Hi Ludovico, On Fri, Apr 17, 2015 at 08:24:43PM -0700, Ludovico Cavedon wrote: > Hi, > > I am trying to find a solution to the following issue. > > I have a client A that sends hundreds of HTTP requests per second to > server B running ha-proxy 1.5.3. > Server B/haproxy forwards them to server C. > > Some of these request are long-polling: server C will receive the > request and hang for a very long time (even hours potentially). > Client A may decide to give up on long-polling request, close the > connection and forget about it. > However, haproxy will keep the connection state half-open, assuming > the client is still there waiting for data. > > On client A the connection state will go to FIN_WAIT2, and will be > forgotten after 60 seconds. > > After a while, client A will happen reuse the source port for a new > connection, and send a SYN packet. On server B, however, the > connection is still there in CLOSE_WAIT. > > I am trying to find a way to avoid this from happening. Ideally I > would like haproxy to close the connection completely if the client > closes its side. > "abortonclose" seemed the right option, but it looks like it works > only if the FIN arrives before the connection to server C has not been > established yet. > > The statement from [1] sounded promising: > > In 1.5 we have even improved that a bit further for users of long-polling > > requests. When "option abortonclose" is set, if the client closes the send > > side, then haproxy forwards this closing event to the server. > > however it does not seem to happen in my case. > > What am I doing wrong? Does it have to do with the fact that > connection to server C are reused/persistent? > Is there maybe a way to tell haproxy to close the client connection if > there is no keepalive?
You're not necessarily doing anything wrong, it's fairly possible there's a bug or an inconsistency between multiple options. I'll have to retry here with your config, because the statement you quoted indeed indicates that we should get a better behaviour. I won't do it immediately because I'm still stuck fixing the things I recently broke before merging that into 1.6. In the mean time I would appreciate it if you could retry with 1.5.11 to verify if you see the close being forwarded or not, because maybe it's due to a bug that was fixed since 1.5.3. Thanks, Willy
