Hello Stephen,
On 2012/06/12 18:59, Stephen Farrell wrote:
Hi Martin,
On 06/12/2012 10:13 AM, "Martin J. Dürst" wrote:
Hello Stephen,
This mail responds to your points on the main technical issue that I
have identified.
On 2012/06/05 20:11, Stephen Farrell wrote:
On 06/05/2012 10:42 AM, "Martin J. Dürst" wrote:
Hello everybody,
Major design issue:
The draft defines two schemes, which differ only slightly, and mostly
just gratuitously (see also editorial issues).
These are the ni: and the nih: scheme. As far as I understand, they
differ as follows:
ni: nih:
authority: optional disallowed
ascii-compatible encoding: base64url base16
check digit: disallowed optional
query part: optional disallowed
decimal presentation of algorithm: disallowed possible
I'll note in passing that the two schemes differ in all those
respects. You may disagree with our design, but basically you're
showing that the two differ in pretty much all possible ways
other than that both include a hash value.
The usability of URIs is strongly influenced by the number of
different
schemes, with the smaller a number, the better. As a somewhat made-up
example, if the original URIs had been separated into httph: for HTML
pages and httpi: for images, or any other arbitrary subdivision
that one
can envision, that would have hurt the growth and extensibility of the
Web. Creating new URI schemes is occasionally necessary, and the ideas
that lead to this draft definitely seem to warrant a new scheme
(*), but
there's no reason for two schemes.
[(*) I know people who would claim the the .well-formed http/https
thing
is completely sufficient, no new scheme needed at all.]
More specifically, if the original URIs had been separated into httpm:
(for machines) and httph: (for humans), the Web for sure wouldn't have
grown at the speed it did (and does) grow. In practice, there are huge
differences in human 'speakability' for URIs (and IRIs, for that
matter); compare e.g. http://google.com with
http://www.google.co.jp/#sclient=psy-ab&hl=en&site=&source=hp&q=hash&oq=hash&aq=f&aqi=g4&aql=
(which I have significantly shortened to hopefully eliminate potential
privacy issues), or compare the average mailto: URI with the average
data: URI. However, what's important is that there never has been a
strong dividing line between machine-only and human-only URIs or
schemes, the division has always been very gradual. Short and mainly
human-oriented URIs have of course been handled by machines, and on
the
other hand, very long URIs have been spoken when really necessary.
"Speakability" has been maintained to some extent by scheme designers,
and to some extent by "survival of the fittest" (URIs that weren't
very
speakable (or spellable/memorizable/guessable/...), and their Web
sites,
might just die out slowly).
It should also be noted that the resistance against multiple URI
schemes
may have been low because there are so many different ways to express
hashes in the draft anyway, and one more (the nih: section is the last
one before the examples section) didn't seem like much of a deal
anymore. But when it comes to URIs, one less is a lot better than one
more.
In the above ni:/nih: distinction, nih: seems to have been added as an
afterthought after realizing that reading an ni: URI aloud over the
phone may be somewhat suboptimal because there is a need for repeated
"upper case" - "lower case" (sure very quickly shortened to "upper" -
"lower" and then to "up" - "low" or something similar). It is not a
bad
idea to try to make sure that IETF technology, and URIs in particular,
are accessible to people with certain kinds of dislexya. (There are
indeed people who have tremendous difficulties with distinguishing
upper- and lower-case letters, and this may or may not be connected
with
other aspects of dislexya.) It is however totally unclear to this
reviewer why this has to lead to two different URI schemes with other
gratuitous differences.
Finding a solution is rather easy (of course, other solutions may also
be possible): Merge the schemes, so that authority, check digit, and
query part are all optional (an authority part and/or a query part may
very well be very useful in human communication, and a check digit
won't
hurt when transmitted electronically) and the decimal presentation of
the algorithm is always allowed, and use base32
(http://tools.ietf.org/html/rfc4648) as the encoding. This leads to a
16.6% less efficient encoding of the value part of the ni: URI, but
given that other URI-related encodings, e.g. the %-encoding resulting
when converting an IRI to an URI, are much less efficient, and that
URI
infrastructure these days can handle URIs with more than 1000 bytes,
this should not be a serious problem. Also, there's a separate binary
format (section 6) that is more compact already.
