- From: Pierre-Antoine Champin <pierre-antoine.champin@liris.cnrs.fr>
- Date: Wed, 19 Oct 2011 09:19:25 +0200
- To: Sandro Hawke <sandro@w3.org>
- CC: Richard Cyganiak <richard@cyganiak.de>, public-rdf-wg <public-rdf-wg@w3.org>
Ok, here is another idea: we keep "graph container" as a rather vague notion, just to point out that many thing that one is tempted to call a "graph" are actually something else. Then we illustrate that notion with two particular kinds of graph containers, more precisely defined: a *graph slot* is a mutable object whose state is (a representation of) an RDF graph. In other words, it provides a function from time to RDF graphs. (e.g. an .ttl file, an instance of jena.rdf.model.Model) a *graph web resource* is a web resource whose representations are graph serializations (fka g-texts) (e.g. my FOAF profile, facebook RDF exports) The definitions overlap: my FOAF profile is managed as a file, and only varies with time, so it is both a graph slot and a graph web resource. They are not identical, because some graph web resources can return different graphs at the same time, depending on other factors (authentication, language negotiation?,...). pa PS: I'm not completely convinced by "graph slot" as a term, but it seems to me that it holds the idea that it contains only a *single* graph at a given time. In addition, it is already used by the DAWG, in a way that seems compatible with this notion. On 10/19/2011 09:03 AM, Pierre-Antoine Champin wrote: > Sandro, > > my point is not to "dismiss" the facebook example, nor to tell facebook > they should do things differently. > > My concern is that we can either: > > * keep a restrictive definition of "graph container", relying on the > notion of *state*, and in that case the facebook example does not fit > the definition (which does not mean it is uninteresting or incorrect!); > > * define "graph container" is broader sense that include the facebook > example, like "anything that, when poked, returns an RDF graph" > > Note that I use the fuzzy term "poke" (suggested by Pat in a previous > conversation) rather than "dereference", because I want the definition > to apply in a "non-web" context (namely, I would like an instance of > rdflib.Graph or jena.rdf.model.Model to be considered a graph container). > > But this fuzziness in terms makes virtually anything a graph container > -- even *I* can produce an RDF graph if you poke me. Does that make me > graph container?? > > So, my concern is that, with too broad a definition, the notion becomes > fuzzy, too general and potentially useless. This is just an intuition, > though. And on the other hand, keeping a restrictive definition for > graph container does not mean that anything that does not fit that > definition is uninteresting. > > pa > > > On 10/15/2011 03:41 AM, Sandro Hawke wrote: >> On Fri, 2011-10-14 at 16:52 +0100, Pierre-Antoine Champin wrote: >>> This is a reaction to Richard's question below... >>> >>> On 10/13/2011 12:16 PM, Richard Cyganiak wrote: >>>> On 12 Oct 2011, at 22:04, Sandro Hawke wrote: >>>>> 1. If a system successfully dereferences URL "L" and obtains a >>>>> representation of an RDF graph, then <L> is a GraphContainer. That >>>>> is, "L" denotes a GraphContainer. Logically, GraphContainer is >>>>> disjoint from foaf:Person (I think!!) so a document that includes "<> >>>>> a foaf:Person" is (by this proposal) logically inconsistent with it >>>>> being served on the Web. >>>> >>>> This seems a logical consequence from httpRange-14. >>> >>> At first I agreed, but depending how you precisely define a graph >>> container, it may not be true. See below... >>> >>>>> 2. So, owl:Ontology heavily overlaps GraphContainer. It might even be >>>>> a subclass of it. (Many OWL ontologies say "<> a owl:Ontology", where >>>>> the <> will be resolved to the address the ontologies is fetched from, >>>>> aka L.) >>>> >>>> Well, sort of. They are orthogonal by definition, but overlap in practice because it can be useful to treat owl:Ontologies as graph containers. >>>> >>>>> 3. Some GraphContainers, "SerialGraphContainers" are functions mapping >>>>> from time to RDF Graphs. We can talk about next & previous & current >>>>> RDF Graphs in a SerialGraphContainer, but not about GraphContainers in >>>>> general. (cf facebook's api for fetching RDF data, which returns >>>>> different RDF data depending on your credentials). >>>> >>>> I don't understand that. If graph containers are mutable, then doesn't that already mean they are functions mapping from time to RDF graphs? >>> >>> Well, as I understood it, we have defined a graph container as something >>> whose state is an RDF graph; for me, "state" reads "a function from time >>> to X", so I concur with Richard: the definition SerialGraphContainers >>> should apply GraphContainer. >>> >>> However, I see where Sandro is coming from: see the facebook example he >>> gives above. However, if Facebook returns different *graphs* for the >>> same resource at the same time, depending on another parameter >>> (credential), then the *state* of the resource is *not* an RDF graph, it >>> is something more (a "graph+ACL" or something like that). >>> >>> Hence, if that resource (in the RESTest sense) has a state which is not >>> a graph, this is not a graph container as we defined it. QED. >>> >>> >>> Of course, we could extend our definition so that it encompasses >>> X-varying graph containers (where X can be "credential", "location", >>> ...) but the open nature of X makes it the definition very generic, may >>> be too much... >>> >>> So my reaction would be to stick to the safe side and restrict "Graph >>> Containers" to vary with time only. Of course, many other things may >>> "hold" graphs in different fashions that depend on other factors, but >>> this WG can/should not define all of them. >> >> Your proof sounds good on paper, but I don't think you can dismiss the >> facebook example so simply. Yes, we can say their thing isn't a >> GraphContainer, via the web that's it's not, since we did a deref and >> got back an RDF graph. :-( >> >> I suppose one could ask them to redirect to parameterized URIs, so I ask >> for: >> http://my.example.com (not a GraphContainer) >> when logged in as sandro and it redirects (302 FOUND) to >> http://example.com/sandro/ (a GraphContainer) >> and then gives me the content. >> >> But I'm guessing facebook (and others) aren't willing to burden their >> users with an extra round trip just for this kind of architectural >> purism. (Some people will like to use the second form URI, but not >> everyone wants to bother.) >> >> Another technical/http solution: >> >> I ask for http://my.example.com and get the response, but it includes a >> "Content-Location: http://example.com/sandro/" header. The client >> could use that header as a flag to indicate the URL that really denotes >> that GraphContainer. And if facebook doesn't include such a header it >> would just be ... "broken". Yeah, that might work. >> >> I don't know how widely Content-Location is used. I know w3.org uses it >> a lot, but I don't know about the web at large. The Google results I >> see for it suggest it's not widely used, but maybe that's okay. It >> seems pretty ReSTful to me. >> >> One final note -- giving a quick look at the facebook API, I don't >> actually see this kind of URL where each person sees a totally different >> thing. Someone said in the meeting they do that; even if they don't, >> surely someone else does or will. What I do see is this: >> >> GET -H 'Accept: text/turtle' https://graph.facebook.com/sandro.hawke >> >> -- Sandro >> >>> pa >>> >>> >>> >>>> >>>>> 4. A ConstantGraphContainer always holds the same RDF Graph. This can >>>>> be used for when you want to attach a dereferenceable URL to a g-snap. >>>>> You put it in a ConstantGraphContainer. >>>> >>>> Seems reasonable. >>>> >>>> Richard >>> >>> >> >> > >
Received on Wednesday, 19 October 2011 07:20:03 UTC