From: Cami (no email)
Date: Thu Jun 03 2004 - 02:24:48 EDT
>>The only thing that is in standy-mode is the mailhosts.. The pop3/imap
>>proxies, aswell as spam layer, aswell as incoming MX layer all user
>>load balancing..
>
> How do you tell a standard POP/IMAP client to what server it should login
> in order to get the mail ? In standard I mean Mozilla, Outlook, Eudora,
> PC-Pine and Pine, Kmail, etc.
All clients simply grab their mail from pop3.mweb.co.za..
That is a load balanced serverfarm which automagically
(transparently) bridges the client's connection to the
mailhost they live on..
>> Where do you get that idea from? Think about what happens
>> if that single mail spool filesystem becomes corrupted..
>
> That is why NetApps and EMCs have snapshots and snapmirror and BCV and
> timefinder. That is also why you pay $$$ for these systems. For the 8
> years that I've worked with Netapp (starting with F200 series back when
> they just started off), I've never had even 1 file corruption on a system
> that holds something like 3.5 million files or more...
To be honest, i've never played with snapshots etc.. How much time
is required to recover from such a corruption? (How long does it
take to recover a snapshot..?)
>>mailhosts in an active/passive configuration allows
>>you to perform drastic changes and maintence without
>>requiring hours of downtime..
>
> If you have servers in a load balancing scheme, you can take any of them
> down, and the loadbalancer will just remove it from the pool. No service
> interruptions.
True.. This is what i do for all the other layers except the mailhosts..
Having all the mailboxes sitting in 1 spool just makes me very nervous
when things go wrong.. (and they always will at some point in time)
> Yes, but it will need to know "which" node failed, and it then would need
> to know what mailstore belongs to that node, and take over for that. It is
That is preconfigured before hand.. The testing i did showed it worked
as intended..
> definetly not trivial to configure. In general, Linux clustering is less
> than "stable and mature", and also, I am not aware (that may be my own
> fault) of any proper Linux Clustering package that is manageable by less
> than the ultimate hardcore sysadmins (which cost alot). However, you may
> not be using Linux at all.
I am using Linux.. I really wouldn't consider the clustering software
we're using to be difficult at all.. Once you have the concepts, its
clear sailing from there..
> In general (but I am not speaking of 350,000 users), using NFS based
> mailstore and a technically sound design works well, is cost effective in
> both hardware and man power, and is scalable (up to a limit of course).
Only reason i'm anti-NFS is that we were using it before and it things
would really slow down as we got bigger, especially during peak times..
Cami
|
|
|