Re: [i2rs] Ignas Bagdonas' Discuss on draft-ietf-i2rs-yang-dc-fabric-network-topology-08: (with DISCUSS and COMMENT)

["Susan Hares" <[e-mail]>]

Ignas: 

I am not trying to clarify the specifics.  This response (as I mentioned), will come from the authors.  As a shepherd/WG chair, I am asking for information regarding the basis of your DISCUSS models for specific points.    

The 2014 document on the IESG discuss criteria is at: 
https://www.ietf.org/blog/discuss-criteria-iesg-review/

What on this list does the following comment refer to: 
"Why DISCUSS? DC fabric is a type of network topology, yes, it has some specifics, but nothing radically different than any purpose built network topology. Developing a separate model for a specific use case at the same time when there is generic and extensible TE model is questionable." 

Perhaps you are not considering the fact this is an I2RS model.  Let me provide 3 comments regaring that point: 
 
1st - I2RS is focusing on models that are capable for the dynamic state and configuration state.  These models have qualitative differences.  The mechanisms of a model which does both dynamic state and configuration state is qualitative different that the basic TE models. This model extends the TE models toward this approach (see   module ietf-dc-fabric-topology reference import of ietf-network-topology).  
  
2nd - During the I2RS process we talked to the TE topology authors and the TE WG.  We agreed that this model has differences.  As a WG Co-chair, I spent time reviewing this interaction. 

3rd - How many of the user community have implemented I2RS dynamic models or the RFC version of the TE models?  

See the comments from Chris Hopps and others on slow pace of the netconf work.  If you are going to restrict to two deployed implementations, you will be joining the IDR camp of requirements and slowing the work further.  The only reason we require 2 implementations for IDR is for the fragile BGP environment and that operators request it due to the global consequences.  Network Management of these early yang models have a much more restricted case. 

Sue Hares 


-----Original Message-----
From: i2rs [mailto:[e-mail]] On Behalf Of Ignas Bagdonas
Sent: Wednesday, April 4, 2018 10:31 AM
To: Susan Hares; 'The IESG'
Cc: [e-mail]; [e-mail]; [e-mail]
Subject: Re: [i2rs] Ignas Bagdonas' Discuss on draft-ietf-i2rs-yang-dc-fabric-network-topology-08: (with DISCUSS and COMMENT)

Hi Sue,


On 03/04/2018 14:59, Susan Hares wrote:
> Ignas:
>
> Yan will answer for the authors but I would like to share some information related to the I2RS working group reviews.  In your response, please specify why each question is a "DISCUSS" quality question rather than a "Comment" question.  The authors and I (as the shepherd) will work to resolve both DISCUSS and comment issues.
>
> Let me review only 5 of your many points because they are pointing in a direction which is different from earlier QA reviews of this document (rtg-dir, ops-dir, yang-doctors) in the 2017-2018 timeframe.

> 1st - Why TE topology model is not sufficient for modelling the representation of DC fabric? Why is DC fabric network topology special compared to any generic fabric based topology?

Why DISCUSS? DC fabric is a type of network topology, yes, it has some specifics, but nothing radically different than any purpose built network topology. Developing a separate model for a specific use case at the same time when there is generic and extensible TE model is questionable.

> This document was reviewed by authors with the TE topology models to make sure there was no conflict or duplication.
>
> Your question implies that only one yang model is appropriate for each type of fabric.

That is exactly opposite. What is special about DC fabric that it has to have a separate model? What is special about fabric type of topology that it has to have a separate model? Why is TE model not suitable?

>     This theory of one yang mode per fabric does not apply to dynamic (ephemeral) datastore versus configuration datastore models.  It is also not true of all models even within the configuration datastore.
> Since there is a yang catalog and selection of yang models is specific to a implemented, there has been no early winnowing of the yang models per type.  If you are insisting on this theory of "one yang model" per fabric type, please provide an RFC reference so that I can help review this DISCUSS criteria with the authors.
>
> This yang model has been implemented by 1 vendor, and there was interest by other vendors.  A deployment target has been identified for this model, and feedback is expected from the users.
Excellent. Please get feedback from user community - even if it is not yet implemented and operations groups will not be able to provide feedback, architecture and engineering groups look into upcoming things and will have what to say.

Speaking of implementations, the ODL faas project (from where the majority of this model seems to be coming from) deals with an instance of overlay that is subsequently treated as an underlay, and that is different that the underlay on top of which that instance is being run. 
If the model focus is on the "fabric as a service" type of topologies then it explicitly needs to state that, and then justify why physical node properties exist together with logical instance properties in that case.


