pipes-4.2.0: Compositional pipelines

Safe HaskellSafe




Conventional Haskell stream programming forces you to choose only two of the following three features:

  • Effects
  • Streaming
  • Composability

If you sacrifice Effects you get Haskell's pure and lazy lists, which you can transform using composable functions in constant space, but without interleaving effects.

If you sacrifice Streaming you get mapM, forM and "ListT done wrong", which are composable and effectful, but do not return a single result until the whole list has first been processed and loaded into memory.

If you sacrifice Composability you write a tightly coupled read, transform, and write loop in IO, which is streaming and effectful, but is not modular or separable.

pipes gives you all three features: effectful, streaming, and composable programming. pipes also provides a wide variety of stream programming abstractions which are all subsets of a single unified machinery:

  • effectful Producers (like generators),
  • effectful Consumers (like iteratees),
  • effectful Pipes (like Unix pipes), and:
  • ListT done right.

All of these are connectable and you can combine them together in clever and unexpected ways because they all share the same underlying type.

pipes requires a basic understanding of monad transformers, which you can learn about by reading either:

  • the paper "Monad Transformers - Step by Step",
  • part III "Monads in the Real World" of the tutorial "All About Monads",
  • chapter 18 of "Real World Haskell" on monad transformers, or:
  • the documentation of the transformers library.

If you want a Quick Start guide to pipes, read the documentation in Pipes.Prelude from top to bottom.

This tutorial is more extensive and explains the pipes API in greater detail and illustrates several idioms.



The pipes library decouples stream processing stages from each other so that you can mix and match diverse stages to produce useful streaming programs. If you are a library writer, pipes lets you package up streaming components into a reusable interface. If you are an application writer, pipes lets you connect pre-made streaming components with minimal effort to produce a highly-efficient program that streams data in constant memory.

To enforce loose coupling, components can only communicate using two commands:

pipes has four types of components built around these two commands:

You can connect these components together in four separate ways which parallel the four above types:

As you connect components their types will change to reflect inputs and outputs that you've fused away. You know that you're done connecting things when you get an Effect, meaning that you have handled all inputs and outputs. You run this final Effect to begin streaming.


Producers are effectful streams of input. Specifically, a Producer is a monad transformer that extends any base monad with a new yield command. This yield command lets you send output downstream to an anonymous handler, decoupling how you generate values from how you consume them.

The following stdinLn Producer shows how to incrementally read in Strings from standard input and yield them downstream, terminating gracefully when reaching the end of the input:

-- echo.hs

import Control.Monad (unless)
import Pipes
import System.IO (isEOF)

--         +--------+-- A 'Producer' that yields 'String's
--         |        |
--         |        |      +-- Every monad transformer has a base monad.
--         |        |      |   This time the base monad is 'IO'.
--         |        |      |  
--         |        |      |  +-- Every monadic action has a return value.
--         |        |      |  |   This action returns '()' when finished
--         v        v      v  v
stdinLn :: Producer String IO ()
stdinLn = do
    eof <- lift isEOF        -- 'lift' an 'IO' action from the base monad
    unless eof $ do
        str <- lift getLine
        yield str            -- 'yield' the 'String'
        stdinLn              -- Loop

yield emits a value, suspending the current Producer until the value is consumed. If nobody consumes the value (which is possible) then yield never returns. You can think of yield as having the following type:

 yield :: Monad m => a -> Producer a m ()

The true type of yield is actually more general and powerful. Throughout the tutorial I will present type signatures like this that are simplified at first and then later reveal more general versions. So read the above type signature as simply saying: "You can use yield within a Producer, but you may be able to use yield in other contexts, too."

Click the link to yield to navigate to its documentation. There you will see that yield actually uses the Producer' (with an apostrophe) type synonym which hides a lot of polymorphism behind a simple veneer. The documentation for yield says that you can also use yield within a Pipe, too, because of this polymorphism:

 yield :: Monad m => a -> Pipe x a m ()

Use simpler types like these to guide you until you understand the fully general type.

for loops are the simplest way to consume a Producer like stdinLn. for has the following type:

 --                +-- Producer      +-- The body of the   +-- Result
 --                |   to loop       |   loop              |
 --                v   over          v                     v
 --                --------------    ------------------    ----------
 for :: Monad m => Producer a m r -> (a -> Effect m ()) -> Effect m r

(for producer body) loops over (producer), substituting each yield in (producer) with (body).

You can also deduce that behavior purely from the type signature:

  • The body of the loop takes exactly one argument of type (a), which is the same as the output type of the Producer. Therefore, the body of the loop must get its input from that Producer and nowhere else.
  • The return value of the input Producer matches the return value of the result, therefore for must loop over the entire Producer and not skip anything.

The above type signature is not the true type of for, which is actually more general. Think of the above type signature as saying: "If the first argument of for is a Producer and the second argument returns an Effect, then the final result must be an Effect."

