Skip to content

suhailshergill/extensible-effects

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Extensible effects (Hackage, GHC)

Build Status Join the chat at https://gitter.im/suhailshergill/extensible-effects Stories in Ready Stories in progress

State and Error benchmarks Nondeterminism benchmarks

Implement effectful computations in a modular way!

The main monad of this package is Eff from Control.Eff. Eff r a is parameterized by the effect-list r and the monadic-result type a similar to other monads. It is the intention that all other monadic computations can be replaced by the use of Eff.

In case you know monad transformers or mtl: This library provides only one monad that includes all your effects instead of layering different transformers. It is not necessary to lift the computations through a monad stack. Also, it is not required to lift every Monad* typeclass (like MonadError) though all transformers.

Quickstart

To experiment with this library, it is suggested to write some lines within ghci.

Recommended Procedure:

  1. get extensible-effects by doing one of the following:
  • add extensible-effects as a dependency to a existing cabal or stack project
  • git clone https://github.com/suhailshergill/extensible-effects.git
  1. start stack ghci or cabal repl inside the project
  2. import Control.Eff and Control.Eff.QuickStart
  3. start with the examples provided in the documentation of the Control.Eff.QuickStart module

Tour through Extensible Effects

This section explains the basic concepts of this library.

The Effect List

import Control.Eff

The effect list r in the type Eff r a is a central concept in this library. It is a type-level list containing effect types.

If r is the empty list, then the computation Eff r (or Eff '[]) does not contain any effects to be handled and therefore is a pure computation. In this case, the result value can be retrieved by run :: Eff '[] a -> a

For programming within the Eff r monad, it is almost never necessary to list all effects that can appear. It suffices to state what types of effects are at least required. This is done via the Member t r typeclass. It describes that the type t occurs inside the list r. If you really want, you can still list all Effects and their order in which they are used (e.g. Eff '[Reader r, State s] a).

Handling Effects

Functions containing something like Eff (x ': r) a -> Eff r a handle effects.

The transition from the longer list of effects (x ': r) to just r is a type-level indicator that the effect x has been handled. Depending on the effect, some additional input might be required or some different output than just a is produced.

The handler functions typically are called run*, eval* or exec*.

Most common Effects

The most common effects used are Writer, Reader, Exception and State.

Writer, Reader and State all provide lazy and strict variants. Each has its own module that exposes a common interface. Importing one or the other controls whether the effect is strict or lazy in its inputs and outputs. It's recommended that you use the lazy variants by default unless you know you need strictness.

In this section, only the core functions associated with an effect are presented. Have a look at the modules for additional details.

The Exception Effect

import Control.Eff.Exception

The exception effect adds the possibility to exit a computation preemptively with an exception. Note that the exceptions from this library are handled by the programmer and have nothing to do with exceptions thrown inside the Haskell run-time.

throwError :: Member (Exc e) r => e -> Eff r a
runError :: Eff (Exc e ': r) a -> Eff r (Either e a)

An exception can be thrown using the throwError function. Its return type is Eff r a with an arbitrary type a. When handling the effect, the result-type changes to Either e a instead of just a. This indicates how the effect is handled: The returned value is either the thrown exception or the value returned from a successful computation.

The State Effect

import Control.Eff.State.{Lazy | Strict}

The state effect provides readable and writable state during a computation.

get :: Member (State s) r => Eff r s
put :: Member (State s) r => s -> Eff r ()
modify :: Member (State s) r => (s -> s) -> Eff r ()
runState :: s -> Eff (State s ': r) a -> Eff r (a, s)

The get function fetches the current state and makes it available within subsequent computation. The put function sets the state to a given value. modify updates the state using a mapping function by combining get and put.

The state-effect is handled using the runState function. It takes the initial state as an argument and returns the final state and effect-result.

The Reader Effect

import Control.Eff.Reader.{Strict | Lazy}

The reader effect provides an environment that can be read. Sometimes it is considered as read-only state.

ask :: Member (Reader e) r => e -> Eff r e
runReader :: e -> Eff (Reader e ': r) a -> Eff r a

ask can be used to retrieve the environment provided to runReader from within a computation which has the Reader effect.

