Auto as Category, Applicative & Arrow (Intro to Machines/Arrows Part 2)

SourceMarkdownLaTeXPosted in Haskell, RamblingsComments


Welcome back! It’s been a while since the last post, admittedly; sorry! In this post we’ll be continuing on from the previous post. In particular, we’re going to be looking at the Auto type as something that is a part of a pretty powerful pattern of abstraction, and try to exploit it to write concise, expressive code using Auto composition and proc notation. We’ll also see first hands the principles of locally stateful composition, and how much more expressive and safe it makes our code.

One motivating factor that I will eventually write about is that we can use this to implement the semantics of Functional Reactive Programming, yay! But through this, I hope you can actually see that it is useful for much, much more!

As always, feel free to leave a comment if you have any questions, or try to find me on twitter, or drop by the #haskell Freenode IRC channel! (I go by jle`)

Note that all of the code in this post can be downloaded (from Auto.hs for the last post, and Auto2.hs for this post’s new material) so you can play along on GHCi, or write your own code using it the concepts and types here :) You can also run it online interactively.

A fair warning: at times this post might feel a bit fragmented; but remember that we really are just going to be exploring and getting very familiar with the Auto type and building an intuition. Everything comes to a mini-climax at the end, and a final satisfying one at the next post — kind of like every Part 2 in every trilogy ever, you know? :)

Recap

We left off in our last post having looked at Auto:

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/machines/Auto.hs#L12-L12
-- interactive: https://www.fpcomplete.com/user/jle/machines

newtype Auto a b = ACons { runAuto :: a -> (b, Auto a b) }

which we saw as a stream that had an influencing input of type a, an internal, opaque state (a function of the input and of the previous state), and an output “head” of type b (also a function of the input and of the previous state).

And we looked at a simple auto which acted like a constantly incrementing stream, but where you could reset the counter by passing in a Just.

Then we took another approach to looking at this — we thought about Autos as functions “with state”. As in, Auto a b was like a function a -> b, but which had an internal state that updated every time it was called.

We saw this in an auto that returns the sum of everything you have given it.

Autos are “function-like things”…they map or “morph” things of type a to things of type b in some form, just like functions. It looks like we stumbled onto some sort of design pattern. Wouldn’t it be cool if we could treat Autos the same way we treat functions? And reason about them the same way, think about them using the same logic?

What is the essence of function-like-ness?

The Essence of Function-like-ness

I’m going to introduce some formality and call things “function-like-things” morphisms (with some laws). Sometimes you’ll see them called “arrows”, but this is a slightly loaded word as there is an Arrow typeclass in Haskell that doesn’t exactly refer to the same thing.

Here is my claim: the “essence” of this functionlikeness is their ability to compose and “chain”, and the nature of that composition process.

That is, if you have a morphism f from a to b, and a morphism g from b to c, then you can “compose” or “chain” them to get a morphism from a to c.

In Haskell we use the (.) operator for this — to say more formally:

f     :: morphism a b
g     :: morphism b c
g . f :: morphism a c

Some important aspects of the nature of this composition is that it must “associate”. That means:

(h . g) . f == h . (g . f)

Composing the composition of h and g to h should be the same as composing h with the composition of g and f.

The final feature is that there must exist some “identity” morphism that leaves other morphisms unchanged under composition:

id :: Morphism b b

id . f  == f
g  . id == g

It doesn’t really matter what id literally “does” — it only matters that it leaves morphisms unchanged.

And…that’s it!

Functions are “morphisms”

We’re just going to take a quick detour verify that normal functions satisfy this new notion of “function-likeness”…so that we aren’t crazy.

In Haskell, our functions are things of type a -> b — a morphism from a to b.

Our composition operator is the familiar (.) from Prelude. You can prove all of the above laws yourself using the definition of (.):

(.) :: (b -> c) -> (a -> b) -> (a -> c)
g . f = \x -> g (f x)

In practice, you can see associativity

ghci> ((< 20) . (+4)) . (^2) $ 4
True
ghci> (< 20) . ((+4) . (^2)) $ 4
True

The identity is just Prelude’s id:

id :: a -> a
id x = x
ghci>      (*3) $ 7
21
ghci> id . (*3) $ 7
21
ghci> (*3) . id $ 7
21

Aside

I mean it, you can prove it yourself if you are bored some time :) I’ll start you off with one of the identity laws:

g . id = \x -> g (id x)     -- definition of (.)
       = \x -> g x          -- definition of id
       = g                  -- eta reduction

So cool…this intuition applies to our actual idea of functions, so we are on a sane track!