Click the link to for to navigate to its documentation. There you will see the fully general type and underneath you will see equivalent simpler types. One of these says that if the body of the loop is a Producer, then the result is a Producer, too:

 for :: Monad m => Producer a m r -> (a -> Producer b m ()) -> Producer b m r

The first type signature I showed for for was a special case of this slightly more general signature because a Producer that never yields is also an Effect:

 data X  -- The uninhabited type

 type Effect m r = Producer X m r

This is why for permits two different type signatures. The first type signature is just a special case of the second one:

 for :: Monad m => Producer a m r -> (a -> Producer b m ()) -> Producer b m r

 -- Specialize 'b' to 'X'
 for :: Monad m => Producer a m r -> (a -> Producer X m ()) -> Producer X m r

 -- Producer X = Effect
 for :: Monad m => Producer a m r -> (a -> Effect     m ()) -> Effect     m r

This is the same trick that all pipes functions use to work with various combinations of Producers, Consumers, Pipes, and Effects. Each function really has just one general type, which you can then simplify down to multiple useful alternative types.

Here's an example use of a for loop, where the second argument (the loop body) is an Effect:

-- echo.hs

loop :: Effect IO ()
loop = for stdinLn $ \str -> do  -- Read this like: "for str in stdinLn"
    lift $ putStrLn str          -- The body of the 'for' loop

-- more concise: loop = for stdinLn (lift . putStrLn)

In this example, for loops over stdinLn and replaces every yield in stdinLn with the body of the loop, printing each line. This is exactly equivalent to the following code, which I've placed side-by-side with the original definition of stdinLn for comparison:

loop = do                      |  stdinLn = do
    eof <- lift isEOF          |      eof <- lift isEOF
    unless eof $ do            |      unless eof $ do
        str <- lift getLine    |          str <- lift getLine
        (lift . putStrLn) str  |          yield str
        loop                   |          stdinLn

You can think of yield as creating a hole and a for loop is one way to fill that hole.

Notice how the final loop only lifts actions from the base monad and does nothing else. This property is true for all Effects, which are just glorified wrappers around actions in the base monad. This means we can run these Effects to remove their lifts and lower them back to the equivalent computation in the base monad:

 runEffect :: Monad m => Effect m r -> m r

This is the real type signature of runEffect, which refuses to accept anything other than an Effect. This ensures that we handle all inputs and outputs before streaming data:

-- echo.hs

main :: IO ()
main = runEffect loop

... or you could inline the entire loop into the following one-liner:

main = runEffect $ for stdinLn (lift . putStrLn)

Our final program loops over standard input and echoes every line to standard output until we hit Ctrl-D to end the input stream:

$ ghc -O2 echo.hs
$ ./echo

The final behavior is indistinguishable from just removing all the lifts from loop:

main = do               |  loop = do
    eof <- isEof        |      eof <- lift isEof
    unless eof $ do     |      unless eof $ do
        str <- getLine  |          str <- lift getLine
        putStrLn str    |          (lift . putStrLn) str
        main            |          loop

This main is what we might have written by hand if we were not using pipes, but with pipes we can decouple the input and output logic from each other. When we connect them back together, we still produce streaming code equivalent to what a sufficiently careful Haskell programmer would have written.

You can also use for to loop over lists, too. To do so, convert the list to a Producer using each, which is exported by default from Pipes:

each :: Monad m => [a] -> Producer a m ()
each as = mapM_ yield as

Combine for and each to iterate over lists using a "foreach" loop:

>>> runEffect $ for (each [1..4]) (lift . print)

each is actually more general and works for any Foldable:

 each :: (Monad m, Foldable f) => f a -> Producer a m ()

So you can loop over any Foldable container or even a Maybe:

>>> runEffect $ for (each (Just 1)) (lift . print)


You might wonder why the body of a for loop can be a Producer. Let's test out this feature by defining a new loop body that creates three copies of every value:

-- nested.hs

import Pipes
import qualified Pipes.Prelude as P  -- Pipes.Prelude already has 'stdinLn'

triple :: Monad m => a -> Producer a m ()
triple x = do
    yield x
    yield x
    yield x

loop :: Producer String IO ()
loop = for P.stdinLn triple

-- This is the exact same as:
-- loop = for P.stdinLn $ \x -> do
--     yield x
--     yield x
--     yield x

This time our loop is a Producer that outputs Strings, specifically three copies of each line that we read from standard input. Since loop is a Producer we cannot run it because there is still unhandled output. However, we can use yet another for to handle this new repeated stream:

-- nested.hs

main = runEffect $ for loop (lift . putStrLn)

This creates a program which echoes every line from standard input to standard output three times:

$ ./nested

But is this really necessary? Couldn't we have instead written this using a nested for loop?

main = runEffect $
    for P.stdinLn $ \str1 ->
        for (triple str1) $ \str2 ->
            lift $ putStrLn str2