I strongly disagree with merging ni& nih. Though that clearly
could be done, it would be an error.
There was no such comment on the uri-review list and the designated
expert was happy. That review was IMO the time for such comments
and second-guessing the designated expert at this stage seems
contrary to the registration requirements. So process-wise I
think your main comment is late.
First, if IETF Last Call is too late to make serious technical comments
on drafts, then I think we have to rename it to IETF Too-Late Call.
Second, designated experts are there to check for minimum requirements
for a registration, and to give advice as they see fit (and have time).
I'm myself a designated expert on "Character Sets", and I have
definitely in the past approved, and would again in the future approve,
registrations for stuff on which I would complain strongly if the
question was "is this a good technical solution".
Graham Klyne, the designated expert for URI scheme registrations, has
confirmed offline that he does not see his role as "expert reviewer" as
judging the technical merit of a URI scheme proposal.
While that's fair enough. Its also fair to note that there was
discussion of the this document on the uri-review list but this
aspect was not raised at all. That list is called "uri-review"
and from its archives it does seem to frequently do more than
just check the paperwork (including quite a few mails from you:-).
Oh well. I'm not sure how familiar you are with the IETF :-), but it's
basically all a volunteer organization. So even more than in an
organization where people are getting paid, some things happen to get a
lot of attention and other things happen to get less attention than
maybe they would deserve. And most people have their day jobs, and
occasionally miss a mail or two, or even a few more. And many people are
on many lists and just skim them until something catches their eye, and
ignore many threads. If you don't do anything like this and survive, I'd
like to know how :-).
So the uri-review list, like any other IETF (or other) list, can get
excited about stuff, sometimes for good reasons, or can just ignore some
aspects of some proposal. It's unfortunate, but it's something we have
to live with. You can't ask everybody on that list to take the same
amount of time to look at all the proposals that e.g. a Security
Directorate or Apps Area Directorate reviewer is using.
But in any case, I also think you're wrong technically in this>> case.
Let's see. I hope we agree that we should come to a conclusion on this
issue on technical merits, rather than on process details.
Sure.
Good.
nih *is* intended for a corner case,
Let me emphasise the above. nih is not intended to be used
broadly, nor often. If you want a hash-based URI scheme for
users to speak that is for broad frequent use then I think
you are free to try design one. But nih is not that and is
not intended to be that.
Then, at the barest minimum, don't sell it as that. The only thing the
draft essentially says is "use this if you want to read/speak it". Now
you are saying "don't use that for users to speak, unless in specific
corner cases", but you still haven't told me or the readers of the draft
what these corner cases are, and why the SHOULD/MUST not use it
otherwise even if they might think it makes sense.
The second, more general advice is of course to not design URIs for
corner cases, because URIs work best if they are used widely. (There may
be a very tiny chance that your case is different, but even after
reading all your mail, I haven't understood that.)
(And I'm not sure such a beast
could really be done well.)
Speaking 256 or so bits of random data aloud is in no way going to be
terribly easy :-).
where humans need to speak these
URIs and was added as a direct result of requirements from the core
WG and not as an afterthought. ni URIs are not intended for that
and so there really are IMO different requirements, (esp. e.g.
checkdigit) that are best met with different schemes.
I agree that the value of a checkdigit is very limited for communication
s/very limited/useless/
among machines (and for communication among humans with the help of
machines, such as in the case of email).
On the other hand, I can't understand why (even assuming we needed a
separate scheme) there is no authority and no query part on nih.
The main intent of nih is to allow entry of something that
confirms something else (e.g. a public key) that is already
present.
At least a little bit of information about the "use case", but this just
leads to more questions:
Do you need an URI for that? (Isn't that what people are supposed to do
with fingerprints these days?)
How/where would a user input that URI e.g. in a browser?
Would copy/paste from an email or a Web page be fine?
Would entry be needed, or could it just be checked? (I understand that
forcing a user to enter the data is way safer than believing her that a
check was made, but then it's also way more tedious, and many users will
just switch to different software.)
Can ni: URIs also be used for the same purpose? If not, why not (they
essentially contain the same information)?
What about other URIs, starting e.g. with http:?
I won't claim that this list of questions is exhaustive, but I hope it's
a good start.