Autos are “morphisms”!

So we see that functions fit this idea. Let’s jump back to what we were trying to show in the first place — that Autos fit this “idea” of function-like things, or morphisms.

Let’s say I had an f :: Auto a b and a g :: Auto b c. I want to “compose” them — g . f. To get an Auto a c, somehow. What would that even mean?

Well…if we think of f like a stateful function that takes in an a and pops out a b…and g like a stateful function that takes in a b and pops out a c…We can think of g . f as a stateful function that takes in an a, feeds it to f, gets the b that f pops out, feeds that to g, and gets the final c out at the end.

Also, Autos spit out both the result (the c) and the “updated Auto”…so the updated Auto of the composition can just be the composition of the updated Autos!

Enough talk, let’s code! We’ll call our composition operator (~.~).

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/machines/Auto2.hs#L67-L70
-- interactive: https://www.fpcomplete.com/user/jle/machines

(~.~) :: Auto b c -> Auto a b -> Auto a c
g ~.~ f = ACons $ \x -> let (y, f') = runAuto f x
                            (z, g') = runAuto g y
                        in  (z, g' ~.~ f')

And…that should be it! We run the input through first f then g, collecting the “modified f and g”, and returning both of those at the end, composed.

Let’s write a useful helper function so that we have more things to test this out on:

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/machines/Auto2.hs#L74-L75
-- interactive: https://www.fpcomplete.com/user/jle/machines

toAuto :: (a -> b) -> Auto a b
toAuto f = ACons $ \x -> (f x, toAuto f)

toAuto basically turns a function a -> b into a stateless Auto a b.

Time to test these out!

ghci> let doubleA  = toAuto (*2)      :: Auto Int Int
ghci> let succA    = toAuto (+1)      :: Auto Int Int

ghci> testAuto_ doubleA [1..10]
[2,4,6,8,10,12,14,16,18,20]

ghci> testAuto_ (succA ~.~ doubleA) [1..10]
[3,5,7,9,11,13,15,17,19,21]

ghci> testAuto_ summer [5,1,9,2,-3,4]
[5,6,15,17,14,18]

ghci> testAuto_ (doubleA ~.~ summer) [5,1,9,2,-3,4]
[10,12,30,34,28,39]

ghci> testAuto_ settableAuto [Nothing,Nothing,Just (-3),Nothing,Nothing]
[1,2,-3,-2,-1]

ghci> testAuto_ summer [1,2,-3,-2,-1]
[1,3,0,-2,-3]

ghci> testAuto_ (summer ~.~ settableAuto)
    |     [Nothing,Nothing,Just (-3),Nothing,Nothing]
[1,3,0,-2,-3]

And it looks like our Autos really can meaningfully compose!

Well, wait. We need one last thing: the identity Auto:

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/machines/Auto2.hs#L78-L79
-- interactive: https://www.fpcomplete.com/user/jle/machines

idA :: Auto a a
idA = ACons $ \x -> (x, idA)
ghci> testAuto_ summer [5,1,9,2,-3,4]
[10,12,30,34,28,39]

ghci> testAuto_ (idA ~.~ summer) [5,1,9,2,-3,4]
[10,12,30,34,28,39]

ghci> testAuto_ (summer ~.~ idA) [5,1,9,2,-3,4]
[10,12,30,34,28,39]

Category

These concepts are actually formalized in the mathematical concept of a “category” — things with objects and morphisms between them, following certain properties (like the ones I mentioned earlier).

In Haskell, we often consider our objects as Haskell types; our usual morphisms is the function arrow, (->)1 — but in this case, it might be interesting to consider a different category — the category of Haskell types and morphisms Auto a b.

In Haskell, we have a typeclass that allows us to do generic operations on all Categories — so now we can basically treat Autos “as if” they were (->). We can literally “abstract over” the idea of a function. Neat, huh?

class Category r where
    id  :: r a a
    (.) :: r b c -> r a b -> r a c

Basically, with Category, we can “abstract over” function composition and id. That is, instead of (.) being only for composing normal functions…we can use to compose morphisms in any Category, like Auto! We can also write “generic code” that works on all morphisms — not just (->)! This is like having functions mapM and sequence — which work for all Monads, not just IO or Maybe or something. We can reason about Monads as things on their own, instead of just as isolated instances.

