Designing APIs for Extensibility

This is a snapshot of some content I wrote for the Commercial Haskell group's Haskell Documentation project. The official home for this content is in that repo. As I often do, I'm copying the content into the blog post itself for ease of viewing.

Every time you make a breaking change in your API, it means that- potentially- one or more of your users will need to change his/her code to adapt. Even if this update is trivial, it adds friction to the code maintenance process. On the other hand, we don't want to be constrained by bad design choices early on in a project, and sometimes a breaking API change is the best option.

The point of this document, however, is to give you a third option: design your APIs from the outset to be extensible. There are common techniques employed in the Haskell world to make APIs that are resilient to changing feature-sets, and by employing them early on in your design process, you can hopefully avoid the painful choices between a better API and happy users.

Almost all techniques start with implementation hiding. Guidelines here are simple: don't expose anything non-public. For example, if you write a number of helper functions, you may not want to start off by exposing them, since you're then telling users that these are good, stable functions to be relied upon. Instead, use explicit export lists on your modules and only include functions that are intended for public consumption.

More important- and more tricky- than functions are data constructors. In many cases, you want to avoid exposing the internals of your data types to users, to allow you to expand on them in the future. A common use case for this is some kind of a data type providing configuration information. Consider that you're going to communicate with some web services, so you write up the following API:

module MyAPI
    ( Settings (..)
    , makeAPICall
    ) where

data Settings = Settings
    { apiKey :: Text
    , hostName :: Text
    }

makeAPICall :: Settings -> Foo -> IO Bar

The way your users will access this will be something like:

makeAPICall Settings
    { apiKey = myAPIKey
    , hostName = "www.example.com"
    } myFoo

Now suppose a user points out that, in some cases, the standard port 80 is not used for the API call. So you add a new field port :: Int to your Settings constructor. This will break your user's code, since the port field will not be set.

Instead, a more robust way of specifying your API will look like:

module MyAPI
    ( Settings
    , mkSettings
    , setHostName
    , makeAPICall
    ) where

data Settings = Settings
    { apiKey :: Text
    , hostName :: Text
    }

-- | Create a @Settings@ value. Uses default value for host name.
mkSettings :: Text -- ^ API Key
           -> Settings
mkSettings key = Settings
    { apiKey = key
    , hostName = "www.example.com"
    }

setHostName :: Text -> Settings -> Settings
setHostName hn s = s { hostName = hn }

makeAPICall :: Settings -> Foo -> IO Bar

Now your user code will instead look like:

makeAPICall (mkSettings myAPIKey) myFoo

This has the following benefits:

• The user is not bothered to fill in default values (in this case, the hostname).
• Extending this API to allow for more fields in the future is trivial: add a new set* function. Internally, you'll add a field to Settings and set a default value in mkSettings.

One thing to note: please do not expose the field accessors directly. If you want to provide getter functions in addition to setters, write them explicitly, e.g.:

getHostName :: Settings -> Text
getHostName = hostName

The reason for this is that by exposing field accessors, users will be able to write code such as:

(mkSettings myAPIKey) { hostName = "www.example.org" }

This ties your hand for future internal improvements, since you are now required to keep a field of name hostName with type Text. By just using set and get functions, you can change your internal representation significantly and still provide a compatibility layer.

For those of you familiar with other languages: this is in fact quite similar to the approach taken in Java or C#. Just because Java does it doesn't mean it's wrong.

Note that this advice is different to, and intended to supersede, the settings type approach. Projects like Warp which previously used that settings type approach are currently migrating to this more extensible approach.

Also, while settings have been used here as a motivating example, the same advice applies to other contexts.

Internal modules

One downside of implementation hiding is that it can make it difficult for users to do things you didn't intend for them to do with your API. You can always add more functionality on demand, but the delay can be a major nuissance for users. A compromise solution in the Haskell community is to provide a .Internal module for your project which exports non-quite-public components. For example, in wai, the Response constructors are exposed in a Network.Wai.Internal module. Normally, users are supposed to use smart constructors like responseFile, but occasionally they'll want more fine-grained control.