[j-nsp] Segment Routing Real World Deployment (was: VPC mc-lag)

[Saku Ytti]

Hey Michael,

On Mon, 9 Jul 2018 at 17:09, Michael Hare <michael.hare at wisc.edu> wrote:

> When I was a bit greener with our MPLS network I would experience the same concern as Alexandre when a dual connected customer lost a PE, in that I would experience loss of service waiting for BGP to timeout.  I briefly went down the wrong path (IMHO) of BFD everywhere, including on directly connected links (yes, I know BFD can help for control plane issues, but 99%+ of the time for us it is 'switch-in-the-middle' problem).  I even put multihop BFD on my iBGP sessions, which I later removed.  The correct configuration (at least in my experience) was to invalidate BGP next hops, as Saku points out, and keep the BGP session up (assuming outage is less than 60s - 90s and normal timers), so I had no delay in re-estalbishing BGP and repopulating RIB in the flap was brief.

This is highly subjective, but I do not like BFD. I would only
consider BFD when L1 liveliness does not work (L2 in between or such).
I think issues undetected by L1 liveliness detection is far smaller
than issues caused by BFD false positives.

I like your solution of limiting resolution, because this is
essentially same as making BGP convergence depend entirely on IGP
convergence, for ~free. Potentially add ADD-PATH and PIC and you can
have redundant route already programmed in HW, requiring no SW
convergence to switch to backup path.

-- 
  ++ytti