Skip to content


Subversion checkout URL

You can clone with
Download ZIP
Generic TCP connection pool for gevent-based applications
Python Shell
Branch: master

Merge pull request #1 from inboxapp/master

add features to retry decorator and allow customization of discard-triggering exceptions
latest commit f7216eb58c
@rasky authored



This package implements a generic TCP connection pool for gevent-based applications. It can be used every time your program needs to connnect to an external service through a TCP-based protocol (including all HTTP protocols like REST APIs), and you want your process to keep and manage a pool of connections to the remote endpoint.

A typical scenario might be a gunicorn-based web application with gevent backend, that accesses remote service through HTTPS APIs. In this case, using the pool shorten the request and minimize the latency because the pool keeps open connections to the remote endpoints, and it is not necessary to do a full SSL handshake every time we need to issue a command.


To install the package, use pip:

$ pip install geventconnpool

or easy_install:

$ easy_install geventconnpool

You need to derive from ConnectionPool and reimplement _new_connection, to specify how to open a connection to the remove site. For instance:

from geventconnpool import ConnectionPool
from gevent import socket

class MyPool(ConnectionPool):
    def _new_connection(self):
        return socket.create_connection(('', 2485))

In this case, we're simply opening a TCP connection to a specified peer.

The pool can be istantiated by specifying how many connections we want to keep open at the same time:

pool = MyPool(20)  # always keep 20 connections open

To access a connection within the pool:

with pool.get() as c:
    if c.recv(5) != "PONG\n":
        raise socket.error("something awful happened")

If the context is quit through a socket.error exception, the connection is discarded and a new open is opened in background, to keep the pool always full of valid connections. Any other exception does not have a special meaning, and the connection will be reinserted into the pool to be reused later.

If your application/library uses exceptions that do not derive from socket.error to signify connection errors (imaplib is one example), you can override which exceptions are treated as triggering discards:

pool = MyPool(20, exc_classes=(socket.error, imaplib.IMAP4.error))

Automatic retrying

If you want to be resilent to temporary network errors, you can use the retry decorator that will re-execute the function if it is quit with a socket.error exception:

from geventconnpool import retry

def senddata(data):
    with pool.get() as c:
        if c.recv(2) != "OK":
            raise socket.error("something awful happened")

Since the pool discards the connections when a socket.error exception is generated, the net effect of retry is that a different connection will be used for each attempt.

The retry decorator has some extra features which are disabled by default. If you pass it a logger, it will log each attempt. You can also specify a maximum number of attempts, an interval to wait between successive attempts, and which specific exception classes to trigger retries on. Log messages and levels are customizable.

import logging
log = logging.getLogger()

from geventconnpool import retry

@retry(exc_classes=(socket.error, imaplib.IMAP4.error), logger=log,
       max_failures=5, interval=2)
def senddata(data):
    with pool.get() as c:
        typ, data ='INBOX')

If you wish to codify a set of options to retry into your code, consider using functools.partial.

Advanced connection examples

When implement a connection pool, it is advisable to perform all the initialization phases of the application protocol within the _new_connection callback. For instance, a protocol might allow to switch to TLS (with a STARTTLS-like) and then require authentication:

from geventconnpool import ConnectionPool
from gevent import socket, ssl

class MyPool(ConnectionPool):
    def _new_connection(self):
        s = socket.create_connection(('', 2485))
        res = s.recv(3)
        if res == "OK\n":
            s = ssl.wrap_socket(s)
        elif res == "NO\n":
            raise socket.error("invalid response to STARTTLS")

        s.send("LOGIN: %s\n" % MY_LOGIN_NAME);
        s.send("PASS: %s\n" % MY_PASS);
        res = s.recv(2)
        if res != "OK":
            raise socket.error("authentication failed")
        return s

As you can see, it is possible to simply raise socket.error if something went wrong. The pool is resistant to temporary connection errors and will retry automatically to estabilish new connections later.

Another common situation might involve the usage of third-party libraries like for instance using boto to connect to Amazon AWS:

from geventconnpool import ConnectionPool
import boto
from boto.exception import NoAuthHanlder

class UsersPool(ConnectionPool):
    def _new_connection(self):
            c = boto.connect_dynamodb(MY_AWS_KEY_ID, MY_AWS_SECRET_KEY)
            return c.get_table("users")
            raise socket.error("error connecting to AWS")

In this case, we don't only connect to AWS and authenticate, but we also open a specific table and return a reference to that table. In fact, it is not necessary for the return value of _new_connection() to be a socket (or socket-like): ConnectionPool treats it as a and return it when get is called.


boto has an internal connection pool, but it is only used to be fully-thread safe, and does not preemptively open the connections, authenticate, and perform initialization. This means that it still makes sense to use ConnectionPool to minimize the latency when communicating to AWS.


Some protocols or networks might require a keepalive mechanism to keep a connection open if it is idle. For instance, the remote peer, a firewall or a load-balancer might close a connection if it is idle for too long.

Sometimes, it is sufficient to rely on the standard TCP-level keeaplive, that can be turned on any TCP socket:

from geventconnpool import ConnectionPool
from gevent import socket

class MyPool(ConnectionPool):
    def _new_connection(self):
        s = socket.create_connection(('', 2485))
        s._sock.setsockopt(socket.SOL_SOCKET, socket.SO_KEEPALIVE, 1)
        return s

The TCP keepalive uses ACK packets to continously communicating with the remote peer. To tune the keepalive parameters (delay between ACKs, number of unanswered ACKs to consider the connnection dropped, etc.), you need to tweak with the proc filesystem (yes, it's a global per-computer configuration).

Alternatively, it is possible to implement an application-level keepalive by implemening the _keepalive method and specifying the keepalive frequency in the constructor:

from geventconnpool import ConnectionPool
from gevent import socket

class MyPool(ConnectionPool):
    def _new_connection(self):
        return socket.create_connection(('', 2485))

    def _keepalive(self, c):
        if c.recv(5) != "PONG\n":
            raise socket.error

pool = MyPool(20, keepalive=30)

The above code uses a keepalive based on an application-level command (PING), and specifies that it should be executed every 30 seconds (per each connection).

_keepalive should raise socket.error to communicate that the connection appears to be broken and should be discarded by the pool.

Something went wrong with that request. Please try again.