Re: Executive summary of: RE: hard questions: request routing

[Francis Zane <francis@research.bell-labs.com>]

Regarding the proposed solutions, I'm agnostic among solutions which
don't do forwarding within the RRS.  I can see cases where small
amounts of forwarding (eg, limited depth, or one-way between certain
classes) could be quite useful, and so I'm happy to explore that as
well.  I'm just very nervous about schemes that allow forwarding in
its complete generality.

I very much like Brad's suggestions, though I would also support
allowing some other trivially-safe cases.  For example, having some
nodes identified as only outgoing or only incoming with forwarding
allowed in the appropriate direction would handle Oliver's situation
with many ISP destinations, but still be safe.  On the other hand,
that proposal seems like (1) or (2), while leaving the door open for
future development of forwarding, so maybe I'm not understanding the
distinctions between all the cases.

My main concern with fully general forwarding mechanisms remains the
ability to deal with different (particularly non-DNS) RRS mechanisms
and all the cases which could arise.  As a (mostly thought-experiment)
example, here's a weird case that concerns me: Suppose I have an RRS
which spans two specific RR mechanisms: DNS (based on local DNS IP)
and some form of L7 (based on client IP).  Is there a way to guarantee
correct (ie, non-looping) behaviour?  (CDN A DNS redirects to CDN B
based on local DNS, which HTTP redirects to CDN A based on client IP).

Francis