Just be sure to use the correct imports so you don’t have name clashes with the Prelude operators:

import Control.Category
import Prelude hiding (id, (.))

First, let’s write the (->) Category instance:

instance Category (->) where
    id    = \x -> x
    g . f = \x -> g (f x)

And then our Auto Category instance:

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/machines/Auto2.hs#L13-L18
-- interactive: https://www.fpcomplete.com/user/jle/machines

instance Category Auto where
    id    = ACons $ \x -> (x, id)
    g . f = ACons $ \x ->
              let (y, f') = runAuto f x
                  (z, g') = runAuto g y
              in  (z, g' . f')

And now… we can work with both (->) and Auto as if they were the “same thing” :)

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/machines/Auto2.hs#L92-L93
-- interactive: https://www.fpcomplete.com/user/jle/machines

doTwice :: Category r => r a a -> r a a
doTwice f = f . f
ghci> doTwice (*2) 5
20
ghci> testAuto_ (doTwice (toAuto (*2))) [5]
[20]
ghci> testAuto_ (doTwice summer) [5,1,9,2,-3,4]
[5,11,26,43,57,61]
ghci> take 6 $ testAuto_ (doTwice summer) (repeat 1)
[1,3,6,10,15,21]

The main cool thing here is that we can now abstract over the “idea” of id and (.), and now our Autos have basically not only captured the idea of function-ness, but can now literally act like normal functions in our code. I mentioned something similar in an earlier post in MonadPlus — the ability to have a “common language” to talk and abstract over many things is powerful not only for expressiveness but for reasoning and maintainability.

More Typeclasses!

Anyways, we love typeclasses so much. Let’s get more familiar with our Auto type and see what other useful typeclasses it can be :) Not only are these good exercises for understanding our type, we’ll also be using these instances later!

Functor

Functor is always a fun typeclass! One use case of Functor is as a “producer transformer” — the f a is some “producer of a”. IO a is a computation that produces an a; Reader r a produces an a when given an r.

So if you have f a, we have a handy function fmap:

fmap :: Functor f => (a -> b) -> f a -> f b

Which says, “If you have an a -> b, I can turn a producer of a’s into a producer of b’s.”

There are some laws associated with fmap — most importantly that fmap id = id.

Can we turn Auto into a Functor?

…well, no, we can’t. Because fmap :: (a -> b) -> Auto a -> Auto b makes no sense…Auto takes two type parameters, not one.

But we can think of Auto r a as a “producer of a”s, when used with runAuto. Our Functor is Auto r:

fmap :: (a -> b) -> Auto r a -> Auto r b

Which says, “Give me any a -> b, and I’ll take an Auto that takes an r and returns an a…and give you an Auto that takes an r and returns a b”.

How can I do that?

Well…for one…I can just “run” the Auto you give me to get the a…and then apply the function to that a to get the b!

For example, if I fmapped show onto summer — if summer was going to output a 1, it will now output a "1". It turns an Auto Int Int into an Auto Int String!

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/machines/Auto2.hs#L20-L23
-- interactive: https://www.fpcomplete.com/user/jle/machines

instance Functor (Auto r) where
    fmap f a = ACons $ \x ->
                 let (y, a') = runAuto a x
                 in  (f y, fmap f a')
ghci> testAuto_ (fmap show summer) [5,1,9,2,-3,4]
["5","6","15","17","14","18"]

Functor, check!

What’s next?

Aside

If you ever have time, try doing some research on the Contravariant Functors and Profunctors. Can you make Auto or Auto r either one of those? Which ones? If not all of them, why not?

Applicative

Everyone knows that the “cool”, hip typeclasses are the classic trio, Functor, Applicative, Monad. Let’s just move on right along and go for Applicative.

If we continue the same sort of pattern that we did with Functor (some Functors being producers-kinda), Applicative gives you two things: the ability to “create a new ‘producer’” producing a given value, and the ability to take something that produces a function and something that produces a value and squash it into something that produces the application of the two.

This stuff…is really better said in types.

class Applicative f where
    pure  :: a -> f a
    (<*>) :: f (a -> b) -> f a -> f b

In pure, give me an a and I’ll give you a “producer” of that very a. In (<*>), give me a producer of a -> b and a producer of a and I’ll give you a producer of b.

