Antw: Re: Speaking about pure fail over - what about pen itself ?

Hi,

we using an existing Veritas Cluster (Suse Linux , 3 Nodes) for our pen installation. I know Veritas is very expensive, but the main Ideas for a cluster are very simple.
1. use shareable medium for the data. (i.e. SAN/NAS or NFS (but don't forget think about "fail over for the NFS-Server ...))
2. use a virtuall IP-adress for the Interface (you can specify mor then one IP-conifguration for a real Interface)

If one node fails - mount the Data volume on the other, an enable the virtual IP-Adress on this host.
Cluster software give you a nice GUI to configure and admin this and has a watchdog process running for automatic switching and so on. Some cool features but this is not the core.

For a not "mission critical and 24/7 H" installation you can simply use another host to check the availability and if this fails start the package (Data Vol. + IP-Addr) on the other clusternode.
You can have a problem if the old node is only partial down (eg.still answers for ICMP Requests) then you can create a difficult Situation.
But that is only a challenge for the availabilitiy check/start script ...

I don't know if there ist a opensource cluster solution, i don't search for it until now.

Wilhard

Wilhard Speck

DIN ITS GmbH
Burggrafenstr.6
10787 Berlin

Tel +49 30 26 01 28 89
Fax +49 30 26 01 42 889
E-mail wilhard.speck_at_dinits.de

>>> Ulric Eriksson <ulric_at_siag.nu> 03/27/06 8:35 pm >>>

On Mon, 27 Mar 2006 fernando_at_fernandoibanez.com wrote:

>> I'm only looking at two boxes, except that they will also balance other
>> services located elsewhere. Should the separate interfaces that pen
>> listens to (not the shared one), be real or can they be virtual ?
>> Also, if going with independent LB boxes (such as the original example),
>> do they need to be powerful boxes ? TIA...
>
> A simple and cheap way to do failover for pen would be using
> failover at the DNS level - with a service like
> Zone Edit (http://zoneedit.com/)
>
> ZoneEdit can test if the server is up and if it is not, it can point
> your domain to a different IP possibly on a different network.

I'm positive that can't work. The old address will be cached on
several levels throughout the Internet and there is no way to
clear all those caches from a central point. If they charge money
for the service, it's a hoax. Otherwise it's just silly.

Multicast in Linux isn't very reliable (or I'm simply
incompetent), så the vrrp solution isn't as good as it seems. The
easiest solution is to point the domain to two addresses, one for
each pen instance.

Actually, the easiest solution is to install pen on a single
server and forget about it, at least as long as the hardware is
reliable.

Ulric
Received on Tue Mar 28 2006 - 11:49:44 CEST