Re: [netmod] upcoming adoptions

[Martin Bjorklund <[e-mail]>]

Robert Wilton <[e-mail]> wrote:
> 
> 
> On 07/09/2017 11:15, Martin Bjorklund wrote:
> > Robert Wilton <[e-mail]> wrote:
> >>
> >> On 07/09/2017 11:05, Martin Bjorklund wrote:
> >>> Robert Wilton <[e-mail]> wrote:
> >>>> On 07/09/2017 03:36, Andy Bierman wrote:
> >>>>> On Wed, Sep 6, 2017 at 10:57 AM, Kent Watsen <[e-mail]
> >>>>> <mailto:[e-mail]>> wrote:
> >>>>>
> >>>>>
> >>>>>
> >>>>>       >> /netconf-state and /restconf-state don't seem to follow the
> >>>>>       >> general
> >>>>>       >> pattern we're correcting with the various NMDA updates.
> >>>>>       Particularly,
> >>>>>       >> these -state trees are NOT for the purpose to providing the
> >>>>>       >> opstate
> >>>>>       >> value for configured nodes.  These modules have the misfortune
> >>>>>       >> of
> >>>>>       >> having "-state" in their name, but they're otherwise fine.
> >>>>>       >
> >>>>>       >
> >>>>>       > This contradicts some details we have been told about NMDA
> >>>>>       >
> >>>>>       > 1) the transition guidelines say otherwise
> >>>>>       >
> >>>>>       > New modules and modules that are not concerned with the
> >>>>>       > operational state of configuration information SHOULD
> >>>>>       > immediately be structured to be NMDA-compatible
> >>>>>
> >>>>>       Yes, I'm suggesting we give ourselves some leeway, by taking
> >>>>>       advantage of the SHOULD keyword above and defer updating these
> >>>>>       two modules to when it makes more sense to do so.
> >>>>>
> >>>>>
> >>>>>
> >>>>> OK -- good.
> >>>>> I think another sentence needs to be added.
> >>>>>
> >>>>>
> >>>>> NMDA compatibility conversion MAY be deferred if the module
> >>>>> does not contain any configuration datastore objects.
> >>>> I agree.
> >>> +1
> >>>
> >>>
> >>>>>       > 2) RD defines operational state to include config=false nodes
> >>>>>       > such as counters, so these modules are properly named.
> >>>>>
> >>>>>       module-name == top-level node name.  Either way, my point is that
> >>>>>       the -state tree in these modules is not trying to provide the
> >>>>>       opstate value for configured nodes (i.e. applied configuration).
> >>>>>
> >>>>>
> >>>>> So a data node naming convention is needed?
> >>>>> And a module naming convention?
> >>>>>
> >>>>> We need a rule that says the suffix "-state" is reserved for NMDA
> >>>>> compatibility
> >>>>> so module names and data nodes SHOULD NOT be named with an identifier
> >>>>> that
> >>>>> ends in this suffix.
> >>>> Also agree.
> >>> -1
> >>>
> >>> There are cases where a -state suffix is natural, e.g. in
> >>> ietf-hardware we have admin-state, oper-state, usage-state etc.
> >>>
> >>> I prefer to have a recommendation that generated modules and top-level
> >>> nodes are called ...-state, but that should not be a reason for making
> >>> -state illegal in general.
> >> Sorry, it was specifically modules and top level data nodes that I
> >> think this restriction should apply to.
> > Even in this case I'd prefer to make a one-way recommendation.  Is
> > there a technical reason for not allowing a normal module or top-level
> > node be called ...-state?  IMO, *if* it is important to mark these
> > modules as being generated, we should use a formal way to convey this
> > information, not rely on a naming convention.  (i.e., use a YANG
> > extension in the module).
> My logic is slightly different:
> 
> I think that new top level containers shouldn't be called ...-state
> because either
> (i) they already contain config nodes, in which case they are
> misnamed, or
> (ii) they might be revised to contain config nodes in the future, in
> which case they would end up being misnamed.
> 
> I basically, think that the suffix "-state" doesn't generally provide
> any useful extra information.

Sure.  But there are many suffixes that don't generally provide useful
extra information.  I just think we should be careful with these kinds
of rules.  Bad names are hopefully catched during the review process.

> Lower down the tree, I guess that using "state" or  "*-state" is OK if
> it can be known that the leaf/container will *always* only be state
> and never potentially be configurable in future.  But I still think
> that it is necessary to be very careful here.
> 
> For example, If I designed a new routing protocol, then in v1 there
> might be no explicit neighbor configuration, and hence I put all of
> the neighbor information into a container call
> neighbor-state. However, if a v2 version of that protocol (or vendor
> extensions) wants to add some per neighbor configuration then it would
> hit the problem that the original container is poorly named to be
> updated to now also hold configuration.  So, generally avoiding
> "-state" in the name of containers seems to be good common sense to
> me.

I agree.  And avoid "*-config".



/martin


>   I would suggest adding something to 6087bis, but my previous
> suggested additions to 6087bis didn't appear to be well received ;-)