Re: [Netconf] comments on draft-ietf-netconf-subscribed-notifications-12

[Martin Bjorklund <[e-mail]>]

"Eric Voit (evoit)" <[e-mail]> wrote:
> > From: Martin Bjorklund, June 14, 2018 2:40 PM
> > 
> > "Eric Voit (evoit)" <[e-mail]> wrote:
> > > > From: Martin Bjorklund, June 14, 2018 4:22 AM
> > > >
> > > > Hi,
> > > >
> > > > Juergen Schoenwaelder <[e-mail]> wrote:
> > > > > On Wed, Jun 13, 2018 at 03:36:01PM +0000, Eric Voit (evoit) wrote:
> > > > > > Each of the terms used are different.  While they all are
> > > > > > defined in the first
> > > > document they are used, let me paraphrase the meanings of the
> > > > definitions...
> > > > > >
> > > > > > Event - something that happened
> > > > > >
> > > > > > Event record - the recorded details of a single event
> > > > > >
> > > > > > Update record - one or more datastore node updates
> > > > > >
> > > > > > <notification> - a structure defined in RFC5277 which is as a
> > > > > > wrapper which
> > > > contains an event record.  A <notification> can exist without any
> > > > active subscription.
> > > > > >
> > > > > > "notification" statement - a structure defined in RFC-7950
> > > > > > section
> > > > > > 7.16
> > > > which allows the definition of event record types specific to a YANG
> > > > module.
> > > > The results of the a YANG "notification" statement are encoded in a
> > > > <notification>.
> > > > >
> > > > > Here is where I am getting lost. The RFC 7950 notification
> > > > > statement (its not a structure btw) does define the content of a
> > notification.
> > > > > And notification used to be defined in RFC 6241 as a
> > > > > "server-initiated message indicating that a certain event has been
> > > > > recognized by the server." Your notion of an event record may come
> > > > > from the RFC 5277 format that adds an eventTime etc. but the
> > > > > relationship of what is a YANG defined notification and how it
> > > > > related to your event record and the <notification> structure is still
> > unclear.
> > > >
> > > > I don't think we should align terminology with 5277.  More important
> > > > is to align with the current set of documents; 7950 and 6241.
> > >
> > > Exactly.
> > >
> > > Note that there was no terms imported from 5277.
> > > Subscribed-notifications does define an umbrella term "notification
> > > message", and uses Section 2.6 to make the minimal connection
> > > necessary to show that an RFC-5277 <notification> is a valid
> > > "notification message".  BTW: We had text in earlier versions of
> > > subscribed-notifications stating the need to support future types of
> > > "notification messages" such as those defined in
> > > draft-ietf-netconf-notification-messages.  However reviewers asked
> > > these evolving references to be removed.
> > >
> > > We do have the option of importing terms from 6241 into the
> > > NETCONF-notif document.  This would be the right place to do it
> > > because in the subscribed-notifications document we want to limit any
> > > introduction of NETCONF dependencies.  (Maybe NETCONF-notif adds text
> > > to say that a "RFC6241 client" maps to subscriber, and "RFC6241
> > > server" maps to publisher?)
> > 
> > 8342 defines the terms "client" and "server" in a transport-agnostic
> > way.  I
> > think we should try to use these terms in new documents, where
> > applicable.
> > In this document, the term "subscriber" is a special "client". 
> 
> Agreed.  Terminology section updated per the last email.
> 
> > I am not sure that a "publisher" is always a "server"; this needs to
> > be decided.
> 
> I think that within YANG-push, a publisher is a special RFC-8342
> "server".  Perhaps what we could add to YANG push in the terminology a
> sentence which says "In this document, the term publisher also
> includes the role of RFC-8342 server."

Can we first agree on the basic terminology in
subscribed-notifications?

First question: are server and publisher potentially different entites
or not?

Looking at restconf-notif, it seems to me that the publisher is an
HTTP2 client, i.e., a different entity than the RESTCONF server.

Is this correct:

NETCONF:
   subscriber is a NETCONF client
   publisher is a NETCONF server
   receiver is a NETCONF client

RESTCONF

   subscriber is a RESTCONF client
   (NOTE 1)
   publisher is a HTTP2 client
   receiver is a HTTP2 server

UDP

   subscriber is a NETCONF/RESTCONF client
   publisher is a UDP client
   receiver is a UDP server

NOTE 1: the RESTCONF server is missing from this picture.  It doesn't
have a special term...


I assume this work is not applicable to CoMI?