> If you are asking this model to cover three-four layer datacenters, this approach is opposite some of the initial feedback to the group to keep the initial model - that is to keep it simple and restricted to 2 layers in order to test the concepts.  If you are asking to provide text (in introduction or appendix) that indicates the initial focus, this can be added.
The document as it is written now tries to cover every possible fabric. 
If the scope is intended to be narrower - it needs to be stated. 
Starting from bounded scope is certainly a right thing to do but that is not how the document reads now.


> 2nd - Multiple layers and multiple roles.
Why DISCUSS? Two stage fabrics and fabrics with a perfectly clean node 
role separation do indeed exist, but that is not necessary a common 
deployment model. The document assumes that those are the only possible 
options.

>   The authors provide slides in several meetings I2RS meeting repository regarding this point.
> The initial feedback suggested reducing the "why" text within the draft.   Again, the initial feedback was to reduce the initial model's text to 2 layers and simple "whys".  See proceedings from IETF 95 forward on I2RS on fabric data model for discussions.
Would users of this model also be required to lookup proceedings of past 
IETF meetings in order to understand whether it may fit their use cases?


> 3rd - The authors will comment on the port restrictions.  Early feedback during the I2RS meetings from vendors may have taken the authors down this path.  In my review, I expect major issues in this area - but I will let the authors comment.
Why DISCUSS? The way how the model specifies port speeds is conflicting 
with common deployment practices.

>   4 - policy is simple.
>
> Again, the initial feedback was to keep initial policies simple and gain feedback from the deployments.

Why DISCUSS? What kind of policy is being discussed here? The assumption 
of one single universal policy fitting all deployments and use cases 
contradicts to operational reality.

"Policy is simple" does not clarify what kind of policy it is.

> 5 - You indicate that the document requires a "major" rewrite clarifying the logic.

Why DISCUSS? Model tries to prescribe a way how all DC networks should 
be built. It intermixes concepts of underlay and overlay. There are 
nodes in the model with unclear purpose and no documented details on 
what and how they are doing.

> Earlier feedback (rtg-dir, ops-dir, yang-doctors) on YANG suggested taking out the lengthy descriptions regarding logic and history.  If we are switching the rules for the YANG models, would you please update the requirements for the YANG models so that shepherds, rtg-dir, ops-dir, and yang-doctors can have rules for review clearly spelled out.

YANG models, and any other deliverables of the IETF, are targetted to 
the users of those deliverables and not necessary to the IETF itself. 
The situation with YANG models is that the main consumer of IETF YANG 
model for a noticeable period was  IETF itself - it was required to 
build the sufficient coverage of models for them to be practically 
useful. We as an industry start to see more practical use of YANG 
modules, and so far the main obstacle for YANG acceptance is the 
difficulty in trying to use it. It is incorrect to assume that outside 
of the IETF WGs that deal with developing the models there is enough of 
understanding of the reasoning behind modelling decisions made. It is 
incorrect to assume that potential users of such models would try to 
lookup proceedings of past IETF meeting trying to get answers - they 
will chose other manageability technologies instead. YANG models need to 
be self-contained from the practical usability perspective - the models 
themselves should contain enough and meaningful descriptions of the 
nodes that it would not raise questions for users trying to deploy those 
models. Descriptions equivalent to those found in command line 
interfaces - if YANG is expected to become a new command line interface, 
it should be no worse than the command line interface. The reasoning 
behind modelling decisions made also need to be documented - at least 
for allowing model users to see whether the model is suitable for 
deployment in the particular environment. As YANG is maturing and 
starting to be deployed, naturally the focus of reviews will change to 
reflect what is required for successful deployment of the technology.

> Summary on Shepherd's comment:
>
> The authors will respond to others specifics, but in order to guide these diligent authors - I need to know what rules you are setting for the 2018 IESG approval of YANG models.  If you are placing a DISCUSS on a YANG model based on a set criteria, the criteria needs to be published on a web page or in an RFC. If I've missed this criteria that the OPS Area has specified,

RFC6087 and draft-ietf-netmod-rfc6087bis.

There are two parts that are important for reviews - the model itself, 
and how the model applies to the managed entities. And there is nothing 
new in the review criteria. The former is rather not that complex, and 
typically can be done within IETF itself. The latter is more complex and 
generally would require feedback from the target users of the model. 
There is a balance between a model being too generic to be practically 
usable and model being too prescriptive to be practically usable. If the 
model puts requirements and restrictions on the managed entities in a 
way that requires to build those managed entities in a specific way, 
predefined by the model authors, the value of such model is 
questionable. Speaking specifically about DC fabric model, it puts 
network design prescriptions that are significantly misaligned with how 
fabric based networks have been and are built. Yes, it is possible to 
find environments where the model would apply directly and with no 
impact, but one would need to look for such deployments quite hard, and 
with a high probability that would be proof of concept or technology 
demonstration type of environments.

