Re: [Idr] WG adoption for draft-haas-flowspec-capability-bits - 3/30 to 4/13

[Jeffrey Haas <[e-mail]>]

Robert,

I think you're mixing some of the points together.

On Wed, Apr 07, 2021 at 12:29:21PM +0200, Robert Raszuk wrote:
> I have a question on how practical this proposal is.
> 
> Fundamental problem with today's BGP capabilities is that the information
> is only known to the peer.
> 
> So if I inject flowspec rule from behind the RR the RR will suppress some
> updates but:
> 
> * sender will have no clue about it
> * any other flow spec BGP speaker which supports request filtering down the
> BGP path will never get the chance to receive and apply the filter.

This is explicitly acknowledged in section 5.

It's also indistinguishable from policy.

> Both are IMHO bad. The latter is in fact directly against flowspec spirit
> to apply filtering as close to the src even if hops on the way are not
> capable of doing so.
> 
> So I am yet to be convinced this proposal is useful.
> 
> Today as a general rule if a router does not support an extension received
> via flowspec it just does not apply it but still can happily propagate the
> update down the road.

Not a single BGP flowspec implementation implemented the "opaque" property
in RFC 5575.  Not one.  And it was the strong motivation to remove that text
in RFC 8955.

This may change in flowspec v2 where we likely will make the NLRI proper
type-length-value rather than implied length as it is currently done.

How well the argument to filter unsupported components stands once we're to
something like flowspec v2 will be the longer question.  For such scenarios,
once we're no longer concerned about propagation characteristics of the
NLRI, it's quite reasoanble for a route reflector to carry NLRI with
filtering components it doesn't understand - but edge devices may not want
it.  But in that scenario, perhaps it's simply better for edge devices to
accept it and simply not install it.

> I think what we have here on the table is an illustration about the growing
> need for domain wide (or set of domains under the same admin) capability
> distribution such that all BGP speakers could advertise their
> capabilities to interested parties. A bit broader than flowspec, but could
> be useful here.

At the day job, we talk about the "flooding domain" of new features that are
gated on their capabilities.  BGP doesn't provide an explicit concept for
this, but it's become a protocol intrinsic behavior.  Unless you configure a
capability between two devices, it does't gain the ability to use the
feature.  A contiguous set of devices using those capabilities becomes that
domain.  That domain may, or may not, have strong overlap with one or more
ASes under the control of one or more parties.

For many BGP features, knowing what those domain boundaries are isn't much
of a problem.  Flowspec implementations with disjoint capabilities is a
place where it could be problematic.  (Again, see section 5.)

---

I don't believe we're likely to come up with a simple answer to BGP
capability domains soon.  That said, operators in a single administrative
domain at least can figure this out themselves because they're the ones
configuring it.

Right now, we are unable to safely deploy new flowspec features. Even when
you have disjoint flooding domains, if your flowspec rules are independent,
you can gain benefit.  So, short term for flowspec v1, I think there's
benefit for this feature.

-- Jeff