/martin


> However my belief is that we shouldn't do a global cut-and-paste of
> "publisher" with some new term for this intersection to use throughout
> the document.  This could end up being confusing.  For example, if we
> define a new term like "YANG-push publisher" a reader of the YANG
> model would need to jump back and forth between the definitions of
> "publisher" and "YANG-push publisher" knowing that both are equally
> valid in this context.
>  
> > > > If subscribed-notifications is transport-independent, it should
> > > > probably not talk too much about <notifcation> etc; this should go
> > > > into the transport docs.
> > >
> > > In general, this is what is done.  In subscribed-notifications, the
> > > only place <notification> is mentioned at all is section 2.6.  If
> > > necessary, we could move this section to NETCONF-notif, but that would
> > > leave no transport independent framing for the notifications.  I guess
> > > it is possible to live without that, but it would leave the
> > > subscribed-notifications feeling incomplete.
> > 
> > If subscribed-notifications is transport independent, it should not
> > have the
> > NETCONF-specific text in 2.6.
> 
> RESTCONF also refers to RFC-5277's <notification>.  See section 6.4.
> So there is precedent and adoption for reuse of the transport element
> beyond NETCONF.
> 
> So while pulling this section out and placing it within NETCONF-notif
> is possible, it would mean either replicating this information
> RESTCONF-notif, or forcing RESTCONF-notif to await the completion of
> draft-ietf-netconf-notification-messages.
> 
> > 2.6 also says:
> > 
> >    In all cases, a single transport session MUST
> >    be capable of supporting the intermixing of RPCs and notifications
> >    from different subscriptions.
> > 
> > This applies to NETCONF, but not for the UDP transport, and I suspect
> > not for
> > the HTTP transport either?
> 
> Agreed.  This statement is now removed. 
>  
> > > I suspect a similar
> > > thought process drove the inclusion of <notification> within RFC-6020
> > > and then RFC-7950.
> > 
> > No; for various reasons YANG was initially positioned as a data
> > modelling
> > language for NETCONF only.  Hence all text about how to map YANG to
> > NETCONF and XML.
> > 
> > > > 7950 says that the "notification" statement defines a notification.
> > > > As Juergen pointed out this term is not defined in the terminology
> > > > section, but nevertheless the term is used.
> > > >
> > > > Does the WG now want to introduce a new term for what the
> > > > "notification"
> > > > statement defines?
> > >
> > > It certainly might make sense to have a future update of RFC-7950 with
> > > something like this.  I spend a bit of time trying to understand the
> > > connection of YANG notification statement with <notification>.  Having
> > > this be better defined would be helpful.
> > >
> > > >  It seems to me that the term "event record" is being proposed for
> > > > this.
> > >
> > > An event record is not necessarily a YANG notification, as the event
> > > record's payload might not be driven by the result of a YANG
> > > statement.
> > 
> > I don't get this.  Can you give an example of when an event record is
> > not
> > defined as a YANG "notification"?
> 
> Another way to put this is that the event record within the
> <notification> might not be encoding YANG data.  For examples, all the
> <notification> within RFC5277 were defined before the availability of
> YANG.
>  
> > > > The answer to this question will have a big impact on the rest of
> > > > the terminology.
> > >
> > > As event record has a larger scope than what can come from a YANG
> > > notification statement, my suggestion would be for the revision of
> > > RFC-7950 to import "event record", and then specify a new subtype term
> > > (maybe "YANG event record"?).  If that term works, a YANG event record
> > > could then be an event record where the contents are populated by the
> > > results of the YANG notification statement.
> > >
> > > > > > Notification message - a message intended for a specific
> > > > > > subscription
> > > > receiver which includes one or more <notification>. A notification
> > > > message will have undergone any security/content filtering on
> > > > embedded <notification> as appropriate for that receiver.
> > > > >
> > > > > So how does this fit Figure 1 of RFC 6241? This figure indicates
> > > > > that <notification> is a message as seen from the messages layer.
> > > > > You are saying a notification message is something else that
> > > > > includes one or more <notification>s. Yes, I know that the diagram
> > > > > in RFC 5277 is different but the diagram in RFC 6241 is the newer one.
> > > >
> > > > This confuses me as well.
> > >
> > > The requirement for the bundling of many events is being driven by
> > > large data center telemetry.  It is unclear at this point whether
> > > NETCONF will be a transport used in this environment.
> > >
> > > If NETCONF does care about this environment, and does want to support
> > > something like draft-ietf-netconf-notification-messages, I do think
> > > tweaks to RFC-6241 will be needed.
> > 
> > Yes, somehow.
> 
> Ok
> 
> Eric
>  
> > > For example what is the definition
> > > of <notification> within 6241, Figure 1 (right now point RFC-5277
> > > isn't explicitly mentioned.).  Must this figure only be interpreted as
> > > a RFC 5277 <notification>?  Can the figure also mean a
> > > draft-ietf-netconf-notification-messages "message"?
> > 
> > RFC 6241 says that <notification> is defined in RFC 5277.  (ok,
> > there's an error
> > in there, and it points to 5717, but there's an errata for that).
> > 
> > > > How much of this do we have to define in this document, and how much
> > > > should go into the transport docs?
> > >
> > > Beyond what I describe above, impacts would be to new/updated
> > > transport drafts.  Plus an update to subscribed-notifications section
> > > 2.6 to indicate that a new transport independent <notification>
> > > construct exists.
> > 
> > Ok.
> > 
> > 
> > /martin
> > 
> > 
> > >
> > > Eric
> > >
> > > > /martin
> > > >
> > > >
> > > > > > Per the discussion below, I see an update record being a
> > > > > > specialized type of
> > > > event record.  For YANG push, the 'event' is driven by the update
> > > > trigger: i.e.,
> > > > either the expiration of a periodic timer (for periodic
> > > > subscriptions), or a change to the datastore (on-change
> > > > subscription).
> > > > > >
> > > > >
> > > > > I am missing a definition what an Update record is. It is surely
> > > > > not in this email. Anyway, if there are changes to architectural
> > > > > concepts, it would be nice to find them in a coherent well explained
> > section.
> > > > >
> > > > > /js
> > > > >
> > > > > --
> > > > > Juergen Schoenwaelder           Jacobs University Bremen gGmbH
> > > > > Phone: [phone]         Campus Ring 1 | 28759 Bremen | Germany
> > > > > Fax:   [phone]         <https://www.jacobs-university.de/>
> > > > >
> > >
>