We can pretty much use this to write our Applicative instance for Auto r.

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/machines/Auto2.hs#L25-L30
-- interactive: https://www.fpcomplete.com/user/jle/machines

instance Applicative (Auto r) where
    pure y    = ACons $ \_ -> (y, pure y)
    af <*> ay = ACons $ \x ->
                  let (f, af') = runAuto af x
                      (y, ay') = runAuto ay x
                  in  (f y, af' <*> ay')

Note that pure gives us a “constant Auto” — an Auto that ignores its input and always just produces the same thing.

The useful thing about Applicative is that it gives us liftA2, which allows us to apply a function “over Applicative”s.

liftA2 :: (a -> b -> c) -> Auto r a -> Auto r b -> Auto r c

That is, it “feeds in” the r input to both the Auto r a and the Auto r b, applies the function to both of the results, and the returns the result.

ghci> testAuto_ summer [5,1,9,2,-3,4]
[5,6,15,17,14,18]

ghci> let addSumDoub = liftA2 (+) doubleA summer
ghci> testAuto_ addSumDoub [5,1,9,2,-3,4]
[15,8,33,21,8,26]
-- [5 + 10, 6 + 2, 15 + 18, 17 + 4, 14 - 6, 18 + 8]

Hopefully by now you’ve seen enough usage of the Auto type and writing Auto combinators that do useful things that you are now Auto experts :) Feel free to press pause here, because we’re going to ramp up to some more unfamiliar abstractions. If you don’t understand some of the examples above, feel free to tinker with them on your own until you are comfortable. And as always, if you have any questions, feel free to leave a comment or drop by the freenode #haskell channel.

Okay, now on to…

Monad

Sykes!! We’re not going to make a Monad instance :) Even though it is possible, a Monad instance for Auto is remarkably useless. We won’t be using a monadic interface when working with Auto, so forget about it! What are Monads, anyway?

Take that, tomtomtom7! :D

Arrow

Okay. As it turns out, Category by itself is nice, but for many of the things we will be playing with function composition, it’s just not going to cut it.

As it turns out, we can actually sort of take advantage of a “do notation-like” syntactical sugar construct to perform complex compositions. But in order to do that, we first need to be able to “side chain” compositions. That is, split off values, perform different compositions on both forks, and recombine them. We require sort of basic set of combinators on top of our Category instance.

The Arrow typeclass was invented for just this — a grab-bag of combinators that allow such side-chaining, forking, merging behavior.

class Category r => Arrow r where
    arr    :: (a -> b) -> r a b
    first  :: r a b -> r (a, c) (b, c)
    second :: r a b -> r (c, a) (c, b)
    (***)  :: r a b -> r c d -> r (a, c) (b, d)
    (&&&)  :: r a b -> r a c -> r a (b, c)

In our case, arr turns any a -> b function into an Auto a b. first turns an Auto a b into an Auto (a, c) (b, c) — an Auto that operates on single values to an Auto that operates only on the first part of a tuple.

(***) chains Autos side-by-side: Auto a b -> Auto c d -> Auto (a, c) (b, d). It basically has each Auto operate on the tuple “in parallel”.

(&&&) “forks”. Give an Auto a b and an Auto a c, and it’ll create a “forking” Auto a (b, c).

Writing the instance is straightforward enough:

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/machines/Auto2.hs#L32-L47
-- interactive: https://www.fpcomplete.com/user/jle/machines

instance Arrow Auto where
    arr f     = ACons $ \x -> (f x, arr f)
    first a   = ACons $ \(x, z) ->
                  let (y, a') = runAuto a x
                  in  ((y, z), first a')
    second a  = ACons $ \(z, x) ->
                  let (y, a') = runAuto a x
                  in  ((z, y), second a')
    a1 *** a2 = ACons $ \(x1, x2) ->
                  let (y1, a1') = runAuto a1 x1
                      (y2, a2') = runAuto a2 x2
                  in  ((y1, y2), a1' *** a2')
    a1 &&& a2 = ACons $ \x ->
                  let (y1, a1') = runAuto a1 x
                      (y2, a2') = runAuto a2 x
                  in  ((y1, y2), a1' &&& a2')

Aside

We can also just take a shortcut and implement these in terms of combinators we have already written from different typeclasses:

instance Arrow Auto where
    arr f     = fmap f id
    first a   = liftA2 (,) (a  . arr fst) (arr snd)
    second a  = liftA2 (,) (arr fst)      (a  . arr snd)
    a1 *** a2 = liftA2 (,) (a1 . arr fst) (a2 . arr snd)
    a1 &&& a2 = (a1 *** a2) . arr (\x -> (x, x))

Remember, id is the identity Auto… and fmap f applies f “after” the identity. So this makes sense.

first is a little trickier; we are using liftA2 (,) on two Autos, kind of like we used before. liftA2 says “run these two Autos in parallel on the same input, and then at the end, (,)-up their results.”