There is no need for an authority for that, for
the use-cases we have. We could speculate about other potential
use-cases but we'd rather not speculate like that when there's
no need to.
There was a lot of stuff that we take for granted today that Tim
Berners-Lee didn't speculate when he designed the first URI schemes. But
if he had designed these schemes with only corner cases in mind, he
wouldn't have invented the Web :-). (sorry about an analogy again)
For the authority, I'd assume that it would be as useful when the URI is
transmitted e.g. over the phone as when it is transmitted e.g. over
email.
We don't have a use for that that I know about.
Do you mean the "over the phone" part or the authority part? (assuming
the later below)
I agree
it could be done, but then I think it'd also impact on
usability, which will be pretty crap no matter what's
done.
Reading "example.com" or some equivalent domain name in addition to a
long hex string shouldn't really make matters much worse.
But making usability worse also seems wrong.
My understanding is that there are some cases where things will be found
with authority, but not without. If that's right, then I'm sure users
will prefer to find what they are looking for, after aurally
transmitting 256 or thereabout bits, even at the cost of an additional
domain name (and a few slashes), which are the easiest parts to get
across the line quickly and correctly.
Not
having an authority also seems to work fine for PGP keys
That the lack of authority works in some cases is perfectly okay. But
it's not an argument for not allowing authority for cases (even those
that maybe you're not seeing yet) where it is or will be useful.
and the lack of an authority does get rid of some threats,
if the nih URI is used for something security-sensitive.
Which would mean that ni: has these threats, yes? Where are they
described? I don't remember seeing them in the security section.
For the query part, there are already various ideas and proposals
floating around,
Where? If you mean draft-hallambaker-decade-params
Correct.
then
we (the authors of that) don't think those are useful for
nih names.
Why not? And what if others think they would be useful?
and at least some of them would be of interest for when
the URI is transmitted e.g. over the phone. Also, even if we currently
didn't have any actual proposals for query parameters, I think it would
be a very bad idea to exclude them a priori for transmission e.g. over
the phone.
I disagree that there is any "very bad idea" here.
Merging ni/nih would also add more complexity for no benefit,
which would be a bad idea.
Can you please explain what kind of complexity would have to be added?
I think its obvious actually. In your table above you highlighted
5 ways in which ni and nih differ. Merging all those yields loads
of combinations, which makes for complexity.
Well, if your implementation is a big switch with 32 cases, one for each
combination, with no code sharing, then indeed you get complexity. But I
hope we agree that that's very bad design, and totally unnecessary. It's
very straightforward to go through the components one-by-one and deal
with them (hint: I'd start with the check digit).
On the other hand, the complexity you get from query parameters is
potentially huge, because, contrary to the basic syntax, we don't know
for sure what combinations of parameters people will dream up, and how
these will interact.
In terms of specification, merging the two schemes doesn't seem to be
difficult or complex at all. Also, in terms of implementation, the only
additions to the ni: scheme that become necessary are the check digit
and the expression of the "suite id" as a decimal. It's very difficult
for me to imagine that this would add significant complexity to an
implementation; if code for nih: exists, that can mostly just be moved
over.
Feel free to look at our code. (With the caveat that I'm a crap
programmer so close your eyes a bit when you look at the 'C' code:-)
Your analogy about httpm/h may appear reasonable, but it is always
unreasonable to draw conclusions from analogies. It is also unwise
to reason from counterfactuals, which we'd also be doing if we
accepted your argument. So I find that speculation utterly useless
to be honest.
It is definitely unreasonable to draw conclusions from analogies *only*.
I only saw the analogy. What in your httpm/h argument is not
couterfactual analogy?
Do you claim that it's not true that URIs are often transferred
electronically, often transferred on paper, and often spoken, and that
these sets for which these activities happen overlap? If you think that
has to be different for ni:/nih:, you have to convince me, not the other
way round.
But if you think that the httpm/h analogy is wrong, and that ni/nih is
different, could you please explain *what* is different?
We have real use cases for ni and nih and we think they differ.
I'd be repeating myself to say why again.
You're still very vague on these "real use cases". Confirmation of PGP
keys by speaking the URI aloud seems to be your main/only use case for
nih:. But I'm only guessing, because you haven't really explained this yet.
In this case, we are dealing with different requirements so this
should stay as-is.
