An open letter

2009 02 24

An open letter to A.A.

In clarification of my hastily-written thoughts on Web applications, and in response to his comments on the same.

First, let me say that you’re right: I’m unlikely to convert you to one particular mode of development, or one way of thinking about development. Nor should I aim to, as we all benefit from a broad ecosystem of models. Though I will make persuasive arguments, this is not meant to be a wholly persuasive document. I’m mostly interested in clarifying my own thoughts. That established…

As software developers in the age of the internet, we’re interested in sharing information between computers over a network. It may mean parallel processing to solve a difficult task, or federation of roles/responsibility among different classes of machines, or providing a view of some data on machine A to a recipient on machine B. Sometimes a human is involved, sometimes not.

The question is: what method do we use to share this data? And the answer should depend on our goals as a developer. In other words, we should choose the technology based on the requirements of the application, rather than adapting the application to suit a given technology.

If my computer is interested in sending a piece of information to another one, how can it be accomplished? At a bare-metal level, we need something to be responsible for taking that information, marshaling it into our transport protocol, de-marshaling it on the receiving end, and passing it up the stack to the recipient. This is the core of distributed computing; any distributed computing mechanism must be able to handle at least this much. It sounds simple, but there are many points to be considered if we expect reliable success:

All these points lead to the inescapable conclusion that “it is a stupid idea to build a distributed application as if it were a non-distributed one.” And it follows that we will design a distributed application differently, depending on the characteristics we value the most.

An online banking application probably values reliable, consistent data and security above speed; a network TV application the opposite. “Web” applications don’t--shouldn’t--get a pass on this sort of fundamental analysis simply because we expect to access them through a browser. If I want a blazingly fast Internet search application, my requirements shouldn’t be compromised because of an underlying assumption that, say, the results have to be provided in XML, because it’s “flexible”. If my searcher is expected first and foremost to be fast (and it is!) then imposing such restrictions when alternatives exist is Just Plain Wrong™.

What does this have to do with Web applications?

Web applications, indeed most applications, overwhelmingly operate on the principle of “provide input to module, get output.” They operate on the basis of function calls or method invocations: input X, output Y. Type search term, get list of URLs. Type bank account number and password, get current balance. Input location from GPS-enabled device, get list of friends in 20 mile radius.

HTTP is a request-response protocol, but one designed to handle specific types of requests and responses: input URI, output document. It is stateless, maintains no session knowledge, and provides value-added features to the effect of advancing this singular purpose: input URI, output document.

That it is possible to map other application domains to this model does not make it prudent. A URI is not a function call. A document is not an interface.

HTTP was not at all designed around the programming language procedure call model abstraction.” “HTTP is not RPC;” it uses “a stream as a parameter [and] requests are directed to resources using a generic interface with standard semantics. … RPC mechanisms, in contrast, are defined in terms of language APIs, not network-based applications.”

When you deviate from the stated and designed purpose of the protocol, you’re objectively Doing It Wrong™. And, pitifully, enough people started wanting to Do It Wrong™ that entire cottage industries sprung up to make it easier. As I mentioned before, the corruption of CSS and things like AJAX and jQuery are among the most visible disturbances in the Force. But even below those, we have these SOAP and REST paradigms that assist the Web application developer in making his mistakes, adding stateful and session-oriented layers to an inherently incompatible transfer protocol. So developers must do one of two things: they must re-invent, poorly, a type of HTTP-as-RPC, repeating mistakes of the past and wasting time and energy on an inescapably inferior product; or, they must mutate their application, and indeed their entire model of application development, to fit the constraints they’ve needlessly (and incorrectly) forced themselves into.

Web applications largely exist in the realm of the latter, and the tools of Web application developers, such as Ruby on Rails, have evolved to encourage and support that paradigm-shift, though I would argue incompetently, without a real understanding of what they’re actually doing. No doubt there are many people getting increasingly skilled at this sort of work. For me it is like a physician who has become extremely proficient at finding the correct vein to let blood.

As an aside, this line of argument is all moot if your application (in the broadest possible sense of the word “application”) is both designed from the ground-up to operate in the stateless REST/SOAP/HTTP paradigm, and interacts with the user through true, pure HTML documents-as-documents. But it’s quite difficult for me to imagine an application in the classic sense that can translate to this model. I’d argue any such thing is more accurately described as a simple content-serving web site, such as Project Gutenberg,, or (very arguably) the original/unadorned Which, of course, is precisely my point: the Web is no place for applications.

What the Web application developer actually wants is something we’ve already had for a long time: actual, desktop-style applications, that you download and run on your own PC. Of course there are huge headaches involved in this model: cross-platform accessibility, operating system version compatibility, library availability, and so on. I neither deny them nor diminish their impact. But in order to avoid those headaches, Web developers co-opted a fundamentally unsuitable technology to do things they have no business doing. And despite the pitfalls of the alternatives, that is Just Plain Wrong™.

Often, what developers really want is a remote procedure call, or RPC. It’s an idea that’s been around since the 70’s, and tons of mental and physical effort has been spent by very intelligent people to make RPC systems reliable, robust, powerful and efficient. Today there are lots of these sorts of middlewares, from CORBA (early 90’s, largely deprecated) to ZeroC’s ICE (which I’ve used extensively) to Google’s internal-cum-external Protocol Buffers and Facebook’s/Apache’s Thrift. These allow developers to define (typically) language-agnostic data types and interfaces, and make (to some degree) network- agnostic method invocations. For machine-to-machine communication, 99% of the time, this is both the most appropriate and most efficient way of sharing data. Plus, they have the added advantage of being designed to do exactly what they’re being used to do: input X, output Y, for any values of X and Y.

This is the extent of my argument against Web applications/Web services, and the source of my antagonism towards Web 2.0, Ruby on Rails, enterprise Java/Tomcat stuff, 37signals, etc. etc. I can advocate for any number of alternative technologies (including RPC, where appropriate) but I guess that’s another thing entirely.