Configuring different TTLs in multiple RRs for the same domain name, TYPE, and CLASS

[Barry Margolin]

In article <mailman.454.1458858570.73610.bind-users at lists.isc.org>,
 Dave Warren <davew at hireahit.com> wrote:

> On 2016-03-24 15:20, Tony Finch wrote:
> > Dave Warren <davew at hireahit.com> wrote:
> >> On 2016-03-24 09:46, Ray Bellis wrote:
> >>> On 24/03/2016 16:41, Tony Finch wrote:
> >>>
> >>>>> When I changed our TTLs from 24h to 1h last year, it didn't have a 
> >>>>> visible
> >>>>> effect on authoritative server query load, much to my surprise.
> >>> I'm not that surprised - there's definitely not a linear correlation
> >>> between the TTL of an RRset and how frequently it's queried.
> >>>
> >>> Unless your TTL is very short, forced expulsion from cache (due to
> >>> cache-size limits) would cause many clients to re-query for a record far
> >>> more frequently than once-per-TTL.
> >> Has anyone ever done any evaluation on this? For average resolvers, what
> >> is the longest TTL that has any utility?
> > There was a great paper published 15 years ago describing a study of DNS
> > cache effectiveness at MIT. http://nms.csail.mit.edu/projects/dns/
> >
> > It concluded (amongst other things) that NS records (and associated
> > address records) are really important, but leaf records that users ask for
> > don't matter so much. (Based on cache hits before TTL expiry, IIRC.)
> >
> > I don't know of a similar study performed more recently.
> 
> The internet was a very different place 15 years ago, in particular, 
> this was before every Windows client machine had it's own DNS cache 
> service and largely before today's connected mobile devices were a thing.

But it was also before the widespread use of CDNs (Akamai was founded 
only 3 years earlier). These days, the most heavily used web sites use 
CDNs, which make heavy use of short TTLs for the leaf CNAME and A 
records.

-- 
Barry Margolin
Arlington, MA