The Writer Effect

import Control.Eff.Writer.{Strict | Lazy}

The writer effect allows one to collect messages during a computation. It is sometimes referred to as write-only state, which gets retrieved at the end of the computation.

tell :: Member (Writer w) r => w -> Eff r ()
runWriter :: (w -> b -> b) -> b -> Eff (Writer w ': r) a -> Eff r (a, b)
runListWriter :: Eff (Writer w ': r) a -> Eff r (a, [w])

Running a writer can be done in several ways. The most general function is runWriter which folds over all written values. However, if you only want to collect the values written, the runListWriter function does that.

Note that compared to mtl, the value written has no Monoid constraint on it and can be collected in any way.

Using multiple Effects

The main benefit of this library is that multiple effects can be included with ease.

If you need state and want to be able exit the computation with an exception, the type of your effectful computation would be the one of myComp below. Then, both the state and exception effect-functions can be used. To handle the effects, both the runState and runError functions have to be provided.

myComp :: (Member (Exc e) r, Member (State s) r) => Eff r a

run1 :: (Either e a, s)
run1 = run . runState initalState . runError $ myComp

run2 :: Either e (a, s)
run2 = run . runError . runState initalState $ myComp

However, the order of the handlers does matter for the final result. run1 and run2 show different executions of the same effectful computation. In run1, the returned state s is the last state seen before an eventual exception gets thrown (similar to the semantics in typical imperative languages), while in run2 the final state is returned only if the whole computation succeeded - transaction style.

Tips and tricks

There are several constructs that make it easier to work with the effects.

If only a part of the result is necessary for further computation, have a look at the eval* and exec* functions which exist for some effects. The exec* functions discard the result of the computation (the a type). The eval* functions discard the final result of the effect.

Instead of writing (Member (Exc e) r, Member (State s) r) => ... it is possible to use the type operator <:: and write [ Exc e, State s ] <:: r => ..., which has the same meaning.

It might be convenient to include the necessary language extensions and disable class-constraint warnings in your project's .cabal file (or package.yaml if you're using stack).

Explanation is a work in progress.

Other Effects

Work in progress.

Integration with IO

IO or any other monad can be used as a base type for the Lift effect. There may be at most one instance of the Lift effect in the effects list, and it must be handled last. Control.Eff.Lift exports the runLift handler and lift function which provide the ability to run arbitrary monadic actions. Also, there are convenient type aliases that allow for shorter type constraints.

f :: IO ()
f = runLift $ do printHello
                 printWorld

-- These two functions' types are equivalent.

printHello :: SetMember Lift (Lift IO) r => Eff r ()
printHello = lift (putStr "Hello")

printWorld :: Lifted IO r => Eff r ()
printWorld = lift (putStrLn " world!")

Note that, since Lift is a terminal effect, you do not need to use run to extract pure values. Instead, runLift returns a value wrapped in whatever monad you chose to use.

Additionally, the Lift effect provides MonadBase, MonadBaseControl, and MonadIO instances that may be useful, especially with packages like lifted-base, lifted-async, and other code that uses those typeclasses.

Integration with Monad Transformers

Work in progress.

Writing your own Effects and Handlers

Work in progress.

Other packages

Some other packages may implement various effects. Here is a rather incomplete list:

Background

extensible-effects is based on the work of Extensible Effects: An Alternative to Monad Transformers. The paper and the followup freer paper contain details. Additional explanation behind the approach can be found on Oleg's website.

Limitations

Ambiguity-Flexibility tradeoff

The extensibility of Eff comes at the cost of some ambiguity. A useful pattern to mitigate this ambiguity is to specialize calls to effect handlers using type application or type annotation. Examples of this pattern can be seen in Example/Test.hs.

Note, however, that the extensibility can also be traded back, but that detracts from some of the advantages. For details see section 4.1 in the paper.

Some examples where the cost of extensibility is apparent:

  • Common functions can't be grouped using typeclasses, e.g. the ask and getState functions can't be grouped in the case of:

    class Get t a where
      ask :: Member (t a) r => Eff r a

    ask is inherently ambiguous, since the type signature only provides a constraint on t, and nothing more. To specify fully, a parameter involving the type t would need to be added, which would defeat the point of having the grouping in the first place.

  • Code requires a greater number of type annotations. For details see #31.