Beautiful Streaming, Concurrent and Reactive Composition

Version on this page:0.5.2
LTS Haskell 22.30:0.10.1@rev:3
Stackage Nightly 2024-07-23:0.10.1@rev:3
Latest on Hackage:0.10.1@rev:3

See all snapshots streamly appears in

BSD-3-Clause licensed by Harendra Kumar
Maintained by [email protected]
This version can be pinned in stack with:streamly-0.5.2@sha256:5a00c20700bfbe37b356dcdca5db8866d41661223bf8593cbbe4906ef553df36,19624

Module documentation for 0.5.2


Streaming Concurrently

Haskell lists express pure computations using composable stream operations like :, unfold, map, filter, zip and fold. Streamly is exactly like lists except that it can express sequences of pure as well as monadic computations aka streams. More importantly, it can express monadic sequences with concurrent execution semantics without introducing any additional APIs.

Streamly expresses concurrency using standard, well known abstractions. Concurrency semantics are defined for list operations, semigroup, applicative and monadic compositions. Programmer does not need to know any low level notions of concurrency like threads, locking or synchronization. Concurrent and non-concurrent programs are fundamentally the same. A chosen segment of the program can be made concurrent by annotating it with an appropriate combinator. We can choose a combinator for lookahead style or asynchronous concurrency. Concurrency is automatically scaled up or down based on the demand from the consumer application, we can finally say goodbye to managing thread pools and associated sizing issues. The result is truly fearless and declarative monadic concurrency.

Where to use streamly?

Streamly is a general purpose programming framwework. It can be used equally efficiently from a simple Hello World! program to a massively concurrent application. The answer to the question, “where to use streamly?” - would be similar to the answer to - “Where to use Haskell lists or the IO monad?”. Streamly generalizes lists to monadic streams, and the IO monad to non-deterministic and concurrent stream composition. The IO monad is a special case of streamly; if we use single element streams the behavior of streamly becomes identical to the IO monad. The IO monad code can be replaced with streamly by just prefixing the IO actions with liftIO, without any other changes, and without any loss of performance. Pure lists too are a special case of streamly; if we use Identity as the underlying monad, streamly streams turn into pure lists. Non-concurrent programs are just a special case of concurrent ones, simply adding a combinator turns a non-concurrent program into a concurrent one.

In other words, streamly combines the functionality of lists and IO, with builtin concurrency. If you want to write a program that involves IO, concurrent or not, then you can just use streamly as the base monad, in fact, you could even use streamly for pure computations, as streamly performs at par with pure lists or vector.

Why data flow programming?

If you need some convincing for using streaming or data flow programming paradigm itself then try to answer this question - why do we use lists in Haskell? It boils down to why we use functional programming in the first place. Haskell is successful in enforcing the functional data flow paradigm for pure computations using lists, but not for monadic computations. In the absence of a standard and easy to use data flow programming paradigm for monadic computations, and the IO monad providing an escape hatch to an imperative model, we just love to fall into the imperative trap, and start asking the same fundamental question again - why do we have to use the streaming data model?

Show me an example

Here is an IO monad code to list a directory recursively:

import Control.Monad.IO.Class (liftIO)
import Path.IO (listDir, getCurrentDir) -- from path-io package

listDirRecursive = getCurrentDir >>= readdir
    readdir dir = do
      (dirs, files) <- listDir dir
      liftIO $ mapM_ putStrLn
             $ map show dirs ++ map show files
      foldMap readdir dirs

This is your usual IO monad code, with no streamly specific code whatsoever. This is how you can run this:

main :: IO ()
main = listDirRecursive

And, this is how you can run exactly the same code using streamly with lookahead style concurrency, the only difference is that this time multiple directories are read concurrently:

import Streamly (runStream, aheadly)

main :: IO ()
main = runStream $ aheadly $ listDirRecursive

Isn’t that magical? What’s going on here? Streamly does not introduce any new abstractions, it just uses standard abstractions like Semigroup or Monoid to combine monadic streams concurrently, the way lists combine a sequence of pure values non-concurrently. The foldMap in the code above turns into a concurrent monoidal composition of a stream of readdir computations.

