[dns-operations] Anycast vs. unicast NS

Michael Sinatra michael at rancid.berkeley.edu
Mon Mar 21 03:16:54 UTC 2011


On 03/18/11 09:57, Dobbins, Roland wrote:
>
> On Mar 18, 2011, at 11:28 PM, Joe Abley wrote:
>
>> While it's philosophically feasible that there is some
>> anycast-specific failure mode that might afflict both together,
>> which might lead people to think that the use of anycast is a
>> single point of failure, I'll observe that I have never heard of
>> such a failure to date.
>
> Concur.  The use of multiple anycast addresses pretty much takes care
> of any scenario I can come up with, at any rate.

Moreover, it is possible to engineer anycast so that if all of the 
anycast routing goop and healthchecks fail, it essentially falls back to 
being unicast.  Unicast can be made to be the worst-case-scenario for 
anycast.

This is why I find the debate between "anycast" and "unicast" to be a 
false dichotomy.  It's really just a debate about design and operational 
procedures.  Anycast can be engineered to avoid common-mode failures** 
as well as anycast+unicast.  We're just incorrectly superimposing 
technology over a design debate.

**Common-mode failure is a more apt concept for what we're discussing 
than single-point of failure, and it is a fairly common concept in the 
literature in complex systems, normal accidents, and nonlinear dynamics.

michael



More information about the dns-operations mailing list