Re: [netmod] All IETF YANG modules MUST include revision-label statements

[tom petch <[e-mail]>]

From: netmod <[e-mail]> on behalf of Andy Bierman <[e-mail]>
Sent: 31 March 2020 17:19

<tp>
Agreeing with Andy that BC/NBC is a bad idea but from a different standpoint.  In general the concept sounds great but turning it into an unambiguous, usable by engineers test is mostly impossible as a number of discussions have shown.  Does it mean I can write a client that copes with old and new  server? or write a server that copes with old and new client? That has been discussed with no resolution.
What if the intended meaning of the WG was clear but someone later finds a loophole and interprets it differently and the module is revised to be unambiguous?  BC for most but NBC for those who found a different meaning.
Much depends on the original specification.  It may say, in a routing protocol,
Unrecognised TLV must be discarded
Unrecognised TLV must be treated as a fatal error
Unrecognised TLV must be forwarded unaltered
but many specifications are not that forward looking.  It is rare for a YANG module to capture that forward thinking.  
I think that BC is a very seductive concept that leads onto the rocks.

Tom Petch

On Tue, Mar 31, 2020 at 8:37 AM Kent Watsen <[e-mail]<mailto:[e-mail]>> wrote:
[replying to Reshad as well]

Hi Rob,

My impression is that Semver 2.0.0 works fine if you can always force clients to move to the latest version of the API whenever any bugfixes are made to the API (whether they are BC or NBC).  This is a natural fit for open source projects, but not so great for long life paid support contracts.

Agreed.


The goal of YANG semver is not to facilitate release branching.  It is to allow vendors to fix YANG modules without forcing clients to update to the latest version of that YANG module (which may contain other unrelated NBC changes and have lots of dependencies on other modules).

This is what Reshad was pointing to as well.  I’m very familiar with the issue, from my Juniper days, where there were all sorts of patch and (gasp) customer special releases, either of which could introduce any number of NBCs.

The background, of course, is that [very important] customers have working/validated infrastructure running a specific release and simply cannot tolerate any change beyond the very specific one they need *NOW*

I get it, truly,  but I feel that the ‘m’ / ‘M’ suffixes are both inconsistent with general understanding and insufficiently to express what is needed.



+1

I also find the granularity of NBC info to be mostly worthless at the module level.
There is no difference between a 1 leaf bugfix and a complete rewrite of the module.
Let's say 1 leaf "type string" needs to be changed to add "length 1..max".
This reduces the value set for 1 leaf by 1 value.

This flags the entire module as NBC and you would bump the major revision number.
The entire premise that one can decide if it is safe to upgrade based on the version string is flawed.


A possible fix might be to allow for <major>.<minor>.<patch>[-<anystring>], thereby enabling vendors to encode any format off a base release…and rely on inspection of the “revision” history indicate if/when NBC changes occurred.

But then I question (again) the need for the simplified format at all, as opposed to just using revision dates.  For instance, if <anysting> represents a long history of NBCs, that they were based on some source M.m.p starts to lose relevance.

Is the expectation that the vendor's module versions will use <major>.<minor>.<patch> values mimicking their release numbers?  For instance, would FooBar OS version 20.1.2 implement YANG module "foobar@20.1.2”?    I can see product mangers pushing for this, but then are companies (like Juniper) that use alternate release name-formatting strategies disadvantaged?  How is that fair?   To thwart this, would the WG be willing to assert that the history MUST start at 0.0.0 and MUST only monotonically increment values?


Note that OpenConfig also hit this problem, but they proposed a different solution.  I..e. ship the base module with another module that contains deviations to fix any bugs in the base module.  Alas this completely decouples the real module history from any revision-date/version number contained in the module, since to really understand the version of the module you also need to know the set of associated patch modules containing any deviations to the base module.

I’d need to see an illustration of this to be sure I understand, but my first impression is that it is yet another attempt to fit a square into a circle.



I don't have a solution proposal, but it would be great if a vendor could issue a patch
to a standard module which says "this is the standard module plus these known Errata ".
OK if this is in the form of deviations

In the end, I see no substitute to relying on “revision” history which 1) perfectly tracks branching history and can flag if/when NBC changes occurred.


Agreed


Kent // contributor




Andy