Ladislav Lhotka <[email protected]> wrote: > > > On 11 Jan 2017, at 13:53, Martin Bjorklund <[email protected]> wrote: > > > > Ladislav Lhotka <[email protected]> wrote: > >> > >>> On 11 Jan 2017, at 13:27, Martin Bjorklund <[email protected]> wrote: > >>> > >>> Ladislav Lhotka <[email protected]> wrote: > >>>> > >>>>> On 11 Jan 2017, at 12:16, Martin Bjorklund <[email protected]> wrote: > >>>>> > >>>>> Ladislav Lhotka <[email protected]> wrote: > >>>>>> > >>>>>>> On 11 Jan 2017, at 10:36, Martin Bjorklund <[email protected]> wrote: > >>>>>>> > >>>>>>> Ladislav Lhotka <[email protected]> wrote: > >>>>>>>> > >>>>>>>>> On 10 Jan 2017, at 09:39, Juergen Schoenwaelder > >>>>>>>>> <[email protected]> wrote: > >>>>>>>>> > >>>>>>>>> On Tue, Jan 10, 2017 at 09:20:36AM +0100, Ladislav Lhotka wrote: > >>>>>>>>>> > >>>>>>>>>> I think we need protocol and YANG specs that are not tied to any > >>>>>>>>>> particular model and that are thus capable of matching unforeseen > >>>>>>>>>> real-world implementations. This is no sci-fi, HTTP and XML schema > >>>>>>>>>> languages work this way. > >>>>>>>>>> > >>>>>>>>> > >>>>>>>>> I disagree that HTTP and XML schema languages do the same thing. Our > >>>>>>>>> goal is interoperable configuration of network devices; the notion > >>>>>>>>> of > >>>>>>>> > >>>>>>>> Even now, a client that's programmed to write straight to running > >>>>>>>> isn't interoperable with a server that has candidate and read-only > >>>>>>>> running. A RESTCONF server that supports only JSON isn't > >>>>>>>> interoperable > >>>>>>>> with a client that supports only XML. > >>>>>>>> > >>>>>>>> We are not in a situation that every pair of a randomly chosen server > >>>>>>>> and client need to be interoperable. It's IMO perfectly fine if IoT > >>>>>>>> and ISP networks use different clients. Yet, both can still use the > >>>>>>>> same RESTCONF, same YANG, and even same YANG modules. > >>>>>>> > >>>>>>> The fact is that that data models are written with a certain set of > >>>>>>> protocol features and datastores in mind (the "meta-model"). Some > >>>>>>> examples: > >>>>>>> > >>>>>>> If we had an "operational-state" datastore like the one proposed, we > >>>>>>> would not see the /foo vs /foo-state split. > >>>>>> > >>>>>> Yes, but I assume this will go away anyway. However, we can still have > >>>>>> YANG modules (and complete schemas) designed for the operational > >>>>>> datastore. The important property of the "meta-model" so far has been > >>>>>> that config and state data are separate, and this is not going to > >>>>>> change. > >>>>>> > >>>>>>> > >>>>>>> If SNMP would have had a CREATE operation, MIBs would not have used > >>>>>>> RowStatus. If NETCONF didn't have a way to create instances, we would > >>>>>>> have seen something similar in YANG models. > >>>>>>> > >>>>>>> If NETCONF had a way to add comments to any node in a datastore, we > >>>>>>> wouldn't have "leaf description" sprinkled throughout the models. > >>>>>>> > >>>>>>> If NETCONF didn't have a generic way to filter retreived data, we'd > >>>>>>> see lots of specific get-* rpcs in YANG models. > >>>>>> > >>>>>> Maybe, but are the last three points relevant to this discussion? > >>>>> > >>>>> The point is that data models are designed with some meta-model in > >>>>> mind. The meta-model includes (some) datastores. You wrote: > >>>> > >>>> But where and how is this reflected in existing YANG modules (except > >>>> for the foo and foo-state split, which is IMO a minor issue)? > >>> > >>> I don't this split is a minor issue. For the openconfig group, this > >>> is one of the major problems with YANG, leading to their design with > >>> duplicate leafs. The reason for adding the operational state > >>> datastore in the form we propose in the draft it to be able to get rid > >>> of this split. > >>> > >>>>> I believe both the protocols and YANG can work with any set of > >>>>> datastores [...] > >>>>> > >>>>> And I don't think that this is true (practically). For example, a > >>>>> YANG module that is designed with the new operational state datastore > >>>>> in mind will be of limited use in a legacy NETCONF server. > >>>> > >>>> Please explain. > >>> > >>> If a YANG module is designed with this new architecture in mind, it > >>> will have a single top-level tree, which can support pre-configuration > >>> and different instances in the config and operational state. > >>> > >>> If such a module is implemented in a legacy NETCONF server, the only > >>> way to get the operational state is to used <get/>. But <get/> will > >>> return the union between running and operational state. The client > >>> can't tell if an instance is really present in the operational state, > >>> or just in the config. > >>> > >>> > >>> My idea what could be done e.g. with ietf-interfaces > >>>> is this: > >>>> > >>>> 1. Split it into two modules, say ietf-interfaces-config and > >>>> ietf-interfaces-state. The former would contain exactly what's now > >>>> inside "interfaces", and the latter will augment it with extra state > >>>> data that are now under "interfaces-state". > >>>> > >>>> 2. The data model for configuration datastores will be defined to > >>>> contain only ietf-interfaces-config whereas for operational-state > >>>> datastore it will be ietf-interfaces-config *and* > >>>> ietf-interfaces-state. > >>> > >>> If we do this for all modules then we haven't gained anything; we > >>> still have duplicate definitions. > >> > >> Show me a single YANG data node definition that's duplicate in my > >> concept above. But then maybe I didn't explain it properly. > > > > The interface's "type" leaf. With the new operational-state > > datastore, /interfaces/interface/type in operational-state and > > /interfaces-state/interface/type are duplicate. > > As I said, ietf-interfaces-state state would consist of augments > containing extra state nodes (i.e. those that are not in > configuration). So "type" won't be there.
So how would a client learn the type of a system-controlled interface? > >> Note also that you slightly misinterpreted my statement that you > >> cited: > >> > >> I believe both the protocols and YANG can work with any set of > >> datastores [...] > >> > >> I didn't say that there cannot be *modules* that are somehow designed > >> for a particular datastore model - I meant YANG the language. > > > > Ok. Yes, you're right, but then we'd probably need some new statement > > in each module that tells which meta-model the YANG module is written > > for. > > I would prefer to have it as state data, basically separate YANG > libraries for configuration datastores and operational-state. But the use case is that a particular module is designed for a certain datastore model (which you wrote above). This is a design-time property, not a rum-time property, so state data (run-time) is not the right solution. If a module is designed for a certain datastore model, that module cannot be implemented in a server with some other datastore model. /martin _______________________________________________ netmod mailing list [email protected] https://www.ietf.org/mailman/listinfo/netmod
