2.  Terminology
   Anycast: the practice of making a particular Service Address
      available in multiple, discrete, autonomous locations, such that
      datagrams sent are routed to one of several available locations.

...routed to the topologically nearest of several...

   Local-Scope Anycast: reachability information for the anycast Service
      Address is propagated through a routing system in such a way that
      a particular anycast node is only visible to a subset of the whole
      routing system.

   Local Node: an Anycast Node providing service using a Local-Scope
      Anycast address.

I think this is Vixie-specificism which just muddies the waters.  To the best of my knowledge, nobody else tries to do this, or uses this terminology.  I've never seen a reasoned argument that there was any reason to do this, nor heard that anyone else had considered doing it.  I'm not sure it needs to be documented outside of ISC tech-notes, for ISC staffers who have to maintain it.

Besides, any particular anycast node is, pretty much by definition, only visible to a subset of the whole routing system.  So that's not a useful definition of something that's been purposely hamstrung.

And you're missing "Anycast Instance."

And if you're going to define "Service Address," you should probably define "Unique Address" as well, just for the sake of thoroughness.

   Anycast is the name given to the practice of making a Service Address
   available to a routing system at Anycast Nodes in two or more
   discrete locations. 

Why do you care whether people do it at two or more discrete locations?  It just needs to be two or more instances, to be anycast.  Who cares where they're located?

       Topological nearness within the
       routing system does not, in general, correlate to round-trip
       performance across a network; in some cases response times may
       see no reduction, and may increase.

It does, in general.  It does not always do so, however.  And while it's technically correct to use "and" between "reduction" and "may increase," I think your purpose would probably be better served by an "or."

4.1.  Protocol Suitability

The general rule is that protocols don't matter, but critical stateful transactions, like online banking sessions, should be handled by unique addresses when possible.

And you may as well take the opportunity to debunk the fallacy about problems with TCP.  I have no idea why there are still idiots, ten years later, who still think there's a problem, but the point of an RFC is to convey information to the clueless, so you might as well avail yourself of the chance.

4.4.5.  Reverse Path Forwarding Checks

Is this necessary?  It doesn't have anything to do with anycast.

4.4.6.  Propagation Scope

Again, this goes into excessive detail on a unique implementation, for which there's no evidence of desirability.

   nodes' internal and external/connecting 
   infrastructure should be scaled to
   support loads far in excess of the average

More specificity here wouldn't hurt.  "Nodes' internal and connecting infrastructure should be scaled in accordance with the "be liberal in what you accept and conservative in what you send" principle.  Specifically, the capacity of each node should be engineered with the expectation that traffic will spill over from other adjacent nodes which fail, are withdrawn, or are overrun by DDoS attacks.  Likewise, nodes should be capable of absorbing potentially much larger volumes of traffic than they can actually serve, without unintentionally failing-over to adjacent nodes, and multiplying the burden which they're already carrying."

4.8.3.  Intra-Node Interior Connectivity
   The IGP in which all nodes participate can be
   viewed as a single point of failure.

How so?  In what case would the IGP override local routes within a node, and bring it down?

4.9.  Node Identification by Clients

You might point out that many people run simple web servers on their anycast servers, which just serve a page with the name of the site on it, to facilitate the work of amateur diagnosticians.

6.3.  Service Hijacking

   It is possible that an unauthorised party might advertise routes
   corresponding to anycast Service Addresses across a network, and by
   doing so capture legitimate request traffic or process requests in a
   manner which compromises the service (or both).  A rogue Anycast Node
   might be difficult to detect by clients or by the operator of the
   service.

That this isn't something to which anycast nodes are particularly vulnerable, and I'm not sure that it makes any sense to say that it's less likely to be noticed if it's done to them...  It might also be worth pointing out that this is potential abuse of the technique.  That is, the attack you're describing is, in fact, anycast.  It's just rogue anycast, whether or not the address is otherwise anycasted.

Otherwise, looks great.  Nice work.

                                       -Bill





Attachment: PGP.sig
Description: This is a digitally signed message part

Reply via email to