About a year ago, Michael Snoyman made a blog post about the ReaderT pattern. Ostensibly, it's a blog post about preferring to encode state as IORef-s when you can. But that's not how we read it. What we saw first and foremost is a story about using extensional type classes describing specifically the effects that functions use, but not how the effects are implemented (see the MonadBalance story in Michael's blog post). We call these extensional type classes capabilities. Here is an excellent blog post from Matt Parsons which takes this aspect to heart.

Capability is a library to replace the MTL with capabilities. In this post, we'll argue why capabalities are important, why you should use them, and tell you about what it took to design a library of capabilities with good ergonomics. It turns out that a brand new language extension that shipped with GHC 8.6, -XDerivingVia, has a crucial role in this story.

The difference with the MTL

How is using capabilities, like in Snoyman's and Parsons's blog posts any different from using the well-trodden MTL? Quite a bit! The MTL's classes, like MonadReader, MonadState and their close relatives, are intensional: they reflect how the monad has been constructed. A monad MyM is a MonadState because it is a stack of monad transformers, one of which is a StateT.

This is because of what Haskell instances mean: if I have an instance,

instance MonadState s m => MonadState s (ReaderT r m)

while it may look like we're saying that it suffices to have MonadState s m for ReaderT r m to be MonadState s, what we are really saying is that MonadState s (ReaderT r m) means that MonadState s m. It defines a computation, rather than a deduction. In particular, we are not permitted to add the following instance:

data St = ...
instance MonadState St (ReaderT (IORef St) IO)

You may want to work around this issue using {-# OVERLAPPING #-} instances. However, in doing so, you are acting against the semantics of instances, and heading for trouble. For an example of issues with overlapping instances, see the warning at the end of the Overlapping instances section of the GHC manual.

In contrast, what the ReaderT pattern is advertising is extensional classes, indicating what an individual function is allowed to do (hence the name "capability"), regardless of how the monad is implemented.

The problem

Irrespective of whether you are willing to twist the arm of the MTL with {-# OVERLAPPING #-} instances, when programming with capabilities you will run into two kind of issues. The first is probably the lesser of the two, but has been a known pain point with the MTL for a while: it's that the MTL uses types (e.g. the state type) to discriminate layers. In other words: with the MTL, you can't have two states of the same type in your capabilities.

This, of course, is not a problem if you are writing your own type classes: you can simply use a different class for each piece of state (e.g. MonadBalance). This leads us to the second, more serious issue: lack of inference. With all its faults, the MTL gives a very comfortable environment: it defines plenty of generic instances, so that when we give a concrete monad stack, then all the instances are automatically computed for us. In contrast, with capabilities and the ReaderT pattern, we collect all the capabilities, and assemble a bespoke type to handle all of these instances. Haskell's instance resolution is simply not equipped for this.

The bottom line is: an insane amount of boilerplate. Custom type class definitions. A slew of instances at each main entry point (where a concrete type is defined for the monad).

Deriving Via

What we would really like is a way to use type class instances the other way around, compared to instance resolution. Instead of reading

instance MonadState s m => MonadState s (ReaderT r m)

as saying that MonadState, on ReaderT means that MonadState s m, and fixing the implementation, we would like to read it as: if I have an implementation of MonadState s m, then this is a possible implementation of MonadState on a ReaderT.

This is made possible by a new language extension available in the freshly released GHC 8.6: -XDerivingVia.

In short, -XDerivingVia is a generalisation of -XGeneralizedNewtypeDeriving that allows you not only to derive an instance for a newtype, but also from a newtype, or, and this is most relevant for us, from a combination of newtypes. For example:

{-# LANGUAGE DerivingVia #-}

import Data.Monoid (Sum (..))

newtype MyInt = MyInt Int
  deriving (Monoid, Semigroup) via Sum Int
  deriving Num via Int

In the above snippet we define MyInt which wraps an Int, and derive two instances for it. The Monoid instance is taken from Sum Int, and the Num instance is taken directly from Int. Note the via keyword in the deriving clauses. (In this example we could also have derived the Num instance using -XGeneralizedNewtypeDeriving.)

You will find a more complete introduction in Baldur Blondal's talk on the subject. If you want all the details, head to the proposal or the paper.

