an indication that these statements have not been subject to cache misses for a long time. Statements that frequently have a low number of “cached since” even if the application has been running a long time may indicate these statements are too frequently subject to cache misses, and that the create_engine.query_cache_size may need to be increased.

Our example program then performs some SELECTs where we can see the same pattern of “generated” then “cached”, for the SELECT of the “a” table as well as for subsequent lazy loads of the “b” table:

INFO sqlalchemy.engine.Engine SELECT a.id AS a_id, a.data AS a_data
FROM a
INFO sqlalchemy.engine.Engine [generated in 0.00009s] ()
INFO sqlalchemy.engine.Engine SELECT b.id AS b_id, b.a_id AS b_a_id, b.data AS b_data
FROM b
WHERE ? = b.a_id
INFO sqlalchemy.engine.Engine [generated in 0.00010s] (1,)
INFO sqlalchemy.engine.Engine SELECT b.id AS b_id, b.a_id AS b_a_id, b.data AS b_data
FROM b
WHERE ? = b.a_id
INFO sqlalchemy.engine.Engine [cached since 0.0005922s ago] (2,)
INFO sqlalchemy.engine.Engine SELECT b.id AS b_id, b.a_id AS b_a_id, b.data AS b_data
FROM b
WHERE ? = b.a_id

From our above program, a full run shows a total of four distinct SQL strings being cached. Which indicates a cache size of four would be sufficient. This is obviously an extremely small size, and the default size of 500 is fine to be left at its default.

How much memory does the cache use?

The previous section detailed some techniques to check if the create_engine.query_cache_size needs to be bigger. How do we know if the cache is not too large? The reason we may want to set create_engine.query_cache_size to not be higher than a certain number would be because we have an application that may make use of a very large number of different statements, such as an application that is building queries on the fly from a search UX, and we don’t want our host to run out of memory if for example, a hundred thousand different queries were run in the past 24 hours and they were all cached.

It is extremely difficult to measure how much memory is occupied by Python data structures, however using a process to measure growth in memory via top as a successive series of 250 new statements are added to the cache suggest a moderate Core statement takes up about 12K while a small ORM statement takes about 20K, including result-fetching structures which for the ORM will be much greater.

Disabling or using an alternate dictionary to cache some (or all) statements

The internal cache used is known as LRUCache, but this is mostly just a dictionary. Any dictionary may be used as a cache for any series of statements by using the Connection.execution_options.compiled_cache option as an execution option. Execution options may be set on a statement, on an Engine or Connection, as well as when using the ORM Session.execute() method for SQLAlchemy-2.0 style invocations. For example, to run a series of SQL statements and have them cached in a particular dictionary:

my_cache = {}
with engine.connect().execution_options(compiled_cache=my_cache) as conn:
    conn.execute(table.select())

The SQLAlchemy ORM uses the above technique to hold onto per-mapper caches within the unit of work “flush” process that are separate from the default cache configured on the Engine, as well as for some relationship loader queries.

The cache can also be disabled with this argument by sending a value of None:

# disable caching for this connection
with engine.connect().execution_options(compiled_cache=None) as conn:
    conn.execute(table.select())

Caching for Third Party Dialects

The caching feature requires that the dialect’s compiler produces SQL strings that are safe to reuse for many statement invocations, given a particular cache key that is keyed to that SQL string. This means that any literal values in a statement, such as the LIMIT/OFFSET values for a SELECT, can not be hardcoded in the dialect’s compilation scheme, as the compiled string will not be reusable. SQLAlchemy supports rendered bound parameters using the BindParameter.render_literal_execute() method which can be applied to the existing Select._limit_clause and Select._offset_clause attributes by a custom compiler, which are illustrated later in this section.

As there are many third party dialects, many of which may be generating literal values from SQL statements without the benefit of the newer “literal execute” feature, SQLAlchemy as of version 1.4.5 has added an attribute to dialects known as Dialect.supports_statement_cache. This attribute is checked at runtime for its presence directly on a particular dialect’s class, even if it’s already present on a superclass, so that even a third party dialect that subclasses an existing cacheable SQLAlchemy dialect such as sqlalchemy.dialects.postgresql.PGDialect must still explicitly include this attribute for caching to be enabled. The attribute should only be enabled once the dialect has been altered as needed and tested for reusability of compiled SQL statements with differing parameters.

For all third party dialects that don’t support this attribute, the logging for such a dialect will indicate dialect does not support caching.

When a dialect has been tested against caching, and in particular the SQL compiler has been updated to not render any literal LIMIT / OFFSET within a SQL string directly, dialect authors can apply the attribute as follows:

from sqlalchemy.engine.default import DefaultDialect


class MyDialect(DefaultDialect):
    supports_statement_cache = True

The flag needs to be applied to all subclasses of the dialect as well:

class MyDBAPIForMyDialect(MyDialect):
    supports_statement_cache = True

Added in version 1.4.5: Added the Dialect.supports_statement_cache attribute.

The typical case for dialect modification follows.

Example: Rendering LIMIT / OFFSET with post compile parameters

As an example, suppose a dialect overrides the SQLCompiler.limit_clause() method, which produces the “LIMIT / OFFSET” clause for a SQL statement, like this:

# pre 1.4 style code
def limit_clause(self, select, **kw):
    text = ""
    if select._limit is not None:
        text += " \n LIMIT %d" % (select._limit,)
    if select._offset is not None:
        text += " \n OFFSET %d" % (select._offset,)
    return text

The above routine renders the Select._limit and Select._offset integer values as literal integers embedded in the SQL statement. This is a common requirement for databases that do not support using a bound parameter within the LIMIT/OFFSET clauses of a SELECT statement. However, rendering the integer value within the initial compilation stage is directly incompatible with caching as the limit and offset integer values of a Select object are not part of the cache key, so that many Select statements with different limit/offset values would not render with the correct value.

