@hackage streaming0.1.3.4

an elementary streaming prelude and a general stream type.

streaming

  1. The freely generated stream on a streamable functor
  2. A freely generated stream of individual Haskell values is a Producer, Generator or Source
  3. Streaming.Prelude
  4. Mother's Prelude v. Streaming.Prelude
  5. How come there's not one of those fancy "ListT done right" implementations in here?
  6. Didn't I hear that free monads are a dog from the point of view of efficiency?
  7. Interoperation with the streaming-io libraries
  8. Where can I find examples of use?
  9. Problems
  10. Implementation and benchmarking notes

Stream can be used wherever FreeT is used. The compiler's standard range of optimizations work better for operations written in terms of Stream. FreeT f m r and Stream f m r are of course extremely general, and many functor-general combinators are exported by the general module Streaming.

But the library is focused on uses of Stream f m r where f is itself in some sense "streaming". This means, very crudely, that it is possible to make strict left folds over it. Where f is complex, and has the form t m, these folds will have the form t m r -> m (a, r), polymorphic in r. In particular, it will be possible, for example, to write the trivial left fold - a drain or runEffects function - t m r -> m r or f r -> m r - polymorphically. Stream f m r preserves this property. In particular, branching and failure are excluded; the latter is always handled in the monad m.

The abstraction is inevitable, though there are many ways of writing it. Once one possesses it, though, one is already in possession of an elementary streaming library, since Stream ((,)a) m r or its equivalent is the type of a producer, generator or source. I try to argue for this more elaborately below, bringing it into connection with the standard streaming io libraries.

  1. The freely generated stream on a streamable functor

(This section is a rather abstract defense of the inevitability of the leading type we are discussing, Stream f m r ; it may be well to skip to the next section.)

As soon as you consider the idea of an effectful stream of any kind whatsoever, for example, a stream of bytes from a handle, however constituted, you will inevitably be forced to contemplate the idea of a streaming succession of just such streams. Thus, for example, however you imagine your bytes streaming from a handle, you will want to consider a succession of such streams divided on newlines.

This is closely related to the fact that, as soon as you contemplate a complex streaming phenomenon, you will want to consider a break in the stream, a function that divides the stream into parts according to some internal characteristic, and allows us to handle the parts separately, making it possible to do one thing with the first part and another with the second. Such a function will not have the form:

splitter :: S -> (S, S)

like the splitting operations we find with lists and the like, e.g.

splitAt 3 :: [a] -> ([a],[a])

Since we can assume an underlying monad m, which may be implicit (in io-streams, for example, IO is implicit in the types of InputStream and Generator), we can write the candidate type thus:

splitter :: S m -> (S m, S m)

These types use ordinary "pure" pairing, and cannot express the fundamental point that I cannot get to the 'second' stream without passing through the 'first'; the features of the 'second half' may depend causally on events in the first half. We do not repair this, but just make it worse, by complicating the type thus

splitter :: S m -> m (S m, S m)

since the effects I must pass through to get to pair, and thus the second element, are precisely the effects putatively contained in the first element in the result type. My idea was to do "one thing with the first half" and "another thing with the second half"; in this type I somehow do the effects of the first half to get the pair, and still have the first half before me, coupled with the second half. If I am not proposing to repeat the action of the first part and I have not lost information, my type must secretly be something like

splitter :: S m -> m (S Identity, S m)

or

splitAccum :: S m -> m ([z], S m)

as we see, e.g. here or, more obscurely in functions like these. (I will return to this difficulty below.)

This point makes it inevitable that a rational stream type will have a return value. It will have the form

S r

or

S m r

and the dividing functions will have the form

splitter :: S r -> S (S r)  

or

splitter :: S m r -> S m (S m r)

Now we can express what we meant by 'doing one thing with the first half and another with the second': we were thinking of applying some sort of polymorphic folds, maybe with types like

folder :: S m x -> m (a, x)

Then we would have

folder . splitter :: S m r -> m (a, S m r)

and can contemplate applying this or another folding operation to the 'second half', e.g.

liftM (fmap folder) . folder . splitter :: S m x -> m (a, m (a,x))

and can reshuffle to get a function S m x -> m ((a,a), x). This function has the form of our original folder function, since it is polymorphic in x.