Yes, we could have! In fact, this is a special case of the following equality, which always holds no matter what:

 -- s :: Monad m =>      Producer a m ()  -- i.e. 'P.stdinLn'
 -- f :: Monad m => a -> Producer b m ()  -- i.e. 'triple'
 -- g :: Monad m => b -> Producer c m ()  -- i.e. '(lift . putStrLn)'

 for (for s f) g = for s (\x -> for (f x) g)

We can understand the rationale behind this equality if we first define the following operator that is the point-free counterpart to for:

 (~>) :: Monad m
      => (a -> Producer b m ())
      -> (b -> Producer c m ())
      -> (a -> Producer c m ())
 (f ~> g) x = for (f x) g

Using (~>) (pronounced "into"), we can transform our original equality into the following more symmetric equation:

 f :: Monad m => a -> Producer b m ()
 g :: Monad m => b -> Producer c m ()
 h :: Monad m => c -> Producer d m ()

 -- Associativity
 (f ~> g) ~> h = f ~> (g ~> h)

This looks just like an associativity law. In fact, (~>) has another nice property, which is that yield is its left and right identity:

-- Left Identity
yield ~> f = f
-- Right Identity
f ~> yield = f

In other words, yield and (~>) form a Category, specifically the generator category, where (~>) plays the role of the composition operator and yield is the identity. If you don't know what a Category is, that's okay, and category theory is not a prerequisite for using pipes. All you really need to know is that pipes uses some simple category theory to keep the API intuitive and easy to use.

Notice that if we translate the left identity law to use for instead of (~>) we get:

for (yield x) f = f x

This just says that if you iterate over a pure single-element Producer, then you could instead cut out the middle man and directly apply the body of the loop to that single element.

If we translate the right identity law to use for instead of (~>) we get:

for s yield = s

This just says that if the only thing you do is re-yield every element of a stream, you get back your original stream.

These three "for loop" laws summarize our intuition for how for loops should behave and because these are Category laws in disguise that means that Producers are composable in a rigorous sense of the word.

In fact, we get more out of this than just a bunch of equations. We also get a useful operator: (~>). We can use this operator to condense our original code into the following more succinct form that composes two transformations:

main = runEffect $ for P.stdinLn (triple ~> lift . putStrLn)

This means that we can also choose to program in a more functional style and think of stream processing in terms of composing transformations using (~>) instead of nesting a bunch of for loops.

The above example is a microcosm of the design philosophy behind the pipes library:

  • Define the API in terms of categories
  • Specify expected behavior in terms of category laws
  • Think compositionally instead of sequentially


Sometimes you don't want to use a for loop because you don't want to consume every element of a Producer or because you don't want to process every value of a Producer the exact same way.

The most general solution is to externally iterate over the Producer using the next command:

 next :: Monad m => Producer a m r -> m (Either r (a, Producer a m r))

Think of next as pattern matching on the head of the Producer. This Either returns a Left if the Producer is done or it returns a Right containing the next value, a, along with the remainder of the Producer.

However, sometimes we can get away with something a little more simple and elegant, like a Consumer, which represents an effectful sink of values. A Consumer is a monad transformer that extends the base monad with a new await command. This await command lets you receive input from an anonymous upstream source.

The following stdoutLn Consumer shows how to incrementally await Strings and print them to standard output, terminating gracefully when receiving a broken pipe error:

import Control.Monad (unless)
import Control.Exception (try, throwIO)
import qualified GHC.IO.Exception as G
import Pipes

--          +--------+-- A 'Consumer' that awaits 'String's
--          |        |
--          v        v
stdoutLn :: Consumer String IO ()
stdoutLn = do
    str <- await  -- 'await' a 'String'
    x   <- lift $ try $ putStrLn str
    case x of
        -- Gracefully terminate if we got a broken pipe error
        Left e@(G.IOError { G.ioe_type = t}) ->
            lift $ unless (t == G.ResourceVanished) $ throwIO e
        -- Otherwise loop
        Right () -> stdoutLn

await is the dual of yield: we suspend our Consumer until we receive a new value. If nobody provides a value (which is possible) then await never returns. You can think of await as having the following type:

 await :: Monad m => Consumer a m a

One way to feed a Consumer is to repeatedly feed the same input using (>~) (pronounced "feed"):

 --                 +- Feed       +- Consumer to    +- Returns new
 --                 |  action     |  feed           |  Effect
 --                 v             v                 v  
 --                 ----------    --------------    ----------
 (>~) :: Monad m => Effect m b -> Consumer b m c -> Effect m c

(draw >~ consumer) loops over (consumer), substituting each await in (consumer) with (draw).