The correction for the above code is to move the literal integer into SQLAlchemy’s post-compile facility, which will render the literal integer outside of the initial compilation stage, but instead at execution time before the statement is sent to the DBAPI. This is accessed within the compilation stage using the BindParameter.render_literal_execute() method, in conjunction with using the Select._limit_clause and Select._offset_clause attributes, which represent the LIMIT/OFFSET as a complete SQL expression, as follows:

# 1.4 cache-compatible code
def limit_clause(self, select, **kw):
    text = ""

    limit_clause = select._limit_clause
    offset_clause = select._offset_clause

    if select._simple_int_clause(limit_clause):
        text += " \n LIMIT %s" % (
            self.process(limit_clause.render_literal_execute(), **kw)
        )
    elif limit_clause is not None:
        # assuming the DB doesn't support SQL expressions for LIMIT.
        # Otherwise render here normally
        raise exc.CompileError(
            "dialect 'mydialect' can only render simple integers for LIMIT"
        )
    if select._simple_int_clause(offset_clause):
        text += " \n OFFSET %s" % (
            self.process(offset_clause.render_literal_execute(), **kw)
        )
    elif offset_clause is not None:
        # assuming the DB doesn't support SQL expressions for OFFSET.
        # Otherwise render here normally
        raise exc.CompileError(
            "dialect 'mydialect' can only render simple integers for OFFSET"
        )

    return text

The approach above will generate a compiled SELECT statement that looks like:

SELECT x FROM y
LIMIT __[POSTCOMPILE_param_1]
OFFSET __[POSTCOMPILE_param_2]

Where above, the __[POSTCOMPILE_param_1] and __[POSTCOMPILE_param_2] indicators will be populated with their corresponding integer values at statement execution time, after the SQL string has been retrieved from the cache.

After changes like the above have been made as appropriate, the Dialect.supports_statement_cache flag should be set to True. It is strongly recommended that third party dialects make use of the dialect third party test suite which will assert that operations like SELECTs with LIMIT/OFFSET are correctly rendered and cached.

See also

Why is my application slow after upgrading to 1.4 and/or 2.x? - in the Frequently Asked Questions section

Using Lambdas to add significant speed gains to statement production

Deep Alchemy

This technique is generally non-essential except in very performance intensive scenarios, and intended for experienced Python programmers. While fairly straightforward, it involves metaprogramming concepts that are not appropriate for novice Python developers. The lambda approach can be applied to at a later time to existing code with a minimal amount of effort.

Python functions, typically expressed as lambdas, may be used to generate SQL expressions which are cacheable based on the Python code location of the lambda function itself as well as the closure variables within the lambda. The rationale is to allow caching of not only the SQL string-compiled form of a SQL expression construct as is SQLAlchemy’s normal behavior when the lambda system isn’t used, but also the in-Python composition of the SQL expression construct itself, which also has some degree of Python overhead.

The lambda SQL expression feature is available as a performance enhancing feature, and is also optionally used in the with_loader_criteria() ORM option in order to provide a generic SQL fragment.

Synopsis

Lambda statements are constructed using the lambda_stmt() function, which returns an instance of StatementLambdaElement, which is itself an executable statement construct. Additional modifiers and criteria are added to the object using the Python addition operator +, or alternatively the StatementLambdaElement.add_criteria() method which allows for more options.

It is assumed that the lambda_stmt() construct is being invoked within an enclosing function or method that expects to be used many times within an application, so that subsequent executions beyond the first one can take advantage of the compiled SQL being cached. When the lambda is constructed inside of an enclosing function in Python it is then subject to also having closure variables, which are significant to the whole approach:

from sqlalchemy import lambda_stmt


def run_my_statement(connection, parameter):
    stmt = lambda_stmt(lambda: select(table))
    stmt += lambda s: s.where(table.c.col == parameter)
    stmt += lambda s: s.order_by(table.c.id)

    return connection.execute(stmt)


with engine.connect() as conn:
    result = run_my_statement(some_connection, "some parameter")

Above, the three lambda callables that are used to define the structure of a SELECT statement are invoked exactly once, and the resulting SQL string cached in the compilation cache of the engine. From that point forward, the run_my_statement() function may be invoked any number of times and the lambda callables within it will not be called, only used as cache keys to retrieve the already-compiled SQL.

Note

It is important to note that there is already SQL caching in place when the lambda system is not used. The lambda system only adds an additional layer of work reduction per SQL statement invoked by caching the building up of the SQL construct itself and also using a simpler cache key.

Quick Guidelines for Lambdas

Above all, the emphasis within the lambda SQL system is ensuring that there is never a mismatch between the cache key generated for a lambda and the SQL string it will produce. The LambdaElement and related objects will run and analyze the given lambda in order to calculate how it should be cached on each run, trying to detect any potential problems. Basic guidelines include:

• Any kind of statement is supported - while it’s expected that select() constructs are the prime use case for lambda_stmt(), DML statements such as insert() and update() are equally usable:

  def upd(id_, newname):
      stmt = lambda_stmt(lambda: users.update())
      stmt += lambda s: s.values(name=newname)
      stmt += lambda s: s.where(users.c.id == id_)
      return stmt
  
  
  with engine.begin() as conn:
      conn.execute(upd(7, "foo"))

• ORM use cases directly supported as well - the lambda_stmt() can accommodate ORM functionality completely and used directly with Session.execute():

  def select_user(session, name):
      stmt = lambda_stmt(lambda: select(User))
      stmt += lambda s: s.where(User.name == name)
  
      row = session.execute(stmt).first()
      return row

• Bound parameters are automatically accommodated - in contrast to SQLAlchemy’s previous “baked query” system, the lambda SQL system accommodates for Python literal values which become SQL bound parameters automatically. This means that even though a given lambda runs only once, the values that become bound parameters are extracted from the closure of the lambda on every run:

  >>> def my_stmt(x, y):
  ...     stmt = lambda_stmt(lambda: select(func.max(x, y)))
  ...     return stmt
  >>> engine = create_engine("sqlite://", echo=True)
  >>> with engine.connect() as conn:
  ...     print(conn.scalar(my_stmt(5, 10)))
  ...     print(conn.scalar(my_stmt(12, 8)))
  SELECT max(?, ?) AS max_1
  [generated in 0.00057s] (5, 10)
  10
  SELECT max(?, ?) AS max_1
  [cached since 0.002059s ago] (12, 8)
  12

  Above, StatementLambdaElement extracted the values of x and y from the closure of the lambda that is generated each time my_stmt() is invoked; these were substituted into the cached SQL construct as the values of the parameters.

• The lambda should ideally produce an identical SQL structure in all cases - Avoid using conditionals or custom callables inside of lambdas that might make it produce different SQL based on inputs; if a function might conditionally use two different SQL fragments, use two separate lambdas:

  # **Don't** do this:
  
  
  def my_stmt(parameter, thing=False):
      stmt = lambda_stmt(lambda: select(table))
      stmt += lambda s: (
          s.where(table.c.x > parameter) if thing else s.where(table.c.y == parameter)
      )
      return stmt
  
  
  # **Do** do this:
  
  
  def my_stmt(parameter, thing=False):
      stmt = lambda_stmt(lambda: select(table))
      if thing:
          stmt += lambda s: s.where(table.c.x > parameter)
      else:
          stmt += lambda s: s.where(table.c.y == parameter)
      return stmt

  There are a variety of failures which can occur if the lambda does not produce a consistent SQL construct and some are not trivially detectable right now.