How does it perform?

Providing monadic streaming and high level declarative concurrency does not mean that streamly compromises with performance in any way. The non-concurrent performance of streamly competes with lists and the vector library. The concurrent performance is as good as it gets, see concurrency benchmarks for detailed performance results and a comparison with the async package.

The following chart shows a summary of the cost of key streaming operations processing a million elements. The timings for streamly and vector are in the 600-700 microseconds range and therefore can barely be seen in the graph. For more details, see streaming benchmarks.

Streaming Operations at a Glance

Streaming Pipelines

The following snippet provides a simple stream composition example that reads numbers from stdin, prints the squares of even numbers and exits if an even number more than 9 is entered.

import Streamly
import qualified Streamly.Prelude as S
import Data.Function ((&))

main = runStream $
       S.repeatM getLine
     & fmap read
     & S.filter even
     & S.takeWhile (<= 9)
     & fmap (\x -> x * x)
     & S.mapM print

Unlike pipes or conduit and like vector and streaming, streamly composes stream data instead of stream processors (functions). A stream is just like a list and is explicitly passed around to functions that process the stream. Therefore, no special operator is needed to join stages in a streaming pipeline, just the standard function application ($) or reverse function application (&) operator is enough. Combinators are provided in Streamly.Prelude to transform or fold streams.

Concurrent Stream Generation

Monadic construction and generation functions e.g. consM, unfoldrM, replicateM, repeatM, iterateM and fromFoldableM etc. work concurrently when used with appropriate stream type combinator (e.g. asyncly, aheadly or parallely).

The following code finishes in 3 seconds (6 seconds when serial):

> let p n = threadDelay (n * 1000000) >> return n
> S.toList $ aheadly $ p 3 |: p 2 |: p 1 |: S.nil

> S.toList $ parallely $ p 3 |: p 2 |: p 1 |: S.nil

The following finishes in 10 seconds (100 seconds when serial):

runStream $ asyncly $ S.replicateM 10 $ p 10

Concurrent Streaming Pipelines

Use |& or |$ to apply stream processing functions concurrently. The following example prints a “hello” every second; if you use & instead of |& you will see that the delay doubles to 2 seconds instead because of serial application.

main = runStream $
      S.repeatM (threadDelay 1000000 >> return "hello")
   |& S.mapM (\x -> threadDelay 1000000 >> putStrLn x)

Mapping Concurrently

We can use mapM or sequence functions concurrently on a stream.

> let p n = threadDelay (n * 1000000) >> return n
> runStream $ aheadly $ S.mapM (\x -> p 1 >> print x) (serially $ repeatM (p 1))

Serial and Concurrent Merging

Semigroup and Monoid instances can be used to fold streams serially or concurrently. In the following example we compose ten actions in the stream, each with a delay of 1 to 10 seconds, respectively. Since all the actions are concurrent we see one output printed every second:

import Streamly
import qualified Streamly.Prelude as S
import Control.Concurrent (threadDelay)

main = S.toList $ parallely $ foldMap delay [1..10]
 where delay n = S.yieldM $ threadDelay (n * 1000000) >> print n

Streams can be combined together in many ways. We provide some examples below, see the tutorial for more ways. We use the following delay function in the examples to demonstrate the concurrency aspects:

import Streamly
import qualified Streamly.Prelude as S
import Control.Concurrent

delay n = S.yieldM $ do
    threadDelay (n * 1000000)
    tid <- myThreadId
    putStrLn (show tid ++ ": Delay " ++ show n)


main = runStream $ delay 3 <> delay 2 <> delay 1
ThreadId 36: Delay 3
ThreadId 36: Delay 2
ThreadId 36: Delay 1


main = runStream . parallely $ delay 3 <> delay 2 <> delay 1
ThreadId 42: Delay 1
ThreadId 41: Delay 2
ThreadId 40: Delay 3

Nested Loops (aka List Transformer)

The monad instance composes like a list monad.

