Adaptive (Self-Adjusting) Computations in JS

This is a toy! Do not use in real projects!

This is a JS/TS reimplementation of the Haskell library described in the paper Monads for Incremental Computations by Magnus Carlsson. Of course, TypeScript does not support monads (higher-kinded types needed), nor do-notation, so a number of compromises were made in the conversion. The Haskell library was in turn inspired by earlier ML implementation, but I have not looked into that code. Also the Haskell improvements (removing explicit writes) do seem desirable.

It has mostly been a fun exercise and a learning aid to understand the ideas behind the adaptive computation literature.

What is Adaptive Computation?

In the programming languages we use today computations are single-usage. We build a data structure, pass it to an algorithm, and receive the result. If at later point of time the data structure changes, the algorithm has to be rerun from scratch.

const data = [1, 2, 3];
const function algo(xs: number[]) {
  let r = 0;
  for (const x of xs) r += x * x;
  return r;
}
const result = algo(data);  // what we call the computation

// some time passes

data[1] = 100;
const newResult = algo(data);  // we have to redo the computation from scratch.

We can design the data structure and the algorithm to be "adaptive", i.e. to support mutations together with an appropriate change to the result. In the example above, this can be done by:

function changeInputAtIdx(
  orignalArr: number[],
  idx: number,
  newVal: number,
  oldResult: number
) {
  return oldResult - originalArr[idx] * originalArr[idx] + newVal * newVal;
}

Notice that the update logic is almost as complicated as the original "computation" and we had to write it by hand. What if we can automatically "derive" if from the original computation, no matter what the computation was.

This is what "adaptive computation" tries to achieve. Some languages like Lustre tackle the problem. Here instead we attempt to solve it by using a library.

How to use (Simplified API)

To make this library a bit more palatable to JS developers that are used to the current set of reactive libraries, I introduced a simple wrapper and different APIs on top of the original ones (see below).

There are only two primitive types:

Signal<T> - a reactive value of type T that can be read using the .value property to simply get the value or using the .read method to create a new signal.
Input<T> - a mutable signal of type T that can be written to.

So an incremental (or reactive) computation looks like.

const a: Input<number> = new Input(1);
const b: Input<number> = new Input(2);
const sum: Signal<number> = a.read(x => b.read(y => x + y)); 
console.log(sum.value);  // 3
a.value = 100;  // sum is syncronously updated.
console.log(sum.value);  // 102

How to use (Original API)

The library is structured around three main classes - Adaptive, Changable and Modifable.

Adaptive can be thought of as the context of an adaptive computation. It holds all the internal data structures needed. Adaptive computation begins by creating an Adaptive:

const a = new Adaptive();

In an adaptable computation we have Modifiable<T> which is like mutable variables of type T in programming languages. We can create them using a.newMod(...). In the ... we pass a Changable<T>, which is like an expression computing a value of type T in familiar programming languages.

The simplest way to produce a Changable is to use the constant function.

const m1 = a.newMod(constant(1));
const m2 = a.newMod(constant(2));

This is the adaptive version of:

const m1 = 1;
const m2 = 2;

The only other way to produce a Changable is to read a Modifiable using a.readMod. It takes a callback with the current value of the Modifiable (very similar to a Promise). At the end of the callback we still have to return a Changable.

For example, the adaptive version of m3 = m1 + 2 will be:

const m3 = a.newMod(a.readMod(m1, x => constant(x + 2)));

And with chaining we can write more complicated expressions like the adaptive version of m3 = m1 + m2.

const m3 = a.newMod(a.readMod(m1, x => a.readMod(m2, y => constant(x + y)));

Simply reading a Modifiable to produce a Changable can be a bit awkward - a.readMod(m, x => constant(x)). To safe some typing we added a.modToC(m) which is just sugar for that same operation.

To summarize the main types here:

Adaptive Type	Classical analog
Modifable	Variable of type T
Changable	Computation resulting in value of type T
Adaptive	Context

At any point the result of the computation can be obtained with Modifiable.get(). Note, that this access is not considered "adaptive" and should be done only outside Changable expressions. Inside Changable one has to use readMod.

console.log(m1.get(), m2.get(), m3.get()); // 1, 2, 3

Finally, we can change the value of any modifiable using a.change(Modifiable, newValue). Once we change all modifiables we would like, all affected computations will be rerun with a.propagate(). Notice we don't need to spell out what will be affected.

a.changeMod(m1, 10);
a.propagate();
console.log(m3.get()); // 12

Note that one should not use a.change inside Changable expressions. Modifiables should be treated as immutable inside Changables, but we can always create new Modifiables using a.newMod.