So the following code replaces every await in stdoutLn with (lift getLine) and then removes all the lifts:

>>> runEffect $ lift getLine >~ stdoutLn

You might wonder why (>~) uses an Effect instead of a raw action in the base monad. The reason why is that (>~) actually permits the following more general type:

 (>~) :: Monad m => Consumer a m b -> Consumer b m c -> Consumer a m c

(>~) is the dual of (~>), composing Consumers instead of Producers.

This means that you can feed a Consumer with yet another Consumer so that you can await while you await. For example, we could define the following intermediate Consumer that requests two Strings and returns them concatenated:

doubleUp :: Monad m => Consumer String m String
doubleUp = do
    str1 <- await
    str2 <- await
    return (str1 ++ str2)

-- more concise: doubleUp = (++) <$> await <*> await

We can now insert this in between (lift getLine) and stdoutLn and see what happens:

>>> runEffect $ lift getLine >~ doubleUp >~ stdoutLn

doubleUp splits every request from stdoutLn into two separate requests and returns back the concatenated result.

We didn't need to parenthesize the above chain of (>~) operators, because (>~) is associative:

-- Associativity
(f >~ g) >~ h = f >~ (g >~ h)

... so we can always omit the parentheses since the meaning is unambiguous:

f >~ g >~ h

Also, (>~) has an identity, which is await!

-- Left identity
await >~ f = f

-- Right Identity
f >~ await = f

In other words, (>~) and await form a Category, too, specifically the iteratee category, and Consumers are also composable.


Our previous programs were unsatisfactory because they were biased either towards the Producer end or the Consumer end. As a result, we had to choose between gracefully handling end of input (using stdinLn) or gracefully handling end of output (using stdoutLn), but not both at the same time.

However, we don't need to restrict ourselves to using Producers exclusively or Consumers exclusively. We can connect Producers and Consumers directly together using (>->) (pronounced "pipe"):

 (>->) :: Monad m => Producer a m r -> Consumer a m r -> Effect m r

This returns an Effect which we can run:

-- echo2.hs

import Pipes
import qualified Pipes.Prelude as P  -- Pipes.Prelude also provides 'stdoutLn'

main = runEffect $ P.stdinLn >-> P.stdoutLn

This program is more declarative of our intent: we want to stream values from stdinLn to stdoutLn. The above "pipeline" not only echoes standard input to standard output, but also handles both end of input and broken pipe errors:

$ ./echo2

(>->) is "pull-based" meaning that control flow begins at the most downstream component (i.e. stdoutLn in the above example). Any time a component awaits a value it blocks and transfers control upstream and every time a component yields a value it blocks and restores control back downstream, satisfying the await. So in the above example, (>->) matches every await from stdoutLn with a yield from stdinLn.

Streaming stops when either stdinLn terminates (i.e. end of input) or stdoutLn terminates (i.e. broken pipe). This is why (>->) requires that both the Producer and Consumer share the same type of return value: whichever one terminates first provides the return value for the entire Effect.

Let's test this by modifying our Producer and Consumer to each return a diagnostic String:

-- echo3.hs

import Control.Applicative ((<$))  -- (<$) modifies return values
import Pipes
import qualified Pipes.Prelude as P
import System.IO

main = do
    hSetBuffering stdout NoBuffering
    str <- runEffect $
        ("End of input!" <$ P.stdinLn) >-> ("Broken pipe!" <$ P.stdoutLn)
    hPutStrLn stderr str

This lets us diagnose whether the Producer or Consumer terminated first:

$ ./echo3
End of input!
$ ./echo3 | perl -e 'close STDIN'
Broken pipe!

You might wonder why (>->) returns an Effect that we have to run instead of directly returning an action in the base monad. This is because you can connect things other than Producers and Consumers, like Pipes, which are effectful stream transformations.

A Pipe is a monad transformer that is a mix between a Producer and Consumer, because a Pipe can both await and yield. The following example Pipe is analagous to the Prelude's take, only allowing a fixed number of values to flow through:

-- take.hs

import Control.Monad (replicateM_)
import Pipes
import Prelude hiding (take)

--              +--------- A 'Pipe' that
--              |    +---- 'await's 'a's and
--              |    | +-- 'yield's 'a's
--              |    | |
--              v    v v
take ::  Int -> Pipe a a IO ()
take n = do
    replicateM_ n $ do                     -- Repeat this block 'n' times
        x <- await                         -- 'await' a value of type 'a'
        yield x                            -- 'yield' a value of type 'a'
    lift $ putStrLn "You shall not pass!"  -- Fly, you fools!

You can use Pipes to transform Producers, Consumers, or even other Pipes using the same (>->) operator:

 (>->) :: Monad m => Producer a m r -> Pipe   a b m r -> Producer b m r
 (>->) :: Monad m => Pipe   a b m r -> Consumer b m r -> Consumer a m r
 (>->) :: Monad m => Pipe   a b m r -> Pipe   b c m r -> Pipe   a c m r

For example, you can compose take after stdinLn to limit the number of lines drawn from standard input:

maxInput :: Int -> Producer String IO ()
maxInput n = P.stdinLn >-> take n
>>> runEffect $ maxInput 3 >-> P.stdoutLn
You shall not pass!

