Stringly Typed

At the first FP Day on Friday, I attended a talk by Don Syme on F# 3.0. There were a number of useful features, but to the functional programmer, the most revolutionary one was type providers. They’re a way to use strong, static typing for data structures beyond those coded into libraries. The most prominent use of these are for data sets so large that the entire set of types necessary to understand all of it is too much to include in your application. However, they’re also useful as an alternative to code generation. Take for example, accessing a page on Wikipedia. If I have a type provider that accesses Wikipedia as I’m writing the code and generates types for each type of article, I can do something like this.

FYI (because most of us don’t know F#), backticks (`) let you put spaces in names, angle brackets (< and >) are used to initialise type providers, and the |> operator pipes the result of one operation into the next.

#r "Wikipedia.dll"

type Wikipedia = Wikipedia<>
Wikipedia().Computing
           .``History of computing``
           .``Computer pioneers``
           .``Dennis Ritchie``
           .Infobox
           .``Known For``
  |> String.concat ", "
  |> (+) "Dennis Ritchie was known for "

// returns "Dennis Ritchie was known for ALGOL, B, BCPL, C, Multics, Unix"

Each of those objects I’m accessing is statically typed. If I misspell the article name, I don’t get an error at runtime. My application fails to compile. How cool is that?

Now, this is excellent, and I hope to see it in Java around 2026 or so. But what’s really interesting is something else Don said (and I’m paraphrasing): if you would hard-code a string, you should instead use type providers. This got me wondering a little. I’ve always thought of strings in my application as simply hard-coded text. But it turns out there’s different kinds of text. There’s the stuff we represent to the user: API documentation, button text—anything you see—and there’s the the stuff we use internally but never show anyone. These latter strings all fall into a category I’m going to call, for lack of a better term, data accessors. They’re things like API URLs, file paths, Windows registry keys… you catch my drift.

We shouldn’t be encoding these accessors as strings. How often have you hunted for a bug for hours before realising you made a typo? Strong typing takes out a whole class of errors for us, and we should be exploiting the concept as much as we possibly can. This means type providers can be used to access file systems and REST calls. And as a bonus, it makes the code so much more readable.

Using the REST API, here’s (roughly) what updating your status on Twitter looks like:

let twitterUrl = "http://twitter.com/"
let newTweet = "status/update"
Http.Request(twitterUrl).Post newTweet, credentials, tweetText

But here’s how it should look:

let twitter = Twitter(credentials)
twitter.Status.Update tweetText

We commonly place user interface strings outside the source code using gettext, Java resource bundles and other such libraries. We’ve come to realise that our applications are more maintainable when we separate out code and data. We should externalise our data accessors in the same way. Type providers are an excellent way of doing this, but they’re not necessary. If an API is sufficiently explorable (like Wikipedia’s), we can generate the code to create these types using all sorts of methods, and simply create a build step that pulls down the latest information for just the types we need. This will give us similar behaviour to type providers in F#, but without IDE support. If it’s an API, we can just create a wrapper around it and use that, as I demonstrated with Twitter above.

I learnt a lot at FP Day, but this was the thing that stuck out the most. Not because it’s shiny and new, but it’s reinforcing something we all know but too often forget on purpose. Strings in your source code are a sign of intermingling data and logic, and make it harder to read and maintain. Some are necessary, but they should be the exception, not the rule. If I put the code let x = 1089 into my code base, the next person to read it would hunt me down and slap me in the face. We shouldn’t tolerate strings in the code either. Pull ’em out and give them names, the same as we do with numbers. And then go one further. Wrap them in a well-tested library that takes care of the details, or even put them in their own resource file.

Data’s not code. Stop treating it like it should be.