[dns-operations] Load balancing DNS queries across many machines

[Graeme Fowler]

On Mon, 2009-07-20 at 21:08 -0700, Joe Abley wrote:
> I know of prominent DNS server clusters built out of individual hosts  
> connected to a dedicated load-balancer device, like an f5 BigIP or a  
> Foundry ServerIron. Those setups provide stateful load-sharing with a  
> variety of origin server selection heuristics. Some of them do address  
> family translation, which somebody mentioned.

I did exactly this some years ago for a large UK hosting/domain
registration company using a cluster of Linux servers which acted as
both load balancer (using LVS) and DNS server (using BIND), the design
ethos being "scale sideways".

We made use of what was then an external patch to BIND, DLZ (Dynamically
Loadable Zones, which is now in the main ISC release train) to hook the
nameservers back to a slaved instance of a SQL database. We went from
between 12-24 hour startup time for a BIND instance with somewhere in
the region of a million zones to a matter of seconds, and were able to
reflect dataset changes and zone addition/removal almost instantly (TTLs
aside).

I think, as I no longer work there, that the system is now frontended by
some vendor hardware load balancing kit.

> For others who don't have UCB's experience with build nameserver  
> clusters using anycast, I have written that experience up several  
> times based (partly or entirely) on work we did at ISC, and maybe some  
> of those contain some useful pointers.

Thanks - useful.

Graeme