On 19/09/2017 10:55, Martin Bjorklund wrote:
Robert Wilton <rwil...@cisco.com> wrote:
On 18/09/2017 19:25, Andy Bierman wrote:
On Mon, Sep 18, 2017 at 11:07 AM, Martin Bjorklund <m...@tail-f.com
<mailto:m...@tail-f.com>> wrote:
Juergen Schoenwaelder <j.schoenwael...@jacobs-university.de
<mailto:j.schoenwael...@jacobs-university.de>> wrote:
> On Mon, Sep 18, 2017 at 05:17:46PM +0100, Robert Wilton wrote:
> >
> > > No. I do not agree that the MUST in RFC 7950 can be removed.
> > > I do not agree the architecture should update YANG at all.
> > OK.
>
> I am with Andy here. <running> has always had the requirement to be
> valid and we are not supposed to change that. Mechanisms for
inactive
> configuration or templating must be designed to be backwards
compatible
> I think.
Ok. If we keep the requirement that <running> in itself must be
valid, it just restricts the usefulness/expressiveness of inactive and
template mechanisms, but it might be ok.
I think that even w/o this requirement, the observable behavior for a
client can be backwards compatible. For example, suppose we have an
inactive access control rule that refers to a non-existing
interface in
<running>. If a client that doesn't know anything about inactive asks
for the contents of <running>, our implementation removes the inactive
nodes from the reply to the client. Only a client that opts-in to
inactive will receive a reply with things that looks invalid if you
don't take the inactive annotation into account.
There are many ways that validation can fail because inactive nodes
are present,
and considered part of the validation.
e,g, min-elements, max-elements, mandatory, unique.
I think we all agree that validation on the enabled nodes is supposed
to continue to work.
Yes.
Here is an attempt at a backwards-compatible solution:
1) current <get-config> and <get> responses only include enabled
nodes.
2) old-style <edit-config> operations do not alter inactive nodes
3) NMDA clients use <get-data>, not <get-config> or <get>. These
responses
include enabled and disabled nodes, so validation does not apply
for <running>
4) new style <edit-config> (i.e. <datastore> parameter used) can alter
any type of data node
//I think that inactive should always be an optional extension. Not
everything needs the additional complexity.
Hence rather than tying this function to specific NETCONF operations,
I would suggest that there should be an extra parameter (like for
with-defaults) to allow a client to indicate to the server that a get
or edit request is using the "with-inactive" extension.
1) The server should also have a capability (or perhaps a leaf defined
in YANG library) to indicate that it supports inactive configuration.
2) If a client doesn't use the extra "with-inactive" parameter during
a get request then only active nodes are returned.
3) If a client doesn't use the extra "with-inactive" parameter during
an edit-data request then the request cannot include an extra inactive
meta-data. The request is processed in a way that is equivalent to an
existing NETCONF implementation, but it may unknowingly remove some
inactive configuration (e.g. via a replace or remove operation on an
inactive node). Operations like create, delete, none, replace should
all treat an inactive target node the same way as in the node doesn't
exist (e.g. delete on an inactive node would return a "data-missing"
error), this ensures that the behaviour that an unaware client
observes is the same as the existing behaviour that it would expect
from a regular 6241 compliant NETCONF implementation.
4) It a client makes a get request including the "with-inactive"
parameter then they also get the inactive nodes as well, marked with a
meta-data annotation.
5) If a client makes an edit request including the "with-inactive"
parameter, then the inactive meta-data annotation can be used to label
inactive nodes. Inactive nodes are regarded as regular data nodes for
create/delete/replace/none operation error checking.
I think that this approach is similar (perhaps even the same) as
Martin described.
This is indeed how our implementation works (except I think we don't
do 5; if the client sends an "inactive" attribute it doesn't have to
also send with-inactive).
Note that the YANG MUST rule still applies, because validation is only
done on enabled nodes.
It is only the response message representations that cannot be
validated, not the contents
of <running> on a server.
So the question is how we can make sure that the text in the NMDA
draft covers this yet-to-be-defined feature w/o having to define it
now? We thought that the current text was sufficient, but do we have
to make any changes to it?
1) Do we also need to illustrate a similar proof of concept templating
implementation to show that templating could work whilst keeping running
valid? I would prefer that this is just deferred to whichever draft
defines templating.
2) I think that we need to reach a decision as to whether the NMDA
architecture needs to explicitly state that <running> is always valid,
or if that can be left to the existing statement in 7950. My thinking
is that if the conclusion is that <running> must always be valid, then
it would be helpful to explicitly state it the descriptions of both
<running> and <startup> in the NMDA architecture.
3) I think that it would be useful to further refine my previous
proposed text for intended, particularly rewriting the text on
validation. This should hopefully also address Balazs' concern about
default values be included in validation.
<Old>
4.4. The Intended Configuration Datastore (<intended>)
The intended configuration datastore (<intended>) is a read-only
configuration datastore. It is tightly coupled to <running>. When
data is written to <running>, the data that is to be validated is
also conceptually written to <intended>. Validation is performed on
the contents of <intended>.
For simple implementations, <running> and <intended> are identical.
<intended> does not persist across reboots; its relationship with
<running> makes that unnecessary.
Currently there are no standard mechanisms defined that affect
<intended> so that it would have different contents than <running>,
but this architecture allows for such mechanisms to be defined.
One example of such a mechanism is support for marking nodes as
inactive in <running>. Inactive nodes are not copied to <intended>,
and are thus not taken into account when validating the
configuration.
Another example is support for templates. Templates are expanded
when copied into <intended>, and the expanded result is validated.
</Old>
<Proposed>
4.1.4. The Intended Configuration Datastore (<intended>)
The intended configuration datastore (<intended>) is a read-only
configuration datastore. It represents the configuration after all
configuration transformations to <running> are performed (e.g.
template expansion, removal of inactive configuration), and is the
configuration that the system attempts to apply.
<intended> is tightly coupled to <running>. Whenever data is written
to <running>, then <intended> is also immediately updated by
performing all necessary transformations to the contents of <running>
and then <intended> is validated.
<intended> may also be updated independently of <running> (e.g., if
one of the configuration transformations is changed), but <intended>
must always be a 'valid configuration data tree' as defined in
Section 8.1 of [RFC7950].
For simple implementations, <running> and <intended> are identical.
The contents of <intended> is also related to the 'config true'
subset of <operational>, and hence a client can determine to what
extent the intended configuration is currently in use by checking
whether the contents of <intended> also appears in <operational>.
<intended> does not persist across reboots; its relationship with
<running> makes that unnecessary.
Currently there are no standard mechanisms defined that affect
<intended> so that it would have different contents than <running>,
but this architecture allows for such mechanisms to be defined.
One example of such a mechanism is support for marking nodes as
inactive in <running>. Inactive nodes are not copied to <intended>.
A second example is support for templates, which can perform
transformations on the configuration from <running> to the
configuration written to <intended>.
</Proposed>
Thanks,
Rob
/martin
.
_______________________________________________
netmod mailing list
netmod@ietf.org
https://www.ietf.org/mailman/listinfo/netmod
