Part of the problem is the definition of the IPv6 anycast service. We
would really want to have a service that works well for UDP as well as
TCP; using an anycast address as a TCP source is problematic; in fact,
even using an anycast address as an UDP source generates interesting
failure cases if the query packet happens to be replicated and the
response happen to come from two servers.
In the short term, we could indeed just use the current IPv4 solution.
In fact, you can probably weasel that into the current IPv6 framework,
explaining that the address is in fact a unicast address to a
"distributed interface" to a "distributed server."
In the mid term, we would want to study a "real" solution. This should
probably be based on some kind of "redirect" scenario: try to reach the
"logical address X:X:X:X", and receive a redirect to a physical
implementation; that could work with TCP as well as UDP, and be robust.
-- Christian Huitema
winmail.dat
