Skip to content

Braiding extensible effects together in a pipeline/workflow of tasks

Notifications You must be signed in to change notification settings

tweag/kernmantle

Repository files navigation

Kernmantle

Kernmantle braids some strands of effects into a Rope that will be the basis of your application. By "effect", understand "any type representing a pure function or an impure (with side effects in the general sense) computation from an input value to an output value". The effects are interlaced, meaning they can be chained like the following:

eff1 >>> eff2 >>> eff1 >>> eff3 >>> eff1 >>> eff1 >>> eff2

where each effect can transmit its result to the next one.

This chain of interlaced effects is the Rope. Each effect is lifted in the Rope with the strand function. Effects can be interpreted (transformed to an actual calculation, possibly with side-effects, or to other effect types) thanks to the entwine function.

Each effect can represent some task in a pipeline. The way each task will be executed completely depends on the final interpretation functions given to entwine. That makes each task of the pipeline fully configurable, and each task or pipeline of tasks very easy to share and reuse. This also (as does any arrow pipeline) gives you the capacity to analyse the pipeline ahead-of-time, before it even runs.

More detailed information about the interest of this paradigm in the context of data science can be found in the ICFP Haskell Symposium 2020 paper we published about Kernmantle, and also in the porcupine package documentation, which uses this paradigm.

To build

Recommended way is to use stack. Nix config might not be up to date.

Why yet another effects library??

Existing effects library model only unary effects (effects parameterized only by their result). Kernmantle models binary effects:

eff a b

where eff is most often contravariant in a (input) and covariant in b (output) like any function (but the type system doesn't enforce that). The simplest binary effect conceivable is therefore... the one that performs no effect, that is, the pure function (->).

How do I create "binary effects"?

Currently, you have roughly four ways to create binary effects (more may exist, please feel free to ping me if I forgot some):

  • model them as (G)ADTs and write interpreter functions,
  • compose them from regular applicative/monadic effects (Reader, Writer, State etc.),
  • reuse existing binary effects (Categories, Arrows, Profunctors etc.),
  • take the Product (from bifunctors) of 2 different binary effects. They are composed in parallel: results from one cannot be passed to the other.

The second option is particularly interesting given the huge amount of libraries in Haskell exposing effects in the form of monads or applicative functors.

Using effects that implement some abstraction (like Category) isn't mandatory, but the advantage of these effects is that you can compose several instances of them before lifting them in the Rope.

Composing regular applicative/monadic effects:

Binary effects can be of the form:

type SomeBinEff f w m a b = f (w a -> m b)

If f is an Applicative, w a Comonad, m a Monad, and if w is distributive over m, then SomeBinEff f w m will be a Category (and an Arrow).

Any of f, w and m can just be Identity:

  • If f, w and m are Identity, that's just a pure function (->),
  • If only f is non-Identity, that's the Cayley profunctor/category,
  • If only m is non-Identity, that's the Kleisli category.
  • If only w is non-Identity, that's the Cokleisli category,

Note that f can be composed out of several applicative functors (with Compose). And of course m can use any usual monadic effect composition (regular monad transformers, capability, polysemy, etc). So reusability is really key here.

To give an example, in the porcupine library which expresses data pipelines, f is a Writer effect that accumulates the requirements of the pipeline (required files and resources, parameters, etc.) and some execution context (namespace for the logger, etc.) and m aggregates some logger, state and reader effects.

The Rope type

The Rope type we introduce here allows to compose effects in a sequential manner: they can pass results to one another. Branching is also supported (via ArrowChoice or Choice classes), but fully dynamic determination of the effect to run next depending on the result of a previous effect is precluded by design (that would require Monad/ArrowApply).

Rope has the following signature:

newtype Rope (record::RopeRec) (mantle::[Strand]) (core::BinEff) a b

-- The kinds used in Rope:
type BinEff = * -> * -> *
type Strand = (Symbol, BinEff) -- A strand is a bin effect with a name
type RopeRec = (Strand -> *) -> [Strand] -> * -- Collects all the methods to run each strand

It's parameterized over all the effects contained in the Rope (that's the mantle part) and an extensible record (from vinyl) which will contain the way to run the effects. Different record offer different trade-offs in terms of extensibility (adding new effects, running existing effects, ie. interpreting them in terms of other effects of the mantle or of the core) and algorithmic complexity (when lifting an effect in the Rope).

The final parameter is the core effect in which the effects of the mantle will end up being interpreted, but most of the code will maintain core polymorphic and just add constraints to it.

Note that the Rope itself is a binary effect, so it can be reused in the contexts we mentioned before.

Do I need to use Arrows to use Kernmantle?

Categories are omnipresent in Haskell, but Arrows get a bad rap for some reason. But effects in Kernmantle aren't required to instanciate any class. The only requirement will be that in the end, you must be able to interpret them in some core binary effect. Rope implements both the Arrow typeclass stack (with the exception of ArrowApply, as we said) and the Profunctor stack, but you don't have to use either one or the other to use Kernmantle. Effects can be represented as pure (G)ADTs and you can manipulate them without resorting to any other abstraction.

Why the name?

A Kernmantle rope is a rope composed of 2 parts, the core (kern) and the mantle. Like any rope, both parts are made out of woven strands. The metaphor seemed quite nice for effects that get interlaced, stranded together, and then interpreted in a core effect.

About

Braiding extensible effects together in a pipeline/workflow of tasks

Resources

Code of conduct

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 4

  •  
  •  
  •  
  •