streamly: Dataflow programming and declarative concurrency

[ array, bsd3, concurrency, dataflow, filesystem, library, list, logic, network, non-determinism, parsing, pipes, reactivity, streaming, streamly, time, unicode ] [ Propose Tags ] [ Report a vulnerability ]

Browse the documentation at https://streamly.composewell.com.

Streamly is a streaming framework to build reliable and scalable software systems from modular building blocks using dataflow programming and declarative concurrency. Stream fusion optimizations in streamly result in high-performance, modular combinatorial programming.

Performance with simplicity:

  • Performance on par with C (Benchmarks)

  • API close to standard Haskell lists (Examples)

  • Declarative concurrency with automatic scaling

  • Filesystem, fsnotify, network, and Unicode support included

  • More functionality provided via many ecosystem packages

Unified and powerful abstractions:

  • Unifies unfolds, arrays, folds, and parsers with streaming

  • Unifies Data.List, list-t, and logict with streaming

  • Unifies concurrency with standard streaming abstractions

  • Provides time-domain combinators for reactive programming

  • Interworks with bytestring and streaming libraries


[Skip to Readme]

Modules

[Index] [Quick Jump]

Flags

Manual Flags

NameDescriptionDefault
fusion-plugin

Use fusion plugin for benchmarks and executables

Disabled
inspection

Enable inspection testing

Disabled
debug

Debug build with asserts enabled

Disabled
dev

Development build

Disabled
has-llvm

Use llvm backend for code generation

Disabled
no-fusion

Disable rewrite rules for stream fusion

Disabled
streamk

Use CPS style streams when possible

Disabled
use-c-malloc

Use C malloc instead of GHC malloc

Disabled
opt

off=GHC default, on=-O2

Enabled
limit-build-mem

Limits memory when building

Disabled
streamly-core

Build only the core modules

Disabled
use-unliftio

Use unliftio-core instead of monad-control

Disabled

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

Downloads

Note: This package has metadata revisions in the cabal description newer than included in the tarball. To unpack the package including the revisions, use 'cabal get'.

Maintainer's Corner

Package maintainers

For package maintainers and hackage trustees

Candidates

Versions [RSS] 0.1.0, 0.1.1, 0.1.2, 0.2.0, 0.2.1, 0.3.0, 0.4.0, 0.4.1, 0.5.0, 0.5.1, 0.5.2, 0.6.0, 0.6.1, 0.7.0, 0.7.1, 0.7.2, 0.7.3, 0.7.3.1, 0.7.3.2, 0.8.0, 0.8.1, 0.8.1.1, 0.8.2, 0.8.3, 0.9.0, 0.10.0, 0.10.1
Change log CHANGELOG.md
Dependencies atomic-primops (>=0.8 && <0.9), base (>=4.9 && <4.18), containers (>=0.5 && <0.7), deepseq (>=1.4.1 && <1.5), directory (>=1.2.2 && <1.4), exceptions (>=0.8 && <0.11), filepath (>=1.2.0.0 && <1.5), fusion-plugin-types (>=0.1 && <0.2), ghc-prim (>=0.2 && <0.10), heaps (>=0.3 && <0.5), lockfree-queue (>=0.2.4 && <0.3), monad-control (>=1.0 && <1.1), mtl (>=2.2 && <2.3), network (>=2.6 && <3.2), primitive (>=0.5.4 && <0.9), transformers (>=0.4 && <0.7), transformers-base (>=0.4 && <0.5), unicode-data (>=0.1 && <0.4), Win32 (>=2.6 && <2.13) [details]
Tested with ghc ==8.0.2, ghc ==8.4.4, ghc ==8.6.5, ghc ==8.8.4, ghc ==8.10.7, ghc ==9.0.1, ghc ==9.2.1
License BSD-3-Clause
Copyright 2017 Composewell Technologies
Author Composewell Technologies
Maintainer streamly@composewell.com
Revised Revision 1 made by adithyaov at 2023-02-28T12:53:41Z
Category Streamly, Concurrency, Streaming, Dataflow, Pipes, Reactivity, List, Logic, Non-determinism, Parsing, Array, Time, Unicode, Filesystem, Network
Home page https://streamly.composewell.com
Bug tracker https://github.com/composewell/streamly/issues
Source repo head: git clone https://github.com/composewell/streamly
Uploaded by adithyaov at 2022-09-13T12:33:16Z
Distributions LTSHaskell:0.10.1, NixOS:0.10.1, Stackage:0.10.1
Reverse Dependencies 34 direct, 4 indirect [details]
Downloads 18334 total (181 in the last 30 days)
Rating 2.5 (votes: 6) [estimated by Bayesian average]
Your Rating
  • λ
  • λ
  • λ
Status Docs uploaded by user
Build status unknown [no reports yet]

