Re: [DNSOP] raising the bar: requiring implementations

[Mukund Sivaraman <[e-mail]>]

On Thu, Mar 29, 2018 at 03:18:45PM -0400, Suzanne Woolf wrote:
> 
> On Mar 28, 2018, at 11:19 AM, Mukund Sivaraman <[e-mail]> wrote:
> 
> > On Wed, Mar 28, 2018 at 10:55:13AM -0400, tjw ietf wrote:
> >> I would say that most things we have adopted in the past few years do have
> >> some implementations to reference.
> >> Not when drafts are adopted, but generally before we head to WGLC I've
> >> always wanted to see someone
> >> who implemented the option in some manner.
> >> 
> >> But yes, agree.
> > 
> > I'd raise the bar even higher, to see complete implementation in a major
> > open source DNS implementation when it applies. Sometimes implementation
> > problems are very revealing (client-subnet should have gone through
> > this).
> 
> This is actually quite a good example of another problem:
> Client-subnet was documented for Informational purposes; it was
> already in wide use in the public internet and had an EDNS opt code
> codepoint allocated from the IANA registry.
> 
> Nothing that happened in DNSOP or the IETF changed that, and I
> strongly suspect nothing that happened in DNSOP or the IETF could have
> changed it.

We found issues in the client-subnet description (in the draft stages)
that were corrected before it became RFC - this involved some behavioral
changes. However, the opportunity was not given to address fundamental
design issues, so what's in the RFC largely documents the
implementations preceding it and still has issues. I didn't think
client-subnet was in wide use when it came to dnsop. Even today it
doesn't look like it's in wide use.

You have asked an interesting question, because what happens if
something is already being used and there's no chance to update/correct
problems because that would make it a different protocol?

IMO, we should not put anything through dnsop that does not allow
reviewing and correcting problems. It is pointless for dnsop to shepherd
a draft to RFC when there's no possibility to influence it.

During later stages of the draft via dnsop, we implemented client-subnet
(resolver side) and found various problems. Some of these were addressed
in the draft, but it was revealing how poor the state of the
design/draft was. This is why I strongly suggest: A code contribution of
a _complete implementation_ of everything described in the draft to one
of the open source DNS implementations should come sometime during
adoption, so that

(a) issues in production implementation are revealed

(b) it can be tried in the real world to understand how it behaves.

As you're a co-chair:

When Bert did that Camel presentation, I felt hooray, here's one of us
who's talking about what the steady churn of ideas and RFCs is doing to
DNS implementations. We finish implementing one draft and at almost
breathless pace, a few more drafts queue up. It's not sustainable,
esp. as all the existing features also need to be maintained for years.

Introducing a new RR type that doesn't involve additional processing is
relatively trivial and nobody objects about it.

Introducing something like TCP pipelining, or rather, out-of-order query
processing, may seem trivial when describing the change, but
implementing it may need fundamental restructuring of the
implementation. Proper implementation of client-subnet needs change
everywhere within a nameserver from the client message parsing code to
the data structures used for cache, and how cache lifetime is managed.

Implementations are being stretched and bent 5 different ways to adapt
to the length and breadth of all these newfangled DNS items that
literally are showing up at an alarming pace.

Bert really hit the spot at the right time. Something needs to be done
to check what becomes an RFC. A good way is to require an open source
implementation. If draft authors cannot supply that or convince an open
source implementation to write support for it, it can go back to where
it came from.

		Mukund