Streamly

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 framework. 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 simplifies streaming and makes it as intuitive as plain lists. Unlike other streaming libraries, no fancy types are required. Streamly is simply a generalization of Haskell lists to monadic streaming optionally with concurrent composition. Streamly stream type can be considered as a list type parameterized by a monad. For example, SerialT IO a is an equivalent of [a] in the IO monad. A stream in an Identity monad, SerialT Identity a, is equivalent to pure lists with equal or better performance. Streams are constructed just like lists are constructed, using nil and cons instead of [] and :. Unlike lists, streams can be constructed from monadic actions, not just pure elements. Streams are processed just like lists are processed. Streamly provides all the list combinators and more, but they are monadic and work in a streaming fashion. In other words streamly just completes what lists lack, you do not need to learn anything new.

Not surprisingly, the monad instance of streamly is a list transformer, with concurrency capability.

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?

Installing and using

Please see INSTALL.md for instructions on how to use streamly with your Haskell build tool or package manager. You may want to go through it before jumping to run the examples below.

The module Streamly provides just the core stream types, type casting and concurrency control combinators. Stream construction, transformation, folding, merging, zipping combinators are found in Streamly.Prelude.

Show me an example

The following code snippet lists a directory tree recursively, reading multiple directories concurrently:

import Control.Monad.IO.Class (liftIO)
import Path.IO (listDir, getCurrentDir) -- from path-io package
import Streamly (AsyncT, adapt)
import qualified Streamly.Prelude as S

listDirRecursive :: AsyncT IO ()
listDirRecursive = getCurrentDir >>= readdir >>= liftIO . mapM_ putStrLn
  where
    readdir dir = do
      (dirs, files) <- listDir dir
      S.yield (map show dirs ++ map show files) <> foldMap readdir dirs

main :: IO ()
main = S.drain $ adapt $ listDirRecursive

AsyncT is a stream monad transformer. If you are familiar with a list transformer, it is nothing but ListT with concurrency semantics. For example, the semigroup operation <> is concurrent. This makes foldMap concurrent too. You can replace AsyncT with SerialT and the above code will become serial, exactly equivalent to a ListT.

Comparative Performance

High performance and simplicity are the two primary goals of streamly. Streamly employs two different stream representations (CPS and direct style) and interconverts between the two to get the best of both worlds on different operations. It uses both foldr/build (for CPS style) and stream fusion (for direct style) techniques to fuse operations. Streamly redefines "blazing fast" for streaming libraries, it competes with lists and vector. Other streaming libraries like "streaming", "pipes" and "conduit" are orders of magnitude slower on most microbenchmarks. See streaming benchmarks for detailed comparison.

The following chart shows a comparison of those streamly and list operations where performance of the two differs by more than 10%. Positive y-axis displays how many times worse is a list operation compared to the same streamly operation, negative y-axis shows where streamly is worse compared to lists.

Streamly uses lock-free synchronization for concurrent operations. It employs auto-scaling of the degree of concurrency based on demand. For CPU bound tasks it tries to keep the threads close to the number of CPUs available whereas for IO bound tasks more threads can be utilized. Parallelism can be utilized with little overhead even if the task size is very small. See concurrency benchmarks for detailed performance results and a comparison with the async package.

File IO

The following code snippet implements some common Unix command line utilities using streamly. To get an idea about IO streaming performance, you can benchmark these against the regular unix utilities using the time command. Make sure to use a big enough input file and compile with ghc -O2 -fspec-constr-recursive=10 when benchmarking. Use +RTS -s flags on the executable to check the space usage, look for maximum residency in the output.

