Re: [OAUTH-WG] Namespacing "type" in RAR

[Dick Hardt <[e-mail]>]

The following are the same URI, but are different strings:

“https://schema.example.org/v1”
“HTTPS://schema.example.org/v1 <https://schema.example.org/v1>”
“https://SCHEMA.EXAMPLE.ORG/v1 <https://schema.example.org/v1>”

Before comparing them to each other, they must be canonicalized so that
they become the same string.

>From earlier in this thread, I am NOT suggesting that it must be a URI, nor
that it is required:

Since the type represents a much more complex object then a JWT claim, a
client developer's tooling could pull down the JSON Schema (or some such)
for a type used in their source code, and provide autocompletion and
validation which would improve productivity and reduce errors. An AS that
is using a defined type could use the schema for input validation. Neither
of these would be at run time. JSON Schema allows comments and examples.

What is the harm in non-normative language around a retrievable URI?


On Tue, Jul 21, 2020 at 9:58 AM Justin Richer <[e-mail]> wrote:

> String comparison works just fine when the strings happen to be URIs, and
> you aren’t treating them as URIs:
>
> “https://schema.example.org/v1”
>
> Is different from
>
> “https://schema.example.org/v2”
>
> And both are different from
>
> “https://schema.example.org:443/v1/“
>
> All of these are strings, and the strings happen to be URIs but that’s
> irrelevant to the comparison process. Can you please help me understand why
> doing a string comparison on these values does not work in exactly the same
> way it would for “foo”, “bar”, and “baz” values? Why would these need to be
> canonicalized to be compared? The definition of a JSON string is an ordered
> set of unicode code points, and this can be compared byte-wise. (Or
> code-point-wise, whatever’s most correct here.) Can you give me
> counter-examples as to where string comparison doesn’t work? And can you
> help me understand how this same worry doesn’t apply to all of the rest of
> the values in the RAR specification, which are also strings and will need
> to be compared?
>
> I’m still very confused as to the URI retrieval issue here, if there even
> is one. It sounds like we’re both saying that it could be useful if type
> values are retrievable when they’re URIs, but that would be something to
> augment a process and not required for the RAR spec. I’m against requiring
> the value to be a URI and against requiring the AS to process that URI *as
> a URI* at runtime. Anything that an AS wants to do with the “type” value,
> including providing additional tooling and validation, is up to the AS and
> outside of the spec.
>
>  — Justin
>
> On Jul 21, 2020, at 12:35 PM, Dick Hardt <[e-mail]> wrote:
>
> This statement:
>
> “compare two strings so that they’re exact”
>
> does not work for either Unicode or URIs. A string, and a canonicalized
> Unicode string are not the same thing. Similar for a URI. I have assumed
> you understand the canonicalization requirement, but it does not sound like
> you do. Would you like examples?
>
>
> wrt. the AS and URI, *you* keep saying that *I* said the AS would retrieve
> the URI. I HAVE NOT SAID THAT!
>
> I am suggesting that the URI MAY be retrievable, and I gave examples on
> how that would be useful for tooling for client developers, and for an AS
> in doing input validation. The URI would NOT be retrieved at run time.
>
>
> On Tue, Jul 21, 2020 at 7:35 AM Justin Richer <[e-mail]> wrote:
>
>> If we treat all the strings as just strings, without any special internal
>> format to be specified or detected, then comparing the strings is a
>> well-understood and well-documented process. I also think that we shouldn’t
>> invent anything here, so if there’s a better way to say “compare two
>> strings so that they’re exact” then that’s what I mean. Sorry if that was
>> unclear.
>>
>> I’m saying the AS should *not* retrieve the URI passed in the “type”
>> value. You brought that up and then described the process that the AS would
