[dns-operations] TCP anycast was Re: DNS query logging

[Rick Jones]

On Thu, 2011-03-17 at 14:44 -0400, Edward Lewis wrote:
> At 10:49 -0700 3/17/11, Rick Jones wrote:
> >So, what is the frequency of "shifts" for anycast anyway?  The NTP folks
> >(as expressed in posts to comp.protocols.time.ntp) seem to not be
> >terribly fond of it even for UDP-based time synchronization.  Those
> >folks are generally pretty detail-oriented so I'd not necessarily
> >suspect they would be easily given to fear uncertainty and doubt, but
> >cannot rule it out entirely.
> >
> >rick jones
> >
> >it is merely coincidental anecdote - the NTP daemon on my workstation at
> >work routinely "discards" time from the internal, anycasted NTP service
> >in favor of servers that are even of higher stratum.  Could be for any
> >number of reasons I suppose besides anycast (perhaps OS choice on the
> >servers) but it does also show the highest jitter of the five servers I
> >have configured to poll for time.
> 
> At the risk of saying the obvious, NTP is more time-sensitive than 
> DNS.  NTP is going to prefer to go to a server where the round trip 
> time is very predictable, so it knows what time delta to add to what 
> it's told.  So I'd say it's fair that NTP is very picky about network 
> latency "rubber banding."  (Keep in mind - it's been a long time 
> since I was into NTP though.)

If there is jitter in the RTT with an anycast NTP server it suggests the
"routing" is changing right?  At least once over an interval of either
64, 128, 256... 1024 seconds.  So, while NTP is more time-sensitive than
DNS, its sensitivities underscore that generic TCP with anycast isn't a
slam dunk in the positive.  At least that is my current interpretation.

> There have been studies done about client "flip flopping" between 
> anycast instances.  The last one I recall (2004 or so) is here
> 
> http://www.nanog.org/meetings/nanog32/presentations/kosters.pdf
> 
> Slide 27 has something on that.
> 
> (It says this though: "DO NOT RUN Anycast with Stateful Transport")
> 
> See also this follow up
> 
> http://www.nanog.org/meetings/nanog39/presentations/larson.pdf
> 
> Courrent "shifts" - low enough that we don't think about it.

Does a shift manifest itself as a failed query, and how is that
generally dealt with?  For example will it cause an EAI_AGAIN or a
different error?

rick