Configuration Design is User Experience Design…and it's hard

December 06, 2016

Join my mailing list…

In exploring the modern front-end ecosystem for my new book, I’ve gotten to experience some truly difficult configuration formats (Webpack) and to work around aspects of Rails that aren’t configurable (Sprockets). Configuration is hard, and it’s an overlooked part of the user experience often designed to make the software library author’s job easier at the expense of ease-of-use.

It doesn’t have to be this way.

Rails is famous for popularizing “convention over configuration”, which boils down to two things:

  • Don’t make everything configurable.
  • Set sensible defaults for everything that is configurable.

This is often referred to as “opinionated software”, but it’s really “software that is easy to use”. This is because I strongly feel developers should be spending time solving problems for their users and not making pointless decisions like where files go, how to name database tables, or what HTML templating language to use. No project failed because the developers didn’t think long and hard enough about the merits of HAML vs Slim.

But, this concern is orthogonal to the user experience of configuring software. Software configuration is, on a whole, a total nightmare. Rails shields itself from some criticism because it doesn’t make you actually configure things very often, but when you do, it’s needlessly complex.

Here’s how to configure the JavaScript minifier in Sprockets:

environment.js_compressor  = :uglify

What does the symbol :uglify mean? What does it do? The documentation for js_compressor is:

Assign a compressor to run on application/javascript assets. The compressor object must respond to compress.

Although you can make a Symbol respond to compress, that’s clearly that’s not what happening here. Turns out there is a lookup table in the code that maps that symbol to the actual compressor, Sprockets::UglifierCompressor (which, strangely, does not respond to compress; the docs are out of date and the object must actually respond to call). So, to specify another value than the default, you have to configure the configuration.

A UI like this isn’t learnable. You can’t simply figure out what other options are available. It’s also not predictable, because you have no way to know what different values will do (without reading the source code and even then it’s not clear—what’s being passed to call?).

Typically, a UI or API has to trade-off learnability with usability. vim is a classic example of leaning toward usability. It’s very difficult to learn, but once you do learn it, you’re more efficient than with most other editors.

Configuration APIs must lean toward learnability, because there is no steady-state where you are editing configuration every day. The usability of configuration doesn’t matter nearly as much as its learnability.

And, because Ruby uses code for configuration, I would argue it makes the framework developer’s life easier. Consider if this were the way to configure the JS compressor:

environment.js_compressor_class = Sprockets::UglifierCompressor

This is eminently more learnable. We know that we’re expected to provide a class. We can also see that the default value is…a class! We can much more easily locate that class than we could a symbol. In fact, if we’re lucky, that class has documentation that we can consult instead of looking at the source.

I should note that you can pass a class to js_compressor=. There’s just all this extra code in Sprockets to handle a Symbol. I can’t understand why. It’s more work for the library maintainer and more difficult in every way for the user.

But, there is a strong culture in Rails for playing code golf at every turn. Anything verbose is “ugly” and not “clean” and therefore must be eliminated. This is not in the interest of users.

The way configuration is used is different than, say, Active Support. I would argue that Rails is right to create APIs like 3.days.ago. But, these opaque and undiscoverable affordances have no place in configuration.

Configuration should be explicit, and if making it more verbose is required to make it explicit, so be it. Note that explicit doesn’t require being verbose.

An explicit configuration is more learnable and readable. It also makes the library author’s job simpler, meaning they can provide more help to the user in properly configuring the library.

For example, what if js_compressor_class= worked like this:

def js_compressor_class=(klass)
  method = klass.instance_method(:call)
  if method && method.arity == 1
    @js_compressor_class = klass
  elsif method.nil?
    raise "#{klass} does not implement `call`"
  else
    raise "#{klass} implements call, but it takes #{method.arity} " +
          "arguments.  It should take exactly 1"
  end
end

How helpful would that be? A lot more so than being able to specify a Symbol to save a few seconds of typing. I know that Ruby and Rails eschew this sort of safety net, but, again, users aren’t in the configuration very often and won’t ever really learn the subtleties of how it works. They need safety checks and helpful error messages far more than in the “regular” Rails API.

Leaving Rails, it’s all downhill, but I think the same lessons could be applied.

Let’s look at Webpack, the current asset-pipeline-du-jour.

Webpack is highly powerful and eminently functional. But its configuration is bordering on performance art. Here’s how the docs recommend you get CSS into your asset bundle:

{
  module: {
    loaders: [
      {
        test: /\.css$/,
        loader: 'style-loader!css-loader' 
      },
      // ...
    ]
  }
}

This is impossible for a newcomer to derive, and pretty difficult to retain. Even if you understand the concept of loaders (which are, roughly, a way to tell Webpack how to handle a file given to require()), this is still incredibly opaque and error-prone. It doesn’t have to be and, honestly, a better format would likely make the code that manages this easier to write and maintain.

So, what is this configuration doing? It says that if someone require()s a file whose name ends in .css, then to run it through some code attached to the string “css-loader” and then run that result through some code attached to the string “style-loader”.

Where did “css-loader” and “style-loader” come from? And how do we know we can combine them in this way? We have only documentation to tell us that this could possibly work and is correct.

That the configuration requires a mini-language embedded in a string is telling us that there’s a lot of room for improvement.

Whatever “style-loader” and “css-loader” mean, they eventualy lead to some piece of code we’ve brought down from NPM. Why not use that piece of code as the configuration? Further, why are we hiding the idea of multiple loaders? And why bury the ordering inside a string? And why order the backwards from how both UNIX and English-speakers read?

Instead, what if we were more honest about what’s being configured. We want to configure a pipeline of loaders, which are functions that we pull in from a dependency:

var CssLoader   = require("css-loader");   # a package from package.json
var StyleLoader = require("style-loader"); # a package from package.json

module: {
  loaders: [
    {
      test: /\.css$/,
      loaderPipeline: [
        CssLoader,
        StyleLoader
      ]
    },

This requires much less context to get right. If we assume some knowledge of what a loader is as well as some basics of JavaScript, we can not only quickly understand this configuration, but we can modify it more easily. I can dig into the css-loader package and see what it exports. I can also easily see the ordering of how my loaders will be used, because we’re using a data structure for ordering things: an array.

Notice how much more clear this actually is. I would also be willing to bet that the code that interprets this configuration would be much simpler, because it’s being handed the actual pieces of code to execute and doesn’t have parse out that ridiculous exclamation point or look up “style-loader” in some internal data structure to find the actual code.

Designing a configuration system is like designing an API, but it’s even harder, because it’s an API that is crucially important to use properly, and something that by definition developers will not be using very frequently. It’s one of the few cases where learnability matters.

But, it could be worse. Here’s how to specify a project dependency in SBT, the Scala Build Tool:

libraryDependencies += "org.apache.derby" % "derby" % "10.4.1.3"
libraryDependencies += "org.scala-tools" %% "scala-stm" % "0.3"

Any ideas what those strings are? Or why the first line uses single percent-signs and the second line uses one double-one? Good luck googling that.