>> take to do so. I have said from the start that the use of a URI is for name
>> spacing and not for addressing content to be fetched, so I’m confused why
>> you think I intend otherwise.
>>
>>  — Justin
>>
>> On Jul 20, 2020, at 2:59 PM, Dick Hardt <[e-mail]> wrote:
>>
>> Canonicalization of URIs and unicode is fairly well specified. I was not
>> suggesting we invent anything there.
>>
>> A byte comparison, as you suggested earlier, will be problematic, as I
>> have pointed out.
>>
>> I'm confused why you are still talking about the AS retrieving a URI.
>>
>> ᐧ
>>
>> On Mon, Jul 20, 2020 at 4:42 AM Justin Richer <[e-mail]> wrote:
>>
>>> Since this is a recommendation for namespace, we could also just say
>>> collision-resistant like JWT, and any of those examples are fine. But that
>>> said, I think there’s something particularly compelling about URIs since
>>> they have somewhat-human-readable portions. But again, I’m saying it should
>>> be a recommendation to API developers and not a requirement in the spec. In
>>> the spec, I argue that “type” should be a string, full stop.
>>>
>>> If documentation is so confusing that developers are typing in the wrong
>>> strings, then that’s bad documentation. And likely a bad choice for the
>>> “type” string on the part of the AS. You’d have the same problem with any
>>> other value the developer’s supposed to copy over.  :)
>>>
>>> I agree that we should call out explicitly how they should be compared,
>>> and I propose we use one of the handful of existing string-comparison RFC’s
>>> here instead of defining our own rules.
>>>
>>> While the type could be a dereferenceable URI, requiring action on the
>>> AS is really getting into distributed authorization policies. We tried
>>> doing that with UMA1’s scope structures and it didn’t work very well in
>>> practice (in my memory and experience). Someone could profile “type" on top
>>> of this if they wanted to do so, with support at the AS for that, but I
>>> don’t see a compelling reason for that to be a requirement as that’s a lot
>>> of complexity and a lot more error states (the fetch fails, or it doesn’t
>>> have a policy, or the policy’s in a format the AS doesn’t understand, or
>>> the AS doesn’t like the policy, etc).
>>>
>>> And AS is always free to implement its types in such a fashion, and that
>>> could make plenty of sense in a smaller ecosystem. And this is yet another
>>> reason that we define “type” as being a string to be interpreted and
>>> understood by the AS — so that an AS that wants to work this way can do so.
>>>
>>>  — Justin
>>>
>>> PS: thanks for pointing out the error in the example in XYZ, I’ll fix
>>> that prior to publication.
>>>
>>> On Jul 18, 2020, at 8:58 PM, Dick Hardt <[e-mail]> wrote:
>>>
>>> Justin: thanks for kindly pointing out which mail list this is.
>>>
>>> To clarify, public JWT claims are not just URIs, but any
>>> collision-resistant namespace:
>>> "Examples of collision-resistant namespaces include: Domain Names,
>>> Object Identifiers (OIDs) as defined in the ITU-T X.660 and      X.670
>>> Recommendation series, and Universally Unique IDentifiers (UUIDs)
>>> [RFC4122]."
>>>
>>> I think letting the "type" be any JSON string and doing a byte-wise
>>> comparison will be problematic. A client developer will be reading
>>> documentation to learn what the types are, and typing it in. Given the wide
>>> set of whitespace characters, and unicode equivalence, different byte
>>> streams will all look the same, and a byte-wise comparison will fail.
>>>
>>> Similarly for URIs. If it is a valid URI, then a byte-wise comparison is
>>> not sufficient. Canonicalization is required.
>>>
>>> These are not showstopper issues, but the specification should call out
>>> how type strings are compared, and provide caveats to an AS developer.
>>>
>>> I have no idea why you would think the AS would retrieve a URL.
>>>
>>> Since the type represents a much more complex object then a JWT claim, a
>>> client developer's tooling could pull down the JSON Schema (or some such)
