Primitive failover. (Not load balancing)

[Barry Margolin]

In article <bv6qr9$ehb$1 at sf1.isc.org>,
 "Aaron Roberts" <aaron at domicilium.com> wrote:

> I figured that a TTL of '0' might be misinterpreted by caches and =
> decided
> that 60s would probably be about right but I simply don't have the =
> experience
> to make a sound judgement myself..  So my question is: What would be the
> lowest 'safe' TTL that I could give the A record?

Some old versions of BIND did indeed misinterpret TTL 0 when recursing 
and caching; they seemed to expire the record before sending it back to 
the client!  But anything non-zero should work OK.

However, I think that making it less than a minute is likely to be 
overkill, especially if you expect to be making the changes manually.  
If the failover isn't so critical that you need to automate it, how can 
you justify needing the caches to clear even faster?  I.e. if it takes 
you 10-15 minutes to update the master, a 1-minute TTL means that the 
total downtime is 11-16 minutes; versus 10-15 minutes with a 1-second 
TTL.  Is the difference between them really significant?

-- 
Barry Margolin, barmar at alum.mit.edu
Arlington, MA