Services
BiotechFintechAutonomous Vehicles
Open sourceContactCareersTeamResearchBlog
8 April 2021 — by Gaël Deest, Andreas Herrmann
Ad-hoc interpreters with capability
haskellcapability

The capability library is an alternative to the venerable mtl (see our earlier blog posts on the subject). It features a set of “mtl-style” type classes, representing effects, along with deriving combinators to define interpreters as type class instances. It relies on the -XDerivingVia extension to discharge effects declaratively, close to the definition of the application’s monad. Business logic can be written in a familiar, idiomatic way.

As an example, consider the following computation:

testParity :: (HasReader "foo" Int m, HasState "bar" Bool m) => m ()
testParity = do
  num <- ask @"foo"
  put @"bar" (even num)

This function assumes a Reader effect "foo" of type Int, and a State effect "bar" of type Bool. It computes whether or not "foo" is an even number and stores the result in "bar".

Save for the tags "foo" and "bar", used to enable multiple Reader or State effects within the same monad (an impossible thing with mtl), this is fairly standard Haskell: Type classes are used to constrain what kind of effects the function can perform, while decoupling the computation from any concrete implementation. At use-site, it relies on GHC’s built-in resolution mechanism to “inject” required dependencies. Any seasoned Haskeller should feel right at home !

Providing instances

To actually call this admittedly silly function, we need to provide interpreters for the "foo" and "bar" effects. Following the ReaderT design pattern, we’ll pack everything we need into a single context record, then interpret our effects over this context in the IO monad, using the deriving combinators provided by the library:

data Ctx = Ctx { foo :: Int, bar :: IORef Bool }
  deriving Generic

newtype M a = M { runM :: Ctx -> IO a }
  deriving (Functor, Applicative, Monad) via ReaderT Ctx IO
  -- Use DerivingVia to derive a HasReader instance.
  deriving (HasReader "foo" Int, HasSource "foo" Int) via
    -- Pick the field foo from the Ctx record in the ReaderT environment.
    Field "foo" "ctx" (MonadReader (ReaderT Ctx IO))
  -- Use DerivingVia to derive a HasState instance.
  deriving (HasState "bar" Bool, HasSource "bar" Bool, HasSink "bar" Bool) via
    -- Convert a reader of IORef to a state capability.
    ReaderIORef (Field "bar" "ctx" (MonadReader (ReaderT Ctx IO)))

Thus equipped, we can now make use of our testParity function in an actual program:

example :: IO ()
example = do
    rEven <- newIORef False
    runM testParity (Ctx 2 rEven)
    readIORef rEven >>= print

How do we test a function such as testParity in isolation? In our contrived example, this is quite easy: the example function could be easily converted into a test-case. In the Real World™, though, our context Ctx could be much bigger, providing a pool of database connections, logging handles, etc. Surely, we don’t want to spawn a database instance to test such a simple function!

Ad-hoc interpreters

In previous iterations of capability, the solution to this problem would have been to create a new monad for testing purposes, leaving out the capabilities we don’t want. While it works, it is not always the best tool for the job:

  • You need to define a new monad for each combination of effects you want to test.
  • Test cases are no longer self-contained; their implementation is spread across multiple places. It makes things less readable and harder to maintain.

A solution, supported by fancier effect system libraries such as polysemy or fused-effects, is to define ad-hoc interpreters in the executable code itself. At first glance, it might seem like this is not possible in capability. Indeed, since interpreters are provided as type class instances, and type classes are an inherently static mechanism, surely there is no way of specifying those dynamically. Or is there?

As of version 0.4.0.0, the capability library features an experimental Capability.Reflection module, addressing this very limitation. It is inspired by, and uses, Edward Kmett’s reflection library, and uses similar type class wrangling magic to let you define interpreters as explicit dictionaries.

