Le 03/03/2017 à 15:44, Jon Simpson a écrit :
Thanks for your response. I’ve attached the requested configuration and
HAProxy trace logs below (TCPdump appeared to be much noisier). I have
an input file which reproduces the 503 case but not predictably, so I’ve
included one log which generated the correct 401 response from HAProxy
and one which returned a 503 response. The 503 response comes back every
time with larger requests than this. I’ve stripped down the
configuration to a minimum which reproduces the issue.
The behaviour only seems to occur with proxy-protocol enabled
(accept-proxy). I initially tested directly sending requests directly to
HAProxy and that works without the 503 responses. We run this HAProxy
configuration behind Amazon ELB - a TCP load balancer with proxy
protocol enabled, where the issue occurs (and it occurred in my test
environment once proxy protocol was enabled and connections went through
the TCP load balancer).
Hi Jon,
Thanks for these info. By checking how you named your trace filters, I
guess you use the HTTP compression. If I'm right, I think I found a bug
with the help of your logs.
The bug is triggered when your HTTP server replies before the end of the
request. I guess it always sends its 401 responses just after reading
requests headers. So it is timing issue from the HAProxy point of view.
This should work of course. If I'm right, the bug only appears when a
filter is used. So a temporary workaround, waiting my fix, is to disable
the HTTP compression.
I'll try to fix it very soon, I just need to talk with Willy to do it
the right way. I will keep you informed.
--
Christopher Faulet
