For some reason GADTs are a particular sticking point for many advocates of Simple Haskell. But sometimes GADTs are the natural form of expressing a concept.

By way of example, here’s a snippet of API I’ve borrowed from JuicyPixels:

data Image a = Image {
    imageWidth :: !Int,
    imageHeight :: !Int,
    imageData :: Vector (PixelBaseComponent a)
    }

data DynamicImage =
    ImageY8 (Image Pixel8) |
    ImageY16 (Image Pixel16) |
    ImageY32 (Image Pixel32) |
    ImageYF (Image PixelF) |
    ImageYA8 (Image PixelYA8) |
    ImageYA16 (Image PixelYA16) |
    ImageRGB8 (Image PixelRGB8) |
    ImageRGB16 (Image PixelRGB16) |
    ImageRGBF (Image PixelRGBF) |
    ImageRGBA8 (Image PixelRGBA8) |
    ImageRGBA16 (Image PixelRGBA16) |
    ImageYCbCr8 (Image PixelYCbCr8) |
    ImageCMYK8 (Image PixelCMYK8) |
    ImageCMYK16 (Image PixelCMYK16)

promoteImage :: ColorConvertible a b => Image a -> Image b 

JuicyPixels was originally released in 2012, and there’s always a good presumption against changing an API, so we can’t blame the developers for not making use of GADTs. But if we were starting a project like this from scratch today, how would we write such code, using the whole language?

The way DynamicImage has been defined makes it a bit difficult to work with. For example, if you want to get the width or height of a dynamic image, you’ll have to write cases for each constructor. And the convertibility from one pixel type to another is encoded at the type level (with ColorConvertible), but there’s no clean way to do it at the value level, if we wanted to provide that functionality dynamically.

We can make this code more expressive, and more elegant, by using a GADT that represents the pixel type of a dynamic image:

data PixelType a where
    PixelType8 :: PixelType Pixel8
    PixelType16 :: PixelType Pixel16
    PixelType32 :: PixelType Pixel32
    PixelTypeF :: PixelType PixelF
    PixelTypeYA8 :: PixelType PixelYA8
    PixelTypeYA16 :: PixelType PixelYA16
    PixelTypeRGB8 :: PixelType PixelRGB8
    PixelTypeRGB16 :: PixelType PixelRGB16
    PixelTypeRGBF :: PixelType PixelRGBF
    PixelTypeRGBA8 :: PixelType PixelRGBA8
    PixelTypeRGBA16 :: PixelType PixelRGBA16
    PixelTypeYCbCr8 :: PixelType PixelYCbCr8
    PixelTypeCMYK8 :: PixelType PixelCMYK8
    PixelTypeCMYK16 :: PixelType PixelCMYK16

data DynamicImage = forall a. DynamicImage (PixelType a) (Image a)

Any code to get properties of the contained Image is now much simpler:

dynamicImageSize :: DynamicImage -> (Int, Int)
dynamicImageSize (DynamicImage _ img) = (imageWidth img, imageHeight img)

We can also express convertibility dynamically, if we want to add that:

canPromotePixel :: PixelType a -> PixelType b -> Maybe (Dict (ColorConvertible a b))
-- provided as appropriate

promoteDynamicImage :: DynamicImage -> PixelType b -> Maybe (Image b)
promoteDynamicImage (DynamicImage pta img) ptb = do
    Dict <- canPromotePixel pta ptb
    return $ promoteImage img

You can see how with Haskell’s fancy shiny GADT feature, the code is now more expressive, and easier to understand for anyone who knows the language.

GADTs are actually not a particularly difficult concept to learn, for someone who already understands rank-n types and existential quantification. And in many situations, they provide a huge benefit in expressibility.

— Ashley Yakeley

The promise of Haskell over other languages is that it allows you to more cleanly and intuitively represent the application domain. This leads to more intelligible and maintainable code. But to take full advantage of what Haskell has to offer, you have to embrace the whole language. This means making use of any appropriate language feature, just as you would with any other language.