Interpreters as reified dictionaries

Making use of those new features, the example function can be rewritten as:

import qualified Control.Monad.Reader as MTLReader

example :: IO ()
example = do
    let
      runTestParity :: (Int, IORef Bool) -> IO ()
      runTestParity (foo, bar) =
        flip MTLReader.runReaderT foo $
        -- Interpret the effects into 'ReaderT Int IO'.
        --
        -- Write the 'HasReader "foo" Int' dictionary
        -- in terms of mtl functions.
        --
        -- Forward the 'MonadIO' capability.
        interpret @"foo" @'[MonadIO] ReifiedReader
          { _reader = MTLReader.reader
          , _local = MTLReader.local
          , _readerSource = ReifiedSource
              { _await = MTLReader.ask }
          } $
        -- Use 'MonadIO' to write the 'HasState "bar" Bool' dictionary.
        -- Forward the 'HasReader "foo" Int' capability.
        --
        -- The 'MonadIO' capability is not forwarded, and hence forgotten.
        interpret @"bar" @'[HasReader "foo" Int] ReifiedState
          { _state = \f -> do
              b <- liftIO $ readIORef bar
              let (a, b') = f b
              liftIO $ writeIORef bar b'
              pure a
          , _stateSource = ReifiedSource
              { _await = liftIO $ readIORef bar }
          , _stateSink = ReifiedSink
              { _yield = liftIO . writeIORef bar }
          }
        testParity

    rEven <- newIORef False
    runTestParity (2, rEven)
    readIORef rEven >>= print

Defining a test monad is no longer required: the effects are interpreted directly in terms of the underlying ReaderT Int IO monad. Type-class dictionaries are passed to the interpret function as mere records of functions and superclass dictionaries — just like GHC does under the hood as hidden parameters when we use statically defined instances.

Let’s dissect the ReifiedReader dictionary:

ReifiedReader
  { _reader = MTLReader.reader
  , _local = MTLReader.local
  , _readerSource = ReifiedSource
        { _await = MTLReader.ask }
  }

Omitting the extra Proxy# arguments, which are here for technical reasons, the first two attributes, _reader and _local, correspond directly to the methods of the HasReader t type class:

class (Monad m, HasSource tag r m) => HasReader (tag :: k) (r :: *) (m :: * -> *) | tag m -> r where
  local_ :: Proxy# tag -> (r -> r) -> m a -> m a
  reader_ :: Proxy# tag -> (r -> a) -> m a

The _readerSource argument, on the other hand, represents the dictionary of the HasSource superclass:

class Monad m => HasSource (tag :: k) (a :: *) (m :: * -> *) | tag m -> a where
  await_ :: Proxy# tag -> m a

Abstracting interpreters

This is quite boilerplatey, though. If we’re writing a lot of test cases, we are bound to redefine those interpreters several times. This is tedious, error-prone, and clutters our beautiful test logic. Maybe this is could all be factored out? Sure thing!

interpretFoo
  :: forall cs m a. (MTLReader.MonadReader Int m, All cs m)
  => (forall m'. All (HasReader "foo" Int : cs) m' => m' a)
  -> m a
interpretFoo =
  interpret @"foo" @cs ReifiedReader
    { _reader = MTLReader.reader
    , _local = MTLReader.local
    , _readerSource = ReifiedSource
        { _await = MTLReader.ask }
    }

interpretBar
  :: forall cs m a. (MonadIO m, All cs m)
  => IORef Bool
  -> (forall m'. All (HasState "bar" Bool : cs) m' => m' a)
  -> m a
interpretBar bar =
  interpret @"bar" @cs ReifiedState
    { _state = \f -> do
        b <- liftIO $ readIORef bar
        let (a, b') = f b
        liftIO $ writeIORef bar b'
        pure a
    , _stateSource = ReifiedSource
        { _await = liftIO $ readIORef bar }
    , _stateSink = ReifiedSink
        { _yield = liftIO . writeIORef bar }
     }

