BGP and anycast

  The anycast hack is pretty well understood, but it has some
serious limitations. As both drafts note, the same AS must be used to
announce the route (either because a single administrative entity
truly controls all instances, or in a shared-control environment).
You also must be wary of deploying services on the shared unicast
address which require TCP. In the current draft-ietf-dnsop-hardie,
that's stated this way:

  One potential problem with using shared unicast addresses is that
  routers forwarding traffic to them may have more than one available
  route, and those routes may, in fact, reach different instances of
  the shared unicast address. Because UDP is self-contained, UDP
  traffic from a single source reaching different instances presents
  no problem. TCP traffic, in contrast, may fail or present
  unworkable performance characteristics in a limited set of
  circumstances. For split-destination failures to occur, the router
  forwarding the traffic must both have equal cost routes to the two
  different instances and use a load sharing algorithm which does
  per-packet rather than per-destination load sharing.

  You don't describe the nature of the services you plan to
deploy, but unless it is the DNS, I would be concerned about your
taking too much guidance from my draft.
          Ted Hardie

That should be a little more precise.

TCP packets can not (for all practical purposes when dealing with "normal"
clients) be self contained.

UDP packets are self contained, from the network view.

But that does not mean that a particular protocol implemented on top of
UDP will necessarily still be self contained, merely that it is possible
for it to be.