If you’re using Haskell in industry, it’s my belief that restricting your team to some “simple” subset of the language for complex problems will make your code more complex and more difficult for competent developers to understand. It will be more likely to have repetitive boilerplate, and more likely to require explicit error calls to handle “impossible” cases. It will be less comprehensible, and so less maintainable.

Here are some claims from “The Simple Haskell Initiative”, which I believe are flawed:

Fancy Haskell is costly to teams because it usually takes more time to understand and limits the pool of people who can effectively contribute.

By taking this attitude, you are not only committing your code to mediocrity, but your team too. Whole, elegant, Haskell, firing on all language cylinders, takes less time for competent developers to understand. And you can raise less-experienced developers to the necessary competence specific to your project.

Things that have been around longer will be more well-tested and understood by a larger group of people. Prefer tried and true techniques over the latest shiny library or language feature. The more foundational something is in your tech stack, the more conservative you should be about adopting new versions or approaches to that thing.

These are two separate things.

For libraries, maturity is certainly a valid concern. But it is unrelated to the structure of the library’s API.

For language features, the GHC team has a high bar for releasing new extensions, often involving formal proof. The implementation of features considered “fancy”, such as GADTs, type families, and polykinds, is understood to be as sound as that of any other part of the language. They are unlikely to be a particular source of compiler defects.

If you adopt a new thing, how much of its complexity will spread throughout the rest of your codebase? You should be more hesitant to adopt something if its complexity is going to spread through a larger portion of your codebase.

Code written in whole Haskell is less complex than “simple” Haskell for the same complex application. That’s the whole point of it.

There is no one definition of what language features count as “Simple Haskell”. Michael Snoyman seeks to define a “boring” subset of Haskell, but his recommended set of “boring” language extensions is quite broad, including most extensions relating to classes and type families, and even PolyKinds. Boring Haskell, in practice, seems to be close to Whole Haskell.

Sam Halliday, by contrast, seeks a vastly more restricted language, rejecting GADTs, type families, multi-parameter classes, and apparently even rank-n types and existential quantification. Such restrictions lead to unnecessarily complicated code, in my view. Here’s a simple example of how GADTs and existential data quantification can improve code generality and intelligibility.

What you should be doing

• Embrace the whole language.
• Set a high quality bar for code within your team.
• Mentor less-experienced developers.

Embrace the whole language. Pretty much every “fancy” feature of the whole GHC Haskell language has a productive purpose. That’s the point of all the academic research. As a competent Haskell developer, you should know when that purpose applies and how to make best use of each feature.

Note that language features are not exactly the same as language extensions. For example, the language gives you the option to either allow or prohibit “incoherent” use of class instances. In most cases, the best use of this language feature is to prohibit, as it can ensure a discipline that leads to more intelligibility and predictability.

Some language features are of relatively specialised use. Polykinded types are very useful for type-oriented applications (such as implementing a typed language interpreter). And, of course, some simpler or more straightforward projects might make use of relatively few language features.

Set a high quality bar for code within your team. Set code expectations as early as you can. Discuss ideas early, and make suggestions for design approaches. Haskell makes refactoring easier than many languages, take advantage of that as appropriate.

Mentor less-experienced developers. When you hire a Java, Python, or C++ developer, you can expect them to be fully competent in each of those languages. You can typically give them development ownership of progressively larger project features, and leave them to get on with it until code review. Given the current state of the Haskell job market, this may not be the case for Haskell developers.

If you hire junior developers who are not yet familiar with the whole language, you will need to mentor them. Take extra time with them to explain how features of the Haskell language work, how they are best used in general, and how you use them in your codebase.

If that sounds like an extra burden, bear in mind that not many developers make the effort to learn Haskell in the first place, and those that do are likely to have more aptitude to learn more about it. Invest in the people you hire. If you’re doing anything worthwhile, you’re in this for the long term.

… Why did you choose Haskell, anyway?

— Ashley Yakeley