Connection Pooling

A connection pool is a standard technique used to maintain long running connections in memory for efficient re-use, as well as to provide management for the total number of connections an application might use simultaneously.

Particularly for server-side web applications, a connection pool is the standard way to maintain a “pool” of active database connections in memory which are reused across requests.

SQLAlchemy includes several connection pool implementations which integrate with the _engine.Engine. They can also be used directly for applications that want to add pooling to an otherwise plain DBAPI approach.

Connection Pool Configuration

The _engine.Engine returned by the create_engine() function in most cases has a QueuePool integrated, pre-configured with reasonable pooling defaults. If you’re reading this section only to learn how to enable pooling - congratulations! You’re already done.

The most common QueuePool tuning parameters can be passed directly to create_engine() as keyword arguments: pool_size, max_overflow, pool_recycle and pool_timeout. For example:

engine = create_engine('postgresql://me@localhost/mydb',
                       pool_size=20, max_overflow=0)

In the case of SQLite, the SingletonThreadPool or NullPool are selected by the dialect to provide greater compatibility with SQLite’s threading and locking model, as well as to provide a reasonable default behavior to SQLite “memory” databases, which maintain their entire dataset within the scope of a single connection.

All SQLAlchemy pool implementations have in common that none of them “pre create” connections - all implementations wait until first use before creating a connection. At that point, if no additional concurrent checkout requests for more connections are made, no additional connections are created. This is why it’s perfectly fine for _sa.create_engine() to default to using a QueuePool of size five without regard to whether or not the application really needs five connections queued up - the pool would only grow to that size if the application actually used five connections concurrently, in which case the usage of a small pool is an entirely appropriate default behavior.

Switching Pool Implementations

The usual way to use a different kind of pool with _sa.create_engine() is to use the poolclass argument. This argument accepts a class imported from the sqlalchemy.pool module, and handles the details of building the pool for you. Common options include specifying QueuePool with SQLite:

from sqlalchemy.pool import QueuePool
engine = create_engine('sqlite:///file.db', poolclass=QueuePool)

Disabling pooling using NullPool:

from sqlalchemy.pool import NullPool
engine = create_engine(
          'postgresql+psycopg2://scott:tiger@localhost/test',
          poolclass=NullPool)

Using a Custom Connection Function

See the section Custom DBAPI connect() arguments / on-connect routines for a rundown of the various connection customization routines.

Constructing a Pool

To use a _pool.Pool by itself, the creator function is the only argument that’s required and is passed first, followed by any additional options:

import sqlalchemy.pool as pool
import psycopg2

def getconn():
    c = psycopg2.connect(user='ed', host='127.0.0.1', dbname='test')
    return c

mypool = pool.QueuePool(getconn, max_overflow=10, pool_size=5)

DBAPI connections can then be procured from the pool using the _pool.Pool.connect() function. The return value of this method is a DBAPI connection that’s contained within a transparent proxy:

# get a connection
conn = mypool.connect()

# use it
cursor_obj = conn.cursor()
cursor_obj.execute("select foo")

The purpose of the transparent proxy is to intercept the close() call, such that instead of the DBAPI connection being closed, it is returned to the pool:

# "close" the connection.  Returns
# it to the pool.
conn.close()

The proxy also returns its contained DBAPI connection to the pool when it is garbage collected, though it’s not deterministic in Python that this occurs immediately (though it is typical with cPython). This usage is not recommended however and in particular is not supported with asyncio DBAPI drivers.

Reset On Return

The pool also includes the a “reset on return” feature which will call the rollback() method of the DBAPI connection when the connection is returned to the pool. This is so that any existing transaction on the connection is removed, not only ensuring that no existing state remains on next usage, but also so that table and row locks are released as well as that any isolated data snapshots are removed. This rollback() occurs in most cases even when using an _engine.Engine object, except in the case when the _engine.Connection can guarantee that a rollback() has been called immediately before the connection is returned to the pool.

For most DBAPIs, the call to rollback() is very inexpensive and if the DBAPI has already completed a transaction, the method should be a no-op. However, for DBAPIs that incur performance issues with rollback() even if there’s no state on the connection, this behavior can be disabled using the reset_on_return option of _pool.Pool. The behavior is safe to disable under the following conditions:

If the database does not support transactions at all, such as using MySQL with the MyISAM engine, or the DBAPI is used in autocommit mode only, the behavior can be disabled.
If the pool itself doesn’t maintain a connection after it’s checked in, such as when using NullPool, the behavior can be disabled.
Otherwise, it must be ensured that:
- the application ensures that all _engine.Connection objects are explicitly closed out using a context manager (i.e. with block) or a try/finally style block
- connections are never allowed to be garbage collected before being explicitly closed.
- the DBAPI connection itself, e.g. connection.connection, is not used directly, or the application ensures that .rollback() is called on this connection before releasing it back to the connection pool.