That folds over streaming types should be polymorphic in their return type is written already into this simple material: we want to 'do one thing with the first half and something else - or the same thing - with the second half'. The thing we 'do with the first half' will have to be something we could do even if the second half doesn't exist, and it must preserve it if it does. In the simplest case, 'what we do with the first half' might be simply to throw it out, or drain it.

Now, to return to the first point, suppose you have the idea the unfolding of some sort of stream from an individual Haskell value, a seed - a file name, as it might be. And suppose you also have some idea of a stream of individual Haskell values - maybe a stream of file names coming from something like du, subjected to some filter. Then you will also have the idea of a streaming succession of such unfoldings linked together end to end in accordance with the initial succession of seed values.

Call the thoughts above the ABCs of streaming. If you understood these ABCs you have a total comprehension of Stream f m r:

  • Stream expresses what the word "succession" meant in the ABCs
  • The general parameter f expresses what was meant by "such streams"
  • m expresses the relevant form of "effect".

General combinators for working with this idea of succession irrespective of the form of succession are contained in the module Stream. They can be used, or example, to organize a succession of io-streams Generators or pipes Producers or the effectful bytestreams of the streaming-bytestring library, or whatever stream-form you can express in a Haskell functor.

  1. A freely generated stream of individual Haskell values is a Producer, Generator or Source

But, of course, as soon as you grasp the general form of succession, you are already in possession of the most basic concrete form: a simple succession of individual Haskell values one after another. This is just Stream ((,) a) m r. Here we prefer Stream (Of a) m r, strictifying the left element of the pair with

data Of a r = !a :> r deriving Functor

Either way, the pairing just links the present element with the rest of the stream. The primitive yield statement just expresses the pairing of the yielded item with the rest of the stream; or rather it is itself the trivial singleton stream.

yield 17  :: Stream (Of Int) IO ()

Streaming.Prelude is focused on the manipulation of this all-important stream-form, which appears in the streaming IO libraries under titles like:

io-streams: Generator a r
pipes:      Producer a m r
conduit:    ConduitM () o m r
streaming:  Stream (Of a) m r

The only difference is that in streaming the simple generator or producer concept is formulated explicitly in terms of the general concept of successive connection. But this is a concept you need and already possess anyway, as your comprehension of the streaming ABCs showed.

The special case of a stream of individual Haskell values that simply comes to an end without a special result is variously expressed thus:

io-streams: InputStream a 
pipes:      Producer a m ()
conduit:    Source m a
machines:   SourceT m a (= forall k. MachineT m k a)
streaming:  Stream (Of a) m ()
  1. Streaming.Prelude

Streaming.Prelude closely follows Pipes.Prelude. But since it restricts itself to use only of the general idea of streaming, it cleverly omits the pipes:

ghci> S.stdoutLn $ S.take 2 S.stdinLn
let's<Enter>
let's
stream<Enter>
stream

Here's a little connect and resume, as the streaming-io experts call it:

ghci> rest <- S.print $ S.splitAt 3 $ S.each [1..10]
1
2
3
ghci> S.sum rest
49

Somehow, we didn't even need a four-character operator for that, nor advice about best practices! - just ordinary Haskell common sense.