>>> for a type used in their source code, and provide autocompletion and
>>> validation which would improve productivity and reduce errors. An AS that
>>> is using a defined type could use the schema for input validation. Neither
>>> of these would be at run time. JSON Schema allows comments and examples.
>>>
>>> What is the harm in non-normative language around a retrievable URI?
>>>
>>> BTW: the example in
>>> https://oauth.xyz/draft-richer-transactional-authz#rfc.section.2 has
>>> not been updated with the "type" field.
>>>
>>>
>>>
>>> On Sat, Jul 18, 2020 at 8:10 AM Justin Richer <[e-mail]> wrote:
>>>
>>>> Hi Dick,
>>>>
>>>> This is a discussion about the RAR specification on the OAuth list, and
>>>> therefore doesn’t have anything to do with alignment with XAuth. In fact, I
>>>> believe the alignment is the other way around, as doesn’t Xauth normatively
>>>> reference RAR at this point? Even though, last I saw, it uses a different
>>>> top-level structure for conveying things, I believe it does say to use the
>>>> internal object structures. I am also a co-author on RAR and we had already
>>>> defined a “type” field in RAR quite some time ago. You did notice that
>>>> XYZ’s latest draft added this field to keep the two in alignment with each
>>>> other, which has always been the goal since the initial proposal of the RAR
>>>> work, but that’s a time lag and not a display of new intent.
>>>>
>>>> In any event, even though I think the decision has bearing in both
>>>> places, this isn’t about GNAP. Working on RAR’s requirements has brought up
>>>> this interesting issue of what should be in the type field for RAR in OAuth
>>>> 2.
>>>>
>>>> I think that it should be defined as a string, and therefore compared
>>>> as a byte value in all cases, regardless of what the content of the string
>>>> is. I don’t think the AS should be expected to fetch a URI for anything. I
>>>> don’t think the AS should normalize any of the inputs. I think that any
>>>> JSON-friendly character set should be allowed (including spaces and
>>>> unicodes), and since RAR already requires the JSON objects to be
>>>> form-encoded, this shouldn’t cause additional trouble when adding them in
>>>> to OAuth 2’s request structures.
>>>>
>>>> The idea of using a URI would be to get people out of each other’s
>>>> namespaces. It’s similar to the concept of “public” vs “private” claims in
>>>> JWT:
>>>>
>>>> https://tools.ietf.org/html/rfc7519#section-4.2
>>>>
>>>> What I’m proposing is that if you think it’s going to be a
>>>> general-purpose type name, then we recommend you use a URI as your string.
>>>> And beyond that, that’s it. It’s up to the AS to figure out what to do with
>>>> it, and RAR stays out of it.
>>>>
>>>>  — Justin
>>>>
>>>> On Jul 17, 2020, at 1:25 PM, Dick Hardt <[e-mail]> wrote:
>>>>
>>>> Hey Justin, glad to see that you have aligned with the latest XAuth
>>>> draft on a type property being required.
>>>>
>>>> I like the idea that the value of the type property is fully defined by
>>>> the AS, which could delegate it to a common URI for reuse. This gets GNAP
>>>> out of specifying access requests, and enables other parties to define
>>>> access without any required coordination with IETF or IANA.
>>>>
>>>> A complication in mixing plain strings and URIs is the
>>>> canonicalization. A plain string can be a fixed byte representation, but a
>>>> URI requires canonicalization for comparison. Mixing the two requires URI
>>>> detection at the AS before canonicalization, and an AS MUST do
>>>> canonicalization of URIs.
>>>>
>>>> The URI is retrievable, it can provide machine and/or human readable
>>>> documentation in JSON schema or some such, or any other content type. Once
>>>> again, the details are out of scope of GNAP, but we can provide examples to
>>>> guide implementers.
>>>>
>>>> Are you still thinking that bare strings are allowed in GNAP, and are
>>>> defined by the AS?
>>>>
>>>>
>>>>
>>>> On Fri, Jul 17, 2020 at 8:39 AM Justin Richer <[e-mail]> wrote:
>>>>
>>>>> The “type” field in the RAR spec serves an important purpose: it
