I wanted to respond to some of the detail in Werner’s article, in addition to ranting about how document exchange was state transfer. So here goes …

The first statement that really gave me pause was this;

The goal of web services is to enable machine-to-machine communication at the same scale and using the same style of protocols as human interface centered World Wide Web.

I don’t believe that’s the case, and it’s certainly not been accomplished IMO. I think that if you asked anybody involved since the early days, that they’d say the goal is just the first part; to enable machine-to-machine communication over the Web (or perhaps Internet). “using the same style of protocols” has never been a requirement or goal of this community that I’ve seen.

Consider what can be done with a Web services identifier versus a Web identifier. Both are URIs, but because Web architecture uses late binding, I know what methods I can invoke when I see a URI (the URI scheme, specifically) and what they mean (because there’s a path back to RFC 2616 from the URI scheme). With an identifier for a Web service, I don’t have sufficient information to know what the interface is and what it means, because Web services are early/statically bound (creating centralization dependencies, ala UDDI). I don’t consider changing the architecture from late/dynamic binding to early/static binding to be “using the same style of protocols”.

I suppose I also take issue with the implicit definition of “distributed objects” as part of Misconception #1, when it says;

An important aspect at the core of distributed object technology is the notion of the object life cycle: objects are instantiated by a factory upon request, a number of operations are performed on the object instance, and sometime later the instance will be released or garbage collected.

I’ll present my definition first; distributed objects are identifiable things encapsulating state and behaviour, which present an interface upon which operations can be invoked remotely. Obviously software objects, like all software, do have a lifecycle. But it’s primarily an implementation detail, and not exposed through the object’s interface. Some systems chose to tie the identifier to some particular in-memory instantiation of an object (rather than to an abstraction for which an object could be instantiated to proxy for, in effect), which created a real mess, but I don’t consider that key to the definition.

Misconception #2 also seems prima facie incorrect to me, at least by my definition of “RPC”; an architectural style where the developer of each component is provided the latitude to define the interface for that component. More concretely, I believe the statement “there are no predefined semantics associated with the content of the XML document sent to the service” to be incorrect because, as I mentioned in my last post, if there is a method name in the document, then that is an explicit request for “predefined semantics”.

I agree with the titles of Misconceptions #3 and #4; Web services don’t need HTTP or Web servers. But I disagree with the explanation provided. That Web services are transport agnostic is fine, but that does not at all imply that they should be application protocol agnostic, although most people use them this way. The core of my disagreement is with this statement in the last paragraph of #4;

The REST principles are relevant for the HTTP binding, and for the web server parsing of resource names, but are useless in the context of TCP or message queue bindings where the HTTP verbs do not apply.

That is incorrect. REST is protocol independent, and has applicability when used with other protocols, be they application or transport. REST is an architectural style, and following its constraints guides the style of use of all technologies you might choose to work with. For example, if you were using FTP RESTfully, then you would do things like identify files with URIs (rather than server/directory/file-name tuples), and interact with them via a single “retrieve” semantic (an abstract “GET”), rather than the chatty and stateful user/pass/binary/chdir/retr process (not-coincidentally, this is how browsers deal with FTP). In essence (though drastically oversimplifying), what REST says is “exchange documents in self-descriptive messages, and interpret them as representations of the state of the resource identified by the associated URI”. That philosophy can be applied to pretty much any protocol you might want to name, especially other transfer protocols (as the Web constitutes an “uber architecture”, of sorts, for data transfer).

That’s most of the big issues as I see them. Anyhow, that was an enjoyable and informative read. Thanks!

Werner posted an article he wrote for IEEE Internet Computing titled Web Services are not Distributed Objects: Common Misconceptions about Service Oriented Architectures.