  1. Mother's Prelude v. Streaming.Prelude

The effort of Streaming.Prelude is to leverage the intuition the user has acquired in mastering Prelude and Data.List and to elevate her understanding into a general comprehension of effectful streaming transformations. Unsurprisingly, it takes longer to type out the signatures. It cannot be emphasized enough, thought, that the transpositions are totally mechanical:

Data.List.Split.chunksOf :: Int -> [a]          -> [[a]]
Streaming.chunksOf       :: Int -> Stream f m r -> Stream (Stream f m) m r

Prelude.splitAt   :: Int -> [a]          -> ([a],[a])
Streaming.splitAt :: Int -> Stream f m r -> Stream f m (Stream f m r)

These concepts are "functor general", in the jargon used in the documentation, and are thus exported by the main Streaming module. Something like break requires us to inspect individual values for their properties, so it is found in the Streaming.Prelude

Prelude.break           :: (a -> Bool) -> [a]               -> ([a],[a])
Streaming.Prelude.break :: (a -> Bool) -> Stream (Of a) m r -> Stream (Of a) m (Stream (Of a) m r)

It is easy to prove that resistance to these types is resistance to effectful streaming itself. I will labor this point a bit more below, but you can also find it developed, with greater skill, in the documentation for the pipes libraries.

  1. How come there's not one of those fancy "ListT done right" implementations in here?

The use of the final return value appears to be a complication, but in fact it is essentially contained in the idea of effectful streaming. This is why this library does not export a _ListT done right/, which would be simple enough - following pipes, as usual:

newtype ListT m a = ListT (Stream (Of a) m ())

The associated monad instance would wrap

yield :: (Monad m)            => a -> Stream (Of a) m ()
for   :: (Monad m, Functor f) => Stream (Of a) m r -> (a -> Stream f m ()) -> Stream f m r

To see the trouble, consider this signature for splitting a ListT very much done right. Here's what becomes of chunksOf. As long as we are trapped in some sort of ListT, however much rightly implemented, these operations can't be made to stream; something like a list must be accumulated. Similarly, try to imagine adding a splitAt or lines function to this API. It would accumulate strict text forever, just as this does and this doesn't and this doesn't The difference is simply that the latter libraries operate with the general concept of streaming, and the whole implementation is governed by it. The attractions of the various "ListT done right" implementations are superficial; the concept belongs to logic programming, not stream programming.

Note similarly that you can write a certain kind of take and drop with the machines library - as you can even with a "ListT done right". But I wish you luck writing splitAt! Similarly you can write a getContents; but I wish you luck dividing the resulting bytestream on its lines. This is - as usual! - because the library was not written with the general concept of effectful succession or streaming in view. Materials for sinking some elements of a stream in one way, and others in other ways - copying each line to a different file, as it might be, but without accumulation - are documented within. So are are myriad other elementary operations of streaming io.

  1. Didn't I hear that free monads are a dog from the point of view of efficiency?

We noted above that if we instantiate Stream f m r to Stream ((,) a) m r or the like, we get the standard idea of a producer or generator. If it is instantiated to Stream f Identity m r then we have the standard _free monad construction/. This construction is subject to certain familiar objections from an efficiency perspective; efforts have been made to substitute exotic cps-ed implementations and so forth. It is an interesting topic.

But in fact, the standard alarmist talk about retraversing binds and quadratic explosions and costly appends, and so on become transparent nonsense with Stream f m r
in its streaming use. The conceptual power needed to see this is basically nil: Where m is read as IO, or some transformed IO, then the dreaded retraversing of the binds in a stream expression would involve repeating all the past actions. Don't worry, to get e.g. the second chunk of bytes from a handle, you won't need to start over and get the first one again! The first chunk has vanished into an unrepeatable past.

