The way it should be is hybrid. For machine management purposes, each machine should be named for the datacenter and rack it's in. For example, dcaa11 is the eleventh machine in rack aa in datacenter dc. But there's also the issue of how to know which machine is running what. For that, we need to remember there can be more than one name for the same thing. The mysql machine should register itself with DNS when it starts up as 0.mysql.datacenter_name.prod.yourdomain.com. The second one should be 1.mysql, etc. Now you have a system where memnonic names and manageable names can coexist.
It's not hard to have DNS CNAME and TXT records for machines, so that you can have both a naming theme for certain classes of servers, as well as descriptive names that indicate function, location, etc.
The way I handle the problem is pretty simple. Functionality gets a subnet 'name' under the root, and there are TXT records and aliases within that to handle description. So, if you have a datacenter in Sacramento, your DB and web machines would be named and alised:
For ambiguous names (like 'mercury' existing in both { planets } and { elements }), only one copy of that name can exist on the network. The people who work with that machine will know it's theirs, and for everybody else, there's CNAME and TXT records for when they're not sure.
Named for the rack it's in??? Oh my god that's stupid! So you seriously thing it's a good idea to have to rename the server if you shuffle around some data center space!?!?
