re: AGIS Route Flaps Interrupting its Peering?
| Since ANS seems to be passing our interface address as the | next-hop directly to some nets (e.g., Digex and Advantis), the failure | as I described above did lead to a loss of connectivity between AGIS | and at least Digex and Advantis. Pending the solution of the MFS | problem, it would have been possible to work around the issue if the | affected nets had routed _through_ their transit provider. Let me reiterate the point that propagating third-party next-hops in the absence of guaranteed fate-sharing is EVIL, or at least very very risky. My opinion is probably at one pole of the spectrum of ideas about NAPs and MAEs, however it's essentially this: do not propagate other people's next-hops at all to your NAP/MAE peers, either using next-hop-self (or the equivalent) or announcing only those prefixes for which you have yourself as a next-hop. Moreover, one should be very cagey about accepting third-party next-hops from one's peers, and either refuse routes with such next-hops, or (with permission only), rewrite the next-hops in question, unless there is a very good reason to do otherwise. On that point, and with respect to "neighbor foo next-hop-self" at an exchange-point, one has to remember that one will be doing multiple hops across the same medium, which is an inefficient way of making use of both one's (or maybe everyone's, depending on the particular fabric) exchange-point bandwidth, and also one's router capacity. One may, given careful thinking, decide to accept and/or propagate third-party next-hops, if this is an issue. However, this should not be the default behaviour at any exchange-point, because the AGIS/DIGEX disconnectivity is a well-known and formerly oft-seen problem. Finally, there is wording in some bilaterals floating out there with respect to third-party transit at exchange points, which may push towards one pole or the other of opinions on third-party next-hops on non-fate-shared (and often *weird*) LISes. On another front, *weird* MAE and NAP setups have caused so much trouble that I sometimes wonder when the next time I get to say "I told you so" about multi-fabric-bridging-from-hell will be, and how bad it will hurt. Sean.
Sean Doran wrote:
My opinion is probably at one pole of the spectrum of ideas about NAPs and MAEs, however it's essentially this: do not propagate other people's next-hops at all to your NAP/MAE peers,
Presumably the other pole being the RA approach where the object you peer with is *never* the next hop (i.e. minimize BGP sessions and packets switched at the possible expense of predictability / reliability / control). Anyone have any idea how much traffic actually *does* shortcut a hop (or more), and thus what impact always using next-hop-self would have? Alex Bligh Xara Networks
participants (2)
-
Alex.Bligh
-
Sean Doran