import Streamly
import qualified Streamly.Prelude as S

loops = do
    x <- S.fromFoldable [1,2]
    y <- S.fromFoldable [3,4]
    S.yieldM $ putStrLn $ show (x, y)

main = runStream loops

Concurrent Nested Loops

To run the above code with, lookahead style concurrency i.e. each iteration in the loop can run run concurrently by but the results are presented in the same order as serial execution:

main = runStream $ aheadly $ loops

To run it with depth first concurrency yielding results asynchronously in the same order as they become available (deep async composition):

main = runStream $ asyncly $ loops

To run it with breadth first concurrency and yeilding results asynchronously (wide async composition):

main = runStream $ wAsyncly $ loops

The above streams provide lazy/demand-driven concurrency which is automatically scaled as per demand and is controlled/bounded so that it can be used on infinite streams. The following combinator provides strict, unbounded concurrency irrespective of demand:

main = runStream $ parallely $ loops

To run it serially but interleaving the outer and inner loop iterations (breadth first serial):

main = runStream $ wSerially $ loops

Magical Concurrency

Streams can perform semigroup (<>) and monadic bind (>>=) operations concurrently using combinators like asyncly, parallelly. For example, to concurrently generate squares of a stream of numbers and then concurrently sum the square roots of all combinations of two streams:

import Streamly
import qualified Streamly.Prelude as S

main = do
    s <- S.sum $ asyncly $ do
        -- Each square is performed concurrently, (<>) is concurrent
        x2 <- foldMap (\x -> return $ x * x) [1..100]
        y2 <- foldMap (\y -> return $ y * y) [1..100]
        -- Each addition is performed concurrently, monadic bind is concurrent
        return $ sqrt (x2 + y2)
    print s

The concurrency facilities provided by streamly can be compared with OpenMP and Cilk but with a more declarative expression.

Rate Limiting

For bounded concurrent streams, stream yield rate can be specified. For example, to print hello once every second you can simply write this:

import Streamly
import Streamly.Prelude as S

main = runStream $ asyncly $ avgRate 1 $ S.repeatM $ putStrLn "hello"

For some practical uses of rate control, see AcidRain.hs and CirclingSquare.hs . Concurrency of the stream is automatically controlled to match the specified rate. Rate control works precisely even at throughputs as high as millions of yields per second. For more sophisticated rate control see the haddock documentation.


From a library user point of view, there is nothing much to learn or talk about exceptions. Synchronous exceptions work just the way they are supposed to work in any standard non-concurrent code. When concurrent streams are combined together, exceptions from the constituent streams are propagated to the consumer stream. When an exception occurs in any of the constituent streams other concurrent streams are promptly terminated. Exceptions can be thrown using the MonadThrow instance.

There is no notion of explicit threads in streamly, therefore, no asynchronous exceptions to deal with. You can just ignore the zillions of blogs, talks, caveats about async exceptions. Async exceptions just don’t exist. Please don’t use things like myThreadId and throwTo just for fun!

Reactive Programming (FRP)

Streamly is a foundation for first class reactive programming as well by virtue of integrating concurrency and streaming. See AcidRain.hs for a console based FRP game example and CirclingSquare.hs for an SDL based animation example.


Streamly, short for streaming concurrently, provides monadic streams, with a simple API, almost identical to standard lists, and an in-built support for concurrency. By using stream-style combinators on stream composition, streams can be generated, merged, chained, mapped, zipped, and consumed concurrently – providing a generalized high level programming framework unifying streaming and concurrency. Controlled concurrency allows even infinite streams to be evaluated concurrently. Concurrency is auto scaled based on feedback from the stream consumer. The programmer does not have to be aware of threads, locking or synchronization to write scalable concurrent programs.

Streamly is a programmer first library, designed to be useful and friendly to programmers for solving practical problems in a simple and concise manner. Some key points in favor of streamly are:

  • Simplicity: Simple list like streaming API, if you know how to use lists then you know how to use streamly. This library is built with simplicity and ease of use as a design goal.
  • Concurrency: Simple, powerful, and scalable concurrency. Concurrency is built-in, and not intrusive, concurrent programs are written exactly the same way as non-concurrent ones.
  • Generality: Unifies functionality provided by several disparate packages (streaming, concurrency, list transformer, logic programming, reactive programming) in a concise API.
  • Performance: Streamly is designed for high performance. It employs stream fusion optimizations for best possible performance. Serial peformance is equivalent to the venerable vector library in most cases and even better in some cases. Concurrent performance is unbeatable. See streaming-benchmarks for a comparison of popular streaming libraries on micro-benchmarks.