... or you can pre-compose take before stdoutLn to limit the number of lines written to standard output:

maxOutput :: Int -> Consumer String IO ()
maxOutput n = take n >-> P.stdoutLn
>>> runEffect $ P.stdinLn >-> maxOutput 3
<Exact same behavior>

Those both gave the same behavior because (>->) is associative:

(p1 >-> p2) >-> p3 = p1 >-> (p2 >-> p3)

Therefore we can just leave out the parentheses:

>>> runEffect $ P.stdinLn >-> take 3 >-> P.stdoutLn
<Exact same behavior>

(>->) is designed to behave like the Unix pipe operator, except with less quirks. In fact, we can continue the analogy to Unix by defining cat (named after the Unix cat utility), which reforwards elements endlessly:

cat :: Monad m => Pipe a a m r
cat = forever $ do
    x <- await
    yield x

cat is the identity of (>->), meaning that cat satisfies the following two laws:

-- Useless use of 'cat'
cat >-> p = p

-- Forwarding output to 'cat' does nothing
p >-> cat = p

Therefore, (>->) and cat form a Category, specifically the category of Unix pipes, and Pipes are also composable.

A lot of Unix tools have very simple definitions when written using pipes:

-- unix.hs

import Control.Monad (forever)
import Pipes
import qualified Pipes.Prelude as P  -- Pipes.Prelude provides 'take', too
import Prelude hiding (head)

head :: Monad m => Int -> Pipe a a m ()
head = P.take

yes :: Monad m => Producer String m r
yes = forever $ yield "y"

main = runEffect $ yes >-> head 3 >-> P.stdoutLn

This prints out 3 'y's, just like the equivalent Unix pipeline:

$ ./unix
$ yes | head -3

This lets us write "Haskell pipes" instead of Unix pipes. These are much easier to build than Unix pipes and we can connect them directly within Haskell for interoperability with the Haskell language and ecosystem.


pipes also provides a "ListT done right" implementation. This differs from the implementation in transformers because this ListT:

  • obeys the monad laws, and
  • streams data immediately instead of collecting all results into memory.

The latter property is actually an elegant consequence of obeying the monad laws.

To bind a list within a ListT computation, combine Select and each:

import Pipes

pair :: ListT IO (Int, Int)
pair = do
    x <- Select $ each [1, 2]
    lift $ putStrLn $ "x = " ++ show x
    y <- Select $ each [3, 4]
    lift $ putStrLn $ "y = " ++ show y
    return (x, y)

You can then loop over a ListT by using every:

 every :: Monad m => ListT m a -> Producer a m ()

So you can use your ListT within a for loop:

>>> runEffect $ for (every pair) (lift . print)
x = 1
y = 3
y = 4
x = 2
y = 3
y = 4

... or a pipeline:

>>> import qualified Pipes.Prelude as P
>>> runEffect $ every pair >-> P.print
<Exact same behavior>

Note that ListT is lazy and only produces as many elements as we request:

>>> runEffect $ for (every pair >-> P.take 2) (lift . print)
x = 1
y = 3
y = 4

You can also go the other way, binding Producers directly within a ListT. In fact, this is actually what Select was already doing:

 Select :: Producer a m () -> ListT m a

This lets you write crazy code like:

import Pipes
import qualified Pipes.Prelude as P

input :: Producer String IO ()
input = P.stdinLn >-> P.takeWhile (/= "quit")

name :: ListT IO String
name = do
    firstName <- Select input
    lastName  <- Select input
    return (firstName ++ " " ++ lastName)

Here we're binding standard input non-deterministically (twice) as if it were an effectful list:

>>> runEffect $ every name >-> P.stdoutLn
Daniel Fischer
Daniel Wagner
Donald Stewart
Donald Duck

Notice how this streams out values immediately as they are generated, rather than building up a large intermediate result and then printing all the values in one batch at the end.

ListT computations can be combined in more ways than Pipes, so try to program in ListT as much as possible and defer converting it to a Pipe as late as possible using loop.

You can combine ListT computations even if their inputs and outputs are completely different:

data In
    = InA A
    | InB B
    | InC C

data Out
    = OutD D
    | OutE E
    | OutF F

-- Independent computations

example1 :: A -> ListT IO D
example2 :: B -> ListT IO E
example3 :: C -> ListT IO F

-- Combined computation

total :: In -> ListT IO Out
total input = case input of
    InA a -> fmap OutD (example1 a)
    InB b -> fmap OutE (example2 b)
    InC c -> fmap OutF (example3 c)

Sometimes you have multiple computations that handle different inputs but the same output, in which case you don't need to unify their outputs:

-- Overlapping outputs

example1 :: A -> ListT IO Out
example2 :: B -> ListT IO Out
example3 :: C -> ListT IO Out

