Hi Rupert, Alexander,
On Mon, Sep 15, 2008 at 04:43:27PM -0700, Rupert Fiasco wrote:
> > Do you see the problem on the HAProxy status page, or in the Mongrel
> > "ps" status? It's my impression that HAProxy health checks ignore the
> > maxconn setting. So we frequently see this "ps" output:
> Precisely, we do not see it on the stats page
OK, this is very important. The "max" value you observed during the bug was what told me it *was* a bug, because that max is reliable since it's computed during the connect() call. The fact that it does not grow above the limit indicates to me that there is really never more than 1 active connection at a time.
> but via that mongrel
> plugin (the 2nd num in the ps output).. So it makes me think that
> haproxy thinks its only sending 1 request to that mongrel, yet the
> mongrel is getting more.
> > It's my impression that HAProxy health checks ignore the
> > maxconn setting.
Yes, currently haproxy cannot queue the health checks. It will be possible after the architecture rework. It will be necessary to always put them at the beginning of the queue though. Another annoying case is for people who have multiple haproxy boxes, because your mongrel can only process one's health-check at a time. This could result in some of the backup box seeing more check failures than the active one.
> Actually we are opting to just use just a TCP health check (versus
> sending a full HTTP HEAD/OPTIONS request to the backend).
> Or at least, please correct me if I am wrong. With this config:
> and in the absence of an "option httpchk GET /foo" it *will* use a TCP
> check vs an HTTP check, right?
Yes you're correct.
> If this is the case, then that HTTP request should never hit Mongrel
> and the ps output should indeed represent full HTTP requests and *not*
> haproxy backend health checks.
> Is this correct?
yes, that's exactly that. However, if you are counting the concurrent requests on the mongrel side, there is another situation which may report more than one request at a time : when the previous one timed out. In this case, haproxy aborts, closes the connection and reports a 504 to the client. But mongrel has no way to know this (because the close in TCP does not exist), and will still maintain this connection until it wants to respond. By that time, another connection can be sent (because there is no more connection on haproxy's side), hence resulting in more than one at a time observed on the mongrel side.
Also, I'm wondering : would there not be any solution to fork several mongrel processes on one port upon startup ? This limit of exactly 1 connection is really annoying, and if the server could call fork() 3 or 4 times instead of only one (akin to haproxy's nbproc), it would provide huge performance and reliability boosts. Just imagining that you cannot even telnet to the port to send a request during traffic would really frustrate me :-/
Willy Received on 2008/09/16 07:09
This archive was generated by hypermail 2.2.0 : 2008/09/16 07:15 CEST