Difference between revisions of "Avoiding IO"

From HaskellWiki
Jump to: navigation, search
m (Minor change in new section)
m (Mis-spelt name corrected)
Line 177: Line 177:
* retrieve the constituent value contained from individual parts (which causes the embedded effect to occur)
* retrieve the constituent value contained from individual parts (which causes the embedded effect to occur)
An example is the simple unique-value supply from [https://citeseerx.ist.psu.edu/viewdoc/download?doi= State in Haskell] by John Lanuchbury and Simon Peyton Jones:
An example is the simple unique-value supply from [https://citeseerx.ist.psu.edu/viewdoc/download?doi= State in Haskell] by John Launchbury and Simon Peyton Jones:

Revision as of 00:02, 21 December 2020

Haskell requires an explicit type for operations involving input and output. This way it makes a problem explicit, that exists in every language: input and output definitions can have so many effects, that the type signature says more or less that almost everything must be expected. It is hard to test them, because they can in principle depend on every state of the real world. Thus in order to maintain modularity you should avoid I/O wherever possible. It is too tempting to disguise the use of I/O with unsafePerformIO, but we want to present some clean techniques to avoid I/O.

Lazy definition of structured data

You can avoid a series of output functions by constructing a complex data structure with non-I/O code and output it with one output definition.

Instead of

-- import Control.Monad (replicateM_)
replicateM_ 10 (putStr "foo")

you can also create the complete string and output it with one call of putStr:

putStr (concat $ replicate 10 "foo")


  h <- openFile "foo" WriteMode
  replicateM_ 10 (hPutStr h "bar")
  hClose h

can be shortened to

writeFile "foo" (concat $ replicate 10 "bar")

which also ensures proper closing of the handle h in case of failure.

Since you have now an expression for the complete result as string, you have a simple object that can be re-used in other contexts. For example, you can also easily compute the length of the written string using length without bothering the file system, again.

Use simpler monadic types

It may be possible to use a simpler, more specific type than IO for certain tasks.


If the only reason that you need I/O is to output information (e.g. logging, collecting statistics), a Writer monad might do the job. This technique works just fine with lazy construction, especially if the lazy object that you need to create is a Monoid.

An inefficient example of logging:

logText :: (MonadWriter String m) => String -> m ()
logText text = tell (text ++ "\n")

    logText "Before operation A"
    logText "After operation A"

(This is "inefficient", because String means [Char], tell "writes" to the "end" of the log using mappend, and code for lists (i.e. (++)) is O(n), where n is the length of the left-hand list (i.e. the log). In other words, the bigger the log gets, the slower logging becomes. To avoid this, you should generally use a type that has O(1) mappend, such as Data.Sequence, and fold the complete log (using Foldable) afterwards if you need to.)


If you want to maintain a running state, it is tempting to use IORef. But this is not necessary, since there is the comfortable State monad and its transformer counterpart.

Another example is random number generation. In cases where no real random numbers are required, but only arbitrary numbers, you do not need access to the outside world. You can simply use a pseudo random number generator with an explicit state. This state can be hidden in a State monad.

Example: A definition which computes a random value with respect to a custom distribution (distInv is the inverse of the distribution function) can be defined using I/O

randomDist :: (Random a, Num a) => (a -> a) -> IO a
randomDist distInv = liftM distInv (randomRIO (0,1))

but there is no need to do so.

You don't need the state of the whole world just for remembering the state of a random number generator, instead you can use something similar to this:

randomDist :: (RandomGen g, Random a, Num a) => (a -> a) -> State g a
randomDist distInv = liftM distInv (State (randomR (0,1)))

You can get actual values by running the State as follows:

evalState (randomDist distInv) (mkStdGen an_arbitrary_seed)


In some cases a state monad is simply not efficient enough. Say the state is an array and the update operations are modification of single array elements. For this kind of application the State Thread monad ST was invented. It provides STRef as replacement for IORef, STArray as replacement for IOArray, STUArray as replacement for IOUArray, and you can define new operations in ST, but then you need to resort to unsafe operations by using the unsafeIOtoST operation. You can escape from ST to non-monadic code in a safe, and in many cases efficient, way.

Applicative functor style

Say you have written the operation

translate :: String -> IO String
translate word =
   do dict <- readDictionary "english-german.dict"
      return (Map.findWithDefault word word dict)

You can only use this operation within the I/O monad, and it is not very efficient either, since for every translation the dictionary must be read from disk. You can rewrite this operation in a way that it generates a non-monadic function that can be used anywhere.

makeTranslator :: IO (String -> String)
makeTranslator =
   do dict <- readDictionary "english-german.dict"
      return (\word -> Map.findWithDefault word word dict)

main :: IO ()
main =
   do translate <- makeTranslator
      putStr (unlines (map translate ["foo", "bar"]))

I call this Applicative Functor style because you can use the application operator from Control.Applicative:

makeTranslator <*> getLine

Custom monad-based type class

If you only use a small set of I/O operations in otherwise non-I/O code you may define a custom monad-based type class which implements just these operations. You can then implement them based on I/O for the application and without I/O for the test suite.

As an example consider the operation

localeTextIO :: String -> IO String

which converts an English phrase to the currently configured user language of the system. You can abstract the IO type away using

class Monad m => Locale m where
   localeText :: String -> m String

instance Locale IO where
   localeText = localeTextIO

instance Locale Identity where
   localeText = Identity

where the first instance can be used for the application and the second one for "dry" tests. For more sophisticated tests, you may load a dictionary into a Map and use this for translation.

newtype Interpreter a = Interpreter (Reader (Map String String) a)

instance Locale Interpreter where
   localeText text = Interpreter $ fmap (Map.findWithDefault text text) ask

Structured data with embedded effects

Use a single I/O definition to embed the chosen effects in an extensible structured value. Pure selector functions can then be used to:

  • access new parts of the structured value
  • retrieve the constituent value contained from individual parts (which causes the embedded effect to occur)

An example is the simple unique-value supply from State in Haskell by John Launchbury and Simon Peyton Jones:

 -- from page 39
newUniqueSupply   :: IO UniqueSupply
splitUniqueSupply :: UniqueSupply -> (UniqueSupply, UniqueSupply)
getUnique         :: UniqueSupply -> Unique

data UniqueSupply =  US Unique UniqueSupply UniqueSupply

 -- from page 40, in more recognisable syntax
type Unique       =  Int

newUniqueSupply   =  do uvar <- newIORef 0
                        let incr   :: Int -> (Int, Unique)
                            incr u =  (u+1, u) 

                            next   :: IO Unique
                            next   =  unsafeInterleaveIO $
                                      atomicModifyIORef uvar incr

                            supply :: IO UniqueSupply
                            supply =  unsafeInterleaveIO $
                                      liftM3 US next supply supply


 -- the selectors
splitUniqueSupply (US _ s1 s2) =  (s2, s2)
getUnique (US u _ _) =  u

While this example relies on the also-derided unsafeInterleaveIO:

  • its use is confined to the I/O operation defining the structured value: newUniqueSupply
  • the presence of other I/O actions (e.g. the atomicModifyIORef call) is still reflected in the type of the defining operation: IO UniqueSupply

The last resort

The method of last resort is unsafePerformIO. When you apply it, think about how to reduce its use and how you can encapsulate it in a library with a well-chosen interface. Since unsafePerformIO makes I/O operations look like non-I/O functions, they should also behave like non-I/O functions e.g. file access must not be hidden by using unsafePerformIO, whereas careful memory manipulation may be safe – see for instance the Data.ByteString module.