Alan McKay wrote:
> My main areas of interest are :
> - DB clustering (PostgreSQL) - yeah, we're looking at commercial stuff
> and skytools
No DB clustering here, though if I had to pick I'd probably go for
Oracle RAC. MySQL clustering doesn't seem good, Postgres sounds
interesting, their EnterpriseDB grid stuff perhaps, haven't looked
too much into it.
> - web server clustering - Apache on CentOS
The web servers I manage serve about 2 billion requests a day on
CentOS 5.2, no clustering, we use load balancing with commercial
load balancers, currently F5 LTM, though will be evaulating Citrix
Netscaler soon. The servers run java/tomcat from original sources,
so despite having an older version of the OS there really isn't
much security risk. Any network-based attacks are thwarted by the
load balancers (such as the recent Apache slow DOS attack).
F5 runs all of their gear on CentOS as well, at least everything
that runs TMOS (which covers GTM, LTM, and a few others as well
I think). It is a stripped down version.
Don't let that make you think you can grab their software and
install it on your own server though, their licensing will
prevent that from happening.
> - storage clustering
Using commercial solutions here too, our NFS cluster from Exanet
does run on top of CentOS 4.4, though they use their own hardware
(IBM X3650) to make it as supportable/stable as possible. These
are NAS head units only they rely on another storage system
in the back end for the raw disk space to serve to clients.
At my previous company I was planning on implementing a GFS cluster
on top of RHEL and VMware with a fiber channel/ raw device map
back end for NFS. Though didn't have time or budget to do it
before I left. Such a solution for my current company would
be torn to shreds as it doesn't scale nearly as well.
Our back end storage from 3PAR runs on top of Debian, you'd never
know it was Debian unless you telnetted to it on port 22 to see
the Debian SSH banner(I like to poke around).
Clustering is a really complex thing to get right, it can often
cause more problems than it would otherwise prevent. Even some
high end clustering is really poor. A couple of jobs ago I had
to use BEA Weblogic application clustering for a massive J2EE
app. Ran us roughly $10 or was it $20k per CPU. We had major,
major outages with that thing. Most of the time we(and BEA)
were able to trace it to the weblogic cluster itself.
So think long and hard about what your trying to accomplish,
and if there is another way to get there without relying on
clustering. When I say clustering I mean pretty tight integration
between the systems in the cluster, where if one box can go
whacky it can take the rest of the cluster with it.
nate