That article is very well written, and Werner makes his point loud and clear as always … but ultimately, it makes some of the same misconceptions as so many others have before it. In this case, I think I’ve boiled it down to one main misconception that I’ve talked about recently;

Web services are based on XML documents and document exchange […]

No, they are not. Just open your wallet and grab a cheque, or a credit card receipt, or your drivers license. These things are what I know a “document” to be; state. If a cop asks me for my drivers license and I hand it to her, I have performed “state transfer”, I haven’t asked her to do anything in particular by transferring this document to her. In contrast, the Web services view of a document includes a “method” which effects the semantics of the movement (aka transport) of that document. So if I had a Web services document which I handed to somebody, I’m not merely submitting that document to them, I’m asking that they perform some explicit action for me as specified by the contained method. This is a very very different thing than what “document exchange” is normally understood to mean.

I suggest that if you made the simple tweak to the big picture Web services vision to require that documents only contain state, then you’d have the Web, or at least a substantial part of it. I consider the Web to be the epitome of large scale document-centric distributed computing architectures.

Ted Leung – whose weblog I just subscribed to a couple of weeks ago and I enjoy reading immensely – just commented on my blog about Adam Bosworth.

First off, I want to be clear that I wasn’t “taking Adam to task”. I was just honestly excited to see that he appeared to closing in on understanding the Web via the seemingly identical path that I took. I think you have to have had the “Web epiphany” before you can appreciate why this excites me so much. 8-)

Ted writes;

Cross off CORBA and replace it with either REST or web services. The Web is already there. The missing piece is OpenDoc or something like it.

I don’t dispute that the browser provides a relatively weak form of compound document framework when compared to OpenDoc and CommonPoint, but my emphasis at the time was in studying the architecture of the system to see if it prevented richer frameworks from being built by extension. And I discovered that no, it didn’t prevent this from happening, and in addition already had some of the architectural features that I felt were required (XML namespaces (well, they came later), serialization-centric (GET), binding of state to behaviour (Content-Type), etc..). And sure enough, we’re finally beginning to see some of these systems being developed now. So I wouldn’t say that we’re missing OpenDoc, I’d just say that we’re working with a primordial-but-extensible version of it.

BTW, I also discovered that just by historical accident, an important part of what I expected to see – client side containers – wasn’t there. Cookies really threw me for a loop for many months, and it wasn’t until I read Roy’s dissertation that I realized that he didn’t like cookies, and that the RESTful solution to the problem they addressed was also a perfectly compound-document-framework friendly solution.

Kudos to Aaron; XML’s deterministic failure model is broken. I agree 100%.

Another sacred cow bites the dust.

Kudos to Kendall Clark for stating XML is Not Self-Describing, as I did last month.

He writes;

Well, I’ve read too much Wittgenstein (not to mention too much Aquinas, Meister Eckhart, and Julian of Norwich) to think that a name is necessarily a self-description

I haven’t read them at all (8-), but I think I have a pretty good understanding of self-description that I developed “Bottom up” during my study of Web architecture over the past few years. As Kendall brought this up again, I’d thought I’d write a few more words about it.

As I see it, description is always with respect to some context. For example, “The sky is blue” is not a self-descriptive statement unless you know;

• ASCII
• English
• Which sky I mean
• Which colour blue I mean

For any bag-o-bits, it seems to me that there exists a finite amount of contextual knowledge which is necessary in order to be able to understand it. “Self-describing” then, should mean that the bag itself contains sufficient information to identify the required contextual knowledge.

Tim Berners-Lee likes to talk a lot about this. Last year in Honolulu at WWW2002, his keynote was Specs Count, and much of it was about the value in the ability to be able to perform successive application of public specifications in order to understand a message. That’s contextual knowledge, and as you can see in his talk, it doesn’t begin with the HTTP message, it goes all the way down to the IP segment and Ethernet frame; even those bits must be considered (see an example of where this issue can show up in practice).

Where the Web fits in here, is with its contribution of an enormously valuable piece of contextual knowledge; RFC 2396 aka URIs. With respect to the example above, I can use URIs instead of strings, where those URIs can be used to provide the specifics of which blue I meant, by relating it to other colours.

There’s lots to be said about XML, RDF, and why SOA based Web services can never be self-descriptive (hint; too many methods). But I’ll leave it at that for now.