IETF is good at developing technology components and fragments, IETF 
typically is not good at dealing with network design and how those 
fragments need to be bound together - that is the reality, and that is 
not necessarily wrong. IETF should be focusing on what it can do best - 
the fragments, and align with users of the fragments on how to improve 
the fragments but not try to direct how users should be building their 
networks. It is important for the reputation of IETF as a credible SDO - 
if IETF manageability mechanisms propose and enforce not necessarily 
right - or just plain broken - network designs, that is a reputation 
problem. This document tends to be proposed standard, and that sets a 
strong message.

Ignas


> Thank you for your review,
> Susan Hares
>
> -----Original Message-----
> From: Ignas Bagdonas [mailto:[e-mail]]
> Sent: Tuesday, April 3, 2018 7:40 AM
> To: The IESG
> Cc: [e-mail]; Susan Hares; [e-mail]; [e-mail]; [e-mail]
> Subject: Ignas Bagdonas' Discuss on draft-ietf-i2rs-yang-dc-fabric-network-topology-08: (with DISCUSS and COMMENT)
>
> Ignas Bagdonas has entered the following ballot position for
> draft-ietf-i2rs-yang-dc-fabric-network-topology-08: Discuss
>
> When responding, please keep the subject line intact and reply to all email addresses included in the To and CC lines. (Feel free to cut this introductory paragraph, however.)
>
>
> Please refer to https://www.ietf.org/iesg/statement/discuss-criteria.html
> for more information about IESG DISCUSS and COMMENT positions.
>
>
> The document, along with other ballot positions, can be found here:
> https://datatracker.ietf.org/doc/draft-ietf-i2rs-yang-dc-fabric-network-topology/
>
>
>
> ----------------------------------------------------------------------
> DISCUSS:
> ----------------------------------------------------------------------
>
> I have concerns about the practical usability of this proposed model as it is specified now.
>
> The intended decoupling of fabric implementation properties (what is termed as "underlay network infrastructure" in the document) and its topology seems to be contradicting to general operational practices of fabric based networks. It is generally true for the context of the overlay but that is not what the document seems to be focusing on. Fabric defines and implements the underlay, not the other way around.
>
> The document does not contain a sufficient description of the logic of the model itself, the reasons for choices made for representation of types and attributes, and at the same time descriptions in modules are single lines that do not add clarification beyond being copies of leaf names. Either there needs to be a section that describes the logic of the model and how it relates to other models, also including examples, or module description fields need to have enough content to be able to have equivalent understanding of model intent and operation. Both are strongly encouraged, as descriptions have value of itself for being a reference for use, and model description is needed for understanding how this particular model fits into the larger hierarchy. Network management does not end at the boundary of the single domain-specific model, it is important to build it into a whole system.
>
> Why TE topology model is not sufficient for modelling the representation of DC fabric? Why is DC fabric network topology special compared to any generic fabric based topology?
>
> How this model could be used for representing more than two stage fabrics that are in wide deployment?
>
> Limiting port bandwidth to a fixed rate is too restrictive. The model as specified already does not cover a set of port speeds that are in deployment.
>
> How would a device that has more than a single role in the fabric be represented?
>
> Service capabilities as they are described belong to the overlay context while they are called device capabilities. Are those the only possible service capabilities? What is the effect of configuring those capabilities?
>
> What is compose-fabric RPC? The document does not define any RPCs.
>
> What is policy driven traffic behavior? Is there the only one policy that fits all possible deployment scenarios?
>
> Looking at the history of the document from the individual submission time and the comments received, it seems that the point fixes to the text went in to cover the specific comments but not to address the broader scope of comments.
> The document would definitely benefit from a major rewrite clarifying the logic behind the decisions made, aligning more with the operational practice of fabric based network design and deployment, and bringing the content in YANG modules to be self-describing.
>
>
> ----------------------------------------------------------------------
> COMMENT:
> ----------------------------------------------------------------------
>
> Fabric and POD are not equivalent terms.
>
> I2RS use case requirements document has expired 11 months ago. Use cases documents are good for tracking the work progress of specification documents, it is questionable whether standalone use cases documents provide value beyond historic record. Is the reference to I2RS use cases document really needed?
>
> What is atomic network?
>
> VLAN is not a fabric building technology as such, while Ethernet is.
>
> What is the need for VNI capacity leaves? What is their effect if configured?
>
> The document intermixes ietf-fabric-* and ietf-dc-fabric-* namespaces.
>
> Serial port-type is present while Infiniband is not - Infiniband based fabrics are widely deployed. What is the extensibility mechanism for adding in new port types?
>
> Is there any deployment experience with this model? The ODL faas project hasn't got much activity over last two years. Are you aware of any other implementations or deployments?
>
>
>

_______________________________________________
i2rs mailing list
[e-mail]
https://www.ietf.org/mailman/listinfo/i2rs