Enter capability

With the above piece of kit in hand, we can write the capability library. This library provides strategies that can be composed to derive capabilities using the -XDerivingVia language extension.

capability defines a set of standard, reusable capability type classes, such as HasReader, or HasState. Contrary to the MTL type classes these are parameterized by a name (aka tag), which makes it possible to refer to, say, multiple different states in the same computation, even if they correspond to the same type.

getAB :: (HasState "a" Int m, HasState "b" Int m) => m (Int, Int)
getAB = do
  a <- get @"a"  -- get state under tag "a"
  b <- get @"b"  -- get state under tag "b"
  pure (a, b)

The library then provides newtypes to derive instances of these capability type-classes in deriving via clauses, similar to Sum in the MyInt example above.

For example, given an MTL MonadReader instance, we can derive a HasReader capability as follows:

data AppData = ...

newtype AppM a = AppM (ReaderT AppData IO a)
  deriving (HasReader "appData" AppData) via
    MonadReader (ReaderT AppData IO)

We can also combine multiple newtypes to derive capability instances. Building on the above example, we can pick a field within AppData as follows:

data AppData = AppData { intRef :: IORef Int, ... }
  deriving Generic

newtype AppM a = AppM (ReaderT AppData IO a)
  deriving (HasReader "intRef" (IORef Int)) via
    Field "intRef" "appData" (MonadReader (ReaderT AppData IO))

The Field combinator takes two tags as arguments. The first specifies the field name, and also the new tag. The second specifies the old tag, which provides the record with the requested field. Note, that the MonadReader newtype can provide an instance for any tag. The Field combinator uses generic-lens under the hood, which is why AppData needs to have a Generic instance.

A worked example: combining writers without guilt

Let's consider a complete example to demonstrate how you could use capability in your own projects. The code is available in the capability repository if you want to follow along.

In this example we will receive a text as input and want to count occurrences of words and letters in the text, ignoring white space. To that end we will use a writer monad. Recall, that writer has the method tell :: w -> m (), which will mappend the given w to the current tally, starting from mempty. This requires a Monoid instance on w.

We start with counting single letters and words.

-- | Count the occurrence of a single letter.
countLetter ::
  HasWriter "letterCount" (Occurrences Char) m
  => Char -> m ()
countLetter letter = tell @"letterCount" (oneOccurrence letter)

-- | Count the occurrence of a single word.
countWord ::
  HasWriter "wordCount" (Occurrences Text) m
  => Text -> m ()
countWord word = tell @"wordCount" (oneOccurrence word)

The type Occurrences k is a newtype around a Map from values k to their count. Its Monoid instance will add occurrences in the expected fashion.

newtype Occurrences k = Occurrences (Map k Int)

instance Ord k => Monoid (Occurrences k)
  -- See repository for instance implementation.

-- | A single occurrence of the given value.
oneOccurrence :: k -> Occurrences k
oneOccurrence k = Occurrences $ Map.singleton k 1

Next, we combine countLetter and countWord to handle one word in the input text.

-- | Count the occurrence of a single word and all the letters in it.
countWordAndLetters ::
  ( HasWriter "letterCount" (Occurrences Char) m
  , HasWriter "wordCount" (Occurrences Text) m )
  => Text -> m ()
countWordAndLetters word = do
  countWord word
  mapM_ countLetter (Text.unpack word)

Finally, we can handle the full input text by first splitting it into its words and then applying the above function to each word.

-- | Count the occurrences of words and letters in a text,
-- excluding white space.
countWordsAndLettersInText ::
  ( HasWriter "letterCount" (Occurrences Char) m
  , HasWriter "wordCount" (Occurrences Text) m )
  => Text -> m ()
countWordsAndLettersInText text =
  mapM_ countWordAndLetters (Text.words text)

In a production setting we might prefer to stream the input, instead of holding he whole text in memory. For simplicity's sake we will omit this here.

With that we have written a program that demands two HasWriter capabilities.

Before we can execute this program we need to define a concrete implementation that provides these capabilities. This is where we make use of the deriving-via strategies that the library offers.