• Don’t use functions inside the lambda to produce bound values - the bound value tracking approach requires that the actual value to be used in the SQL statement be locally present in the closure of the lambda. This is not possible if values are generated from other functions, and the LambdaElement should normally raise an error if this is attempted:

  >>> def my_stmt(x, y):
  ...     def get_x():
  ...         return x
  ...
  ...     def get_y():
  ...         return y
  ...
  ...     stmt = lambda_stmt(lambda: select(func.max(get_x(), get_y())))
  ...     return stmt
  >>> with engine.connect() as conn:
  ...     print(conn.scalar(my_stmt(5, 10)))
  Traceback (most recent call last):
    # ...
  sqlalchemy.exc.InvalidRequestError: Can't invoke Python callable get_x()
  inside of lambda expression argument at
  <code object <lambda> at 0x7fed15f350e0, file "<stdin>", line 6>;
  lambda SQL constructs should not invoke functions from closure variables
  to produce literal values since the lambda SQL system normally extracts
  bound values without actually invoking the lambda or any functions within it.

  Above, the use of get_x() and get_y(), if they are necessary, should occur outside of the lambda and assigned to a local closure variable:

  >>> def my_stmt(x, y):
  ...     def get_x():
  ...         return x
  ...
  ...     def get_y():
  ...         return y
  ...
  ...     x_param, y_param = get_x(), get_y()
  ...     stmt = lambda_stmt(lambda: select(func.max(x_param, y_param)))
  ...     return stmt

• Avoid referring to non-SQL constructs inside of lambdas as they are not cacheable by default - this issue refers to how the LambdaElement creates a cache key from other closure variables within the statement. In order to provide the best guarantee of an accurate cache key, all objects located in the closure of the lambda are considered to be significant, and none will be assumed to be appropriate for a cache key by default. So the following example will also raise a rather detailed error message:

  >>> class Foo:
  ...     def __init__(self, x, y):
  ...         self.x = x
  ...         self.y = y
  >>> def my_stmt(foo):
  ...     stmt = lambda_stmt(lambda: select(func.max(foo.x, foo.y)))
  ...     return stmt
  >>> with engine.connect() as conn:
  ...     print(conn.scalar(my_stmt(Foo(5, 10))))
  Traceback (most recent call last):
    # ...
  sqlalchemy.exc.InvalidRequestError: Closure variable named 'foo' inside of
  lambda callable <code object <lambda> at 0x7fed15f35450, file
  "<stdin>", line 2> does not refer to a cacheable SQL element, and also
  does not appear to be serving as a SQL literal bound value based on the
  default SQL expression returned by the function.  This variable needs to
  remain outside the scope of a SQL-generating lambda so that a proper cache
  key may be generated from the lambda's state.  Evaluate this variable
  outside of the lambda, set track_on=[<elements>] to explicitly select
  closure elements to track, or set track_closure_variables=False to exclude
  closure variables from being part of the cache key.

  The above error indicates that LambdaElement will not assume that the Foo object passed in will continue to behave the same in all cases. It also won’t assume it can use Foo as part of the cache key by default; if it were to use the Foo object as part of the cache key, if there were many different Foo objects this would fill up the cache with duplicate information, and would also hold long-lasting references to all of these objects.

  The best way to resolve the above situation is to not refer to foo inside of the lambda, and refer to it outside instead:

  >>> def my_stmt(foo):
  ...     x_param, y_param = foo.x, foo.y
  ...     stmt = lambda_stmt(lambda: select(func.max(x_param, y_param)))
  ...     return stmt

  In some situations, if the SQL structure of the lambda is guaranteed to never change based on input, to pass track_closure_variables=False which will disable any tracking of closure variables other than those used for bound parameters:

  >>> def my_stmt(foo):
  ...     stmt = lambda_stmt(
  ...         lambda: select(func.max(foo.x, foo.y)), track_closure_variables=False
  ...     )
  ...     return stmt

  There is also the option to add objects to the element to explicitly form part of the cache key, using the track_on parameter; using this parameter allows specific values to serve as the cache key and will also prevent other closure variables from being considered. This is useful for cases where part of the SQL being constructed originates from a contextual object of some sort that may have many different values. In the example below, the first segment of the SELECT statement will disable tracking of the foo variable, whereas the second segment will explicitly track self as part of the cache key:

  >>> def my_stmt(self, foo):
  ...     stmt = lambda_stmt(
  ...         lambda: select(*self.column_expressions), track_closure_variables=False
  ...     )
  ...     stmt = stmt.add_criteria(lambda: self.where_criteria, track_on=[self])
  ...     return stmt

  Using track_on means the given objects will be stored long term in the lambda’s internal cache and will have strong references for as long as the cache doesn’t clear out those objects (an LRU scheme of 1000 entries is used by default).

Cache Key Generation

In order to understand some of the options and behaviors which occur with lambda SQL constructs, an understanding of the caching system is helpful.

SQLAlchemy’s caching system normally generates a cache key from a given SQL expression construct by producing a structure that represents all the state within the construct:

>>> from sqlalchemy import select, column
>>> stmt = select(column("q"))
>>> cache_key = stmt._generate_cache_key()
>>> print(cache_key)  # somewhat paraphrased
CacheKey(key=(
  '0',
  <class 'sqlalchemy.sql.selectable.Select'>,
  '_raw_columns',
  (
    (
      '1',
      <class 'sqlalchemy.sql.elements.ColumnClause'>,
      'name',
      'q',
      'type',
      (
        <class 'sqlalchemy.sql.sqltypes.NullType'>,
      ),
    ),
  ),
  # a few more elements are here, and many more for a more
  # complicated SELECT statement
),)

The above key is stored in the cache which is essentially a dictionary, and the value is a construct that among other things stores the string form of the SQL statement, in this case the phrase “SELECT q”. We can observe that even for an extremely short query the cache key is pretty verbose as it has to represent everything that may vary about what’s being rendered and potentially executed.

The lambda construction system by contrast creates a different kind of cache key:

>>> from sqlalchemy import lambda_stmt
>>> stmt = lambda_stmt(lambda: select(column("q")))
>>> cache_key = stmt._generate_cache_key()
>>> print(cache_key)
CacheKey(key=(
  <code object <lambda> at 0x7fed1617c710, file "<stdin>", line 1>,
  <class 'sqlalchemy.sql.lambdas.StatementLambdaElement'>,
),)

Above, we see a cache key that is vastly shorter than that of the non-lambda statement, and additionally that production of the select(column("q")) construct itself was not even necessary; the Python lambda itself contains an attribute called __code__ which refers to a Python code object that within the runtime of the application is immutable and permanent.

