On Thursday 10 July 2008 00:11:12 Willy Tarreau wrote:
> I've just found the cause of the problem you're experiencing. I could even
> partially reproduce it. It is caused by a bug introduced in 1.3.8 which is
> in fact quite hard to trigger : if the client is streaming data into
> haproxy faster than the server reads, and manages to always maintain data
> in the request buffer, the server timeout is not updated. Then when the
> client stops sending data after it has been streaming for at least as long
> as the server timeout, the timeout is seen as expired.
when I read your answer for the first time I was thinking that this is is very unlikely to be the case in my scenario because haproxy has a fast connection to the servers but requests are coming in via a (relatively) slow internet connections.
But reading your mail again, I noticed that requests might indeed arrive at high speed because there is an apache proxy sitting between the client and haproxy. Apache likely buffers the requests and then streams the data at once over to haproxy which is located on the same machine.
But I find it quite odd that neither me nor any of my colleagues cold trigger this situation, only remote users with a relatively slow Internet connections reported problems with large post requests.
-Rainer Received on 2008/07/11 14:12
This archive was generated by hypermail 2.2.0 : 2008/07/11 14:15 CEST