The first of those two autos is a . arr fst — get the first thing in the tuple, and then chain the a auto onto it. The second of those two autos just simply extracts out the second part of the tuple.

a           :: Auto a b
arr fst     :: Auto (a, c) a

a . arr fst :: Auto (a, c) b
arr snd     :: Auto (a, c) c

liftA2 (,) (a . arr fst) (arr snd) :: Auto (a, c) (b, c)

What does this show? Well, that Arrow really isn’t anything too special…it’s really just what we already had — a Category with Applicative. But we are able to define more efficient instances, and also sort of look at the problem in a “different way”.

What we have here isn’t really anything too mystical. It’s just some basic combinators. And like the aside says, we didn’t introduce any “new power” — anything we could “do” with Auto using Arrow, we could already do with Category and Applicative.

The main point is just that we have these neat combinators to chain things in more useful and expressive ways — something very important when we eventually go into AFRP.

ghci> let sumDoub = summer &&& doubleA
ghci> testAuto_ sumDoub [5,1,9,2,-3,4]
[(5, 10), (6, 2), (15, 18), (17, 4), (14, -6), (18, 8)]

As we’ll see, the real benefit of Arrow will be in the syntactical sugar it provides, analogous to Monad’s do-blocks.

ArrowChoice

Another useful set of combinators is the ArrowChoice typeclass, which provides:

left  :: Auto a b -> Auto (Either a c) (Either b c)
right :: Auto a b -> Auto (Either c a) (Either c b)
(+++) :: Auto a b -> Auto c d -> Auto (Either a c) (Either b d)
(|||) :: Auto a c -> Auto b c -> Auto (Either a b) c

If you look really closely…left is kinda like first; right is kinda like second(+++) is kinda like (***), and (|||) is like a backwards (&&&).

If Arrow allows computations to be “side-chained”, ArrowChoice allows computations to be “skipped/ignored”.

We’ll instance left, which applies the given Auto on every Left input, and passes any Right input along unchanged; the Auto isn’t stepped or anything. The rest of the methods can be implemented in terms of left and arr.

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/machines/Auto2.hs#L49-L56
-- interactive: https://www.fpcomplete.com/user/jle/machines

instance ArrowChoice Auto where
    left a = ACons $ \x ->
                 case x of
                   Left l  ->
                     let (l', a') = runAuto a l
                     in  (Left l', left a')
                   Right r ->
                     (Right r, left a)

We’ll see ArrowChoice used in the upcoming syntactic sugar construct, enabling for if/then/else’s and case statements. Don’t worry about it for now if you don’t understand it.

Proc Notation

So finally, here is the real reason Arrow is useful. It’s actually a pretty well-kept secret, but…just like Monad enables do notation syntactical sugar, Arrow enables proc notation syntactical sugar. Which is probably cooler.

Not gonna lie.

A lot of AFRP and a lot of what we’re going to be doing will pretty much rely on proc notation to be able to express complex compositions…rather elegantly.

Proc notation consists of lines of “arrows”:

arrow -< x

which says “feed x through the Arrow arrow”.

Like in monadic do-blocks, you can also “bind” the result, to be used later in the block:

y <- arrow -< x

Which says “feed x through the Arrow arrow, and name the result y”.

Hey! It looks like a little ASCII arrow! Cute, huh?

The last line of a proc do block is the “return”/result of the block, like in monadic do-blocks.

Let’s write our first proc block; one that emulates our liftA2 (+) doubleA summer:

doubSummer :: Auto Int Int
doubSummer = proc x -> do
    summed  <- summer  -< x
    doubled <- doubleA -< x
    id -< summed + doubled

In the last line, we want to “return” summed + double; we have to put an Arrow command there, so we can just feed summed + double through id, to have it pop out at the end.

