I personally don't name machines by "what they are", for very good reason.
Plus, no one but the admin group uses the "real" machine names, also for
very good reason. Here's some illustrative examples...
If you call an HP K370 "k370.somedomain.com", and then upgrade it to a K570
by just adding a few more cpu cards, do you really want to still call it
K370.somedomain.com? If not, you have to retrain your user community and
this is a pain, and kinda defeats the whole idea of using meaningful names
so people don't need to know IP addresses. As a result, machine names
generally indicate what they are used for... so a machine that processes
orders might be "orders.somedomain.com". However, this can cause issues
unless the second point below is heeded...
If you give machines names that indicate "what they do" rather than "what
they are", you can then run into problems if you do clustering, or if a
machine goes down for a while (hardware issues), or more commonly, if for
load reasons you need to move an application to a different machine. Thus -
(IMHO), machines should have a "real" name that has nothing to do with what
it is or what it does. These real hostnames are typically what admin people
will use. Then, define CNAME's for the applications that the machine serves
up...this is what "users" use.
Here's a realworld example. A particular webserver is named "joe" (at this
place, the naming scheme is baseball related, so machines get names like
hank, cy, willie, mark, sandy, ozzie, etc.). It serves up images via the web
for bulk email that is sent out to subscribers that has links in it back to
images on "joe". However, users inside the company who put up these images
via sftp go to the cname for the machine which is "image". The benefit of
this is, if joe ever gets too heavily loaded, or goes down or whatever, I
can easily change the "image" cname to point to a different webserver here
and move the content to it (in the case of a web server cluster, the content
is already there). Thus, the admin group knows "joe" is down, it doesn't
respond at "joe". But the emails sent out reference
http://image.somedomain.com, and that is what the users inside the company
sftp to, so to them, everything still works as it did before while some
other machine is doing that task.
The big benefit is admins know the machines by their real name, as their job
focus is machine-centric. However, we can shuffle applications and services
around to other machines whilly-nilly, and we don't have to tell the users
anything because they still use the other name that is "service" specific.
After all, admins should be spending their time watching system resources
generally. If one machine starts getting sluggish, they need to be able to
move one or a few of the services that machine provides to a less loaded
machine - and they should be able to do this virtually instantaneously,
without the users having to do anything different or know a different
machine name. Heck, most users have the "machine name" stored in their
terminal emulation software, and dont even know what it's called". And half
of them don't know how to change it themselves.
This becomes especially important (and required) when you do clustering. The
most common example is web farms, but it applies equally to any service. If
you have 30 machines that ALL serve up content for the single URL
http://www.bigcorp.com you certainly can't name them all the same name. And
if one of the machines in the cluster dies, you also can't have a 1/30th
"hole" in your services. The above paradigm addresses this nicely.
To take it one step further, we use "wackamole" - a free unix program that
assures certain IP addresses are always available. It will shuffle the group
of addresses amongst multiple machines on the fly. If a machine goes down,
it moves it's ip addresses equally among the remaining machines in the
cluster. Most people accomplish "clustering" with round robin DNS, which is
great, but you run into the problem of if one machine in the round robin
list dies, anyone who gets the address of the down machine resolved gets
nothing. wackamole solves this problem very nicely. And as you might guess,
doing things like this doesn't work well if you are using machine names
unique to each machine.
Jay West
---
[This E-mail scanned for viruses by Declude Virus]
Received on Tue Oct 22 2002 - 10:31:00 BST