Introducing: the Auto library!

SourceMarkdownLaTeXPosted in Auto, Haskell, Projects, RamblingsComments


Auto: README (with examples) / hackage / tutorial / examples / github

(Before anything, maybe take a quick look at the detailed description in the README for a quick motivating example and explanation of the library)

Today I’m announcing and beginning promotion of my auto library, a denotative and locally stateful programming DSL and platform, now on hackage. auto is suitable when your program involves an input or output that is a discrete stream of things — events, views, etc., like turn based games, GUI’s, numerical computations…; it allows you to state (possibly cyclic) complex relationships between streams of values by composing simple, primitive ones. You can read the README too for a detailed buzz-word laden exposition with nice well-commented short demos and examples, get started with the tutorial, check out the directory of sample projects, and even see a live running todoMVC (source) example!

Over the next week or two I’m going to be breaking down real-world projects written on auto, and even be talking about the design processes of programs written using auto. You can follow along on the series page, follow me on twitter, or just subscribe to the rss feed feed; expect a post on designing, from start to finish,

  1. A fully running chat bot
  2. A GUI-based todo app on ghcjs
  3. A text-based adventure game a la the classic rogue
  4. A numerical computation DSL

But enough of that…what is auto, why does it exist, and what are its design principles?

Auto

on State

I designed auto because there really aren’t any good solutions in Haskell for declaratively describing locally stateful programs in a compositional way. And a lack of denotational semantics to reason with them.

The go-to implementation for a turn-based game is to have a “giant state monad”. It is a clever “hack”, but really, all we’ve done is began programming a game with global mutable state. People have tried getting over this by using lenses and zoomers, but this processes doesn’t quite scale.

Even now many games, GUI’s, numerical computations, etc. are written as folds or state compositions over a giant state. Surely there is a better way?

auto provides denotational semantics for the composition and transformation of stream transformers. At a high level, it is an example of “locally stateful programming”. Each component and stream transformer really operates as its own entity, separate from the world, in a composable way. Typical programs involve building a (possibly cyclic) graph of relationships between quantities over a stream.

Composing two transformers side-by-side or end-to-end creates a new transformer…and the state of each trasnformer is “closed off” from the other and the rest of the world.

sumAndProduct = proc x -> do
    sums  <- sumFrom 0     -< x
    prods <- productFrom 1 -< x
    id -< sums + prods

sumFrom 0 denotes a relationship between x and sums such that sums is the cumulative sum of all x’s seen. productFrom 1 denotes a relationship between x and prods that prods is the cumulative product of all xs seen. With sumAndProduct, we built a new relationship — the output is the sum of the cumulative sum and the cumulative product of the inputs — by composing two primitives.

-- running our Autos over the stream [1..10] to get a new stream
ghci> streamAuto' (sumFrom 0) [1..10]
[1,3, 6,10, 15, 21,  28,   36,    45,     55]
ghci> streamAuto' (productFrom 1) [1..10]
[1,2, 6,24,120,720,5040,40320,362880,3628800]
ghci> streamAuto' sumAndProduct [1..10]
[2,5,12,34,135,741,5068,40356,362925,3628855]

Each of them maintain their own “state”…and even sumAndProduct will maintain its own internal state as you compose it with other things.

You build complex programs. For games, you might have an enemy monster, or a player character…why should a player character’s update be able to access the state of the enemy monster? Why should the enemy monster be able to access the state of the player? Now, the enemy monster’s state is only accessible to the enemy monster itself…nobody else can touch it.

In my opinion, this really is the only sane way to have stateful loops. Separate out your update into the composition of primitives, and describe it using denotative, pure relationships.

on Architecture

Using auto, your entire program is structured as a stream transformer, built from the composition of smaller primitives, and using the various combinators and semantic streams of the library.

In the end, “running” it is simply gathering the input one at a time, pushing it through, and acting on the result.

For example, in the todo example, the entire application is just one Auto that takes a stream of input commands and outputs a stream of output GUI views. It’s launch so that the Auto waits on a Chan queue to consume inputs as they come in. All the javascript front-end has to do is render the output gui view, and hook up the DOM elements to trigger events that add new inputs to the queue. That’s it!

How do we build that Auto? By composing smaller, simpler ones. Each denoting their own relationship, each keeping track of their state locally. The “full GUI state” never exists anywhere…it is aggregated during composition.

By the way, auto does allow you to take “snap shots” of the actual states of Autos as they are run, as a binary…so you can serialize, freeze, and resume Autos from any previous state at-will. Free undos, free save states. And this serialization composes, so the combination of two serialized Autos with internal state will also be serialized appropriately.

on Comparisons

Throughout its development auto has been compared to FRP libraries like netwire. A full address of this comparison is offered on the readme. The main difference is that FRP offers an expressive language for working with, manipulating, and transforming continuous-time behaviors. auto borrows some aspects of FRP as well as some practical API aspects in order to build something separate, manipulating and transforming causal (discrete) streams. There are many situations where FRP is not quite suited — it’d be like using a vector art program to describe a bitmap. There are domains where auto is not suited — that is, the semantic model doesn’t allow you to say anything meaningful about continuous time behaviors.

auto has also been compared to pipes and conduit, but there are some major differences in design and philosophy. pipes is a more general-purpose co-routine library with an emphasis on being able to apply equational reasoning to many different domains. conduit focuses around the problem of effective streaming with resource management. Both work “sources” that come from underlying monads like IO; auto discourages this except for disciplined exceptions, and it’s definitely not the norm. auto works as value stream transformers, working with “pure” one-by-one transformations on streams of values; pipes and conduit provide effect stream manipulators, managing streams of effects in constant space, with resource management, etc…and often involving output effects as a result (“consumers”).1

on the Future

So, I’ll be using this blog to post complete walk-throughs on designing specific apps like the ones in the examples repo. Along the way hopefully some general design principles will become apparent, too. I’ve been working on this library and have been writing “real world” code for it for almost a year now, so I have some experience with design practices and patterns that work and don’t work.

For now, you can try reading over the tutorial if you want, or peruse the examples repo!

Also, I definitely welcome any criticism on the design of the library or the semantic model, or of its use cases. I’ve been more or less working on this alone for almost a year, so now is the time for any delusions of mine to be vetted out in public!

on Support

For now, the official support channel is #haskell-auto on freenode (I’m jle`), but you can always use the issue tracker too, email me, or find me on twitter as mstk.

All this being said, auto is still kind of technically in a sorta pre-release state, because not all of the tests are written yet. But the API should be stable and updates before 0.3.x are going to all be backwards compatible (API-wise) bug fixes or filling in holes.


  1. One could still use a subset of pipes that does not stream effects, but merely values, and that does somewhat fill a similar role; this is used in the mvc library to build similar applications that auto tries to build. However, due to mvc’s “global state” nature, you lose many of the local statefulness idioms in auto, and a lot of auto’s benefits and design philosophies go away, for other reasons as well.↩︎

Comments powered by Disqus