Subject: Re: [vserver] cluster file system and vserver
From: "Edward Capriolo" <edlinuxguru@gmail.com>
Date: Tue, 15 Jul 2008 17:49:47 -0400

On Tue, Jul 15, 2008 at 5:25 PM, Martin Fick <mogulguy@yahoo.com> wrote:
> --- On Tue, 7/15/08, Edward Capriolo <edlinuxguru@gmail.com> wrote:
>
>> >
>> I have setup glusterfs between vservers. It seems to work
>> well under very limited testing. I do not understand why you would
>> need to run gluster ontop of  DRBD, but you probably can prove me
>> wrong. I am sure there are good reasons to :)
>
> Simple, glusterfs AFR does not prevent split brain operation.  AFR is well suited
for clusters where nodes rarely go down and can be expected to come back up before any
others go down and where network outages are very unlikely.  Glusterfs AFR really is
a very simplistic node replication algorithm that is meant more for providing high availability
but it does not yet ensure consistent data.  If you have lots of disks and can do AFR
with several copies (more than 2) than it might be safer, but that might also kill you
performance.  Search or ask on the glusterfs mailing list if you want to know more why
AFR might not be a good idea for this.
>
> -Martin
>
>
>
>
>

Hey don't sigh over java. I notice that usually people will run any
type of interpreted or scripting language but you say JAVA and they
are like 'Java No Way!'. It is a debate for another place but I always
say 'Don't knock it until you tried it'.

As it applies to this circumstance, with Hadoop and JAVA I was able to
get a running solution up in about 2 hours after I followed a simple
tutorial. It took A LOT of digging to get gluster working, probably
about 3 days of research on my part poking around in IRC rooms
Luster,Openvz,vserver, piecing cryptic mailserver posts together. It
took me two days of research before I figured out LUSTER was just not
going to work with vserver. I had a nightmare experience or two with
DRBD. For kicks 'google linux-ha, drbd' and my email to see me trying
to setup a two node DRBD cluster and running into every problem but
the kitchen sink, disk formatting, kernel updates, lockups.

You can not compare the three. HADOOP is not a general purpose
filesystem, more like filesystem+grid computing. Yahoo is using it and
apache foundation is developing it, so like before you pass it up
think on that.

GlusterFS seems really impressive. I wish I had more time for it.
Seems like you can do just about anything with it. Network mirrors/
network stripes/ file replication. I never benchmarked it. Once I
figured out the install it seems to be pretty easy.

DRBD has its place as a more ACTIVE/PASSIVE ACTIVE/ACTIVE fail over
solution with very low level replication. It is not going to scale
past three servers.