All of the difficulties a streaming library is attempting to avoid are concentrated in the deep irrationality of

sequence :: (Monad m, Traversable t) => t (m a) -> m (t a)

In the streaming context, this becomes

sequence :: Monad m, Functor f => Stream f m r -> Stream f m r
sequence = id

It is of course easy enough to define

accumulate :: Monad m, Functor f => Stream f m r -> m (Stream f Identity r)

or reifyBindsRetraversingWherePossible or _ICan'tTakeThisStreamingAnymore, as you might call it. The types themselves teach the user how to avoid or control the sort of accumulation characteristic of sequence in its various guises e.g. mapM f = sequence . map f and traverse f = sequence . fmap f and replicateM n = sequence . replicate n. See for example the types of

Control.Monad.replicateM :: Int -> m a -> m [a]
Streaming.Prelude.replicateM :: Int -> m a -> Stream (Of a) m ()

If you want to tempt fate and replicate the irrationality of Control.Monad.replicateM, then sure, you can define the hermaphroditic chimera

accumulate . Streaming.Prelude.replicateM :: Int -> m a -> m (Stream (Of a) Identity ())

which is what we find in our diseased base libraries. But once you know how to operate with a stream directly you will see less and less point in what is called extracting the (structured) value from IO. Consider the apparently innocent distinction between

"getContents" :: String

and

getContents :: IO String 

Omitting consideration of eof, we might define getContents thus

getContents = sequence $ repeat getChar

There it is again! The very devil! By contrast there is no distinction between

"getContents" :: Stream (Of Char) m ()  -- the IsString instance is monad-general

and

getContents :: MonadIO m => Stream (Of Char) m ()

They unify just fine. That is, if I make the type synonym

type String m r = Stream (Of Char) m r

I get, for example:

"getLine"                              :: String m  ()
getLine                                :: String IO ()
"getLine" >> getLine                   :: String IO ()
splitAt 20 $ "getLine" >> getLine      :: String IO (String IO ())
length $ "getLine" >> getLine          :: IO Int

and can dispense with half the advice they will give you on #haskell. It is only a slight exaggeration to say that a stream should never be "extracted from IO".

With sequence and traverse, we accumulate a pure succession of pure values from a pure succession of monadic values. Why bother if you have intrinsically monadic conception of succession or traversal? Stream f m r gives you an immense body of such structures and a simple discipline for working with them. Spinkle id freely though your program, under various names, if you get homesick for sequence and company.

  1. Interoperation with the streaming-io libraries

The simplest form of interoperation with pipes is accomplished with this isomorphism:

Pipes.unfoldr Streaming.next        :: Stream (Of a) m r   -> Producer a m r
Streaming.unfoldr Pipes.next        :: Producer a m r      -> Stream (Of a) m r                     

Of course, streaming can be mixed with pipes wherever pipes itself employs Control.Monad.Trans.Free; speedups are frequently appreciable. (This was the original purpose of the main Streaming module, which just mechanically transposes a simple optimization employed in Pipes.Internal.) Interoperation with io-streams is thus:

Streaming.reread IOStreams.read     :: InputStream a       -> Stream (Of a) IO ()
IOStreams.unfoldM Streaming.uncons  :: Stream (Of a) IO () -> IO (InputStream a)

A simple exit to conduit would be, e.g.:

Conduit.unfoldM Streaming.uncons    :: Stream (Of a) m ()  -> Source m a

These conversions should never be more expensive than a single >-> or =$=.

At a much more general level, we also of course have interoperation with free:

Free.iterTM  Stream.wrap              :: FreeT f m a -> Stream f m a
Stream.iterTM Free.wrap               :: Stream f m a -> FreeT f m a 
  1. Where can I find examples of use?

For some simple ghci examples, see the commentary throughout the Prelude module. For slightly more advanced usage see the commentary in the haddocks of streaming-bytestring and e.g. these replicas of shell-like programs from the io-streams tutorial. Here's a simple streaming GET request with intrinsically streaming byte streams. Here is a comically simple 'high - low' game

  1. Problems

Questions about this library can be put as issues through the github site or on the pipes mailing list. (This library understands itself as part of the pipes "ecosystem.")

