[sclug] Mirror disks across machines like distributed RAID?

[Dickon Hood]

On Wed, Jan 06, 2010 at 16:35:13 +0000, John Stumbles wrote:
: Dickon Hood wrote:
: >On Tue, Jan 05, 2010 at 18:42:35 +0000, Keith Edmunds wrote:

: >: In the case, for example, of an HA web server, putting the websites on a
: >: DRBD device seems straightforward and reliable. How would you propose
: >: doing that in the application layer (and saving the aforesaid bundle)?

: >HA webservers can usually be managed happily with the usual tricks:
: >clusters of machines, load balancers (if you insist), HSRP-alike failover,
: >virtual IP addresses and whatnot.  Or you can simply do what the BBC has
: >done for many years, and have a five-minute TTL on the DNS and ifconfig an
: >alias on another machine should one of the farm go down.  This sort of
: >thing isn't rocket science.

: Even with a cluster of webservers they have to have access to the same 
: data so surely you've still got to have some sort of HA disc system?

No point.  Replicate.  It's cheaper than a traditional HA disc system,
arguably less complex than a DRBD setup, and almost certainly more
performant: disc RTTs become important if you're cobbling a page together
from multiple objects..

: I know most of the data can be sourced from a master machine and
: replicated across local discs on the slave webservers but there's still
: got to be some flow of data back from the slaves from web forms, cookies
: and what-have-you.

This is where people go wrong: maintaining state server-side.  Make things
stateless on the server, and get the clients to maintain all that.  Solves
a bunch of problems: synchronising across the server farms, stale session
garbage collection, and sundry others.  Again, the BBC's model works quite
well here, although it obviously doesn't work in all cases.  Also, that
sort of processing shouldn't be done on the webserver anyway.  It should
be handled by an application server layer behind the webservers, for any
reasonably large site.

: And when you start talking about general purpose filestores with
: people's home directories on them there's surely no way you can engineer
: each and every application the users might run to get round failure of
: one disc?

No, at that point you need something reasonably available.  But you will
probably find that a cheap(ish) box exporting over NFS will work well for
that sort of thing, and a full-blown multi-master, multi-head SAN setup
overkill.

Don't get me wrong: there are applications for which HA disc is absolutely
required.  It's just it mostly isn't.

-- 
Dickon Hood

Due to digital rights management, my .sig is temporarily unavailable.
Normal service will be resumed as soon as possible.  We apologise for the
inconvenience in the meantime.

No virus was found in this outgoing message as I didn't bother looking.