The “reset on return” step may be logged using the logging.DEBUG log level along with the sqlalchemy.pool logger, or by setting echo_pool='debug' with _sa.create_engine().

Pool Events

Connection pools support an event interface that allows hooks to execute upon first connect, upon each new connection, and upon checkout and checkin of connections. See _events.PoolEvents for details.

Dealing with Disconnects

The connection pool has the ability to refresh individual connections as well as its entire set of connections, setting the previously pooled connections as “invalid”. A common use case is allow the connection pool to gracefully recover when the database server has been restarted, and all previously established connections are no longer functional. There are two approaches to this.

Disconnect Handling - Pessimistic

The pessimistic approach refers to emitting a test statement on the SQL connection at the start of each connection pool checkout, to test that the database connection is still viable. Typically, this is a simple statement like “SELECT 1”, but may also make use of some DBAPI-specific method to test the connection for liveness.

The approach adds a small bit of overhead to the connection checkout process, however is otherwise the most simple and reliable approach to completely eliminating database errors due to stale pooled connections. The calling application does not need to be concerned about organizing operations to be able to recover from stale connections checked out from the pool.

It is critical to note that the pre-ping approach does not accommodate for connections dropped in the middle of transactions or other SQL operations. If the database becomes unavailable while a transaction is in progress, the transaction will be lost and the database error will be raised. While the _engine.Connection object will detect a “disconnect” situation and recycle the connection as well as invalidate the rest of the connection pool when this condition occurs, the individual operation where the exception was raised will be lost, and it’s up to the application to either abandon the operation, or retry the whole transaction again. If the engine is configured using DBAPI-level autocommit connections, as described at Setting Transaction Isolation Levels including DBAPI Autocommit, a connection may be reconnected transparently mid-operation using events. See the section How Do I “Retry” a Statement Execution Automatically? for an example.

Pessimistic testing of connections upon checkout is achievable by using the :paramref:`_pool.Pool.pre_ping` argument, available from _sa.create_engine() via the :paramref:`_sa.create_engine.pool_pre_ping` argument:

engine = create_engine("mysql+pymysql://user:pw@host/db", pool_pre_ping=True)

The “pre ping” feature will normally emit SQL equivalent to “SELECT 1” each time a connection is checked out from the pool; if an error is raised that is detected as a “disconnect” situation, the connection will be immediately recycled, and all other pooled connections older than the current time are invalidated, so that the next time they are checked out, they will also be recycled before use.

If the database is still not available when “pre ping” runs, then the initial connect will fail and the error for failure to connect will be propagated normally. In the uncommon situation that the database is available for connections, but is not able to respond to a “ping”, the “pre_ping” will try up to three times before giving up, propagating the database error last received.

Note

the “SELECT 1” emitted by “pre-ping” is invoked within the scope of the connection pool / dialect, using a very short codepath for minimal Python latency. As such, this statement is not logged in the SQL echo output, and will not show up in SQLAlchemy’s engine logging.

New in version 1.2: Added “pre-ping” capability to the _pool.Pool class.

Custom / Legacy Pessimistic Ping

Before :paramref:`_sa.create_engine.pool_pre_ping` was added, the “pre-ping” approach historically has been performed manually using the _events.ConnectionEvents.engine_connect() engine event. The most common recipe for this is below, for reference purposes in case an application is already using such a recipe, or special behaviors are needed:

from sqlalchemy import exc
from sqlalchemy import event
from sqlalchemy import select

some_engine = create_engine(...)

@event.listens_for(some_engine, "engine_connect")
def ping_connection(connection, branch):
    if branch:
        # "branch" refers to a sub-connection of a connection,
        # we don't want to bother pinging on these.
        return

    try:
        # run a SELECT 1.   use a core select() so that
        # the SELECT of a scalar value without a table is
        # appropriately formatted for the backend
        connection.scalar(select(1))
    except exc.DBAPIError as err:
        # catch SQLAlchemy's DBAPIError, which is a wrapper
        # for the DBAPI's exception.  It includes a .connection_invalidated
        # attribute which specifies if this connection is a "disconnect"
        # condition, which is based on inspection of the original exception
        # by the dialect in use.
        if err.connection_invalidated:
            # run the same SELECT again - the connection will re-validate
            # itself and establish a new connection.  The disconnect detection
            # here also causes the whole connection pool to be invalidated
            # so that all stale connections are discarded.
            connection.scalar(select(1))
        else:
            raise

The above recipe has the advantage that we are making use of SQLAlchemy’s facilities for detecting those DBAPI exceptions that are known to indicate a “disconnect” situation, as well as the _engine.Engine object’s ability to correctly invalidate the current connection pool when this condition occurs and allowing the current _engine.Connection to re-validate onto a new DBAPI connection.

