Re: [v6ops] PI heresy [draft-ietf-v6ops-ula-usage-recommendations - work or abandon?]

[[e-mail]]

Gert,

>>> Networks break further upstream, without the actual upstream being
>>> noticeable
>>> broken.  Networks break certain paths, while others are working fine.
>>> 
>>> We can do better than all-or-nothing.  We should.
>> 
>> I don't think a host should try to figure out which paths work and which
>> don't.
> 
> I disagree, and strongly so.  Maybe not "probe all paths at the same time",
> but if one is not working, it would be silly not to try alternatives if
> they exist.
> 
> [..]
>> I propose that routers just inform hosts which source prefixes are unusable
>> according to unspecified monitoring code running on the router. (Obviously,
>> routers should also propate that knowledge over internal routing protocols in
>> setups with multiple routers).
> 
> So, if you have upstreams A and B, and targets C and D, of which "C" cannot
> be reached over "A" and "D" not via "B" (not all that unlikely given the
> direction in which low-price SoHo ISPs and peering wars are developing),
> would you rather declare both broken, or just use B to reach C, and A to
> reach D?
> 
> Defining "the upstream is broken" is actually fairly hard - the router might
> be perfectly happy about upstream A, because 99% of its monitoring targets
> work perfectly well - but if the user's 1% that he wants to talk to do not
> work (routing hickup, anycast cluster broken, peering overloaded), it is
> the wrong conclusion to declare "it is working, no need to try anything
> else"...

exactly. the end-points has much more information about the end to end path than the first-hop router has.
unless you want it to be a DPI/layer violating/state-hogging middlebox.
even having every router participate in global routing wouldn't help you very much. global routing has convergence times measured in many minutes.

cheers,
Ole