Readme for streamly-0.8.3

[back to package description]

Streamly: Idiomatic Haskell with the Performance of C

Gitter chat Hackage

Streamly is a Haskell library that provides the building blocks to build safe, scalable, modular and high performance software. Streamly offers:

  • The type safety of Haskell.
  • The performance of C programs.
  • Powerful abstractions for structuring your code.
  • Idiomatic functional programming.
  • Declarative concurrency for the seamless use of multiprocessing hardware.

About This Document

This guide introduces programming with Streamly using a few practical examples:

The guide then looks at how Streamly achieves its performance. It concludes with a brief discussion about Streamly's design philosophy, and with suggestions for further reading.

Getting Started

Installing Streamly

If you wish to follow along with this guide, you will need to have Streamly installed.

Please see the Getting Started With The Streamly Package guide for instructions on how to install Streamly.

If you wish to run benchmarks, please be sure to build your application using the instructions in the Build Guide.

An overview of the types used in these examples

As an expository device, we have indicated the types at the intermediate stages of stream computations as comments in the examples below. The meaning of these types are:

  • A SerialT IO a is a serial stream of values of type a in the IO Monad.
  • An AsyncT IO a is a concurrent (asynchronous) stream of values of type a in the IO Monad.
  • An Unfold IO a b is a representation of a function that converts a seed value of type a into a stream of values of type b in the IO Monad.
  • A Fold IO a b is a representation of a function that converts a stream of type a to a final accumulator of type b in the IO Monad.

A Note on Module Naming

Some of the examples below use modules from the Internal Streamly package hierarchy. These are not really internal to the library. We classify Streamly modules into two categories:

  • Released modules and APIs: These modules and APIs are stable. Significant changes to these modules and APIs will cause Streamly's version number to change according to the package versioning policy.
  • Pre-release modules and APIs: These modules and APIs have not been formally released yet. They may change in the near future, and such changes will not necessarily be reflected in Streamly's package version number. As yet unreleased modules and APIs reside in the Internal namespace.

Please use a minor release upper bound to adhere to the Haskell PVP when using the pre-release (internal) modules.

The Examples

Modular Word Counting

A Fold in Streamly is a composable stream consumer. For our first example, we will use Folds to count the number of bytes, words and lines present in a file. We will then compose individual Folds together to count words, bytes and lines at the same time.

Please see the file WordCountModular.hs for the complete example program, including the imports that we have omitted here.

Count Bytes (wc -c)

We start with a code fragment that counts the number of bytes in a file:

import qualified Streamly.Data.Fold as Fold
import qualified Streamly.Internal.FileSystem.File as File
import qualified Streamly.Prelude as Stream

wcb :: String -> IO Int
wcb file =
    File.toBytes file        -- SerialT IO Word8
  & Stream.fold Fold.length  -- IO Int

Count Lines (wc -l)

The next code fragment shows how to count the number of lines in a file:

-- ASCII character 10 is a newline.
countl :: Int -> Word8 -> Int
countl n ch = if ch == 10 then n + 1 else n

-- The fold accepts a stream of `Word8` and returns a line count (`Int`).
nlines :: Monad m => Fold m Word8 Int
nlines = Fold.foldl' countl 0

wcl :: String -> IO Int
wcl file =
    File.toBytes file  -- SerialT IO Word8
  & Stream.fold nlines -- IO Int