Disconnect Handling - Optimistic

When pessimistic handling is not employed, as well as when the database is shutdown and/or restarted in the middle of a connection’s period of use within a transaction, the other approach to dealing with stale / closed connections is to let SQLAlchemy handle disconnects as they occur, at which point all connections in the pool are invalidated, meaning they are assumed to be stale and will be refreshed upon next checkout. This behavior assumes the _pool.Pool is used in conjunction with a _engine.Engine. The _engine.Engine has logic which can detect disconnection events and refresh the pool automatically.

When the _engine.Connection attempts to use a DBAPI connection, and an exception is raised that corresponds to a “disconnect” event, the connection is invalidated. The _engine.Connection then calls the _pool.Pool.recreate() method, effectively invalidating all connections not currently checked out so that they are replaced with new ones upon next checkout. This flow is illustrated by the code example below:

from sqlalchemy import create_engine, exc
e = create_engine(...)
c = e.connect()

try:
    # suppose the database has been restarted.
    c.execute(text("SELECT * FROM table"))
    c.close()
except exc.DBAPIError, e:
    # an exception is raised, Connection is invalidated.
    if e.connection_invalidated:
        print("Connection was invalidated!")

# after the invalidate event, a new connection
# starts with a new Pool
c = e.connect()
c.execute(text("SELECT * FROM table"))

The above example illustrates that no special intervention is needed to refresh the pool, which continues normally after a disconnection event is detected. However, one database exception is raised, per each connection that is in use while the database unavailability event occurred. In a typical web application using an ORM Session, the above condition would correspond to a single request failing with a 500 error, then the web application continuing normally beyond that. Hence the approach is “optimistic” in that frequent database restarts are not anticipated.

Setting Pool Recycle

An additional setting that can augment the “optimistic” approach is to set the pool recycle parameter. This parameter prevents the pool from using a particular connection that has passed a certain age, and is appropriate for database backends such as MySQL that automatically close connections that have been stale after a particular period of time:

from sqlalchemy import create_engine
e = create_engine("mysql://scott:tiger@localhost/test", pool_recycle=3600)

Above, any DBAPI connection that has been open for more than one hour will be invalidated and replaced, upon next checkout. Note that the invalidation only occurs during checkout - not on any connections that are held in a checked out state. pool_recycle is a function of the _pool.Pool itself, independent of whether or not an _engine.Engine is in use.

More on Invalidation

The _pool.Pool provides “connection invalidation” services which allow both explicit invalidation of a connection as well as automatic invalidation in response to conditions that are determined to render a connection unusable.

“Invalidation” means that a particular DBAPI connection is removed from the pool and discarded. The .close() method is called on this connection if it is not clear that the connection itself might not be closed, however if this method fails, the exception is logged but the operation still proceeds.

When using a _engine.Engine, the _engine.Connection.invalidate() method is the usual entrypoint to explicit invalidation. Other conditions by which a DBAPI connection might be invalidated include:

a DBAPI exception such as OperationalError, raised when a method like connection.execute() is called, is detected as indicating a so-called “disconnect” condition. As the Python DBAPI provides no standard system for determining the nature of an exception, all SQLAlchemy dialects include a system called is_disconnect() which will examine the contents of an exception object, including the string message and any potential error codes included with it, in order to determine if this exception indicates that the connection is no longer usable. If this is the case, the _ConnectionFairy.invalidate() method is called and the DBAPI connection is then discarded.
When the connection is returned to the pool, and calling the connection.rollback() or connection.commit() methods, as dictated by the pool’s “reset on return” behavior, throws an exception. A final attempt at calling .close() on the connection will be made, and it is then discarded.
When a listener implementing _events.PoolEvents.checkout() raises the DisconnectionError exception, indicating that the connection won’t be usable and a new connection attempt needs to be made.

All invalidations which occur will invoke the _events.PoolEvents.invalidate() event.

