Hi all, Just for been sure before push the fix are we agree on the next statements:
1) Increase socket buffers (probably, strong yes) According the latest comment on COUCHDB-1986: https://issues.apache.org/jira/browse/COUCHDB-1986?focusedCommentId=13955355#comment-13955355 32KiB may be not enough for better performance. Uncommenting the config line to increase buffers value up to 256KiB sounds more reasonable in this particular case. Also, this fix is very easy to apply on older releases that are also affected to the same issue on the same OS's. 2) Disable Nagle's algorithm This is unrelated to the our issue, but if we're going to change the socket options, why not to improve communication for small data chunks in the same time? Should be helpful for replication too. -- ,,,^..^,,, On Sun, Mar 30, 2014 at 12:20 AM, Jan Lehnardt <[email protected]> wrote: > > On 28 Mar 2014, at 14:39 , Dave Cottlehuber <[email protected]> wrote: > >> Hi folks, >> >> In this week’s IRC meeting we discussed what’s up with 1.6.0. While the >> decision >> rests with the Release Manager djc@ it’s worth discussing. >> >> # Blockers >> >> We have 1 blocker, https://issues.apache.org/jira/browse/COUCHDB-1986 and the >> fix is simply setting the tcp receive buffer size to 32786, or 4x the >> previous >> value. >> >> Exactly why this is a now a problem and wasn’t before is not clear, but in >> a couple of weeks of fiddling we’ve not come up with a better solution, and >> nobody’s found other issues as a result of it. >> >> In the meeting we put forward 3 options for how to include this fix in the >> source: >> >> 1. uncomment the equivalent line in default.ini >> >> socket_options = [{recbuf, 262144}, {sndbuf, 262144}, {nodelay, true}] >> >> 2. update the default value directly in /src/mochiweb/internal.hrl >> >> -define(RECBUF_SIZE, 8192 * 4). > > +1 to either one of 1. or 2. > > And to re-cutting 1.6.x from master. > > Best > Jan > -- > > >> >> 3. change the specific setting only for the replication tests >> >> This boils down to “have people been seeing this in the real world?” and the >> answer is yes, with pretty severe impact. >> >> Please pick one! >> >> # Patch 1.6.x branch or update to current master >> >> Personally I’m undecided on this, master has 180 further commits , esp >> Fauxton, and also a few other important ones. These could be cherry picked >> but is there any reason not to cut a new 1.6.x off current master instead? >> I’ve no real view of the actual work involved BTW, so please educate me. >> >> https://github.com/apache/couchdb/compare/1.6.x...master >> >> Comments? >> >> -- >> Dave Cottlehuber >> Sent from my PDP11 >> >> >> >