You can think of id like return in normal do notation.

Simple useful example

How about an Auto (Either Int Int) (Int, Int), which maintains two internal counters. You increment the first one with an input of Left x, and you increment the second one with an input of Right x. The output is the state of both counters.

We could write this “from scratch”, using explicit recursion:

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/machines/Auto2.hs#L102-L109
-- interactive: https://www.fpcomplete.com/user/jle/machines

dualCounterR :: Auto (Either Int Int) (Int, Int)
dualCounterR = dualCounterWith (0, 0)
  where
    dualCounterWith (x, y) = ACons $ \inp ->
                               let newC = case inp of
                                            Left i  -> (x + i, y)
                                            Right i -> (x, y + 1)
                               in  (newC, dualCounterWith newC)

But we all know in Haskell that explicit recursion is usually a sign of bad design and is best avoided whenever possible. So many potential places for bugs!

Let’s try writing the same thing using Auto composition:

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/machines/Auto2.hs#L112-L116
-- interactive: https://www.fpcomplete.com/user/jle/machines

dualCounterC :: Auto (Either Int Int) (Int, Int)
dualCounterC = (summer *** summer) . arr wrap
  where
    wrap (Left i)  = (i, 0)
    wrap (Right i) = (0, i)

That’s a bit more succinct, but I think the proc notation is much nicer!

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/machines/Auto2.hs#L119-L127
-- interactive: https://www.fpcomplete.com/user/jle/machines

dualCounterP :: Auto (Either Int Int) (Int, Int)
dualCounterP = proc inp -> do
    let (add1, add2) = case inp of Left i  -> (i, 0)
                                   Right i -> (0, i)

    sum1 <- summer -< add1
    sum2 <- summer -< add2

    id -< (sum1, sum2)

It’s a bit more verbose…but I think it’s much clearer what’s going on, right?

ghci> testAuto_ dualCounterP [Right 1, Left 2, Right (-4), Left 10, Right 3]
[(0, 1), (2, 1), (2, -3), (12, -3), (12, 0)]

Proc shines

And let’s say we wanted another constraint. Let’s say that…for the Left case, every other time it’s a Left, ignore the value and don’t add anything. That is, every second, fourth, sixth Left i input should ignore the i and not add anything.

How would we do this in the explicit recursive case? Why — well, adding another component to the “explicit state” tuple, and dealing with that when necessary.

I don’t even know how to begin writing it in a readable way using arrow composition.

But the proc notation? Piece of cake!

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/machines/Auto2.hs#L148-L160
-- interactive: https://www.fpcomplete.com/user/jle/machines

dualCounterSkipP :: Auto (Either Int Int) (Int, Int)
dualCounterSkipP = proc inp -> do
    (add1, add2) <- case inp of
                      Left i -> do
                        count <- summer -< 1
                        id -< (if odd count then i else 0, 0)
                      Right i ->
                        id -< (0, i)

    sum1 <- summer -< add1
    sum2 <- summer -< add2

    id -< (sum1, sum2)
ghci> testAuto_ dualCounterP [Right 1, Left 2, Right (-4), Left 10, Right 3]
[(0, 1), (2, 1), (2, -3), (2, -3), (2, 0)]

And that’s something to write home about :)

Locally Stateful Composition

The last example highlights something very significant about Autos and their Arrow-based composition: Autos with composition allow you to make locally stateful compositions.

What if we had done the above using some sort of state monad, or doing the explicit recursion?

We’d have carried the “entire” state in the parameter:

-- source: https://github.com/mstksg/inCode/tree/master/code-samples/machines/Auto2.hs#L136-L145
-- interactive: https://www.fpcomplete.com/user/jle/machines

dualCounterSkipR :: Auto (Either Int Int) (Int, Int)
dualCounterSkipR = counterFrom ((0, 0), 1)
  where
    counterFrom ((x, y), s) =
      ACons $ \inp ->
        let newCS = case inp of
                      Left i  | odd s     -> ((x + i, y), s + 1)
                              | otherwise -> ((x    , y), s + 1)
                      Right i             -> ((x, y + i), s    )
        in  (fst newCS, counterFrom newCS)

Not only is it a real mess and pain — and somewhere where bugs are rife to pop up — note the entire state is contained in one thing. That means everything has access to it; all access is sort of haphazard and ad-hoc, as well. Note that the Right case can do whatever it wants with the s. It has access to it, can read it, act on it, modify it…anything it wants! We can’t really “enforce” that the Right case can’t touch the s, without putting in more complicated work/overhead.

