On Mon, Jun 26, 2017 at 01:58:40PM +0000, [email protected] wrote:
> Hi Job, all
> 
> > From: Job Snijders [mailto:[email protected]]  > Sent: Thursday, June 22, 2017 
> > 10:47 PM
> [...]
> 
> > I think that the neighbor ASBR should _not_ strip the GSHUT
> > well-known community.
> 
> I'm personally open to both options.
> Discussing this further:
> 
> - First of all, stripping the g-shut well-known community fits the
> goal of the document and its requirements (RFC 6198). The main
> motivation is to have a g-shut solution with no required BGP protocol
> extension, and where the backup path is known to the AS. i.e. we are
> not looking for an Internet wide convergence/g-shut. We are primarily
> interested in a g-shut within the responsibility of the AS. IOW, when
> it's "my fault" if my customer experiences a packet loss. In this
> case, there is no need to propagate the g-shut community further.
> 
> - removing the g-shut community reduces Internet wide churns,
> including compared to when g-shut is not used.
> Here are the 3 cases, focusing on the first step of the BGP
> convergence (searching for the backup path)
>   - No g-shut: a WITHDRAW is propagated to other ASes
>   - g-shut propagating the community: an UDPDATE is propagated to
>   others ASes (same path, adding the community)
>   - g-shut removing the community: no BGP messages propagated to
>   others ASes (same path, same communities/attribute)
> 
> In general, reducing the BGP churn is considered as a feature. By
> removing the g-shut community, we are at the same time:
> a) g-shut in both affected ASes
> b) reducing Internet churn.
> 
> Now if we choose to not remove the community, we improve "a" by
> covering the cases where the backup path is unknown to the AS. And we
> keep "b" unchanged compared to today (but degrade it compare to
> current g-shut draft).

I don't think we really have a say in whether BGP Communities are
removed or not. This is outside our control. If we remove the hard
requirement to remove the community, the behaviour falls back to
https://tools.ietf.org/html/rfc7454#section-11 which imho is fine: leave
the choice with the operator. Perhaps the draft does not need to define
whether to strip or not.

I have to admit that the first time I used a a 'g-shut -06 compliant'
automated process, I was surprised that that I didn't see the community
on the IBGP outbound announcements. Removing the community somewhat
reduces visibility to those involved with the operation.

> As for my requirements, I'm considering that our ASes have the
> knowledge of the backup path. Hence I don't need for the extra
> coverage. Regarding the extra cost, I agree that one can hardly
> consider this unacceptable since this is the current behavior.
> 
> TL;DR: it's a tradeoff between 2 secondary objectives:
> - reducing Internet churned (compared to today)
> - improving the g-shut coverage when the AS does not know the backup path

+ improve visibility into the operation

> Draft can possibly discuss both options, at the cost of additional
> reading complexity. But this possibly could be discussed in a
> different section.
>
> I'd welcome more opinion, before choosing the main text.

Bruno, perhaps my interpretation of the above sentence's phrasing is
somewhat mangled because neither of us are a native speaker, but keep in
mind this is a GROW working group document. :-)

Kind regards,

Job

_______________________________________________
GROW mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/grow

Reply via email to