When the lambda also includes closure variables, in the normal case that these variables refer to SQL constructs such as column objects, they become part of the cache key, or if they refer to literal values that will be bound parameters, they are placed in a separate element of the cache key:

>>> def my_stmt(parameter):
...     col = column("q")
...     stmt = lambda_stmt(lambda: select(col))
...     stmt += lambda s: s.where(col == parameter)
...     return stmt

The above StatementLambdaElement includes two lambdas, both of which refer to the col closure variable, so the cache key will represent both of these segments as well as the column() object:

>>> stmt = my_stmt(5)
>>> key = stmt._generate_cache_key()
>>> print(key)
CacheKey(key=(
  <code object <lambda> at 0x7f07323c50e0, file "<stdin>", line 3>,
  (
    '0',
    <class 'sqlalchemy.sql.elements.ColumnClause'>,
    'name',
    'q',
    'type',
    (
      <class 'sqlalchemy.sql.sqltypes.NullType'>,
    ),
  ),
  <code object <lambda> at 0x7f07323c5190, file "<stdin>", line 4>,
  <class 'sqlalchemy.sql.lambdas.LinkedLambdaElement'>,
  (
    '0',
    <class 'sqlalchemy.sql.elements.ColumnClause'>,
    'name',
    'q',
    'type',
    (
      <class 'sqlalchemy.sql.sqltypes.NullType'>,
    ),
  ),
  (
    '0',
    <class 'sqlalchemy.sql.elements.ColumnClause'>,
    'name',
    'q',
    'type',
    (
      <class 'sqlalchemy.sql.sqltypes.NullType'>,
    ),
  ),
),)

The second part of the cache key has retrieved the bound parameters that will be used when the statement is invoked:

>>> key.bindparams
[BindParameter('%(139668884281280 parameter)s', 5, type_=Integer())]

For a series of examples of “lambda” caching with performance comparisons, see the “short_selects” test suite within the Performance performance example.

“Insert Many Values” Behavior for INSERT statements

Added in version 2.0: see Optimized ORM bulk insert now implemented for all backends other than MySQL for background on the change including sample performance tests

Tip

The insertmanyvalues feature is a transparently available performance feature which typically requires no end-user intervention in order for it to take place as needed. This section describes the architecture of the feature as well as how to measure its performance and tune its behavior in order to optimize the speed of bulk INSERT statements, particularly as used by the ORM.

As more databases have added support for INSERT..RETURNING, SQLAlchemy has undergone a major change in how it approaches the subject of INSERT statements where there’s a need to acquire server-generated values, most importantly server-generated primary key values which allow the new row to be referenced in subsequent operations. In particular, this scenario has long been a significant performance issue in the ORM, which relies on being able to retrieve server-generated primary key values in order to correctly populate the identity map.

With recent support for RETURNING added to SQLite and MariaDB, SQLAlchemy no longer needs to rely upon the single-row-only cursor.lastrowid attribute provided by the DBAPI for most backends; RETURNING may now be used for all SQLAlchemy-included backends with the exception of MySQL. The remaining performance limitation, that the cursor.executemany() DBAPI method does not allow for rows to be fetched, is resolved for most backends by foregoing the use of executemany() and instead restructuring individual INSERT statements to each accommodate a large number of rows in a single statement that is invoked using cursor.execute(). This approach originates from the psycopg2 fast execution helpers feature of the psycopg2 DBAPI, which SQLAlchemy incrementally added more and more support towards in recent release series.

Current Support

The feature is enabled for all backend included in SQLAlchemy that support RETURNING, with the exception of Oracle Database for which both the python-oracledb and cx_Oracle drivers offer their own equivalent feature. The feature normally takes place when making use of the Insert.returning() method of an Insert construct in conjunction with executemany execution, which occurs when passing a list of dictionaries to the Connection.execute.parameters parameter of the Connection.execute() or Session.execute() methods (as well as equivalent methods under asyncio and shorthand methods like Session.scalars()). It also takes place within the ORM unit of work process when using methods such as Session.add() and Session.add_all() to add rows.

For SQLAlchemy’s included dialects, support or equivalent support is currently as follows:

• SQLite - supported for SQLite versions 3.35 and above

• PostgreSQL - all supported Postgresql versions (9 and above)

• SQL Server - all supported SQL Server versions [1]

• MariaDB - supported for MariaDB versions 10.5 and above

• MySQL - no support, no RETURNING feature is present

• Oracle Database - supports RETURNING with executemany using native python-oracledb / cx_Oracle APIs, for all supported Oracle Database versions 9 and above, using multi-row OUT parameters. This is not the same implementation as “executemanyvalues”, however has the same usage patterns and equivalent performance benefits.

Changed in version 2.0.10:

[1]

”insertmanyvalues” support for Microsoft SQL Server is restored, after being temporarily disabled in version 2.0.9.

Disabling the feature

To disable the “insertmanyvalues” feature for a given backend for an Engine overall, pass the create_engine.use_insertmanyvalues parameter as False to create_engine():

engine = create_engine(
    "mariadb+mariadbconnector://scott:tiger@host/db", use_insertmanyvalues=False
)

The feature can also be disabled from being used implicitly for a particular Table object by passing the Table.implicit_returning parameter as False:

t = Table(
    "t",
    metadata,
    Column("id", Integer, primary_key=True),
    Column("x", Integer),
    implicit_returning=False,
)

The reason one might want to disable RETURNING for a specific table is to work around backend-specific limitations.

Batched Mode Operation

The feature has two modes of operation, which are selected transparently on a per-dialect, per-Table basis. One is batched mode, which reduces the number of database round trips by rewriting an INSERT statement of the form:

INSERT INTO a (data, x, y) VALUES (%(data)s, %(x)s, %(y)s) RETURNING a.id

into a “batched” form such as:

INSERT INTO a (data, x, y) VALUES
    (%(data_0)s, %(x_0)s, %(y_0)s),
    (%(data_1)s, %(x_1)s, %(y_1)s),
    (%(data_2)s, %(x_2)s, %(y_2)s),
    ...
    (%(data_78)s, %(x_78)s, %(y_78)s)
RETURNING a.id

where above, the statement is organized against a subset (a “batch”) of the input data, the size of which is determined by the database backend as well as the number of parameters in each batch to correspond to known limits for statement size / number of parameters. The feature then executes the INSERT statement once for each batch of input data until all records are consumed, concatenating the RETURNING results for each batch into a single large rowset that’s available from a single Result object.