The basic streaming functionality of streamly is equivalent to that provided by streaming libraries like vector, streaming, pipes, and conduit. In addition to providing streaming functionality, streamly subsumes the functionality of list transformer libraries like pipes or list-t, and also the logic programming library logict. On the concurrency side, it subsumes the functionality of the async package, and provides even higher level concurrent composition. Because it supports streaming with concurrency we can write FRP applications similar in concept to Yampa or reflex.

See the Comparison with existing packages section at the end of the tutorial.

Further Reading

For more information, see:


If you require professional support, consulting, training or timely enhancements to the library please contact [email protected].


The code is available under BSD-3 license on github. Join the gitter chat channel for discussions. You can find some of the todo items on the github wiki. Please ask on the gitter channel or contact the maintainer directly for more details on each item. All contributions are welcome!

This library was originally inspired by the transient package authored by Alberto G. Corona.



Bug Fixes

  • Cleanup any pending threads when an exception occurs.
  • Fixed a livelock in ahead style streams. The problem manifests sometimes when multiple streams are merged together in ahead style and one of them is a nil stream.
  • As per expected concurrency semantics each forked concurrent task must run with the monadic state captured at the fork point. This release fixes a bug, which, in some cases caused an incorrect monadic state to be used for a concurrent action, leading to unexpected behavior when concurrent streams are used in a stateful monad e.g. StateT. Particularly, this bug cannot affect ReaderT.


  • Performance improvements, especially space consumption, for concurrent streams


Bug Fixes

  • Leftover threads are now cleaned up as soon as the consumer is garbage collected.
  • Fix a bug in concurrent function application that in certain cases would unnecessarily share the concurrency state resulting in incorrect output stream.
  • Fix passing of state across parallel, async, wAsync, ahead, serial, wSerial combinators. Without this fix combinators that rely on state passing e.g. maxThreads and maxBuffer won’t work across these combinators.


  • Added rate limiting combinators rate, avgRate, minRate, maxRate and constRate to control the yield rate of a stream.
  • Add foldl1', foldr1, intersperseM, find, lookup, and, or, findIndices, findIndex, elemIndices, elemIndex, init to Prelude


  • The Streamly.Time module is now deprecated, its functionality is subsumed by the new rate limiting combinators.


Bug Fixes

  • foldxM was not fully strict, fixed.


Breaking changes

  • Signatures of zipWithM and zipAsyncWithM have changed
  • Some functions in prelude now require an additional Monad constraint on the underlying type of the stream.


  • once has been deprecated and renamed to yieldM


  • Add concurrency control primitives maxThreads and maxBuffer.
  • Concurrency of a stream with bounded concurrency when used with take is now limited by the number of elements demanded by take.
  • Significant performance improvements utilizing stream fusion optimizations.
  • Add yield to construct a singleton stream from a pure value
  • Add repeat to generate an infinite stream by repeating a pure value
  • Add fromList and fromListM to generate streams from lists, faster than fromFoldable and fromFoldableM
  • Add map as a synonym of fmap
  • Add scanlM', the monadic version of scanl’
  • Add takeWhileM and dropWhileM
  • Add filterM


Breaking changes

  • Some prelude functions, to whom concurrency capability has been added, will now require a MonadAsync constraint.

Bug Fixes

  • Fixed a race due to which, in a rare case, we might block indefinitely on an MVar due to a lost wakeup.
  • Fixed an issue in adaptive concurrency. The issue caused us to stop creating more worker threads in some cases due to a race. This bug would not cause any functional issue but may reduce concurrency in some cases.


