Use nodelay socket option for couchdb httpd server

Some python libraries that provide api for couchdb access will take much longer than using plain @requests@ lib:

# put 100 docs using requests, couchdb and couchdbreq libs
# couchdb server doesn't have nodelay option set in this run
requests
  finished in 0.578608036041 secs
  mean: 0.005786
couchdb
  finished in 4.32773804665 secs
  mean: 0.043277
couchdbreq
  finished in 4.60177111626 secs
  mean: 0.046018

Note that using plain @requests@ is in this case almost 10 times faster than using the other libs. It is still unclear why this is the case.

If you do add the @nodelay@ flag in @/etc/couchdb/default.ini@:

[httpd]
(...)
socket_options = [(...), {nodelay, true}]

Then it's much faster:

# put 100 docs using requests, couchdb and couchdbreq libs
# couchdb server *does* have nodelay option set in this run
requests
  finished in 0.586463928223 secs
  mean: 0.005865
couchdb
  finished in 0.275610923767 secs
  mean: 0.002756
couchdbreq
  finished in 0.53059887886 secs
  mean: 0.005306

Note that it doesn't make much a difference when using plain @requests@ lib, but it makes a huge difference when using the other libraries. We are currently using the @couchdb@ library (the second one in the measurements), so this actually makes a big difference for our case.

See: https://stackoverflow.com/questions/13807706/couchdbkit-10x-slower-than-requests

In my tests running soledad client, server and couchdb in localhost, i have the following results:

# upload 100 documents, without nodelay socket option set in couchdb
real    0m31.011s
# upload 100 documents, with nodelay socket option set in couchdb
real    0m12.493s

That is a factor of 2.5 gain in speed. Note, however, that i'm performing tests in localhost connections so when using a real network the gain might be lower.

(from redmine: created on 2016-07-21, closed on 2016-08-03, relates #6154)