Haskell Design Patterns: .Extended Modules

A clean way of extending third-party modules
Published on January 20, 2015 under the tag haskell

Introduction

For a long time, I have wanted to write a series of blogposts about Design Patterns in Haskell. This never really worked out. It is hard to write about Design Patterns.

First off, I have been writing Haskell for a long time, so mostly things feel natural and I do not really think about code in terms of Design Patterns.

Additionaly, I think there is a very, very thin line between what we call “Design Patterns” and what we call “Common Sense”. Too much on one side of the line, and you sound like a complete idiot. Too much on the other side of the line, and you sound like a pretentious fool who needs five UML diagrams in order to write a 100-line program.

However, in the last year, I have both been teaching more Haskell, and I have been reading even more code written by other people. The former made me think harder about why I do things, and the latter made me notice patterns I hadn’t thought of before, in particular if they were formulated in another way.

This has given me a better insight into these patterns, so I hope to write a couple of blogposts like this over the next couple of months. We will see how it goes – I am not exactly a prolific blogger.

The first blogpost deals with what I call .Extended Modules. While the general idea has probably been around for a while, the credit for this specific scheme goes to Bas van Dijk, Simon Meier, and Thomas Schilling.

.Extended Modules: the problem

This problem mainly resolves around organisation of code.

Haskell allows for building complex applications out of small functions that compose well. Naturally, if you are building a large application, you end up with a lot of these small functions.

Imagine we are building some web application, and we have a small function that takes a value and then sends it to the browser as JSON:

json :: (MonadSnap m, Aeson.ToJSON a) => a -> m ()
json x = do
    modifyResponse $ setContentType "application/json"
    writeLBS $ Aeson.encode x

The question is: where do we put this function? In small projects, these seem to inevitably end up inside the well-known Utils module. In larger, or more well-organised projects, it might end up in Foo.Web or Foo.Web.Utils.

However, if we think outside of the box, and disregard dependency problems and libraries including every possible utility function one can write, it is clearer where this function should go: in Snap.Core.

Putting it in Snap.Core is obviously not a solution – imagine the trouble library maintainers would have to deal with in order to include all these utility functions.

The basic scheme

The scheme we use to solve this is simple yet powerful: in our own application’s non-exposed modules list, we add Snap.Core.Extended.

src/Snap/Core/Extended.hs:

{-# LANGUAGE OverloadedStrings #-}
module Snap.Core.Extended
    ( module Snap.Core
    , json
    ) where

import qualified Data.Aeson as Aeson
import           Snap.Core

json :: (MonadSnap m, Aeson.ToJSON a) => a -> m ()
json x = do
    modifyResponse $ setContentType "application/json"
    writeLBS $ Aeson.encode x

The important thing to notice here is the re-export of module Snap.Core. This means that, everywhere in our application, we can use import Snap.Core.Extended as a drop-in replacement for import Snap.Core.

This also makes sharing code in a team easier. For example, say that you are looking for a catMaybes for Data.Vector.

Before, I would have considered either defining this in a where clause, or locally as a non-exported function. This works for single-person projects, but not when different people are working on different modules: you end up with five implementations of this method, scattered throughout the codebase.

With this scheme, however, it’s clear where to look for such a method: in Data.Vector.Extended. If it’s not there, you add it.

Aside from utility functions, this scheme also works great for orphan instances. For example, if we want to serialize a HashMap k v by converting it to [(k, v)], we can add a Data.HashMap.Strict.Extended module.

src/Data/HashMap/Strict/Extended.hs:

{-# OPTIONS_GHC -fno-warn-orphans #-}
module Data.HashMap.Strict.Extended
    ( module Data.HashMap.Strict
    ) where

import           Data.Binary         (Binary (..))
import           Data.Hashable       (Hashable)
import           Data.HashMap.Strict

instance (Binary k, Binary v, Eq k, Hashable k) => Binary (HashMap k v) where
    put = put . toList
    get = fmap fromList get

A special case of these .Extended modules is Prelude.Extended. Since you will typically import Prelude.Extended into almost all modules in your application, it is a great way to add a bunch of (very) common imports from base, so import noise is reduced.

This is, of course, quite subjective. Some might want to add a few specific functions to Prelude (as illustrated below), and others might prefer to add all of Control.Applicative, Data.List, Data.Maybe, and so on.

src/Prelude/Extended.hs:

module Prelude.Extended
    ( module Prelude
    , foldl'
    , fromMaybe
    ) where

import           Data.List  (foldl')
import           Data.Maybe (fromMaybe)
import           Prelude

Scaling up

The basic scheme breaks once our application consists of several cabal packages.

If we have a package acmecorp-web, which depends on acmecorp-core, we would have to expose Data.HashMap.Strict.Extended from acmecorp-core, which feels weird.

A simple solution is to create an unordered-containers-extended package (which is not uploaded to the public Hackage for obvious reasons). Then, you can export Data.HashMap.Strict.Extended from there.

This solution creates quite a lot of overhead. Having many modules is fine, since they are easy to manage – they are just files after all. Managing many packages, however, is harder: every package introduces a significant amount of overhead: for example, repos need to be maintained, and dependencies need to be managed explicitly in the cabal file.

An alternative solution is to simply put all of these modules together in a hackage-extended package. This solves the maintenance overhead and still gives you a very clean module hierarchy.

Conclusion

After using this scheme for over year in a large, constantly evolving Haskell application, it is clear to me that this is a great way to organise and share code in a team.

A side-effect of this scheme is that it becomes very convenient to consider some utility functions from these .Extended modules for inclusion in their respective libraries, since they all live in the same place. If they do get added, just remove the originals from hackage-extended, and the rest of your code doesn’t even break!

Thanks to Alex Sayers for proofreading!

ce0f13b2-4a83-4c1c-b2b9-b6d18f4ee6d2