  • Added a concurrent lookahead stream type Ahead
  • Added fromFoldableM API that creates a stream from a container of monadic actions
  • Monadic stream generation functions consM, |:, unfoldrM, replicateM, repeatM, iterateM and fromFoldableM can now generate streams concurrently when used with concurrent stream types.
  • Monad transformation functions mapM and sequence can now map actions concurrently when used at appropriate stream types.
  • Added concurrent function application operators to run stages of a stream processing function application pipeline concurrently.
  • Added mapMaybe and mapMaybeM.


Bug Fixes

  • Fixed a bug that caused some transformation ops to return incorrect results when used with concurrent streams. The affected ops are take, filter, takeWhile, drop, dropWhile, and reverse.


Breaking changes

  • Changed the semantics of the Semigroup instance for InterleavedT, AsyncT and ParallelT. The new semantics are as follows:

    • For InterleavedT, <> operation interleaves two streams
    • For AsyncT, <> now concurrently merges two streams in a left biased manner using demand based concurrency.
    • For ParallelT, the <> operation now concurrently meges the two streams in a fairly parallel manner.

    To adapt to the new changes, replace <> with serial wherever it is used for stream types other than StreamT.

  • Remove the Alternative instance. To adapt to this change replace any usage of <|> with parallel and empty with nil.

  • Stream type now defaults to the SerialT type unless explicitly specified using a type combinator or a monomorphic type. This change reduces puzzling type errors for beginners. It includes the following two changes:

    • Change the type of all stream elimination functions to use SerialT instead of a polymorphic type. This makes sure that the stream type is always fixed at all exits.
    • Change the type combinators (e.g. parallely) to only fix the argument stream type and the output stream type remains polymorphic.

    Stream types may have to be changed or type combinators may have to be added or removed to adapt to this change.

  • Change the type of foldrM to make it consistent with foldrM in base.

  • async is renamed to mkAsync and async is now a new API with a different meaning.

  • ZipAsync is renamed to ZipAsyncM and ZipAsync is now ZipAsyncM specialized to the IO Monad.

  • Remove the MonadError instance as it was not working correctly for parallel compositions. Use MonadThrow instead for error propagation.

  • Remove Num/Fractional/Floating instances as they are not very useful. Use fmap and liftA2 instead.


  • Deprecate and rename the following symbols:
    • Streaming to IsStream
    • runStreaming to runStream
    • StreamT to SerialT
    • InterleavedT to WSerialT
    • ZipStream to ZipSerialM
    • ZipAsync to ZipAsyncM
    • interleaving to wSerially
    • zipping to zipSerially
    • zippingAsync to zipAsyncly
    • <=> to wSerial
    • <| to async
    • each to fromFoldable
    • scan to scanx
    • foldl to foldx
    • foldlM to foldxM
  • Deprecate the following symbols for future removal:
    • runStreamT
    • runInterleavedT
    • runAsyncT
    • runParallelT
    • runZipStream
    • runZipAsync


  • Add the following functions:
    • consM and |: operator to construct streams from monadic actions
    • once to create a singleton stream from a monadic action
    • repeatM to construct a stream by repeating a monadic action
    • scanl' strict left scan
    • foldl' strict left fold
    • foldlM' strict left fold with a monadic fold function
    • serial run two streams serially one after the other
    • async run two streams asynchronously
    • parallel run two streams in parallel (replaces <|>)
    • WAsyncT stream type for BFS version of AsyncT composition
  • Add simpler stream types that are specialized to the IO monad
  • Put a bound (1500) on the output buffer used for asynchronous tasks
  • Put a limit (1500) on the number of threads used for Async and WAsync types



  • Add iterate, iterateM stream operations

Bug Fixes

  • Fixed a bug that casued unexpected behavior when pure was used to inject values in Applicative composition of ZipStream and ZipAsync types.



  • Make cons right associative and provide an operator form .: for it
  • Add null, tail, reverse, replicateM, scan stream operations
  • Improve performance of some stream operations (foldl, dropWhile)

Bug Fixes

  • Fix the product operation. Earlier, it always returned 0 due to a bug
  • Fix the last operation, which returned Nothing for singleton streams


  • Initial release