This “batched” form allows INSERT of many rows using much fewer database round trips, and has been shown to allow dramatic performance improvements for most backends where it’s supported.

Correlating RETURNING rows to parameter sets

Added in version 2.0.10.

The “batch” mode query illustrated in the previous section does not guarantee the order of records returned would correspond with that of the input data. When used by the SQLAlchemy ORM unit of work process, as well as for applications which correlate returned server-generated values with input data, the Insert.returning() and UpdateBase.return_defaults() methods include an option Insert.returning.sort_by_parameter_order which indicates that “insertmanyvalues” mode should guarantee this correspondence. This is not related to the order in which records are actually INSERTed by the database backend, which is not assumed under any circumstances; only that the returned records should be organized when received back to correspond to the order in which the original input data was passed.

When the Insert.returning.sort_by_parameter_order parameter is present, for tables that use server-generated integer primary key values such as IDENTITY, PostgreSQL SERIAL, MariaDB AUTO_INCREMENT, or SQLite’s ROWID scheme, “batch” mode may instead opt to use a more complex INSERT..RETURNING form, in conjunction with post-execution sorting of rows based on the returned values, or if such a form is not available, the “insertmanyvalues” feature may gracefully degrade to “non-batched” mode which runs individual INSERT statements for each parameter set.

For example, on SQL Server when an auto incrementing IDENTITY column is used as the primary key, the following SQL form is used [2]:

INSERT INTO a (data, x, y)
OUTPUT inserted.id, inserted.id AS id__1
SELECT p0, p1, p2 FROM (VALUES
    (?, ?, ?, 0), (?, ?, ?, 1), (?, ?, ?, 2),
    ...
    (?, ?, ?, 77)
) AS imp_sen(p0, p1, p2, sen_counter) ORDER BY sen_counter

A similar form is used for PostgreSQL as well, when primary key columns use SERIAL or IDENTITY. The above form does not guarantee the order in which rows are inserted. However, it does guarantee that the IDENTITY or SERIAL values will be created in order with each parameter set [3]. The “insertmanyvalues” feature then sorts the returned rows for the above INSERT statement by incrementing integer identity.

For the SQLite database, there is no appropriate INSERT form that can correlate the production of new ROWID values with the order in which the parameter sets are passed. As a result, when using server-generated primary key values, the SQLite backend will degrade to “non-batched” mode when ordered RETURNING is requested. For MariaDB, the default INSERT form used by insertmanyvalues is sufficient, as this database backend will line up the order of AUTO_INCREMENT with the order of input data when using InnoDB [4].

For a client-side generated primary key, such as when using the Python uuid.uuid4() function to generate new values for a Uuid column, the “insertmanyvalues” feature transparently includes this column in the RETURNING records and correlates its value to that of the given input records, thus maintaining correspondence between input records and result rows. From this, it follows that all backends allow for batched, parameter-correlated RETURNING order when client-side-generated primary key values are used.

The subject of how “insertmanyvalues” “batch” mode determines a column or columns to use as a point of correspondence between input parameters and RETURNING rows is known as an insert sentinel, which is a specific column or columns that are used to track such values. The “insert sentinel” is normally selected automatically, however can also be user-configuration for extremely special cases; the section Configuring Sentinel Columns describes this.

For backends that do not offer an appropriate INSERT form that can deliver server-generated values deterministically aligned with input values, or for Table configurations that feature other kinds of server generated primary key values, “insertmanyvalues” mode will make use of non-batched mode when guaranteed RETURNING ordering is requested.

See also

[2]

• Microsoft SQL Server rationale

  “INSERT queries that use SELECT with ORDER BY to populate rows guarantees how identity values are computed but not the order in which the rows are inserted.” https://learn.microsoft.com/en-us/sql/t-sql/statements/insert-transact-sql?view=sql-server-ver16#limitations-and-restrictions

[3]

• PostgreSQL batched INSERT Discussion

  Original description in 2018 https://www.postgresql.org/message-id/29386.1528813619@sss.pgh.pa.us

  Follow up in 2023 - https://www.postgresql.org/message-id/be108555-da2a-4abc-a46b-acbe8b55bd25%40app.fastmail.com

[4]

• MariaDB AUTO_INCREMENT behavior (using the same InnoDB engine as MySQL)

  https://dev.mysql.com/doc/refman/8.0/en/innodb-auto-increment-handling.html

  https://dba.stackexchange.com/a/72099

Non-Batched Mode Operation

For Table configurations that do not have client side primary key values, and offer server-generated primary key values (or no primary key) that the database in question is not able to invoke in a deterministic or sortable way relative to multiple parameter sets, the “insertmanyvalues” feature when tasked with satisfying the Insert.returning.sort_by_parameter_order requirement for an Insert statement may instead opt to use non-batched mode.

In this mode, the original SQL form of INSERT is maintained, and the “insertmanyvalues” feature will instead run the statement as given for each parameter set individually, organizing the returned rows into a full result set. Unlike previous SQLAlchemy versions, it does so in a tight loop that minimizes Python overhead. In some cases, such as on SQLite, “non-batched” mode performs exactly as well as “batched” mode.

Statement Execution Model

For both “batched” and “non-batched” modes, the feature will necessarily invoke multiple INSERT statements using the DBAPI cursor.execute() method, within the scope of single call to the Core-level Connection.execute() method, with each statement containing up to a fixed limit of parameter sets. This limit is configurable as described below at Controlling the Batch Size. The separate calls to cursor.execute() are logged individually and also individually passed along to event listeners such as ConnectionEvents.before_cursor_execute() (see Logging and Events below).

Configuring Sentinel Columns

In typical cases, the “insertmanyvalues” feature in order to provide INSERT..RETURNING with deterministic row order will automatically determine a sentinel column from a given table’s primary key, gracefully degrading to “row at a time” mode if one cannot be identified. As a completely optional feature, to get full “insertmanyvalues” bulk performance for tables that have server generated primary keys whose default generator functions aren’t compatible with the “sentinel” use case, other non-primary key columns may be marked as “sentinel” columns assuming they meet certain requirements. A typical example is a non-primary key Uuid column with a client side default such as the Python uuid.uuid4() function. There is also a construct to create simple integer columns with a a client side integer counter oriented towards the “insertmanyvalues” use case.

