On Mon 2016-04-11 08:52:44 -0400, Shane Kerr <[email protected]> wrote:
> People who know more can correct me if I'm wrong, but my understanding
> is that this sort of RRset rotation is intended to work around clients
> that would always pick use first RR from an RRset. Back in the day,
> that meant that one gopher server would take all the load. ;)
>
> Obviously a DNS server would prefer not to have to do the work of
> ordering queries. If you use cyclic order, then you have to maintain
> state about the last answer. If you randomize it, then you need some
> source of entropy.

randomized doesn't require a lot of entropy -- it just requires some
small chunk at system startup to seed a CSRPNG that you can use
indefinitely.

> My own recommendation would be that a fixed order be used in all
> on-the-wire communication, and that we rely on client-side libraries
> to either understand that multiple answers may appear to a question or
> at alternatively to randomize the one answer they use. (Note that this
> is independent of the privacy concerns, as it results in simpler code
> and more deterministic behavior all around.) :)

I do like the simplicity and clarity of your proposal, but relying on
client-side libraries to be fixed seems like a non-starter,
unfortunately.  This is the classic IETF problem for any protocol with a
deployed base; the folks who have the incentive and capacity to fix some
endpoints aren't necessarily the folks who control the "right" endpoints
to fix :/

   --dkg

_______________________________________________
dns-privacy mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/dns-privacy

Reply via email to