-- Combined computation

total :: In -> ListT IO Out
total input = case input of
    InA a -> example1 a
    InB b -> example2 b
    InC c -> example3 c

Other times you have multiple computations that handle the same input but produce different outputs. You can unify their outputs using the Monoid and Functor instances for ListT:

-- Overlapping inputs

example1 :: In -> ListT IO D
example2 :: In -> ListT IO E
example3 :: In -> ListT IO F

-- Combined computation

total :: In -> ListT IO Out
total input =
       fmap OutD (example1 input)
    <> fmap OutE (example2 input)
    <> fmap OutF (example3 input)

You can also chain ListT computations, feeding the output of the first computation as the input to the next computation:

-- End-to-end

aToB :: A -> ListT IO B
bToC :: B -> ListT IO C

-- Combined computation

aToC :: A -> LIstT IO C
aToC = aToB >=> bToC

... or you can just use do notation if you prefer.

However, the Pipe type is more general than ListT and can represent things like termination. Therefore you should consider mixing Pipes with ListT when you need to take advantage of these extra features:

-- Mix ListT with Pipes

example :: In -> ListT IO Out

pipe :: Pipe In Out IO ()
pipe = Pipes.takeWhile (not . isC) >-> loop example
    isC (InC _) = True
    isC  _      = False

So promote your ListT logic to a Pipe when you need to take advantage of these Pipe-specific features.


pipes is more powerful than meets the eye so this section presents some non-obvious tricks you may find useful.

Many pipe combinators will work on unusual pipe types and the next few examples will use the cat pipe to demonstrate this.

For example, you can loop over the output of a Pipe using for, which is how map is defined:

map :: Monad m => (a -> b) -> Pipe a b m r
map f = for cat $ \x -> yield (f x)

-- Read this as: For all values flowing downstream, apply 'f'

This is equivalent to:

map f = forever $ do
    x <- await
    yield (f x)

You can also feed a Pipe input using (>~). This means we could have instead defined the yes pipe like this:

yes :: Monad m => Producer String m r
yes = return "y" >~ cat

-- Read this as: Keep feeding "y" downstream

This is equivalent to:

yes = forever $ yield "y"

You can also sequence two Pipes together. This is how drop is defined:

drop :: Monad m => Int -> Pipe a a m r
drop n = do
    replicateM_ n await

This is equivalent to:

drop n = do
    replicateM_ n await
    forever $ do
        x <- await
        yield x

You can even compose pipes inside of another pipe:

customerService :: Producer String IO ()
customerService = do
    each [ "Hello, how can I help you?"        -- Begin with a script
         , "Hold for one second."
    P.stdinLn >-> P.takeWhile (/= "Goodbye!")  -- Now continue with a human

Also, you can often use each in conjunction with (~>) to traverse nested data structures. For example, you can print all non-Nothing elements from a doubly-nested list:

>>> runEffect $ (each ~> each ~> each ~> lift . print) [[Just 1, Nothing], [Just 2, Just 3]]

Another neat thing to know is that every has a more general type:

 every :: (Monad m, Enumerable t) => t m a -> Producer a m ()

Enumerable generalizes Foldable and if you have an effectful container of your own that you want others to traverse using pipes, just have your container implement the toListT method of the Enumerable class:

class Enumerable t where
    toListT :: Monad m => t m a -> ListT m a

You can even use Enumerable to traverse effectful types that are not even proper containers, like MaybeT:

input :: MaybeT IO String
input = do
    str <- lift getLine
    guard (str /= "Fail")
    return str
>>> runEffect $ every input >-> P.stdoutLn
>>> runEffect $ every input >-> P.stdoutLn


This tutorial covers the concepts of connecting, building, and reading pipes code. However, this library is only the core component in an ecosystem of streaming components. Derived libraries that build immediately upon pipes include:

  • pipes-concurrency: Concurrent reactive programming and message passing
  • pipes-parse: Minimal utilities for stream parsing
  • pipes-safe: Resource management and exception safety for pipes
  • pipes-group: Grouping streams in constant space

These libraries provide functionality specialized to common streaming domains. Additionally, there are several libraries on Hackage that provide even higher-level functionality, which you can find by searching under the "Pipes" category or by looking for packages with a pipes- prefix in their name. Current examples include:

  • pipes-extras: Miscellaneous utilities
  • pipes-network/pipes-network-tls: Networking
  • pipes-zlib: Compression and decompression
  • pipes-binary: Binary serialization
  • pipes-attoparsec: High-performance parsing
  • pipes-aeson: JSON serialization and deserialization

Even these derived packages still do not explore the full potential of pipes functionality, which actually permits bidirectional communication. Advanced pipes users can explore this library in greater detail by studying the documentation in the Pipes.Core module to learn about the symmetry of the underlying Proxy type and operators.