Sentinel columns may be indicated by adding Column.insert_sentinel to qualifying columns. The most basic “qualifying” column is a not-nullable, unique column with a client side default, such as a UUID column as follows:

import uuid

from sqlalchemy import Column
from sqlalchemy import FetchedValue
from sqlalchemy import Integer
from sqlalchemy import String
from sqlalchemy import Table
from sqlalchemy import Uuid

my_table = Table(
    "some_table",
    metadata,
    # assume some arbitrary server-side function generates
    # primary key values, so cannot be tracked by a bulk insert
    Column("id", String(50), server_default=FetchedValue(), primary_key=True),
    Column("data", String(50)),
    Column(
        "uniqueid",
        Uuid(),
        default=uuid.uuid4,
        nullable=False,
        unique=True,
        insert_sentinel=True,
    ),
)

When using ORM Declarative models, the same forms are available using the mapped_column construct:

import uuid

from sqlalchemy.orm import DeclarativeBase
from sqlalchemy.orm import Mapped
from sqlalchemy.orm import mapped_column


class Base(DeclarativeBase):
    pass


class MyClass(Base):
    __tablename__ = "my_table"

    id: Mapped[str] = mapped_column(primary_key=True, server_default=FetchedValue())
    data: Mapped[str] = mapped_column(String(50))
    uniqueid: Mapped[uuid.UUID] = mapped_column(
        default=uuid.uuid4, unique=True, insert_sentinel=True
    )

While the values generated by the default generator must be unique, the actual UNIQUE constraint on the above “sentinel” column, indicated by the unique=True parameter, itself is optional and may be omitted if not desired.

There is also a special form of “insert sentinel” that’s a dedicated nullable integer column which makes use of a special default integer counter that’s only used during “insertmanyvalues” operations; as an additional behavior, the column will omit itself from SQL statements and result sets and behave in a mostly transparent manner. It does need to be physically present within the actual database table, however. This style of Column may be constructed using the function insert_sentinel():

from sqlalchemy import Column
from sqlalchemy import Integer
from sqlalchemy import String
from sqlalchemy import Table
from sqlalchemy import Uuid
from sqlalchemy import insert_sentinel

Table(
    "some_table",
    metadata,
    Column("id", Integer, primary_key=True),
    Column("data", String(50)),
    insert_sentinel("sentinel"),
)

When using ORM Declarative, a Declarative-friendly version of insert_sentinel() is available called orm_insert_sentinel(), which has the ability to be used on the Base class or a mixin; if packaged using declared_attr(), the column will apply itself to all table-bound subclasses including within joined inheritance hierarchies:

from sqlalchemy.orm import declared_attr
from sqlalchemy.orm import DeclarativeBase
from sqlalchemy.orm import Mapped
from sqlalchemy.orm import mapped_column
from sqlalchemy.orm import orm_insert_sentinel


class Base(DeclarativeBase):
    @declared_attr
    def _sentinel(cls) -> Mapped[int]:
        return orm_insert_sentinel()


class MyClass(Base):
    __tablename__ = "my_table"

    id: Mapped[str] = mapped_column(primary_key=True, server_default=FetchedValue())
    data: Mapped[str] = mapped_column(String(50))


class MySubClass(MyClass):
    __tablename__ = "sub_table"

    id: Mapped[str] = mapped_column(ForeignKey("my_table.id"), primary_key=True)


class MySingleInhClass(MyClass):
    pass

In the example above, both “my_table” and “sub_table” will have an additional integer column named “_sentinel” that can be used by the “insertmanyvalues” feature to help optimize bulk inserts used by the ORM.

Configuring Monotonic Functions such as UUIDV7

Using a monotonic function such as uuidv7 is supported by the “insertmanyvalues” feature most easily by establishing the function as a client side callable, e.g. using Python’s built-in uuid.uuid7() call by providing the callable to the Connection.default parameter:

import uuid

from sqlalchemy import UUID, Integer

t = Table(
    "t",
    metadata,
    Column("id", UUID, default=uuid.uuid7, primary_key=True),
    Column("x", Integer),
)

In the above example, SQLAlchemy will invoke Python’s uuid.uuid7() function to create new primary key identifiers, which will be batchable by the “insertmanyvalues” feature.

However, some databases like PostgreSQL provide a server-side function for uuid7 called uuidv7(); in SQLAlchemy, this would be available from the func namespace as func.uuidv7(), and may be configured on a Column using either Connection.default to allow it to be called as needed, or Connection.server_default to establish it as part of the table’s DDL. However, for full batched “insertmanyvalues” behavior including support for sorted RETURNING (as would allow the ORM to most effectively batch INSERT statements), an additional directive must be included indicating that the function produces monotonically increasing values, which is the monotonic=True directive. This is illustrated below as a DDL server default using Connection.server_default:

from sqlalchemy import func, Integer

t = Table(
    "t",
    metadata,
    Column("id", UUID, server_default=func.uuidv7(monotonic=True), primary_key=True),
    Column("x", Integer),
)

Using the above form, a batched INSERT…RETURNING on PostgreSQL with UpdateBase.returning.sort_by_parameter_order set to True will look like:

INSERT INTO t (x) SELECT p0::INTEGER FROM
(VALUES (%(x__0)s, 0), (%(x__1)s, 1), (%(x__2)s, 2),   ...)
AS imp_sen(p0, sen_counter) ORDER BY sen_counter
RETURNING t.id, t.id AS id__1

Similarly if the function is configured as an ad-hoc server side function using Connection.default:

t = Table(
    "t",
    metadata,
    Column("id", UUID, default=func.uuidv7(monotonic=True), primary_key=True),
    Column("x", Integer),
)

The function will then be rendered in the SQL statement explicitly:

INSERT INTO t (id, x) SELECT uuidv7(), p1::INTEGER FROM
(VALUES (%(x__0)s, 0), (%(x__1)s, 1), (%(x__2)s, 2), ...)
AS imp_sen(p1, sen_counter) ORDER BY sen_counter
RETURNING t.id, t.id AS id__1

Added in version 2.1: Added support for explicit monotonic server side functions using monotonic=True with any Function.

See also

postgresql_monotonic_functions

Controlling the Batch Size

A key characteristic of “insertmanyvalues” is that the size of the INSERT statement is limited on a fixed max number of “values” clauses as well as a dialect-specific fixed total number of bound parameters that may be represented in one INSERT statement at a time. When the number of parameter dictionaries given exceeds a fixed limit, or when the total number of bound parameters to be rendered in a single INSERT statement exceeds a fixed limit (the two fixed limits are separate), multiple INSERT statements will be invoked within the scope of a single Connection.execute() call, each of which accommodate for a portion of the parameter dictionaries, known as a “batch”. The number of parameter dictionaries represented within each “batch” is then known as the “batch size”. For example, a batch size of 500 means that each INSERT statement emitted will INSERT at most 500 rows.