  1. Implementation and benchmarking notes

This library defines an optimized FreeT with an eye to use with streaming libraries, namely:

data Stream f m r
     = Return r
     | Step !(f (Stream f m r))
     | Effect (m (Stream f m r))

in place of the standard FreeT that we find in the free library, which is approximately:

newtype FreeT f m r = FreeT {runFreeT :: m (Either r (f (FreeT f m r)))}

Rather than wrapping each step in a monadic 'layer', such a layer is put alongside separate 'pure' constructors for a functor 'layer' and a final return value. The maneuver is very friendly to the compiler, but requires a bit of subtlety to protect a sound monad instance. Just such an optimization is adopted internally by the pipes library. As in pipes, the constructors are here left in an Internal module; the main Streaming module exporting the type itself and various operations and instances.

I ran a simple benchmark (adjusting a script of John Weigly) using a very simple composition of functions:

toList 
. filter (\x -> x `mod` 2 == 0) 
. map (+1) 
. drop 1000 
. map (+1) 
. filter even 
. each

as it interpreted by various libraries - streaming, conduit, (Weigley's) simple-conduit, io-streams and machines.

The the results were fairly pleasing:

benchmarking basic/stream
time                 85.45 ms   (81.63 ms .. 89.32 ms)
                     0.994 R²   (0.982 R² .. 0.999 R²)
mean                 86.53 ms   (84.16 ms .. 90.51 ms)
std dev              4.987 ms   (2.301 ms .. 7.906 ms)
variance introduced by outliers: 18% (moderately inflated)

benchmarking basic/conduit
time                 101.3 ms   (88.77 ms .. 111.3 ms)
                     0.976 R²   (0.911 R² .. 0.996 R²)
mean                 95.56 ms   (84.90 ms .. 103.6 ms)
std dev              13.76 ms   (8.210 ms .. 21.79 ms)
variance introduced by outliers: 43% (moderately inflated)

benchmarking basic/simple-conduit
time                 199.2 ms   (174.1 ms .. 215.6 ms)
                     0.993 R²   (0.978 R² .. 1.000 R²)
mean                 198.4 ms   (190.0 ms .. 202.2 ms)
std dev              7.091 ms   (1.565 ms .. 10.000 ms)
variance introduced by outliers: 14% (moderately inflated)

benchmarking basic/pipes
time                 211.7 ms   (180.8 ms .. 232.7 ms)
                     0.991 R²   (0.974 R² .. 1.000 R²)
mean                 207.7 ms   (199.1 ms .. 218.7 ms)
std dev              12.34 ms   (5.989 ms .. 17.67 ms)
variance introduced by outliers: 15% (moderately inflated)

benchmarking basic/data-list
time                 202.7 ms   (186.5 ms .. 225.5 ms)
                     0.990 R²   (0.970 R² .. 1.000 R²)
mean                 199.3 ms   (188.4 ms .. 207.4 ms)
std dev              11.67 ms   (6.966 ms .. 15.11 ms)
variance introduced by outliers: 15% (moderately inflated)

benchmarking basic/iostreams
time                 265.7 ms   (247.2 ms .. 284.8 ms)
                     0.997 R²   (0.990 R² .. 1.000 R²)
mean                 265.6 ms   (261.9 ms .. 272.8 ms)
std dev              7.094 ms   (146.8 μs .. 8.387 ms)
variance introduced by outliers: 16% (moderately inflated)

benchmarking basic/machines
time                 1.123 s    (NaN s .. 1.206 s)
                     0.999 R²   (0.999 R² .. 1.000 R²)
mean                 1.134 s    (1.114 s .. 1.145 s)
std dev              17.29 ms   (0.0 s .. 19.07 ms)
variance introduced by outliers: 19% (moderately inflated)

This sequence of pre-packaged combinators is, I think, as friendly as it could possibly be to the more recent conduit fusion framework. That framework of course doesn't apply to user-defined operations; there we should expect times like those shown for pipes. Since the combinators from streaming are defined with naive recursion, more or less as the user might, we have reason to think this result is characteristic, but much more benchmarking is needed before anything can be said with certainty.