Re: [Idr] Bug in RFC 7911 (add-paths) and tie-breaking

[Jeffrey Haas <[e-mail]>]

Robert,

On Sat, Jun 26, 2021 at 04:41:33PM +0200, Robert Raszuk wrote:
> If there are common scenarios where a given BGP speaker will locally
> > originate multiple paths without having gotten them from somewhere else,
> > I'm
> > not currently aware of it.

It's peculiar that your mail software consistently eats the first part of
the quote above.  Anyway...

> I think it is pretty common to connect PE (for simplicity assume non VPN
> case) to two customer CEs.
> 
> Those CEs are over static + bfd.
> 
> Further those may be connected with different link bw.
> 
> So if you redistribute those two static routes to different CE's next hops
> you will likely end up with two paths in BGP. Normally sending one will be
> cool as you set next hop self, normally repair any failure, locally load
> balance across links to CEs etc ... life is good.
> 
> But the moment you enable add-paths, adventure (not to say pollution)
> starts.
> 
> I think this could be a pretty common case this discussion could get
> triggered off.

Agreed.

For my part, I see it as an example of the draft.  The nexthops could
arguably be considered the peer-id for purposes of that proposal.

> And thx for bringing up the  conn-restore draft. I guess it matured
> sufficiently to perhaps refresh it and start WG adoption call ? Either
> Pradosh or myself should have the latest xml for it.

I won't make a specific argument for maturity.  I have personally brought it
out repeatedly over the years because it summarizes this problem well.

It has issues with incremental deployment that would need to be addressed in
the text.  How should a domain that has a mix of BGP speakers that
understands this or not work together?  Probably fine since as John notes
it's basically "good enough".

The reason I haven't pressed on advancing the draft in my day job capacity
over the years is largely the observation John ends up making with one
possible refinement:  Locally, do whatever you like to pick a consistent
active route.  As long as you're getting to the endpoint via tunneling
rather than hop-by-hop along the way it's likely fine.

Depending on how this conversation advances, it might be worth bringing the
document back for adoption consideration.

-- Jeff