It’s potentially important to be able to adjust the batch size, as a larger batch size may be more performant for an INSERT where the value sets themselves are relatively small, and a smaller batch size may be more appropriate for an INSERT that uses very large value sets, where both the size of the rendered SQL as well as the total data size being passed in one statement may benefit from being limited to a certain size based on backend behavior and memory constraints. For this reason the batch size can be configured on a per-Engine as well as a per-statement basis. The parameter limit on the other hand is fixed based on the known characteristics of the database in use.

The batch size defaults to 1000 for most backends, with an additional per-dialect “max number of parameters” limiting factor that may reduce the batch size further on a per-statement basis. The max number of parameters varies by dialect and server version; the largest size is 32700 (chosen as a healthy distance away from PostgreSQL’s limit of 32767 and SQLite’s modern limit of 32766, while leaving room for additional parameters in the statement as well as for DBAPI quirkiness). Older versions of SQLite (prior to 3.32.0) will set this value to 999. MariaDB has no established limit however 32700 remains as a limiting factor for SQL message size.

The value of the “batch size” can be affected Engine wide via the create_engine.insertmanyvalues_page_size parameter. Such as, to affect INSERT statements to include up to 100 parameter sets in each statement:

e = create_engine("sqlite://", insertmanyvalues_page_size=100)

The batch size may also be affected on a per statement basis using the Connection.execution_options.insertmanyvalues_page_size execution option, such as per execution:

with e.begin() as conn:
    result = conn.execute(
        table.insert().returning(table.c.id),
        parameterlist,
        execution_options={"insertmanyvalues_page_size": 100},
    )

Or configured on the statement itself:

stmt = (
    table.insert()
    .returning(table.c.id)
    .execution_options(insertmanyvalues_page_size=100)
)
with e.begin() as conn:
    result = conn.execute(stmt, parameterlist)

Logging and Events

The “insertmanyvalues” feature integrates fully with SQLAlchemy’s statement logging as well as cursor events such as ConnectionEvents.before_cursor_execute(). When the list of parameters is broken into separate batches, each INSERT statement is logged and passed to event handlers individually. This is a major change compared to how the psycopg2-only feature worked in previous 1.x series of SQLAlchemy, where the production of multiple INSERT statements was hidden from logging and events. Logging display will truncate the long lists of parameters for readability, and will also indicate the specific batch of each statement. The example below illustrates an excerpt of this logging:

INSERT INTO a (data, x, y) VALUES (?, ?, ?), ... 795 characters truncated ...  (?, ?, ?), (?, ?, ?) RETURNING id
[generated in 0.00177s (insertmanyvalues) 1/10 (unordered)] ('d0', 0, 0, 'd1',  ...
INSERT INTO a (data, x, y) VALUES (?, ?, ?), ... 795 characters truncated ...  (?, ?, ?), (?, ?, ?) RETURNING id
[insertmanyvalues 2/10 (unordered)] ('d100', 100, 1000, 'd101', ...

...

INSERT INTO a (data, x, y) VALUES (?, ?, ?), ... 795 characters truncated ...  (?, ?, ?), (?, ?, ?) RETURNING id
[insertmanyvalues 10/10 (unordered)] ('d900', 900, 9000, 'd901', ...

When non-batch mode takes place, logging will indicate this along with the insertmanyvalues message:

...

INSERT INTO a (data, x, y) VALUES (?, ?, ?) RETURNING id
[insertmanyvalues 67/78 (ordered; batch not supported)] ('d66', 66, 66)
INSERT INTO a (data, x, y) VALUES (?, ?, ?) RETURNING id
[insertmanyvalues 68/78 (ordered; batch not supported)] ('d67', 67, 67)
INSERT INTO a (data, x, y) VALUES (?, ?, ?) RETURNING id
[insertmanyvalues 69/78 (ordered; batch not supported)] ('d68', 68, 68)
INSERT INTO a (data, x, y) VALUES (?, ?, ?) RETURNING id
[insertmanyvalues 70/78 (ordered; batch not supported)] ('d69', 69, 69)

...

See also

Configuring Logging

Upsert Support

The PostgreSQL, SQLite, and MariaDB dialects offer backend-specific “upsert” constructs insert(), insert() and insert(), which are each Insert constructs that have an additional method such as on_conflict_do_update()` or ``on_duplicate_key(). These constructs also support “insertmanyvalues” behaviors when they are used with RETURNING, allowing efficient upserts with RETURNING to take place.

Engine Disposal

The Engine refers to a connection pool, which means under normal circumstances, there are open database connections present while the Engine object is still resident in memory. When an Engine is garbage collected, its connection pool is no longer referred to by that Engine, and assuming none of its connections are still checked out, the pool and its connections will also be garbage collected, which has the effect of closing out the actual database connections as well. But otherwise, the Engine will hold onto open database connections assuming it uses the normally default pool implementation of QueuePool.

The Engine is intended to normally be a permanent fixture established up-front and maintained throughout the lifespan of an application. It is not intended to be created and disposed on a per-connection basis; it is instead a registry that maintains both a pool of connections as well as configurational information about the database and DBAPI in use, as well as some degree of internal caching of per-database resources.

However, there are many cases where it is desirable that all connection resources referred to by the Engine be completely closed out. It’s generally not a good idea to rely on Python garbage collection for this to occur for these cases; instead, the Engine can be explicitly disposed using the Engine.dispose() method. This disposes of the engine’s underlying connection pool and replaces it with a new one that’s empty. Provided that the Engine is discarded at this point and no longer used, all checked-in connections which it refers to will also be fully closed.

Valid use cases for calling Engine.dispose() include:

• When a program wants to release any remaining checked-in connections held by the connection pool and expects to no longer be connected to that database at all for any future operations.

• When a program uses multiprocessing or fork(), and an Engine object is copied to the child process, Engine.dispose() should be called so that the engine creates brand new database connections local to that fork. Database connections generally do not travel across process boundaries. Use the Engine.dispose.close parameter set to False in this case. See the section Using Connection Pools with Multiprocessing or os.fork() for more background on this use case.

• Within test suites or multitenancy scenarios where many ad-hoc, short-lived Engine objects may be created and disposed.

