Re: [netmod] js review of draft-ietf-netmod-schema-mount-09

[Juergen Schoenwaelder <[e-mail]>]

On Thu, Apr 05, 2018 at 02:33:40PM +0200, Martin Bjorklund wrote:
> Hi,
> 
> Thank you for this review!  Comments inline.
> 
> Juergen Schoenwaelder <[e-mail]> wrote:
> > Here is my review of draft-ietf-netmod-schema-mount-09.
> > 
> > * Abstract
> > 
> >    This document defines a mechanism to combine YANG modules into the
> >    schema defined in other YANG modules.
> > 
> >   I do not know what this says - I think this text is confusing. What
> >   does it mean to 'combine' YANG modules? What is the notion of
> >   'schema' used here?
> 
> Howabout:
> 
>   This document defines a mechanism to add the schema trees defined by
>   a set of YANG modules into the schema tree defined in another YANG
>   module.
>

This is better.

> (see more below on terminology)
> 
> >   Does the text help someone to decide whether
> >   this mechanisms is something worth to study in order to solve a
> >   given modeling problem?  (A good abstract would IMHO do that.)
> > 
> >   Note that the mount mechanisms has serious limitations as well that
> >   perhaps need to spelled out right up-front, i.e., it only works with
> >   pre-defined mount-points (augments are much more flexible in this
> >   regard, the schema mount defined here is by its very design not
> >   very flexible.
> 
> I don't agree that this is a "serious limitation", and I don't think
> an abstract should list things that the document doesn't do.

OK. Remove 'serious' but clearly this schema mount mechanism is not as
flexible as it could me. What about this:

   This document defines a mechanism to add the schema trees defined
   by a set of YANG modules onto mount points defined in another YANG
   module.

This at least hints at the fact that mount points are rather static
citizens.
 
> >   While I think I understand the difference made between
> >   implementation-time and run-time, the description is somewhat
> >   confusing since the run-time mount will also be exposed via YANG
> >   library and hence defining implementation-time by 'defined by a
> >   server implementor and is as stable as YANG library information of
> >   the server' is somewhat fuzzy. I assume what you mean is that in the
> >   case 2. the mounted schema is fixed at implementation time while in
> >   the case 3. the mounted schema may vary and be discovered at
> >   run-time. However, you do not define things this way but rather talk
> >   about properties that do however not define things.
> 
> Howabout:
> 
> OLD:
> 
> + Run-time: the mounted schema is defined by instance data that is
>   part of the mounted data model. If there are multiple instances of
>   the same mount point (e.g., in multiple entries of a list), the
>   mounted data model may be different for each instance.
> 
> NEW:
> 
> + Run-time: the mounted schema can vary at run time and is defined by
>   instance data that is part of the mounted data model. If there are
>   multiple instances of the same mount point (e.g., in multiple
>   entries of a list), the mounted data model may be different for each
>   instance.

Better.
 
> > * Glossary of New Terms
> > 
> >      o  top-level schema: a schema according to [RFC7950] in which schema
> >       trees of each module (except augments) start at the root node.
> > 
> >   You do not import 'schema' from RFC 7950 since, well, it is not
> >   defined in RFC 7950. I think you often mean a schema tree (as
> >   defined in RFC 7950) when you use 'schema'. Well, even this is not
> >   true since a 'schema tree' according to RFC 7950 is scoped to a
> >   module. RFC 8342 defines a 'datastore schema' but then I am not sure
> >   this corresponds to 'schema' as used in this draft. In fact, the
> >   mounted schema may be considered part of the 'datastore schema'.  I
> >   think we are handwaving with our terminology here but then perhaps I
> >   am the only one who cares...
> 
> I have imported "schema tree" from 7950, and changed teh definition of
> top-level schema to:
> 
> - top-level schema: a set of modules in which the schema
>   trees of each module (except augments) start at the root node.

Still sounds complicated and not quite clear. Do you mean this:

- top-level schema: a set of schema trees of a set of modules starting
  at the root node

> >   What we actually have are schema tree (of a module per RFC 7950) and
> >   a collection of schema trees sharing a common root (this is likely
> >   what is meant with "schema" in this document). And then schema mount
> >   simply provides a mechanism to have additional (statically defined)
> >   roots in a schema.