Count Words (wc -w)

Our final code fragment counts the number of whitespace-separated words in a stream:

countw :: (Int, Bool) -> Word8 -> (Int, Bool)
countw (n, wasSpace) ch =
    if isSpace $ chr $ fromIntegral ch
    then (n, True)
    else (if wasSpace then n + 1 else n, False)

-- The fold accepts a stream of `Word8` and returns a word count (`Int`).
nwords :: Monad m => Fold m Word8 Int
nwords = fst <$> Fold.foldl' countw (0, True)

wcw :: String -> IO Int
wcw file =
    File.toBytes file   -- SerialT IO Word8
  & Stream.fold nwords  -- IO Int

Counting Bytes, Words and Lines Together

By using the Tee combinator we can compose the three folds that count bytes, lines and words individually into a single fold that counts all three at once. The applicative instance of Tee distributes its input to all the supplied folds (Fold.length, nlines, and nwords) and then combines the outputs from the folds using the supplied combiner function ((,,)).

import qualified Streamly.Internal.Data.Fold.Tee as Tee

-- The fold accepts a stream of `Word8` and returns the three counts.
countAll :: Fold IO Word8 (Int, Int, Int)
countAll = Tee.toFold $ (,,) <$> Tee Fold.length <*> Tee nlines <*> Tee nwords

wc :: String -> IO (Int, Int, Int)
wc file =
    File.toBytes file    -- SerialT IO Word8
  & Stream.fold countAll -- IO (Int, Int, Int)

This example demonstrates the excellent modularity offered by Streamly's simple and concise API. Experienced Haskellers will notice that we have not used bytestrings—we instead used a stream of Word8 values, simplifying our program.

The Performance of Word Counting

We compare two equivalent implementations: one using Streamly, and the other using C.

The performance of the Streamly word counting implementation is:

$ time WordCount-hs gutenberg-500MB.txt
11242220 97050938 574714449 gutenberg-500MB.txt

real    0m1.825s
user    0m1.697s
sys     0m0.128s

The performance of an equivalent wc implementation in C is:

$ time WordCount-c gutenberg-500MB.txt
11242220 97050938 574714449 gutenberg-500MB.txt

real    0m2.100s
user    0m1.935s
sys     0m0.165s

Concurrent Word Counting

In our next example we show how the task of counting words, lines, and bytes could be done in parallel on multiprocessor hardware.

To count words in parallel we first divide the stream into chunks (arrays), do the counting within each chunk, and then add all the counts across chunks. We use the same code as above except that we use arrays for our input data.

Please see the file WordCountParallel.hs for the complete working code for this example, including the imports that we have omitted below.

The countArray function counts the line, word, char counts in one chunk:

import qualified Streamly.Data.Array.Foreign as Array

countArray :: Array Word8 -> IO Counts
countArray arr =
      Stream.unfold Array.read arr            -- SerialT IO Word8
    & Stream.decodeLatin1                     -- SerialT IO Char
    & Stream.foldl' count (Counts 0 0 0 True) -- IO Counts

Here the function count and the Counts data type are defined in the WordCount helper module defined in WordCount.hs.

When combining the counts in two contiguous chunks, we need to check whether the first element of the next chunk is a whitespace character in order to determine if the same word continues in the next chunk or whether the chunk starts with a new word. The partialCounts function adds a Bool flag to Counts returned by countArray to indicate whether the first character in the chunk is a space.

partialCounts :: Array Word8 -> IO (Bool, Counts)
partialCounts arr = do
    let r = Array.getIndex arr 0
    case r of
        Just x -> do
            counts <- countArray arr
            return (isSpace (chr (fromIntegral x)), counts)
        Nothing -> return (False, Counts 0 0 0 True)

addCounts then adds the counts from two consecutive chunks:

addCounts :: (Bool, Counts) -> (Bool, Counts) -> (Bool, Counts)
addCounts (sp1, Counts l1 w1 c1 ws1) (sp2, Counts l2 w2 c2 ws2) =
    let wcount =
            if not ws1 && not sp2 -- No space between two chunks.
            then w1 + w2 - 1
            else w1 + w2
     in (sp1, Counts (l1 + l2) wcount (c1 + c2) ws2)

