Haskell Typeclasses versus Records

In the ogmarkup library, the output generation is driven by a set of templates and functions. In the earlier versions of ogmarkup, the configuration was the GenConf record (module Text.Ogmarkup.Private.Config). However, a few days ago, we merged a Pull Request in dev that changes things a bit.

Why Change GenConf?

Commit after commit, the definition of the GenConf record has grown in complexity until it gets more than ten fields. That made the declaration of a GenConf instance a little tricky. See for yourself:

It was a real mess. It was hard to write, hard to read and, as a consequence, very error prone. In these conditions, GHC is not a great assistance to the developers. Its error messages basically consist of a “too few arguments” complaint.

Switching To A Typeclass GenConf

A while ago, I read an article on typeclass-based libraries. We made the switch in the commit 888a86dc

Typeclass benefits

I do not regret our choice, for several reasons. The GenConf typeclass definition brings several default implementations. Another benefit rises when the time comes to declare a new instance. You can override several configuration functions without worrying about the declaration order. As a side effect, the functions name is visible, so it improve the declaration readability.

We kept the record field names, so the ogmarkup library update was pretty straightforward. I just updated the function signatures.


Yet, I found one major drawback to this change: GenConf is a multi-parameter type class. You have to enable at least two GHC extensions:

  • FlexibleInstances
  • MultiParamTypeClasses

Without them, you won’t be able to define a new GenConf instance. I am not very comfortable with the idea to oblige our potential future users to enable specific GHC extensions. At least, we will have to make it very clear in the documentation1.

All in all, I think this update was a good idea, especially because we are not immune to see the GenConf definition changes again in the future and becomes even more complex.

Side Note: Functional Dependency

Finally, let’s have a look at the typeclass definition.

I had already seen such strange declaration before, but I didn’t know what it meant before. c -> o means if you know which type is c, you can guess o. In other words, there is only one o for one c. It is called a functional dependency (see FunctionalDependencies GHC extension). The main thing to understand is that, without this dependency, for one c, GHC cannot guess which o to choose.

  1. It is still not the case, but it will.