So what are you planning to do about the term 'schema' used throughout
the module? We kind of define what a top-level schema is but leave
schema undefined and perhaps we would also benefit from having a term
like mounted schema. I am thinking along these lines:

  schema = collection of schema trees with a common root
  top-level schema = schema rooted at the root node
  mounted schema = schema rooted at a mount point
  parent schema (of a mounted schema) = schema containing the mount point

> > * Multiple Levels of Schema Mount
> > 
> >   What is a 'subschema'? What is a 'schema level'? Is a subschema the
> >   same as a schema, i.e. a collection of schema trees with a common
> >   root? If we need terms such as 'subschema' or 'schema level', then
> >   we should define them. But perhaps just some tweaking the text to
> >   avoid new terms can solve the issue.
> 
> I have changed "subschema" to "schema", and removed "schema level":
> 
> NEW:
> 
>   YANG modules in a mounted schema MAY again contain mount points
>   under which other schemas can be mounted.  Consequently, it is
>   possible to construct data models with an arbitrary number of
>   mounted schemas.  A schema for a mount point contained in a mounted
>   module can be specified by implementing "ietf-yang-library" and
>   "ietf-yang-schema-mount" modules in the mounted schema, and
>   specifying the schemas exactly as it is done in the top-level
>   schema.

OK 

> >   Why are parent-references only useful for the 'shared-schema' case?
> >   An 'inline' mount can't refer to stuff outside the mount jail?
> 
> Correct.  We have debated if this makes sense for inline or not.  As
> it is, the model is designed so that this can be added in the future,
> if it turns out that this is needed.

OK

> >   Looking at the YANG definition of 'parent-reference', I am left
> >   somewhat clueless in which situations these xpath expressions are
> >   evaluations and when the nodesets are merged with other xpath
> >   expression evaluation results.
> 
> The YANG module says:
> 
>              When XPath expressions in the mounted schema
>              are evaluated, the 'parent-reference' leaf-list is taken
>              into account.
> 
> and
> 
>                For the purposes of evaluating XPath
>                expressions whose context nodes are defined in the
>                mounted schema, the union of all these nodesets
>                together with ancestor nodes are added to the
>                accessible data tree.

OK. I probably did not read carefully enough. My understanding is now
that the set of parent-reference xpath expressions yields a union of
nodesets that are becoming part of the context for the evaluation of
any xpath expression appearing in a mounted schema. And these parent
references are specific to a mount point instance. May make sense.

> >   It seems that these parent references
> >   are the only actual difference between 'inline' and 'shared-schema'
> >   mounts.
> > 
> > * Data Model
> > 
> >   I have not really understood what the difference between 'inline'
> >   and 'shared-schema' is. I understand that the later can have
> >   'parent-references' but it is unclear why the other can't and if
> >   there is not strong architectural reason why there have to be two
> >   choices. It also seems that the 'namespace' list is only meaningful
> >   if there are parent references, no? So why is this then global, i.e.
> >   also provided for 'inline' mounts?
> 
> As you note, the 'namespace' list is global, so there is just one list
> that covers all mount-points.  It's not really correct to state that
> the 'namespace' list is "provided for 'inline' mounts".

OK. It seems that for a client capable to support a 'shared schema',
the 'inline' schema really is just a special case without parent
references. (I wonder whether anything should be said to YANG library
version numbers, whether they are always scoped by the mount point
or have meaning across mount points; if two YANG library instances
in mounted schemas have the same version number, does this imply
that these YANG library instances are identical or is this just a
version number clash? But then this probably goes across the spirit
of not talking about YANG library too much.)
 
> >   I guess I do not really
> >   understand the distinction. If there are no parent-references, what
> >   is the difference between 'shared-schema' and 'inline'?
> 
> The difference is that shared-schema can have parent-references, and
> that all instances of such a mount point have exactly the same
> schema.

See comment above about YANG library version numbers. I am trying to
understand which assumption a client can make to be efficient in both
cases (and whether it is possible to be efficient while essentially
handling inline and shared-schema using a common approach).

> > * Security Considerations
> > 
> >   I agree with others that something needs to be said how NACM applies
> >   to mounted schemas.
> 
> I have added the following (short) section:
> 
> 7.  Interaction with the Network Configuration Access Control Model
>     (NACM)
> 
>    If NACM [RFC8341] is implemented on a server, it can be used to
>    control access to nodes defined by the mounted schema in the same way
>    as for nodes defined by the top-level schema.
> 
>    For example, suppose the module "ietf-interfaces" is mounted in the
>    "root" container in the "logical-network-element" list defined in
>    [I-D.ietf-rtgwg-lne-model].  Then the following NACM path can be used
>    to control access to the "interfaces" container (where the character
>    '\' is used where a line break has been inserted for formatting
>    reasons):
> 
>      <path xmlns:lne=
>              "urn:ietf:params:xml:ns:yang:ietf-logical-network-element"
>            xmlns:if="urn:ietf:params:xml:ns:yang:ietf-interfaces">
>        /lne:logical-network-elements\
>          /lne:logical-network-element/lne:root/if:interfaces
>      </path>

This helps. Can I also mount NACM into a mount point? If yes, what if
both are present?

/js

-- 
Juergen Schoenwaelder           Jacobs University Bremen gGmbH
Phone: [phone]         Campus Ring 1 | 28759 Bremen | Germany
Fax:   [phone]         <https://www.jacobs-university.de/>