To count in parallel we now only need to divide the stream into arrays, apply our counting function to each array, and then combine the counts from each chunk.

wc :: String -> IO (Bool, Counts)
wc file = do
      Stream.unfold File.readChunks file -- AheadT IO (Array Word8)
    & Stream.mapM partialCounts          -- AheadT IO (Bool, Counts)
    & Stream.maxThreads numCapabilities  -- AheadT IO (Bool, Counts)
    & Stream.fromAhead                   -- SerialT IO (Bool, Counts)
    & Stream.foldl' addCounts (False, Counts 0 0 0 True) -- IO (Bool, Counts)

Please note that the only difference between a concurrent and a non-concurrent program lies in the use of the Stream.fromAhead combinator. If we remove the call to Stream.fromAhead, we would still have a perfectly valid and performant serial program. Notice how succinctly and idiomatically we have expressed the concurrent word counting problem.

A benchmark with 2 CPUs:

$ time WordCount-hs-parallel gutenberg-500MB.txt
11242220 97050938 574714449 gutenberg-500MB.txt

real    0m1.284s
user    0m1.952s
sys     0m0.140s

These example programs have assumed ASCII encoded input data. For UTF-8 streams, we have a concurrent wc implementation with UTF-8 decoding. This concurrent implementation performs as well as the standard wc program in serial benchmarks. In concurrent mode Streamly's implementation can utilise multiple processing cores if these are present, and can thereby run much faster than the standard binary.

Streamly provides concurrency facilities similar to OpenMP and Cilk but with a more declarative style of expression. With Streamly you can write concurrent programs with ease, with support for different types of concurrent scheduling.

A Concurrent Network Server

We now move to a slightly more complicated example: we simulate a dictionary lookup server which can serve word meanings to multiple clients concurrently. This example demonstrates the use of the concurrent mapM combinator.

Please see the file WordServer.hs for the complete code for this example, including the imports that we have omitted below.

import qualified Streamly.Data.Fold as Fold
import qualified Streamly.Network.Inet.TCP as TCP
import qualified Streamly.Network.Socket as Socket
import qualified Streamly.Unicode.Stream as Unicode

-- Simulate network/db query by adding a delay.
fetch :: String -> IO (String, String)
fetch w = threadDelay 1000000 >> return (w,w)

-- Read lines of whitespace separated list of words from a socket, fetch the
-- meanings of each word concurrently and return the meanings separated by
-- newlines, in same order as the words were received. Repeat until the
-- connection is closed.
lookupWords :: Socket -> IO ()
lookupWords sk =
      Stream.unfold Socket.read sk               -- SerialT IO Word8
    & Unicode.decodeLatin1                       -- SerialT IO Char
    & Stream.wordsBy isSpace Fold.toList         -- SerialT IO String
    & Stream.fromSerial                          -- AheadT  IO String
    & Stream.mapM fetch                          -- AheadT  IO (String, String)
    & Stream.fromAhead                           -- SerialT IO (String, String)
    & Stream.map show                            -- SerialT IO String
    & Stream.intersperse "\n"                    -- SerialT IO String
    & Unicode.encodeStrings Unicode.encodeLatin1 -- SerialT IO (Array Word8)
    & Stream.fold (Socket.writeChunks sk)        -- IO ()

serve :: Socket -> IO ()
serve sk = finally (lookupWords sk) (close sk)

-- | Run a server on port 8091. Accept and handle connections concurrently. The
-- connection handler is "serve" (i.e. lookupWords).  You can use "telnet" or
-- "nc" as a client to try it out.
main :: IO ()
main =
      Stream.unfold TCP.acceptOnPort 8091 -- SerialT IO Socket
    & Stream.fromSerial                   -- AsyncT IO ()
    & Stream.mapM serve                   -- AsyncT IO ()
    & Stream.fromAsync                    -- SerialT IO ()
    & Stream.drain                        -- IO ()

Merging Incoming Streams

In the next example, we show how to merge logs coming from multiple nodes in your network. These logs are merged at line boundaries and the merged logs are written to a file or to a network destination. This example uses the concatMapWith combinator to merge multiple streams concurrently.