Connections that are checked out are not discarded when the engine is disposed or garbage collected, as these connections are still strongly referenced elsewhere by the application. However, after Engine.dispose() is called, those connections are no longer associated with that Engine; when they are closed, they will be returned to their now-orphaned connection pool which will ultimately be garbage collected, once all connections which refer to it are also no longer referenced anywhere. Since this process is not easy to control, it is strongly recommended that Engine.dispose() is called only after all checked out connections are checked in or otherwise de-associated from their pool.

An alternative for applications that are negatively impacted by the Engine object’s use of connection pooling is to disable pooling entirely. This typically incurs only a modest performance impact upon the use of new connections, and means that when a connection is checked in, it is entirely closed out and is not held in memory. See Switching Pool Implementations for guidelines on how to disable pooling.

See also

Connection Pooling

Using Connection Pools with Multiprocessing or os.fork()

Working with Driver SQL and Raw DBAPI Connections

The introduction on using Connection.execute() made use of the text() construct in order to illustrate how textual SQL statements may be invoked. When working with SQLAlchemy, textual SQL is actually more of the exception rather than the norm, as the Core expression language and the ORM both abstract away the textual representation of SQL. However, the text() construct itself also provides some abstraction of textual SQL in that it normalizes how bound parameters are passed, as well as that it supports datatyping behavior for parameters and result set rows.

Invoking SQL strings directly to the driver

For the use case where one wants to invoke textual SQL directly passed to the underlying driver (known as the DBAPI) without any intervention from the text() construct, the Connection.exec_driver_sql() method may be used:

with engine.connect() as conn:
    conn.exec_driver_sql("SET param='bar'")

Added in version 1.4: Added the Connection.exec_driver_sql() method.

Working with the DBAPI cursor directly

There are some cases where SQLAlchemy does not provide a genericized way at accessing some DBAPI functions, such as calling stored procedures as well as dealing with multiple result sets. In these cases, it’s just as expedient to deal with the raw DBAPI connection directly.

The most common way to access the raw DBAPI connection is to get it from an already present Connection object directly. It is present using the Connection.connection attribute:

connection = engine.connect()
dbapi_conn = connection.connection

The DBAPI connection here is actually a “proxied” in terms of the originating connection pool, however this is an implementation detail that in most cases can be ignored. As this DBAPI connection is still contained within the scope of an owning Connection object, it is best to make use of the Connection object for most features such as transaction control as well as calling the Connection.close() method; if these operations are performed on the DBAPI connection directly, the owning Connection will not be aware of these changes in state.

To overcome the limitations imposed by the DBAPI connection that is maintained by an owning Connection, a DBAPI connection is also available without the need to procure a Connection first, using the Engine.raw_connection() method of Engine:

dbapi_conn = engine.raw_connection()

This DBAPI connection is again a “proxied” form as was the case before. The purpose of this proxying is now apparent, as when we call the .close() method of this connection, the DBAPI connection is typically not actually closed, but instead released back to the engine’s connection pool:

dbapi_conn.close()

While SQLAlchemy may in the future add built-in patterns for more DBAPI use cases, there are diminishing returns as these cases tend to be rarely needed and they also vary highly dependent on the type of DBAPI in use, so in any case the direct DBAPI calling pattern is always there for those cases where it is needed.

See also

How do I get at the raw DBAPI connection when using an Engine? - includes additional details about how the DBAPI connection is accessed as well as the “driver” connection when using asyncio drivers.

Some recipes for DBAPI connection use follow.

Calling Stored Procedures and User Defined Functions

SQLAlchemy supports calling stored procedures and user defined functions several ways. Please note that all DBAPIs have different practices, so you must consult your underlying DBAPI’s documentation for specifics in relation to your particular usage. The following examples are hypothetical and may not work with your underlying DBAPI.

For stored procedures or functions with special syntactical or parameter concerns, DBAPI-level callproc may potentially be used with your DBAPI. An example of this pattern is:

connection = engine.raw_connection()
try:
    cursor_obj = connection.cursor()
    cursor_obj.callproc("my_procedure", ["x", "y", "z"])
    results = list(cursor_obj.fetchall())
    cursor_obj.close()
    connection.commit()
finally:
    connection.close()

Note

Not all DBAPIs use callproc and overall usage details will vary. The above example is only an illustration of how it might look to use a particular DBAPI function.

Your DBAPI may not have a callproc requirement or may require a stored procedure or user defined function to be invoked with another pattern, such as normal SQLAlchemy connection usage. One example of this usage pattern is, at the time of this documentation’s writing, executing a stored procedure in the PostgreSQL database with the psycopg2 DBAPI, which should be invoked with normal connection usage:

connection.execute("CALL my_procedure();")

This above example is hypothetical. The underlying database is not guaranteed to support “CALL” or “SELECT” in these situations, and the keyword may vary dependent on the function being a stored procedure or a user defined function. You should consult your underlying DBAPI and database documentation in these situations to determine the correct syntax and patterns to use.

Multiple Result Sets

Multiple result set support is available from a raw DBAPI cursor using the nextset method:

connection = engine.raw_connection()
try:
    cursor_obj = connection.cursor()
    cursor_obj.execute("select * from table1; select * from table2")
    results_one = cursor_obj.fetchall()
    cursor_obj.nextset()
    results_two = cursor_obj.fetchall()
    cursor_obj.close()
finally:
    connection.close()

Registering New Dialects

The create_engine() function call locates the given dialect using setuptools entrypoints. These entry points can be established for third party dialects within the setup.py script. For example, to create a new dialect “foodialect://”, the steps are as follows:

1. Create a package called foodialect.

2. The package should have a module containing the dialect class, which is typically a subclass of sqlalchemy.engine.default.DefaultDialect. In this example let’s say it’s called FooDialect and its module is accessed via foodialect.dialect.

3. The entry point can be established in setup.cfg as follows:

   [options.entry_points]
   sqlalchemy.dialects =
       foodialect = foodialect.dialect:FooDialect

If the dialect is providing support for a particular DBAPI on top of an existing SQLAlchemy-supported database, the name can be given including a database-qualification. For example, if FooDialect were in fact a MySQL dialect, the entry point could be established like this:

[options.entry_points]
sqlalchemy.dialects
    mysql.foodialect = foodialect.dialect:FooDialect

The above entrypoint would then be accessed as create_engine("mysql+foodialect://").

Registering Dialects In-Process

SQLAlchemy also allows a dialect to be registered within the current process, bypassing the need for separate installation. Use the register() function as follows:

from sqlalchemy.dialects import registry


registry.register("mysql.foodialect", "myapp.dialect", "MyMySQLDialect")

The above will respond to create_engine("mysql+foodialect://") and load the MyMySQLDialect class from the myapp.dialect module.

Connection / Engine API

Result Set API