The Dynamic Dependency Graph

Underneath the hood the library is tracking which Modifiables were read to produce the other ones. This mechanism is dynamic, which means that it can change from one execution to another.

const switchMod = a.newMod(
  a.readMod(booleanMod, b => {
    return b ? a.readMod(modTrue, x => x + 1) : a.readMod(modFalse, x => x + 2);
  })
);

In this example, the dynamic dependency graph will track which one of modTrue or modFalse was read in the previous recomputation. If b is True and later modFalse changes, no spurious recalculation of switchMod will occur. If at a later point b changes the dependency graph will automatically adjust.

Why all the callbacks

The pervasive use of closures allows for the right scope of reevaluation to occur.

const result = a.newMod(a.readMod(aMod, a => {
  const intermed1 = a * 2;
  return a.readMod(bMod, b => {
    const intermed2 = b * 2;
    return constant(intermed1 + intermed2);
  })
});

In this example if bMod changes, only the second callback will be evaluated. Hence, the result a * 2 will be immediately reused. Note, however, that if a changes the b * 2 result will not be reused. Because a is accessible inside the b reading closure, the library cannot know that b's computation was not dependent on a.

In order to achieve the more efficient computation reuse one has to rewrite the computation as below. Having to think about these scenarios is a major downside of this approach to adaptive computations.

const a2Mod = a.newMod(a.readMod(aMod, a => constant(a * 2));
const b2Mod = a.newMod(a.readMod(aMod, b => constant(b * 2));
const result = a.newMod(a.readMod(a2Mod, a2 => a.readMod(b2Mod, b2 => constant(a2 + b2))));

Implementation details

The original Haskell implementation has at least four different monads. I have removed all, but one - Changable and replaced the rest with simple mutable operations.

As observed in the Carlsson paper the Changable monad is nothing more than the continuation monad with some callback caching based on causality ordering. A nice gentle intro to the continuation monad can be found in Dan Piponi's blog here,and here.

The original Haskell implementation uses a data structure by Deitz and Sleator, which I skipped implementing and instead used a trivial replacement (which would likely fail to scale).

Syntax sugar

Undoubtedly the current syntax is odd, because of the explicit closures. Similarly to Promises and async/await, one can invent a syntactic sugar to make it look like writing regular imperative code with adapt/read keywords.

const result = adapt with a { return (read a2Mod) + (read b2Mod); }
// transpiles to
const result = a.newMod(a.readMod(a2Mod, a2 => a.readMod(b2Mod, b2 => constant(a2 + b2))));

Connections with Reactive programming, FRP and dataflow languages

I think adaptive computations are smaller and more restrictive versions of the same underlying ideas. As such this library likely can be fully backed by libraries like Rxjs, but that might be an overkill. Adaptive computations as defined by this library are necessarily smaller scoped than general dataflow, because one cannot express operations like take(3), which would mean adapt to changes in other Modifiables only up to 3 times. Being more restrictive, of course results in simpler model of how this system works, compared to a general dynamic graph of streams.

This area needs more exploration.

Examples

See examples/ directory for non-trivial adaptive algorithm examples.

Next Steps

add support for pluggable equality checking.
add "demand-driven" propagation, i.e. if I am reading signal X, and I know one of its inputs in the
add more examples of non-trivial algorithms that can be turned adaptive.
build a JS framework around this (jk).

Open Quesions

Can native JS arrays and objects be turned into their Modifable equivalents. The best current answer is use custom data strucutres like the linked list in aqsort example, which of course is a non-starter for any general use.
Are the runtime performance and memory consumption of this acceptable?
What runtime checks should be added to get back some of the guarantees that the Haskell implentation had using monads?
what happens when adaptable computations throw.

Developing

Run tsc -w and jest --watchAll in two different terminals.

References

Magnus Carlsson. Monads for Incremental Computing. ICFP '02 Proceedings of the seventh ACM SIGPLAN international conference on Functional programming link
U. Acar, G. Blelloch, and R. Harper. Adaptive functional programming. In Principles of Programming Languages (POPL02), Portland, Oregon, January 2002. ACM. link
P.F.Dietz and D.D.Sleator. Two algorithms for mainitaining order in a list. In Proceedings. 19th ACM Symposium. Theory of Computing, 1987. link

rkirov / adapt-comp