DNS round robin for WWW

[Kevin Darcy]

Mark.Andrews at nominum.com wrote:

> >
> > Isn't that working for you?
> >
> > The only thing I'd add is that "round-robin" is a bit of a misnomer for
> > later versions of BIND: the default "cyclic" sort order of BIND picks the
> > first address of the answer at random; for a name with only 2 A records,
> > there is functionally no difference between "cyclic" and "random". The
> > official line from ISC is that the consequences of true round-robin,
> > "cyclic" and "random" are identical over the long haul. But this glosses
>
>   "cyclic" and "round-robin"
>
> > over the fact that randomness in any form can produce temporary "spikes".
> >
> >
> > - Kevin
> >
>
>         Kevin, feel free to send patches, but please do not grow the size
>         of a databuf.  It is possible to recover two bits from d_rcode as
>         there are only three states this can be in and 4 bits are used.
>         You only need one bit to mark where the last cycle started.

Struct databuf, what's that? :-)

Seriously, I think it's probably a waste to add this to BIND 8. I was thinking
the positional information might fit nicely in the dns_rdataset structure of
BIND 9, though. I'm also thinking that "permuted" order
(ABC/BCA/CAB/ACB/BAC/CBA) is probably more useful than true round-robin anyway
-- that way each entry gets equal time even in a failover scenario. I just
need to come up with an efficient algorithm for calculating the permutations
incrementally.


- Kevin