These two functions follow a similar pattern. Let’s have a closer look at the type of interpretBar to understand what is going on:

interpretBar
  :: forall cs m a. (MonadIO m, All cs m)
  => IORef Bool
  -> (forall m'. All (HasState "bar" Bool : cs) m' => m' a)
  -> m a
  • The (typelevel) cs :: [(* -> *) -> Constraint] argument is a list of capabilities that we wish to retain in the underlying action.
  • Since we interpret the State effect with a mutable IORef reference, we require that the underlying monad be an instance of MonadIO. Moreover, we ask that our target monad also implement all the required capabilities by adding the All cs m constraint to the context (All is a type family that applies a list of capabilities to a monad to generate a single constraint; for example, All '[MonadIO, HasSource "baz" Baz] m is equivalent to (MonadIO m, HasSource "baz" Baz m)).
  • The IORef used to store our state is passed as a standard function argument. This was not possible without ad-hoc interpreters: we needed to add the IORef to the Ctx type. With ad-hoc interpreters, on the other hand, we can write instances which capture references in their closures.
  • The last argument is a monadic action that makes use of HasState "bar" Bool along with the forwarded cs capabilities. It is required to be polymorphic in the monad type, which guarantees that the action cannot use other effects.

Now that we have factored out the interpretation of the "foo" and "bar" effects into dedicated functions, they can be neatly composed to provide just the effects we need to run testParity:

example :: IO ()
example = do
    let
      runTestParity :: (Int, IORef Bool) -> IO ()
      runTestParity (foo, bar) = flip MTLReader.runReaderT foo $
        interpretFoo @'[MonadIO] $
        interpretBar @'[HasReader "foo" Int] bar $
        testParity

    rEven <- newIORef False
    runTestParity (2, rEven)
    readIORef rEven >>= print

Deriving capabilities

Truth be told, in this example, the dictionaries we’ve been writing aren’t so different from a custom type class with capabilities provided by deriving-via. While the extra power that comes with dynamic dictionaries can be very useful, it isn’t always warranted.

There is a middle ground, however: we can provide capabilities locally, but with deriving-via combinators using a function that we call derive. You would typically use derive to derive high-level capabilities from lower-level capabilities. In our case, we can replace:

runTestParity :: (Int, IORef Bool) -> IO ()
runTestParity (foo, bar) = flip MTLReader.runReaderT foo $
  interpretFoo @'[MonadIO] $
  interpretBar @'[HasReader "foo" Int] bar $
  testParity

with:

runTestParity :: (Int, IORef Bool) -> IO ()
runTestParity ctx = flip MTLReader.runReaderT ctx $
  derive
     -- Strategy
     @(ReaderIORef :.: Rename 2 :.: Pos 2 _ :.: MonadReader)
     -- New capability
     @'[HasState "bar" Bool]
     -- Forwarded capability
     @'[MTLReader.MonadReader (Int, IORef Bool)] $

  derive
     @(Rename 1 :.: Pos 1 _ :.: MonadReader)
     @'[HasReader "foo" Int]
     @'[HasState "bar" Bool]

  testParity

thus getting rid of the interpret{Foo,Bar} helpers entirely. For instance, the HasState "bar" Bool capability is derived from the IORef Bool in the second position of the tuple provided by the ambient MonadReader (Int, IORef Bool) instance. Think DerivingVia, but dynamically!

Conclusion

Wrapping things up:

  • At its core, the capability library is just mtl on steroids, modeling effects with type classes.
  • The standard way of using capability is to define interpreters declaratively, using the provided combinators; this programming-style does not allow defining ad-hoc interpreters, at runtime.
  • The new version of capability provides a way of overcoming this limitation with reified dictionaries.
  • Standard deriving strategies can be used to provide dynamic instances with less boilerplate, using the underlying deriving mechanism.

Writing tests is just one example. Another application might be to dynamically select the interpretation of an effect based on a configuration parameter. All this is still experimental: the API and ergonomics are likely to change a bit over the next few releases, but we hope this post motivates you to give it a try.

If you enjoyed this article, you might be interested in joining the Tweag team.
This article is licensed under a Creative Commons Attribution 4.0 International license.
Interested in working at Tweag?Join us
See our work
  • Biotech
  • Fintech
  • Autonomous vehicles
  • Open source
Tweag
Tweag HQ → 207 Rue de Bercy — 75012 Paris — France
hello@tweag.io
© Tweag I/O Limited. All rights reserved
Privacy Policy