-----邮件原件-----
发件人: Juergen Schoenwaelder [mailto:[email protected]] 
发送时间: 2020年7月27日 18:42
收件人: Qin Wu <[email protected]>
抄送: [email protected]
主题: Re: [netmod] The NETMOD WG has placed draft-tao-netmod-yang-node-tags in 
state "Candidate for WG Adoption"

On Mon, Jul 27, 2020 at 10:13:43AM +0000, Qin Wu wrote:
> 
> On the technical side, it is not clear why a central list of tags providing 
> meta information is preferrable over metadata annotations that can be shipped 
> with the values.
> 
> 
> [Qin]: If we take metadata annotation approach, we need to define new module 
> for each device level modules such as interface module, IP management module, 
> BGP module, QoS module deployed in the device or revise these modules that 
> has already been published and update devices with these new modules 
> corresponding to deployed device level modules. Suppose the device level 
> modules that has been deployed in the device is huge, such metadata 
> annotation solution require update on each device for the new modules, the 
> number of new modules is same as the number of existing deployed device level 
> modules, which doesn't scale well.
> 

I do not think this answer is technically correct. For a counter example, see 
the definition of md:annotation origin in RFC 8342.

[Qin]:umm, have a second thought, I think this is the design choice we made, we 
did investigate metadata annotation approach, but the essence of metadata 
annotation is to add attribute to the model, we think adding attributes still 
have limitation, since you can not add too many attributes one time to the same 
xml element.
Secondly, we need to know which data is the characteristics data or telemetry 
data category the client want to subscribe before sending subscription request 
to the server. If we take metadata annotation approach, we need to have at 
least two times subscriptions before finding targeted data objects, or first 
retrieve telemetry data tag from the server and then subscribe targeted data 
object. It is hard to automate subscription with metadata annotation approach.

> It does not seem harmful to adopt this work if people think this is needed 
> and they are willing to work on this. The others can safely ignore this with. 
> Personally, I am not convinced yet, but then I do not work on 'multiple 
> dimensional network visibility analysis'.
> Perhaps a good convincing concrete example would help, the tables in Fig. 2 
> do not help me much.
>
> [Qin]: Suppose we can classify telemetry data from different angles, 
> we can tag the telemetry data with various different data object tag, 
> e.g., whether the data object is performance metric data, which category the 
> performance metric data belong, if the data object Is performance metric data 
> or KPI data, we can further classify KPI data from operation type, e.g., 
> whether KPI data is average value, minimum value, maximum value, whether KPI 
> data allow set threshold for it, if the KPI data can come from multiple 
> source, We can indicate whether KPI data is aggregated data or not, so 
> collector can use these telemetry data tag to subscribe targeted data object 
> or these telemetry data tag associated with subscribed data can be further 
> consumed by big data analytics platform/open observability platform such as 
> InfluxDB/grafana and provide different outputs based on the telemetry tag or 
> telemetry data classification that is selected.
> 
> The goal seems to add tags to node instances. I do not understand what they 
> are called 'self-explanation-node-tags' and not simply 'node-tags'.

You assume that the device manufacturer implementing YANG modules can predict 
what a management application wants to do with a certain piece of information. 
I doubt this is generally true. And if you look at basic counters, they often 
serve many different purposes. Even if the WG believes this to be true, there 
likely needs an overwrite mechanism in case the manufacturer predicted things 
wrongly.
[Qin]: no, it is not always necessary for device manufacturer to do such 
prediction, it can be put into offline document together with module schema for 
NMS to retrieve. In addition, it allows the device adding or removing tag in 
the same as module tag does. These tag changes can be advertised Using server 
capability defined in draft-ietf-netconf-notification-capabilities to inform 
the client about it.

> [Qin]: The reason to call it as 'self-explanation-node-tag' is to make it 
> better reflect what it is really aimed for, if you have better name, I am 
> happy to accept it.

Well, to me these look like node instance tags. If anything, I would be 
interested in a generic solution and not something taylored to one specific use 
case (telemetry) and this is why I would prefer to have the specific semantics 
in the tags and not in the set of leafs carrying the tags.

[Qin]:We did target generic solution, telemetry is not specific case, but it is 
the main case. If you think there are other cases, we are happy to cover it. 
The semantics for each node tag has been defined within the module, semantics 
can be polished step by step.
Leaf carrying the tag helps data consumer such as influxdb/grafana to further 
classify the telemetry data and provide better network visbility.
Leaf carrying the tag approach has been widely used by many IETF work such as 
draft-ietf-netmod-module-tags, draft-ietf-netmod-yang-module-versioning-01.
For other reasons, please see the clarification mentioned above.

/js

-- 
Juergen Schoenwaelder           Jacobs University Bremen gGmbH
Phone: +49 421 200 3587         Campus Ring 1 | 28759 Bremen | Germany
Fax:   +49 421 200 3103         <https://www.jacobs-university.de/>
_______________________________________________
netmod mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/netmod

Reply via email to