extensible-effects: An Alternative to Monad Transformers

This is a package candidate release! Here you can preview how this package release will appear once published to the main package index (which can be accomplished via the 'maintain' link below). Please note that once a package has been published to the main package index it cannot be undone! Please consult the package uploading documentation for more information.

[maintain] [Publish]


This package introduces datatypes for typeclass-constrained effects, as an alternative to monad-transformer based (datatype-constrained) approach of multi-layered effects. Any help is appreciated!

[Skip to Readme]


Versions 1.0, 1.1.0, 1.2.0, 1.2.1, 1.4.1, 1.5.0, 1.6.0, 1.7.1,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,
Change log None available
Dependencies base (>=4.7 && <4.12), monad-control (==1.0.*), transformers (>=, transformers-base (==0.4.*) [details]
License MIT
Author Oleg Kiselyov, Amr Sabry, Cameron Swords, Ben Foppa
Maintainer suhailshergill@gmail.com
Category Control, Effect
Home page https://github.com/suhailshergill/extensible-effects
Source repo head: git clone https://github.com/suhailshergill/extensible-effects.git
Uploaded by shergill at 2018-06-24T18:23:15Z



Manual Flags


Force usage of OpenUnion51.hs implementation

Automatic Flags

Use -f <flag> to enable a flag, or -f -<flag> to disable that flag. More info


Maintainer's Corner

For package maintainers and hackage trustees

Readme for extensible-effects-

[back to package description]

Extensible effects (Hackage)

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

Implement effectful computations in a modular way!

The main and only monad is built upon 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.


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:
  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.

For the Writer, Reader and State, there are lazy and a strict variants. Each has its own module that provide the same interface. By importing one or the other, it can be controlled if the effect is strict or lazy in its inputs and outputs. Unless required otherwise, it is suggested to use the lazy variants.

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 functions accesses the current state and makes it usable within the further computation. The put function sets the state to the 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

The environment given to the handle the reader effect is the one given during the computation if asked for.

The Writer Effect

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

The writer effect allows to output 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 that folds over all written values. However, if you only want to collect the 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 the 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 the disabling of the class-constriant warnings in the cabal-file of your project. Explanation is work in progress

Other Effects

work in progress

Integration with IO

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

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

-- 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 value. Instead, runLift returns a value wrapped in whatever monad you chose to use.

In addition, 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:


extensible-effects is based on the work 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.


Ambiguity-Flexibility tradeoff

The extensibility of Eff comes at the cost of some ambiguity. A useful pattern to mitigate the ambiguity is to specialize the call to the handler of effects 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:

Current implementation only supports GHC version 7.8 and above

This is not a fundamental limitation of the design or the approach, but there is an overhead with making the code compatible across a large number of GHC versions. If this is needed, patches are welcome :)