It is well-known that the writer monad provided by MTL has a space leak. In capability, we can derive a writer capability from a state capability instead, to avoid this issue. In fact, we don't even provide a way to derive a writer capability from a writer monad. Following the ReaderT pattern we derive the state capabilities from reader capabilities on IORefs.

First, we define the application context. A record holding two IORefs - one for each counter.

-- | Counter application context.
data CounterCtx = CounterCtx
  { letterCount :: IORef (Occurrences Char)
    -- ^ Counting letter occurrences.
  , wordCount :: IORef (Occurrences Text)
    -- ^ Counting word occurrences.
  } deriving Generic

Next, we define our application monad.

-- | Counter application monad.
newtype Counter a = Counter { runCounter :: CounterCtx -> IO a }
  deriving (Functor, Applicative, Monad) via (ReaderT CounterCtx IO)

Note that we use ReaderT in the deriving via clause as a strategy to derive the basic Functor, Applicative, and Monad instances.

Deriving the writer capabilities makes use of a large set of newtypes provided by the capability library. Each line after the via keyword corresponds to one newtype. Comments explain the purpose of the respective newtype. Read these from bottom to top.

  deriving (HasWriter "letterCount" (Occurrences Char)) via
    (WriterLog  -- Generate HasWriter using HasState of Monoid
    (ReaderIORef  -- Generate HasState from HasReader of IORef
    (Field "letterCount" "ctx"  -- Focus on the field letterCount
    (MonadReader  -- Generate HasReader using mtl MonadReader
    (ReaderT CounterCtx IO)))))  -- Use mtl ReaderT newtype

The "wordCount" writer is derived in the same way:

  deriving (HasWriter "wordCount" (Occurrences Text)) via
    WriterLog (ReaderIORef
    (Field "wordCount" "ctx" (MonadReader (ReaderT CounterCtx IO))))

The only thing left is to combine all these pieces into an executable program. We will take the text as an argument, and return an IO action that executes countWordsAndLettersInText using our Counter monad, and prints the resulting word and letter counts to standard output.

-- | Given a text count the occurrences of all words and letters in it,
-- excluding white space, and print the outcome to standard output.
wordAndLetterCount :: Text -> IO ()
wordAndLetterCount text = do

First, we setup the required IORefs and the counter context:

  lettersRef <- newIORef Map.empty
  wordsRef <- newIORef Map.empty
  let ctx = CounterCtx
        { letterCount = lettersRef
        , wordCount = wordsRef
        }

Then, we call countWordsAndLettersInText on the input text, and instantiate it using our Counter application monad:

  let counter :: Counter ()
      counter = countWordsAndLettersInText text

Finally, we run counter and print the results:

  runCounter counter ctx
  let printOccurrencesOf name ref = do
        putStrLn name
        occurrences <- readIORef ref
        ifor_ occurrences $ \item num ->
          putStrLn $ show item ++ ": " ++ show num
  printOccurrencesOf "Letters" lettersRef
  printOccurrencesOf "Words" wordsRef

Executing this program in GHCi should produce the following output:

>>> wordAndLetterCount "ab ba"
Letters
'a': 2
'b': 2
Words
"ab": 1
"ba": 1

This concludes the example. We invite you to experiment with this library. It is still in an early stage and the API is subject to change. However, your feedback will help to evolve it in a better direction.

A word on free monads

Another solution to many of the same problems has been known for a while: free monads and extensible effects. As it happens, capability and free monads can be formally compared. In this paper, Mauro Jaskelioff and Russell O'Connor, prove that free monads are a special case of capabilities (it's not phrased in these terms, of course, but that's what the paper amounts to).

So another way of looking at capability is that it is a library of extensible effects. It makes it possible to write effects which are not available to free monads: free monads can only model algebraic effects, while capabilities do not have such a restriction. For instance the HasCatch capability, giving a function the ability to catch errors, is not algebraic, hence not available to free monads.

However, the most important reason for us to develop capability is that we find this style of programming quite manageable and idiomatic, whereas free-monad programming quickly becomes unwieldy. This is an entirely subjective judgement of course, but we believe that it has slowed the adoption of extensible effects. Absolutely wonderful though they are! As a bonus, capabilities should be more efficient than free-monad-style programming because it doesn't rely on a tagged encoding.