Problems with varnish backend health

[Hugo Cisneiros (Eitch)]

On Thu, May 24, 2012 at 3:03 AM, Winkelmann, Thomas (Radio/Tele FFH -
Online) <t.winkelmann at ffh.de> wrote:
> Hello,
>
> we have a setup with two varnish servers (version 3.0.2) and two backend
> servers. This setup works fine, if we split the requests with an
> loadbalancer to the two varnish servers. Around 500req/s 20MBit/s bandwidth
> on each server.
>
> In Addition we have defined a backend probe which will tests the backend
> each 10s. This also works fine.
>
> Now, if we switch all requests to one varnish server, after some time
> (around 20 seconds) the backends will be marked as sick, but they are both
> up and answering all requests. Also without any backend probe defined in
> VCL, varnish gets in trouble serving content. Some months ago, we had no
> problems with this and served more than 400Mbit/s from one single server.

I don't really know exactly what could be, but sometimes when I get
backend errors (missconfiguration or overload on servers), and begin
debugging with varnishlog:

varnishlog -i Backend_health
( | "grep Backend_health" to show only the probes)

This usually gives me the return code from the probes and any problems
with it. If the server begins to return a different code than  HTTP
200, you can see exactly when this happens... It's a start :)

-- 
[]'s
Hugo
www.devin.com.br