Ray,

On Fri, 12 Jun 2015 11:52:44 +0100
Ray Bellis <[email protected]> wrote:

> On 03/06/2015 17:22, Joe Abley wrote:
> > On 3 Jun 2015, at 17:17, Shane Kerr wrote:
> > 
> >> On Wed, 03 Jun 2015 13:57:39 +0100
> >> Ray Bellis <[email protected]> wrote:
> >>
> >>> Whilst discussing 5966-bis with my co-authors connection-close with the
> >>> co-authors, we were reminded of this point I made in
> >>> draft-bellis-dnsop-connection-close in relation to ยง7 of RFC 6891:
> >>>
> >>> "  TODO: note - the constraint in RFC 6891 appears unnecessarily strict
> >>>  - it appears to mandate that the EDNS(0) support indication is on a
> >>>  per-request basis, but it would be reasonable on a connection-
> >>>  orientated transport to assume that ANY preceding request on that
> >>>  connection with an OPT RR is sufficient to indicate that the client
> >>>  supports EDNS(0)."
> >>>
> >>> Is this something that the WG believes needs to be fixed?
> >>
> >> The potential benefit is that a client could omit the OPT RR on
> >> subsequent messages? Seems a relatively small benefit, and there are
> >> costs. (Are there other benefits?)
> > 
> > I agree.
> >
> > I think there's a baked-in expectation that OPT pseudo-RR is included in
> > every DNS message, not on every connection (where the transport is
> > connection-oriented).
> 
> The use case we are considering is that absent an OPT RR in each
> request, RFC 6891 doesn't permit the server to unilaterally send back an
> OPT RR in a response (e.g. for connection signalling purposes) even it
> one was previously seen on the same persistent connection.
> 
> Furthermore if the specification of a particular OPT code requires that
> it only be sent when received from the client then that must also apply
> *per message*.
> 
> We're not saying this is wrong, but in the context of a persistent
> connection that could be quite an overhead and if necessary we should
> explicitly document this with full awareness of the potential consequences.
> 
> It seems that EDNS is not an entirely suitable mechanism for
> *connection* related signalling and it's this issue that we're trying to
> find a solution to.

Okay, perhaps I misunderstood the original question.

I was taking the question to be "in our current clarification work,
should we consider omitting OPT RR from every message on a session". In
that case, I think the answer is "no, this is not a clarification, but
a protocol change".

If the question is, "should we change the protocol so that EDNS over
connection-oriented transports works differently", then I think the
answer is, "hm... interesting. It probably won't provide a big win,
but lets explore it."

:) 

Cheers,

--
Shane

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

Reply via email to