Note that grep -c counts the number of lines where the pattern occurs whereas the snippet below counts the total number of occurrences of the pattern, therefore, the output may differ.

import qualified Streamly.Prelude as S
import qualified Streamly.Fold as FL
import qualified Streamly.Mem.Array as A
import qualified Streamly.FileSystem.File as File

import Data.Char (ord)
import System.Environment (getArgs)
import System.IO (openFile, IOMode(..), stdout)

cat src = File.writeArrays stdout $ File.readArraysUpto (256*1024) src
cp src dst = File.writeArrays dst $ File.readArraysUpto (256*1024) src
wcl src = print =<<
    ( S.length
    $ FL.splitSuffixBy (== fromIntegral (ord '\n')) FL.drain
    $ File.read src)
grepc pat src = print . (subtract 1) =<<
    ( S.length
    $ FL.splitOn (A.fromList (map (fromIntegral . ord) pat)) FL.drain
    $ File.read src)
avgll src = print =<<
    ( FL.foldl' avg
    $ FL.splitSuffixBy (== fromIntegral (ord '\n')) FL.length
    $ File.read src)
    where avg = (/) <$> toDouble FL.sum <*> toDouble FL.length
          toDouble = fmap (fromIntegral :: Int -> Double)
llhisto src = print =<< 
    ( FL.foldl' (FL.classify FL.length)
    $ S.map bucket
    $ FL.splitSuffixBy (== fromIntegral (ord '\n')) FL.length
    $ File.read src)
    where
    bucket n = let i = n `div` 10 in if i > 9 then (9,n) else (i,n)

main = do
    name <- fmap head getArgs
    src <- openFile name ReadMode
    -- cat src          -- Unix cat program
    -- wcl src          -- Unix wc -l program
    -- grepc "aaaa" src -- Unix grep -c program

    -- dst <- openFile "dst.txt" WriteMode
    -- cp src dst       -- Unix cp program

    -- avgll src        -- get average line length
    llhisto src      -- get line length histogram

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 = S.drain $
       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.

Concurrent Stream Generation

consM or its operator form |: can be used to construct a stream from monadic actions. A stream constructed with consM can run the monadic actions in the stream 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), note the order of elements in the resulting output, the outputs are consumed as soon as each action is finished (asyncly):

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

Use aheadly if you want speculative concurrency i.e. execute the actions in the stream concurrently but consume the results in the specified order:

> S.toList $ aheadly $ p 3 |: p 2 |: p 1 |: S.nil
[3,2,1]

Monadic stream generation functions e.g. unfoldrM, replicateM, repeatM, iterateM and fromFoldableM etc. can work concurrently.

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

S.drain $ asyncly $ S.replicateM 10 $ p 10

Concurrency Auto Scaling

Concurrency is auto-scaled i.e. more actions are executed concurrently if the consumer is consuming the stream at a higher speed. How many tasks are executed concurrently can be controlled by maxThreads and how many results are buffered ahead of consumption can be controlled by maxBuffer. See the documentation in the Streamly module.

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 = S.drain $
      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
> S.drain $ 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)

Serial

main = S.drain $ delay 3 <> delay 2 <> delay 1

ThreadId 36: Delay 3
ThreadId 36: Delay 2
ThreadId 36: Delay 1

Parallel

main = S.drain . 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 = S.drain loops

(1,3)
(1,4)
(2,3)
(2,4)

Concurrent Nested Loops

To run the above code with speculative concurrency i.e. each iteration in the loop can run concurrently but the results are presented to the consumer of the output in the same order as serial execution:

main = S.drain $ aheadly $ loops

Different stream types execute the loop iterations in different ways. For example, wSerially interleaves the loop iterations. There are several concurrent stream styles to execute the loop iterations concurrently in different ways, see the Streamly.Tutorial module for a detailed treatment.

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 = S.drain $ 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.

Exceptions

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.

Conclusion

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:

• Detailed tutorial
• Reference documentation
• Examples
• Guides
• Streaming benchmarks
• Concurrency benchmarks

Support

Please feel free to ask questions on the streamly gitter channel. If you require professional support, consulting, training or timely enhancements to the library please contact support@composewell.com.

Contributing

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.