Please see the file MergeServer.hs for the complete working code, including the imports that we have omitted below.

import qualified Streamly.Data.Unfold as Unfold
import qualified Streamly.Network.Socket as Socket

-- | Read a line stream from a socket.
-- Note: lines are buffered, and we could add a limit to the
-- buffering for safety.
readLines :: Socket -> SerialT IO (Array Char)
readLines sk =
    Stream.unfold Socket.read sk                 -- SerialT IO Word8
  & Unicode.decodeLatin1                         -- SerialT IO Char
  & Stream.splitWithSuffix (== '\n') Array.write -- SerialT IO String

recv :: Socket -> SerialT IO (Array Char)
recv sk = Stream.finally (liftIO $ close sk) (readLines sk)

-- | Starts a server at port 8091 listening for lines with space separated
-- words. Multiple clients can connect to the server and send streams of lines.
-- The server handles all the connections concurrently, merges the incoming
-- streams at line boundaries and writes the merged stream to a file.
server :: Handle -> IO ()
server file =
      Stream.unfold TCP.acceptOnPort 8090        -- SerialT IO Socket
    & Stream.concatMapWith Stream.parallel recv  -- SerialT IO (Array Char)
    & Stream.unfoldMany Array.read               -- SerialT IO Char
    & Unicode.encodeLatin1                       -- SerialT IO Word8
    & Stream.fold (Handle.write file)            -- IO ()

main :: IO ()
main = withFile "output.txt" AppendMode server

Listing Directories Recursively/Concurrently

Our next example lists a directory tree recursively, reading multiple directories concurrently.

This example uses the tree traversing combinator iterateMapLeftsWith. This combinator maps a stream generator on the Left values in its input stream (directory names in this case), feeding the resulting Left values back to the input, while it lets the Right values (file names in this case) pass through to the output. The Stream.ahead stream joining combinator then makes it iterate on the directories concurrently.

Please see the file ListDir.hs for the complete working code, including the imports that we have omitted below.

import Streamly.Internal.Data.Stream.IsStream (iterateMapLeftsWith)

import qualified Streamly.Prelude as Stream
import qualified Streamly.Internal.FileSystem.Dir as Dir (toEither)

-- Lists a directory as a stream of (Either Dir File).
listDir :: String -> SerialT IO (Either String String)
listDir dir =
      Dir.toEither dir               -- SerialT IO (Either String String)
    & Stream.map (bimap mkAbs mkAbs) -- SerialT IO (Either String String)

    where mkAbs x = dir ++ "/" ++ x

-- | List the current directory recursively using concurrent processing.
main :: IO ()
main = do
    hSetBuffering stdout LineBuffering
    let start = Stream.fromPure (Left ".")
    Stream.iterateMapLeftsWith Stream.ahead listDir start
        & Stream.mapM_ print

Rate Limiting

For bounded concurrent streams, a stream yield rate can be specified easily. For example, to print "tick" once every second you can simply write:

main :: IO ()
main =
      Stream.repeatM (pure "tick")  -- AsyncT IO String
    & Stream.timestamped            -- AsyncT IO (AbsTime, String)
    & Stream.avgRate 1              -- AsyncT IO (AbsTime, String)
    & Stream.fromAsync              -- SerialT IO (AbsTime, String)
    & Stream.mapM_ print            -- IO ()

Please see the file Rate.hs for the complete working code.

The concurrency of the stream is automatically controlled to match the specified rate. Streamly's rate control works precisely even at throughputs as high as millions of yields per second.

For more sophisticated rate control needs please see the Streamly reference documentation.