>>>>> defines what goes in the rest of the object, including what other fields
>>>>> are available and what values are allowed for those fields. It provides an
>>>>> API-level definition for requesting access based on multiple dimensions,
>>>>> and that’s really powerful and flexible. Each type can use any of the
>>>>> general-purpose fields like “actions” and/or add its own fields as
>>>>> necessary, and the “type” parameter keeps everything well-defined.
>>>>>
>>>>> The question, then, is what defines what’s allowed to go into the
>>>>> “type” field itself? And what defines how that value maps to the
>>>>> requirements for the rest of the object? The draft doesn’t say anything
>>>>> about it at the moment, but we should choose the direction we want to go.
>>>>> On the surface, there are three main options:
>>>>>
>>>>> 1) Require all values to be registered.
>>>>> 2) Require all values to be collision-resistant (eg, URIs).
>>>>> 3) Require all values to be defined by the AS (and/or the RS’s that it
>>>>> protects).
>>>>>
>>>>> Are there any other options?
>>>>>
>>>>> Here are my thoughts on each approach:
>>>>>
>>>>> 1) While it usually makes sense to register things for
>>>>> interoperability, this is a case where I think that a registry would
>>>>> actually hurt interoperability and adoption. Like a “scope” value, the RAR
>>>>> “type” is ultimately up to the AS and RS to interpret in their own context.
>>>>> We :want: people to define rich objects for their APIs and enable
>>>>> fine-grained access for their systems, and if they have to register
>>>>> something every time they come up with a new API to protect, it’s going to
>>>>> be an unmaintainable mess. I genuinely don’t think this would scale, and
>>>>> that most developers would just ignore the registry and do what they want
>>>>> anyway. And since many of these systems are inside domains, it’s completely
>>>>> unenforceable in practice.
>>>>>
>>>>> 2) This seems reasonable, but it’s a bit of a nuisance to require
>>>>> everything to be a URI here. It’s long and ugly, and a lot of APIs are
>>>>> going to be internal to a given group, deployment, or ecosystem anyway.
>>>>> This makes sense when you’ve got something reusable across many
>>>>> deployments, like OIDC, but it’s overhead when what you’re doing is tied to
>>>>> your environment.
>>>>>
>>>>> 3) This allows the AS and RS to define the request parameters for
>>>>> their APIs just like they do today with scopes. Since it’s always the
>>>>> combination of “this type :AT: this AS/RS”, name spacing is less of an
>>>>> issue across systems. We haven’t seen huge problems in scope value overlap
>>>>> in the wild, though it does occur from time to time it’s more than
>>>>> manageable. A client isn’t going to just “speak RAR”, it’s going to be
>>>>> speaking RAR so that it can access something in particular.
>>>>>
>>>>> And all that brings me to my proposal:
>>>>>
>>>>> 4) Require all values to be defined by the AS, and encourage
>>>>> specification developers to use URIs for collision resistance.
>>>>>
>>>>> So officially in RAR, the AS would decide what “type” means, and
>>>>> nobody else. But we can also guide people who are developing
>>>>> general-purpose interoperable APIs to use URIs for their RAR “type”
>>>>> definitions. This would keep those interoperable APIs from stepping on each
>>>>> other, and from stepping on any locally-defined special “type” structure.
>>>>> But at the end of the day, the URI carries no more weight than just any
>>>>> other string, and the AS decides what it means and how it applies.
>>>>>
>>>>> My argument is that this seems to have worked very, very well for
>>>>> scopes, and the RAR “type” is cut from similar descriptive cloth.
>>>>>
>>>>> What does the rest of the group think? How should we manage the RAR
>>>>> “type” values and what they mean?
>>>>>
>>>>>  — Justin
>>>>> _______________________________________________
>>>>> OAuth mailing list
>>>>> [e-mail]
>>>>> https://www.ietf.org/mailman/listinfo/oauth
>>>>>
>>>>
>>>>
>>>
>>
>