Hi Jason,
On 13/10/2017 19:43, Sterne, Jason (Nokia - CA/Ottawa) wrote:
Hi all,
There are a few threads on the mailing list that touch on the concept
of system-controlled resources (mostly list entries):
https://mailarchive.ietf.org/arch/msg/netmod/3fTSHIh_MfHzmuDCoicAGiXA2E0
https://mailarchive.ietf.org/arch/msg/netmod/KIsSgKByQWpqYzA4i6Bwc8fuH3w
https://mailarchive.ietf.org/arch/msg/netmod/mjLJdiYErtNG41dJ5bJ5ji07cz0
A few drafts & RFCs also refer to the concept:
https://tools.ietf.org/html/draft-ietf-netmod-revised-datastores-04
https://tools.ietf.org/html/rfc7223
Several vendor implementations have list entries (instance data) that
are populated by the server and can be referenced (leafref) from other
places in the configuration. These system entries are useful
pre-created policies, interfaces, etc that can then be used (and
referred-to) by operators in their explicit configuration.
If those entries are only expected to exist in the <operational>
datastore, then in theory any references to them in user created
configuration will cause a validation problem in the candidate/running
(missing leafref target).
One solution discussed in the mailing lists is to change every
reference to lists that could contain a system created entry to a
“require-instance false” leafref. But then some useful validation is
lost. In many cases the model is more correctly “require-instance
true” but the set of targets includes the system create entries.
I agree that this is not a good general solution for system created
configuration that is always expected to exist on the device because
some of the useful validation is lost.
But I think that this solution does work well were the system created
entries are truly dynamic in nature, e.g. it seems to work well for the
topology YANG module where the topology may be explicitly configured,
but would more normally be learned dynamically from protocol
interactions, or perhaps be configured via a dynamic configuration protocol.
Another solution discussed is to have the system created entries
appear in the <intended> datastore (as part of template/expansion).
That would make validation pass on the intended datastore, but then
the candidate/running/startup datastores would not be valid (would be
missing leafref targets if any part of the config refers to system
created entries). THis sounds similar to the problem that has been
discussed in the past about the fact that templates (in the running)
basically mean the running/candidate aren’t necessarily valid (until
after template expansion, which means only the intended would be valid).
I think that this solution is OK, but not necessarily ideal.
As you say, it means that <running>/<candidate>/<startup> may not be
valid, which I see as quite a big down side. Longer term, I think that
it would be a good aim to allow the configuration to be validated off
the device, if a client desired to do so.
Another approach could be to actually have those system created
entries show up in running/candidate. That would ensure that
references to those entries are valid. But if the whole concept of
templates just cause the running/candidate to not be valid anyways
maybe we wouldn’t worry about the invalid aspect of references to
system created list entries ?
I know that quite a few implementations do this today, but I'm generally
not a fan of the system modifying <running>. It seems that an overall
architecture is much cleaner if <running> has a single source of truth
and hence can be exclusively controlled by the client.
But I also like the approach where the client (rather than the device)
explicitly writes these default entries into the configuration, if they
are referenced elsewhere by the configuration, to make the configuration
"complete". E.g. if part of the configuration references the loopback0
interface then also explicitly add the necessary loopback0 configuration
to instantiate the "loopback0" interface. When this configuration is
pushed to the device (i.e. using merge or replace operation semantics)
then the system should silently accept/ignore the explicit configuration
to create the loopback0 interface if it already exists on the system.
At the moment, IETF, and other SDOs are busy defining standard YANG
models, but for those models to end up being truly generic they also
need to have consistency about which bits of configuration are always
expected to implicitly exist on the device. E.g. considering the
example above of configuration referencing loopback0: if some systems
automatically create a loopback0 interface and others do not, then a
generic configuration needs to handle both scenarios.
If IETF standardizes YANG configuration templates, then perhaps it would
be good to investigate whether some of these "useful default system
properties" could instead be embodied into one or more standard device
templates? These templates could then be explicitly referenced in the
<running> configuration. This may allow <running> to be small, but
still allow it to be "complete" and able to be validated off the box.
Thanks,
Rob
Rgds,
Jason
_______________________________________________
netmod mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/netmod
_______________________________________________
netmod mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/netmod
