Hi Willy, Liang:
Please forgive me for my poor English.
In my experience of using HA Proxy 1.4.8 in 2009, there are some 50x
errors when Nginx is used to be backend server.
The errors can take over 0.5% of total requests.
However, the error rate can be reduced to 0.1% when increasing the size
of 'tune.bufsize'.
Moreover, even though 'tune.bufsize' is not changed, this error rate can
also be reduced to 0.02% when we use Apache instead of Nginx.
Maybe 'epoll' is related to 'select'?
PS:
The above is for your reference only.
The static content has been separated and then Apache can only handle
PHP requests.
Regards,
Yuxans Yao
于 2012年11月07日 17:02, Willy Tarreau 写道:
> Hi Liang,
>
> On Wed, Nov 07, 2012 at 04:03:47PM +0800, ?????? wrote:
>> Hi Willy/Yuxans,
>>
>> Appreciate for your quick reply. Another question is our haproxy log shows
>> many 503 errors. Could you please take a look the attached configuration and
>> log files and give us some suggestions?
> There's something strange there. All of them show the same pattern :
>
> Aug 20 09:50:04 cache195 haproxy[5722]: 113.134.48.103:60269
> [20/Aug/2012:09:50:01.771] www-pool optimus_deliver_server/web116
> 2237/0/-1/-1/2237 503 212 - - CC-- 22272/22249/76/11/0 0/0 {Mozilla/4.0
> (compatible; MSIE 6.0; Windows NT 5.1; SV1)} "GET
> /static/userback163.gif?jsType=6&meStationId=1&clUrl=http%3A%2F%2Fnews.163.com%2F12%2F0820%2F01%2F89AK98C600011229.html&referUrl=http%3A%2F%2Fwww.163.com%2F
> HTTP/1.1"
>
> An apparent immediate client close while connecting to the server. Having
> one such from time to time would not surprize me, but here you have a lot.
> I suspect that haproxy is improperly logging the origin of the issue ('C'
> for client close while it might be another reason). The queues are empty,
> there was no connect retry, so I think that the connection was not even
> attempted.
>
> Maybe the request came in immediately followed by a FIN packet. It could
> be a proxy causing this to many of your users. Could you please try to
> remove "option abortonclose" and see if the problem persists ? You might
> be observing the side effect of an external misbehaviour.
>
> BTW, what version is this ?
>
> Willy
>