In the proc example…the s is a summer that is “locked inside” the Left branch. Right branch stuff can’t touch it.

In fact, all of the summers keep their own state, independently from each other. Nothing can really modify their state except for themselves, if they chose to. Less room for bugs, too, in adding, because you already know that summer works.

This property — that every single component maintains its own internal state — is, in my opinion, one of the most significant aspects of this whole game with Auto and Category and Arrow etc. Every component minds its own state.

And also — if we wanted to “add a new component” to the state, like we did, we don’t have to really change anything besides just plopping it on. In the explicit recursion example, we needed to go in and change the state type to “make room” for the new state. We needed to pretty much refactor the entire thing!

This really demonstrates the core principles of what composability and modularity even really mean.

A Quick Gotcha

Remember that with proc notation, you are really just composing and building up a giant Auto. Each individual Auto that you compose has to already be known at “composition time”. (That is, before you ever “run” it, the structure of the Auto is known and fixed).

This means that you can’t use bindings from proc blocks to form the Autos that you are composing:

foo = proc x -> do
    y <- auto1 -< x
    auto2 y -< y

This won’t work. That’s because this is really supposed to be a composition of auto1 and auto2 y. But what is auto2 y? y doesn’t even exist when you are making the compositions! y is just a name we gave to the output of auto1, in the process of our stepping it. y doesn’t exist until we “step” foo…so can’t use auto2 y in the process of composing foo.

To see more clearly, see what we’d do if we tried to write foo as a compositino:2

foo = auto2 y . auto1

Where does the y come from?!

Hopefully from this it is clear to see that it doesn’t make sense to use what you bind/name in proc notation to actually “create” the Arrow you are using.

Remember, proc notation is result <- arrow -< input. The arrow part has to already be known before everything even starts, so you can’t use things you bind to determine it :)

Moving on

Welp, hopefully by now you are experts at working with the Auto machine, and understanding it as “function-like things”. You’ve gotten deep and intimate by instancing some common typeclasses.

Then you saw Arrow, and understood how Auto fits into the Arrow abstraction. And then you learned about proc notation, and how…everything just…fits together. And you can declare some nice computations/compositions in a way that looks a lot like monadic do blocks.

We saw how complex compositions — and complex recursions — now look really pretty and nice in proc-do notation.

And then we saw an extension of the “opaque state” concept we learned last time — locally stateful compositions. Using Auto composition and the Arrow instance, we can now combine Autos with local state together…and they all maintain and keep track of their own state. No more “global state”, like before — every individual component only minds what it knows, and nothing else. And that this is really what “composability” really is all about.

Up next, we will transition from Auto to the Wire abstraction, which is sort of like an Auto with more features.

And then we will be on our way! :D

Exercises

Yeah, I know that a lot of this post was pretty abstract…finding ways to make this post immediately useful with applications was one of the reasons why it took so long for me to get it out, after the last one.

That being said, there are some things you can try out test your understanding before Part 3 :)

  1. Write the Profunctor instance mentioned above; look at the Functor instance we wrote as a reference. And hey, how about Strong and Choice, too?

  2. Try writing the various Autos we wrote last time at the end using composition and proc notation instead of explicit recursion. Feel free to define your own “primitives” if you find that you must.

    Some of these might be trickier than others!

    Note that some of these can be done with just a straight-up autoFold, for the entire thing. While this is neat and all, it might be more useful to practice the principles of local statefulness, and try to break things up into as many primitives as possible; always try to avoid keeping every part of your state in one giant autoFold parameter.

    • Rolling average: You should be able to do this with just autoFold and the right proc block. You can even do it with straight up composition, but it’s a bit less clean.

    • onFor: You should be able to do this with settableAuto (or something like that), and some nice proc routing with if/then/elses.

    • autoMap: This should also be doable with autoFold; although there isn’t much state to separate out, so this example isn’t as interesting. It might be more fun to use this one as a component of a larger Auto, and see what you can use it for!


  1. Remember, we can write a -> b as (->) a b; like other operators, (->) can be used both infix and prefix.↩︎

  2. This was originally a typo but I like the word so much that I’m just going to leave it in here.↩︎

Comments powered by Disqus