Hi Willy, Cyril It works very well.
Thanks for your helps. Regards, Seri -----Original Message----- From: "Willy Tarreau"<w...@1wt.eu> To: "Seri"<seri0...@naver.com>; Cc: "Cyril Bonté"<cyril.bo...@free.fr>; "Lukas Tribus"<luky...@hotmail.com>; "HAProxy"<haproxy@formilux.org>; Sent: 2014-05-01 (목) 03:12:42 Subject: Re: in uri balance, http-keep-alive broken On Wed, Apr 30, 2014 at 12:02:29PM +0200, Willy Tarreau wrote: > On Wed, Apr 30, 2014 at 11:49:20AM +0200, Willy Tarreau wrote: > > On Wed, Apr 30, 2014 at 06:21:40PM +0900, Seri wrote: > > > Hi, > > > > > > This issue is reproduced in this environment( frontend(http-keep-alive), > > > backend(http-server-close) ). > > > In this environment(frontend(http-server-close), > > > backend(http-server-close)), this works well. > > > > Even then, still no luck for me :-( > > > > Could you take the network capture that Lukas requested ? > > OK with Cyril's off-list help, I now can also reproduce it, so don't > waste your time with the traces. It seems extremely timing-dependant! And now here comes the fix : http://git.1wt.eu/web?p=haproxy.git;a=commitdiff;h=644c101e2ddf9 It was really nasty, and used to work by chance (timing-dependant) even if the chance was always on my side during all the tests. I did something stupid, make a request analyser's progress possibly depend on the response analyser. It's fixed now. I'll push other updates, we don't need anymore to hold on the request parsing if the LB algorithm only needs the headers, so in most cases we won't have to do that except for POST requests. Also, I found that we don't use the check_post parameter anymore, we always check the body when the url_param is not found in the URL and the request is a POST. Another thing that can probably be removed to simplify configurations... Please test the fix. Emeric has some minor updates pending for the server-side SSL, and I'm about to fix the bind-process issue for listeners (at least I hope so) so I intend to issue dev25 and the only missing thing for -final will be the changes on the agent-check (part of which was already started long ago). Willy