Cisco Distributed Director

[Kevin Darcy]

IMHO, if one's load-balancing needs are so rigorous that they require sub-second
volatility, then perhaps one shouldn't be using a DNS-based load-balancing
mechanism at all. DNS is already burdened enough without having scads of TTL 0
RR's forcing servers to work overtime. I see that Cisco also sells
LocalDirector, which load-balances transparently without relying on DNS. (Note
that I've never used either product, so don't take this as an endorsement).


- Kevin

Werner Wiethege wrote:

> On Fri, 16 Jun 2000, Barry Margolin wrote:
>
> > Since your TTL is 0 seconds, the caching nameserver shouldn't actually
> > cache the record.  It should forward it to the client machine, and then
> > discard it.  The next time a client tries to look it up, the caching server
> > should go back to the DD.  If you have even load balancing configured on
> > the DD, it should alternate which address it gives out each time.  AFAIK,
> > the fact that all the queries are coming from the same nameserver shouldn't
> > matter.
>
> The BIND 8.2.3T5B code checks for staleness of a record with the
> following comparison (in function stale in ns_eq.c):
>                   dp->d_ttl >= (u_int32_t)tt.tv_sec
> where d_ttl is the time when a record expires and tv_sec the current
> time. I assume older versions have the same kind of check.
> Considering records to be valid as long as the times in seconds are
> the same can explain the behaviour that a caching nameserver keeps
> returning the same value for almost a second when the TTL is 0.
> Tests done by Pete Taylor where he left out the equal sign have
> shown the expected behaviour for TTL 0.
> Does TTL have inclusive or exclusive interpretation in the standards
> or is it an implementation issue?
>
> Werner