add default backlog to socket.listen()

[neologix]

BPO	21455
Nosy	@pitrou, @vstinner
Files	socket_listen.diff socket_listen-1.diff

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2014-05-22.20:45:13.362>
created_at = <Date 2014-05-08.19:13:27.028>
labels = ['type-feature', 'library']
title = 'add default backlog to socket.listen()'
updated_at = <Date 2014-05-22.20:45:13.361>
user = 'https://bugs.python.org/neologix'

bugs.python.org fields:

activity = <Date 2014-05-22.20:45:13.361>
actor = 'neologix'
assignee = 'none'
closed = True
closed_date = <Date 2014-05-22.20:45:13.362>
closer = 'neologix'
components = ['Library (Lib)']
creation = <Date 2014-05-08.19:13:27.028>
creator = 'neologix'
dependencies = []
files = ['35186', '35188']
hgrepos = []
issue_num = 21455
keywords = ['patch']
message_count = 11.0
messages = ['218121', '218127', '218135', '218278', '218279', '218369', '218374', '218401', '218411', '218910', '218919']
nosy_count = 5.0
nosy_names = ['pitrou', 'vstinner', 'neologix', 'python-dev', 'sbt']
pr_nums = []
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue21455'
versions = ['Python 3.5']

[neologix]

Having to pass an explicit backlog value to listen() is a pain: most people don't know which value to pass, and often end up using a value too small which can lead to connections being rejected.
For example, if you search throughout the standard library, you'll see many different hard-coded values.

The patch attached uses a default value of SOMAXCONN for socket.listen(): it should give a good default behavior (that's what golang does for example: in fact they go even further by checking the runtime value from net.core.somaxconn).

A second step would be to update the codebase to remove explicit backlogs (unless specific case).

[pitrou]

Is there a risk of SOMAXCONN being huge and therefore allocating a large amount of resources?

[neologix]

Is there a risk of SOMAXCONN being huge and therefore allocating a large amount of resources?

On a sensible operating system, no, but better safe than sorry: the
patch attached caps the value to 128 (a common SOMAXCONN value).
It should be high enough to avoid connection drops in common
workloads, and still guard against non-sensical SOMAXCONN values.

[neologix]

Any objection to the last version?

If not, I'll commit it within a few days.

[pitrou]

Looks good to me!

[vstinner]

Py_MIN(SOMAXCONN, 128)

On Windows, it's not the best choice to use this hardcoded limit. socket.SOMAXCONN is 2^31-1.

listen() documentation says that Windows chooses a reasonable backlog value for you if you pass SOMAXCONN:
http://msdn.microsoft.com/en-us/library/windows/desktop/ms739168%28v=vs.85%29.aspx

You should maybe use SOMAXCONN by default on Windows, and Py_MIN(SOMAXCONN, 128) by default on other platforms (UNIX).

[vstinner]

An article suggests to use max(1024, socket.SOMAXCONN) (to "listen() backlog as large as possible") instead of socket.SOMAXCONN because the OS maximum can be larger than SOMAXCONN (and that it's not possible in Python to get the OS value):
http://utcc.utoronto.ca/~cks/space/blog/python/AvoidSOMAXCONN?showcomments#comments

The following article tries to explain why the default limit is 128 on Linux:
https://derrickpetzold.com/p/somaxconn/

Article giving the value chosen by Windows when SOMAXCONN is passed: it depends on the Windows version (between 5 and 50, hard limit of 200 on Windows 7):
http://stackoverflow.com/questions/4709756/listen-maximum-queue-size-per-windows-version

[neologix]

> Py_MIN(SOMAXCONN, 128)

On Windows, it's not the best choice to use this hardcoded limit. socket.SOMAXCONN is 2^31-1.

I don't see what this would bring: 128 *is* a reasonable limit.

listen() documentation says that Windows chooses a reasonable backlog value for you if you pass SOMAXCONN:
http://msdn.microsoft.com/en-us/library/windows/desktop/ms739168%28v=vs.85%29.aspx

You should maybe use SOMAXCONN by default on Windows, and Py_MIN(SOMAXCONN, 128) by default on other platforms (UNIX).

Trying to come up with a good heuristic is a lost battle, since it
depends not only on the OS but also the application. The goal is to
have a default value that works, is large enough to avoid connection
drops in common use cases, and not too large to avoid resource
consumption.

An article suggests to use max(1024, socket.SOMAXCONN) (to "listen() backlog as large as possible")
instead of socket.SOMAXCONN because the OS maximum can be larger than SOMAXCONN (and that
it's not possible in Python to get the OS value):

In theory we could use net.core.somaxconn sysctl & Co (that's what go
does), but IMO it's overkill.

[vstinner]

Hum ok, thanks for your explanation Charles-François. socket_listen-1.diff looks good to me.

[python-dev]

New changeset 09371221e59d by Charles-François Natali in branch 'default':
Issue bpo-21455: Add a default backlog to socket.listen().
http://hg.python.org/cpython/rev/09371221e59d

[neologix]

Committed, thanks!