To learn more about pipes, ask questions, or follow pipes development, you can subscribe to the haskell-pipes mailing list at:


... or you can mail the list directly at:


Additionally, for questions regarding types or type errors, you might find the following appendix on types very useful.

Appendix: Types

pipes uses parametric polymorphism (i.e. generics) to overload all operations. You've probably noticed this overloading already:

This overloading is great when it works, but when connections fail they produce type errors that appear intimidating at first. This section explains the underlying types so that you can work through type errors intelligently.

Producers, Consumers, Pipes, and Effects are all special cases of a single underlying type: a Proxy. This overarching type permits fully bidirectional communication on both an upstream and downstream interface. You can think of it as having the following shape:

Proxy a' a b' b m r

Upstream | Downstream
    |         |
a' <==       <== b'  -- Information flowing upstream
    |         |
a  ==>       ==> b   -- Information flowing downstream
    |    |    |

The four core types do not use the upstream flow of information. This means that the a' and b' in the above diagram go unused unless you use the more advanced features provided in Pipes.Core.

pipes uses type synonyms to hide unused inputs or outputs and clean up type signatures. These type synonyms come in two flavors:

  • Concrete type synonyms that explicitly close unused inputs and outputs of the Proxy type
  • Polymorphic type synonyms that don't explicitly close unused inputs or outputs

The concrete type synonyms use () to close unused inputs and X (the uninhabited type) to close unused outputs:

type Effect = Proxy X () () X

 Upstream | Downstream
    |         |
X  <==       <== ()
    |         |
() ==>       ==> X
    |    |    |
type Producer b = Proxy X () () b

Upstream | Downstream
    |         |
X  <==       <== ()
    |         |
() ==>       ==> b
    |    |    |
  • Consumer: explicitly closes the downstream end, forbidding yields
type Consumer a = Proxy () a () X

Upstream | Downstream
    |         |
() <==       <== ()
    |         |
a  ==>       ==> X
    |    |    |
type Pipe a b = Proxy () a () b

Upstream | Downstream
    |         |
() <==       <== ()
    |         |
a  ==>       ==> b
    |    |    |

When you compose Proxys using (>->) all you are doing is placing them side by side and fusing them laterally. For example, when you compose a Producer, Pipe, and a Consumer, you can think of information flowing like this:

       Producer                Pipe                 Consumer
    +-----------+          +----------+          +------------+
    |           |          |          |          |            |
X  <==         <==   ()   <==        <==   ()   <==          <== ()
    |  stdinLn  |          |  take 3  |          |  stdoutLn  |
() ==>         ==> String ==>        ==> String ==>          ==> X
    |     |     |          |    |     |          |      |     |
    +-----|-----+          +----|-----+          +------|-----+
          v                     v                       v
          ()                    ()                      ()

Composition fuses away the intermediate interfaces, leaving behind an Effect:

    |                                   |
X  <==                                 <== ()
    |  stdinLn >-> take 3 >-> stdoutLn  |
() ==>                                 ==> X
    |                                   |

pipes also provides polymorphic type synonyms with apostrophes at the end of their names. These use universal quantification to leave open any unused input or output ends (which I mark using *):

  • Producer': marks the upstream end unused but still open
type Producer' b m r = forall x' x . Proxy x' x () b m r

Upstream | Downstream
    |         |
 * <==       <== ()
    |         |
 * ==>       ==> b
    |    |    |
  • Consumer': marks the downstream end unused but still open
type Consumer' a m r = forall y' y . Proxy () a y' y m r

Upstream | Downstream
    |         |
() <==       <== * 
    |         |
a  ==>       ==> *
    |    |    |
  • Effect': marks both ends unused but still open
type Effect' m r = forall x' x y' y . Proxy x' x y' y m r

Upstream | Downstream
    |         |
 * <==       <== * 
    |         |
 * ==>       ==> *
    |    |    |

Note that there is no polymorphic generalization of a Pipe.

Like before, if you compose a Producer', a Pipe, and a Consumer':

       Producer'               Pipe                 Consumer'
    +-----------+          +----------+          +------------+
    |           |          |          |          |            |
 * <==         <==   ()   <==        <==   ()   <==          <== *
    |  stdinLn  |          |  take 3  |          |  stdoutLn  |
 * ==>         ==> String ==>        ==> String ==>          ==> *
    |     |     |          |     |    |          |      |     |
    +-----|-----+          +-----|----+          +------|-----+
          v                      v                      v
          ()                     ()                     ()

... they fuse into an Effect':

    |                                   |
 * <==                                 <== *
    |  stdinLn >-> take 3 >-> stdoutLn  |
 * ==>                                 ==> *
    |                                   |

Polymorphic type synonyms come in handy when you want to keep the type as general as possible. For example, the type signature for yield uses Producer' to keep the type signature simple while still leaving the upstream input end open:

 yield :: Monad m => a -> Producer' a m ()