Reactive Programming

Streamly supports reactive (time domain) programming because of its support for declarative concurrency. Please see the Streamly.Prelude module for time-specific combinators like intervalsOf, and folds like takeInterval in Streamly.Internal.Data.Fold. Please also see the pre-release sampling combinators in the Streamly.Internal.Data.Stream.IsStream.Top module for throttle and debounce like operations.

The examples AcidRain.hs and CirclingSquare.hs demonstrate reactive programming using Streamly.

More Examples

If you would like to view more examples, please visit the Streamly Examples web page.

Further Reading

Performance

As you have seen in the word count example above, Streamly offers highly modular abstractions for building programs while also offering the performance close to an equivalent (imperative) C program.

Streamly offers excellent performance even for byte-at-a-time stream operations using efficient abstractions like Unfolds and terminating Folds. Byte-at-a-time stream operations can simplify programming because the developer does not have to deal explicitly with chunking and re-combining data.

Streamly exploits GHC's stream fusion optimizations (case-of-case and spec-constr) aggressively to achieve C-like speed, while also offering highly modular abstractions to developers.

Streamly will usually perform very well without any compiler plugins. However, we have fixed some deficiencies that we had noticed in GHC's optimizer using a compiler plugin. We hope to fold these optimizations into GHC in the future; until then we recommend that you use this plugin for applications that are performance sensitive.

Benchmarks

We measured several Haskell streaming implementations using various micro-benchmarks. Please see the streaming benchmarks page for a detailed comparison of Streamly against other streaming libraries.

Our results show that Streamly is the fastest effectful streaming implementation on almost all the measured microbenchmarks. In many cases it runs up to 100x faster, and in some cases even 1000x faster than some of the tested alternatives. In some composite operation benchmarks Streamly turns out to be significantly faster than Haskell's list implementation.

Note: If you can write a program in some other way or with some other language that runs significantly faster than what Streamly offers, please let us know and we will improve.

Notes

Streamly comes equipped with a very powerful set of abstractions to accomplish many kinds of programming tasks: it provides support for programming with streams and arrays, for reading and writing from the file system and from the network, for time domain programming (reactive programming), and for reacting to file system events using fsnotify.

Please view Streamly's documentation for more information about Streamly's features.

Concurrency

Streamly uses lock-free synchronization for achieving concurrent operation with low overheads. The number of tasks performed concurrently are determined automatically based on the rate at which a consumer is consuming the results. In other words, you do not need to manage thread pools or decide how many threads to use for a particular task. For CPU-bound tasks Streamly will try to keep the number of threads close to the number of CPUs available; for IO-bound tasks it will utilize more threads.

The parallelism available during program execution can be utilized with very little overhead even where the task size is very small, because Streamly will automatically switch between serial or batched execution of tasks on the same CPU depending on whichever is more efficient. Please see our concurrency benchmarks for more detailed performance measurements, and for a comparison with the async package.

Design Goals

Our goals for Streamly from the very beginning have been:

  1. To achieve simplicity by unifying abstractions.
  2. To offer high performance.

These goals are hard to achieve simultaneously because they are usually inversely related. We have spent many years trying to get the abstractions right without compromising performance.

Unfold is an example of an abstraction that we have created to achieve high performance when mapping streams on streams. Unfold allows stream generation to be optimized well by the compiler through stream fusion. A Fold with termination capability is another example which modularizes stream elimination operations through stream fusion. Terminating folds can perform many simple parsing tasks that do not require backtracking. In Streamly, Parsers are a natural extension to terminating Folds; Parsers add the ability to backtrack to Folds. Unification leads to simpler abstractions and lower cognitive overheads while also not compromising performance.

Credits

The following authors/libraries have influenced or inspired this library in a significant way:

Please see the credits directory for a full list of contributors, credits and licenses.

Licensing

Streamly is an open source project available under a liberal BSD-3-Clause license

Contributing to Streamly

As an open project we welcome contributions:

Getting Support

Professional support is available for Streamly: please contact support@composewell.com.

You can also join our community chat channel on Gitter.