Re: strange logging line/statistics weirdness - plus 'timeout queue' weirdness?

On Wed, Sep 3, 2008 at 4:23 AM, Willy Tarreau <w#1wt.eu> wrote:
> Hi Jeffrey,
>
> On Wed, Aug 27, 2008 at 12:44:07PM +0800, Jeffrey 'jf' Lim wrote:
>> hi folks, I've got a situation where I have maxconn set on a frontend,
>> and it seems that when this happens and the maxconn is hit, a
>> connection can stay stuck indefinitely in some queue until a
>> connection slot is freed up? Initially, I was thinking that perhaps
>> the 'timeout queue' would cover this, but apparently not. I've got
>> 'timeout queue 5s' in the defaults section, but my connection will
>> stay stuck forever, until a connection slot is freed up.
>>
>> In addition to 'timeout queue' not being effective (yes, I know
>> 'timeout queue' is not applicable for a frontend, but then... should
>> we be able to set a timeout for this still?)
>
> There is no queue in frontends. They simply stop accepting requests when
> their maxconn is reached. However, it is possible that you have been waiting
> in the queue on the backend, but the log should reflect that.
>

gotcha. Thanks for clarifying that.

>> The strange log line I see (questionable value is Tt - since I
>> obviously spent a minute and a half waiting for a reply)
>>
>> (extra cruft cut) [27/Aug/2008:12:30:53.081] front-xxx
>> back-xxx/server1 0/0/0/0/17 200 362 - - --VN 0/0/0/0/0 0/0 "GET /
>> HTTP/1.1"
>>
>> Is this correct? Tt = 0?
>
> The total time here is 17 ms and the log indicate that you spent no time
> in the queue nor anywhere BTW (everything ran instantly). While it's possible,
> it seems very unlikely, especially considering the time you waited. What
> version is this ? Wouldn't this be one before 1.3.14.6 or 1.3.15.2 ? In
> earlier versions, the queue management was buggy but IMHO it would not
> end up doing nasty things like this.
>

:) yeah, i would have considered that. Unfortunately, the version i'm talking about here is 1.3.15.2 ...

> Something you can check is the difference in time between what syslog days
> and what haproxy says. syslogd indicates the time its received the log,
> which generally corresponds to the end of session. haproxy logs the time
> it received the connection. Most often, the difference between both is
> approximately equal to the total time (more or less rounding errors due
> to syslog having a granularity of one second).
>

hm? I'll have to take a look at it later. Tt = 0 still doesnt look right, though - i definitely spent time waiting for something to happen...

-jf Received on 2008/09/03 05:56