Re: Scope of Communities [was: Re: Last call for bgp-redistribution]

On Thu, Jul 25, 2002 at 01:45:54PM -0500, Tom Barron wrote:
> Yes, common practice is that the community is *to* the AS *and* that it's
> just one AS-hop away.

I'm curious if the "one AS-hop away" observation is true.
To arbitrarily pick on some ISPs, lets say I buy transit from
UUNet and FooISP.  I know that FooISP buys service from UUNet.  Why
should I be able to put communities that say "deprefer this route"
that are UUNet specific and inject them out FooISP?

> That's why I considered it, ahem, innovative to
> suggest that one would leave communities in place on egress to mark *from*.

The question is, do we want to preclude doing this?

> But now I think I see that your point is really congruent with my saying
> that the base spec underdetermines the BCP.

More likely that this "novelty" is something that someone hasn't
found a good use for, and maybe just qualify that with a "yet".

Let me give another example that might be useful from an old NANOG
discussion.  AboveNet was blackholing a certain IP address in a netblock
that they were otherwise cleanly routing.  This cause people who
were using said route themselves grief if they were trying to get to
said address.  AboveNet *could* have tacked on a community that said
"this route is unclean, use at your own risk."  6461:1313 or some such. :-)

> (I distinguish between global significance and global scope because as 
> my colleague Hal Peterson has taught me, only so many communities will fit
> in a single update message.)

Hopefully, the implementation allows for a reasonable number of them.
Ideally, the number stays small due to a couple of factors:
o Operators clean up the ones of local significance.
o AS Paths are short.  Each AS in the middle should only need to
  add a few communities.

> Today extended communities are rarely used except for stuff like PPVPN (of
> course the draft is only now at last call).  You may be suggesting that BCP
> would be to migrate to extended communities where transitivity and
> nontransitivity are explicit, and where (as in redistribution communities)
> well-known values would be used to encode signalling that is today done on
> a more ad-hoc basis.  If that is what your are suggesting, I tend to agree.

According to the recent implementation report, extended communities
are pretty well supported in upcoming releases.

> What about this variant (Andrew?)

I'm not Andrew, but this still sounds good.

> 
>        - strip all communities on input except
>           * communities you tell your customers they may use
>           * well-known communities 
> 
>        - act on well-known communities and communities you tell your
>          customers they may use
> 
>        - strip all communities on output except
>           * well-known communities
>           * communities you have deliberately added to signal
>             to the next AS

-- 
Jeff Haas 
NextHop Technologies