This type signature lets us use yield within a Pipe, too, because the Pipe type synonym is a special case of the polymorphic Producer' type synonym:

 type Producer' b m r = forall x' x . Proxy x' x () b m r
 type Pipe    a b m r =               Proxy () a () b m r

The same is true for await, which uses the polymorphic Consumer' type synonym:

 await :: Monad m => Consumer' a m a

We can use await within a Pipe because a Pipe is a special case of the polymorphic Consumer' type synonym:

 type Consumer' a   m r = forall y' y . Proxy () a y' y m r
 type Pipe      a b m r =               Proxy () a () b m r

However, polymorphic type synonyms cause problems in many other cases:

  • They usually give the wrong behavior when used as the argument of a function (known as the "negative" or "contravariant" position) like this:
f :: Producer' a m r -> ...  -- Wrong

f :: Producer  a m r -> ...  -- Right

The former function only accepts polymorphic Producers as arguments. The latter function accepts both polymorphic and concrete Producers, which is probably what you want.

  • Even when you desire a polymorphic argument, this induces a higher-ranked type, because it translates to a forall which you cannot factor out to the top-level to simplify the type signature:
f :: (forall x' x y' . Proxy x' x y' m r) -> ...

These kinds of type signatures require the RankNTypes extension.

  • Even when you have polymorphic type synonyms as the result of a function (i.e. the "positive" or "covariant" position), recent versions of ghc such still require the RankNTypes extension. For example, the fromHandle function from Pipes.Prelude requires RankNTypes to compile correctly on ghc-7.6.3:
fromHandle :: MonadIO m => Handle -> Producer' String m ()
  • You can't use polymorphic type synonyms inside other type constructors without the ImpredicativeTypes extension:
io :: IO (Producer' a m r)  -- Type error without ImpredicativeTypes
  • You can't partially apply polymorphic type synonyms:
stack :: MaybeT (Producer' a m) r  -- Type error

In these scenarios you should fall back on the concrete type synonyms, which are better behaved. If concrete type synonyms are unsatisfactory, then ask ghc to infer the most general type signature and use that.

For the purposes of debugging type errors you can just remember that:

 Input --+    +-- Output
         |    |
         v    v
Proxy a' a b' b m r
      ^    ^
      |    |
      +----+-- Ignore these

For example, let's say that you try to run the stdinLn Producer. This produces the following type error:

>>> runEffect P.stdinLn
    Couldn't match expected type `X' with actual type `String'
    Expected type: Effect m0 r0
      Actual type: Proxy X () () String IO ()
    In the first argument of `runEffect', namely `P.stdinLn'
    In the expression: runEffect P.stdinLn

runEffect expects an Effect, which is equivalent to the following type:

Effect          IO () = Proxy X () () X      IO ()

... but stdinLn type-checks as a Producer, which has the following type:

Producer String IO () = Proxy X () () String IO ()

The fourth type variable (the output) does not match. For an Effect this type variable should be closed (i.e. X), but stdinLn has a String output, thus the type error:

   Couldn't match expected type `X' with actual type `String'

Any time you get type errors like these you can work through them by expanding out the type synonyms and seeing which type variables do not match.

You may also consult this table of type synonyms to more easily compare them:

type Effect             = Proxy X  () () X
type Producer         b = Proxy X  () () b
type Consumer    a      = Proxy () a  () X
type Pipe        a    b = Proxy () a  () b

type Server        b' b = Proxy X  () b' b 
type Client   a' a      = Proxy a' a  () X

type Effect'            m r = forall x' x y' y . Proxy x' x y' y m r
type Producer'        b m r = forall x' x      . Proxy x' x () b m r
type Consumer'   a      m r = forall      y' y . Proxy () a y' y m r

type Server'       b' b m r = forall x' x      . Proxy x' x b' b m r
type Client'  a' a      m r = forall      y' y . Proxy a' a y' y m r

Appendix: Time Complexity

There are three functions that give quadratic time complexity when used in within pipes:

For example, the time complexity of this code segment scales quadratically with n:

import Control.Monad (replicateM)
import Pipes

quadratic :: Int -> Consumer a m [a]
quadratic n = replicateM n await

These three functions are generally bad practice to use, because all three of them correspond to "ListT done wrong", building a list in memory instead of streaming results.

However, sometimes situations arise where one deliberately intends to build a list in memory. The solution is to use the "codensity transformation" to transform the code to run with linear time complexity. This involves:

  • wrapping the code in the Codensity monad transformer (from Control.Monad.Codensity module of the kan-extensions package) using lift
  • applying sequence / replicateM / mapM
  • unwrapping the code using lowerCodensity

To illustrate this, we'd transform the above example to:

import Control.Monad.Codensity (lowerCodensity)

linear :: Monad m => Int -> Consumer a m [a]
linear n = lowerCodensity $ replicateM n $ lift await

This will produce the exact same result, but in linear time.