If "different requirements" is your main (or only) real argument,
That would be a valid argument.
It would only be a valid argument if these are requirements that can't
be consolidated.
could
you at least explain exactly how they are different?
I did that above. Asking for an "exact" explanation seems
like asking the same thing again.
Here are the pieces of text from above that I can at least in some way
link to actual requirements:
"nih is not intended to be used broadly, nor often."
"The main intent of nih is to allow entry of something that
confirms something else (e.g. a public key) that is already
present."
Is that all there is? Or did I miss something? Maybe it would help this
discussion if you wrote a few paragraphs of free-standing text
explaining the use cases (and restrictions, and reasons for these) for
both schemes. It may turn out to be a valid addition to the document at
the very minimum. (I'm not asking for an independent "use cases and
requirements" draft :-)
Just that one
requirement came from the core WG and others from other WGs or other
parties doesn't help me to understand how the actual requirements
differ. (Please note that even if the requirements differ, that doesn't
mean that we need different technology to address them.)
Perhaps not. But that was the design choice we made and its
a valid one.
Given the more than 20 years of experience we have of how URIs are
handled on the Web and around it, I have very good reasons to strongly
doubt that this is a good design choice.
Why do you say that ni: URIs are not intended for humans to speak?
So phone me up and say this:
ni:///sha-256;UyaQV-Ev4rdLoHyJJWCi11OHfrYv9E1aGQAlMO2X_-Q
So apart from the upper/lower problem, would the equivalent nih: URI be
any better?
What
am I supposed to do if I got an ni: URI in a mail message and call you
on the phone to tell you about that?
Not my problem actually, but I guess most people might
say "remember that mail you sent me with all that gobbledygook
nonsense - what the hell was that about?" :-)
I think "tell you about it" should have been "speak it out to you".
If I want to send somebody the
information in an ni: URI by mail, should I use only the ni: version or
only the nih: version, or both, if I can't exclude that the recipient
may want to relay this information via voice?
You can try either and let me know what works. This seems like
a very artificial use-case for ni.
Why artificial? Long URIs get spoken over the phone once in a while. And
I really don't understand why this is "very artificial" for ni:, but
"indispensable" for nih:.
Finally, we have (some, early,) running code that matches the
current draft and that ought also count for something
How much?
Feel free to go look and see. [1] I've not counted lines of
code, but we have c, python, ruby and clojure library
implementations and some apps and other bits and pieces.
[1] http://sourceforge.net/projects/netinf/
The boiler plate on every ID is pretty clear that they are not
set in stone. Also, the changes needed to merge the two schemes are not
rocket science, quite to the contrary. (I herewith volunteer to fix the
Ruby version, just to show)
I didn't say our code is set in stone. I said that running code
counts.
I didn't say a merge would require rocket science.
It somehow sounded like this.
I said it'd be
a bad idea and would produce a worse result.
Can you substantiate that? The discussion on use cases is still open,
and the argument with all the 32 combinations also didn't fly.
when compared
to a change that would be a gratuitous dis-improvement
In what sense would merging the two schemes be a dis-improvement? Can
you please explain?
I believe I did that above.
Very marginally, perhaps. More information would definitely help.
based it
seems upon dubious argument
If you think that my arguments are dubious, please explain exactly why.
I believe I did that above. (To be clear: not all your argument is
dubious
Thanks for some encouragement :-).
but the httph/m part is IMO.)
I'm sorry that I started out with that analogy, but I thought it would
be easy to understand. You seem to be caught in the specific use/corner
cases, but I'm trying to look at the long term big picture.
that is also offered at the wrong
point in the process.
See above. If there's something wrong with IETF Last Call, or with the
fact that the Apps Area Directorate does reviews (which I don't think),
then that should be addressed separately. For this discussion, I hope we
can concentrate on technical issues.
Right. But let's not ignore the fact that the uri-review
list had sight of this at the end of April.
[If you keep insisting, why don't you check your MUA and look at all the
messages I sent on the draft. You may be surprised, but I seriously hope
you will stop bringing up "uri-review" again.]
Bottom line - we have use-cases and a valid design and running
code that as far as we know works and I see no reason to make
the change you'd like, which would make thing worse IMO.
It may look like some more work in the short time, but it will make
things more general and create more potential for the future.
Regards, Martin.