Using FIFO vs. LIFO

The QueuePool class features a flag called :paramref:`.QueuePool.use_lifo`, which can also be accessed from _sa.create_engine() via the flag :paramref:`_sa.create_engine.pool_use_lifo`. Setting this flag to True causes the pool’s “queue” behavior to instead be that of a “stack”, e.g. the last connection to be returned to the pool is the first one to be used on the next request. In contrast to the pool’s long- standing behavior of first-in-first-out, which produces a round-robin effect of using each connection in the pool in series, lifo mode allows excess connections to remain idle in the pool, allowing server-side timeout schemes to close these connections out. The difference between FIFO and LIFO is basically whether or not its desirable for the pool to keep a full set of connections ready to go even during idle periods:

engine = create_engine(
    "postgreql://", pool_use_lifo=True, pool_pre_ping=True)

Above, we also make use of the :paramref:`_sa.create_engine.pool_pre_ping` flag so that connections which are closed from the server side are gracefully handled by the connection pool and replaced with a new connection.

Note that the flag only applies to QueuePool use.

New in version 1.3.

Using Connection Pools with Multiprocessing or os.fork()

It’s critical that when using a connection pool, and by extension when using an _engine.Engine created via _sa.create_engine(), that the pooled connections are not shared to a forked process. TCP connections are represented as file descriptors, which usually work across process boundaries, meaning this will cause concurrent access to the file descriptor on behalf of two or more entirely independent Python interpreter states.

Depending on specifics of the driver and OS, the issues that arise here range from non-working connections to socket connections that are used by multiple processes concurrently, leading to broken messaging (the latter case is typically the most common).

The SQLAlchemy _engine.Engine object refers to a connection pool of existing database connections. So when this object is replicated to a child process, the goal is to ensure that no database connections are carried over. There are three general approaches to this:

Disable pooling using NullPool. This is the most simplistic, one shot system that prevents the _engine.Engine from using any connection more than once:
```
from sqlalchemy.pool import NullPool
engine = create_engine("mysql://user:pass@host/dbname", poolclass=NullPool)
```
Call _engine.Engine.dispose() on any given _engine.Engine as soon one is within the new process. In Python multiprocessing, constructs such as multiprocessing.Pool include “initializer” hooks which are a place that this can be performed; otherwise at the top of where os.fork() or where the Process object begins the child fork, a single call to _engine.Engine.dispose() will ensure any remaining connections are flushed. This is the recommended approach:
```
engine = create_engine("mysql://user:pass@host/dbname")

def run_in_process():
    # process starts.  ensure engine.dispose() is called just once
    # at the beginning
    engine.dispose()

    with engine.connect() as conn:
        conn.execute(text("..."))

p = Process(target=run_in_process)
p.start()
```
An event handler can be applied to the connection pool that tests for connections being shared across process boundaries, and invalidates them. This approach, when combined with an explicit call to dispose() as mentioned above, should cover all cases:
```
from sqlalchemy import event
from sqlalchemy import exc
import os

engine = create_engine("...")

@event.listens_for(engine, "connect")
def connect(dbapi_connection, connection_record):
    connection_record.info['pid'] = os.getpid()

@event.listens_for(engine, "checkout")
def checkout(dbapi_connection, connection_record, connection_proxy):
    pid = os.getpid()
    if connection_record.info['pid'] != pid:
        connection_record.dbapi_connection = connection_proxy.dbapi_connection = None
        raise exc.DisconnectionError(
                "Connection record belongs to pid %s, "
                "attempting to check out in pid %s" %
                (connection_record.info['pid'], pid)
        )
```
Above, we use an approach similar to that described in Disconnect Handling - Pessimistic to treat a DBAPI connection that originated in a different parent process as an “invalid” connection, coercing the pool to recycle the connection record to make a new connection.

When using the above recipe, ensure the dispose approach from #2 is also used, as if the connection pool is exhausted in the parent process when the fork occurs, an empty pool will be copied into the child process which will then hang because it has no connections.

The above strategies will accommodate the case of an _engine.Engine being shared among processes. However, for the case of a transaction-active Session or _engine.Connection being shared, there’s no automatic fix for this; an application needs to ensure a new child process only initiate new _engine.Connection objects and transactions, as well as ORM Session objects. For a Session object, technically this is only needed if the session is currently transaction-bound, however the scope of a single Session is in any case intended to be kept within a single call stack in any case (e.g. not a global object, not shared between processes or threads).

API Documentation - Available Pool Implementations

Connection Pooling — SQLAlchemy 2.0.0b1 documentation

Contents