[wplug] Thoughts on a Storage System

[Michael Semcheski]

On Tue, May 19, 2009 at 3:26 PM, Michael Skowvron
<michaels at penguincentral.org> wrote:
> Why are you looking for a distributed data solution? I'm curious why a traditional or
> monolithic shared or clustered storage system does not fit your needs.

High volume non-distributed systems tend to require a big investment
in hardware upfront.  I would much rather buy relatively cheap
hardware as we need it.  If you buy something big and expensive,
you're paying for tomorrow's technology today.  I'd rather buy
tomorrow's technology the day after when its on sale.  That's the
general principal.  I'm open to ideas, but things like Lustre or GFS
aren't technologies I have experience with.  I have been meaning to
read up on Lustre to see how it would fit the bill...

So far, we've been using 1U servers with 4 SATA disks in a RAID5 or 2U
servers with 8 SATA disks in RAID6.  We fill it up and buy another
server.  Wash, rinse, repeat.  It works, but its not efficient.  Data
tends to be clumped together.  So, the new server gets all the new
data.  When you pull out yesterday's data, its all on the same server.
 That's slow.  Also, for about half the time that server's in use,
half of its disk is totally unused.  It would be better if things were
more balanced.

Also, to some extent we'd be willing to sacrifice a bit of space for
redundancy.  Obviously, it hurts the bottom line, but we still come
out well ahead of buying a SAN or something really nifty.