[go: up one dir, main page]

Skip to content

syncpoint/signal

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

92 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Signal

Time-varying values with acyclic static dependencies and synchronous glitch-free updates. Signal is heavily inspired by and in many aspects quite similar to flyd. Our thanks and appreciation go out to the people who provided this neat piece of software. Thank you guys!

Table of content

Expand

Quickstart

npm install @syncpoint/signal
npm install ramda # optional

Combining two input signal is as simple as

import Signal from '@syncpoint/signal'
const a = Signal.of(3)
const b = Signal.of(2)
const c = Signal.link((a, b) => a * b, [a, b])
b(4) //=> 12

BTW: Signal has no runtime dependencies.

Motivation

Signal falls in the broad spectrum of Functional Reactive Programming (FRP). We think FRP is not a fancy hammer (sorry: technology) which solves each and every problem. It's rather a different way to think about certain problems and derive solutions which posses properties common to functional programming in general. We are not here to argue whether or not FP might have its benefits over imperative paradigm. We simply offer a small abstraction which might be useful to you in certain circumstances.

Signal is one of many packages out there to choose from. Some of them are Methusalems pathing the FRP way starting nearly a decade ago. Some are feature-packed Swiss Army knives for nearly every conceivable use case. Some are heavy, some are light. Some are abandoned, some are actively maintained. It's hard to layout a map of FRP-Land and it's even harder to pin-point the exact location where package XY sits.

But we can say what Signal is not: Signal is not a Web/GUI framework like Solid, Vue or the likes. Signal is not RxJS. Signal is not an application state management framework like Redux or MobX. Signal is not a framework at all. Signal is just signals.

History of Origins

For one of our projects we had to extent a rather complex OpenLayers interaction to support custom geometries. Three attempts to extent the existing code failed. We figured that encapsulating a ton of mutable state in a single class was not the way we wanted to do business. We looked into alternative approaches to structure code and route and transform incoming mouse and keyboard events until they eventually updated the view model. The first implementation was based on @most/core which worked out really nice. But the older I the author get, the more paranoid about project dependencies I become. @most/core seemed a little dormant, it is written in TypeScript (we are not huge fans of) and has some inherent complexity for features we don't need (async scheduling, time tracking). Maintaining @most/core if necessary was out of the question. Next flyd. Yes, an old-timer too, but simple enough. Migrating to flyd was as simple as it gets. No scheduler, synchronous updates and a relatively small JavaScript code base. In the end, flyd had some rough edges we liked to iron out et voilĂ , Signal came into being.

Introduction

Signal provides two primitives: Input signals Signal.of and linked signals Signal.link. An input signal is just a container for a current value. In general, signals are only updated if the new value is not the same value as its current value. One or more input signals can be linked to one output signal. The link function derives the output value from the input values. The output signal's value is automatically updated when at least one input signal's value has changed.

const sum = (a, b) => a + b
const a = Signal.of(39)
const b = Signal.of(3)
const c = Signal.link(sum, [a, b])
c() //=> 42
a(3); c() //=> 6

A signal is basically a function where the 0-ary form returns the signal's current value, and 1-ary form updates the current value. Linked signals are read-only and cannot be updated explicitly. Naturally, linked signals can be used as input signals for other linked signals.

const a = Signal.of(1)
const b = Signal.link(a => a + 1, a)
const c = Signal.link(b => b * 2, b)
c() //=> 4

Signals can be initialized asynchronously. Signal.deferred accepts either a value or a promise or a function which returns a value or a promise. Note: If the signal should asynchronously be initialized with a function value, it must be provided as a promise and cannot be given directly as a value.

const a = Signal.deferred(() => Promise.resolve(42))
console.log(a()) //=> undefined
setImmediate(() => console.log(a())) //=> 42

Side-effects can be triggered with Signal.on. The side-effect function is invoked every time the signal value changes. Signal.on returns a function which, when called, removes the effect from the signal's dependent list, so it is no longer called.

const a = Signal.of()
const acc = []
const push = x => acc.push(x)

// Hint: a.on(fn) is the same as Signal.on(fn, a)
const dispose = a.on(push)
a(1); a(2); acc //=> [1, 2]
dispose()
a(3); acc //=> [1, 2] (unchanged)

The constructors Signal.of, Signal.link and Signal.deferred take all an optional equals option, which controls when two consecutive values are considered equal. equals option overrides the default of Object.is.

const a = Signal.of(0, { equals: R.F }) // always different
const b = Signal.scan(R.flip(R.append), [], a)
R.range(0, 5).map(() => a(0))
b() //=> [ 0, 0, 0, 0, 0, 0 ]

Other operators currently don't take an equals option, but it's always possible to explicitly set equals option on (derived) signals with Signal.options.

const a = Signal.of(0)
const b = a.map(R.add(1))
Signal.options(b, { equals: (a, b) => (a % 2) === (b % 2) })
const c = Signal.scan(R.flip(R.append), [], b)
;[5, 3, 10, 0].map(a)
c() //=> [ 1, 6, 11 ]

And really, that's all there is to know. Except...

Signals are Monads

Signal conforms to Fantasy Land specification for algebraic structures in the following ways:

  • Filterable: filter :: Signal s => (a -> boolean) -> s a -> s a
  • Functor: map :: Signal s => (a -> b) -> s a -> s b
  • Apply: ap :: Signal s => s (a -> b) -> s a -> s b
  • Applicative: of :: Signal s => a -> s a
  • Monad: chain :: Signal s => (a -> s b) -> s a -> s b
  • Setoid: equals :: Signal s => s a -> s b -> Boolean
  • Ord: lte :: Signal s => s a -> s b -> Boolean

These operations can be used directly on signals (fluent interface) or preferably in point-free notation for improved composability.

const a = Signal.of(3)
  .map(x => x + 1)
  .filter(x => x % 2 === 0)
  .map(x => x * 2)
  .chain(Signal.of)
  .ap(Signal.of(x => x > 4))
a() //=> true

Although filter, map, etc. are exposed as curried functions under the Signal namespace, e.g. Signal.map, we prefer using a third-party library like Ramda for point-free style.

import * as R from 'ramda'
const fn = R.compose(
  R.ap(Signal.of(x => x > 4)),
  R.chain(Signal.of),
  R.map(x => x * 2),
  R.filter(x => x % 2 === 0),
  R.map(x => x + 1)
)

const a = fn(Signal.of(3))
a() //=> true

Signal supports transducer protocol

A transducer (not the device) is a powerful tool for transformation of data in an efficient way. What's more transducers like map and filter can be applied to any suitable collection or container type like array, sets, streams, channels, CSP and so on. The transducer protocol proposed by James Long assures that transducers are compatible between different libraries like Ramda, which is used in the following examples. There is a bunch of excellent write-ups about what transducers are, how they work and what possible implementations might look like. Attempting to explain this ourselves would result in failure. Here is a small list of posts for the curious:

General transducers expand Signal's functionality beyond its built-in operators. For Ramda's R.drop or R.reject for example there are no direct equivalents in Signal. But we can use them through Signal.transduce.

const a = Signal.of()
const b = Signal.transduce(R.drop(3), a)
const c = Signal.scan(R.flip(R.append), [], b)
R.range(1, 7).map(a)
c() //=> [4, 5, 6]

Composing transducers is more efficient than introducing intermediate signal in a long chain of transformations.

const xf = R.compose(
  R.map(R.add(-1)),
  R.filter(x => x % 2 === 0),
  R.map(R.multiply(3))
)

const a = Signal.of()
const b = Signal.transduce(xf, a)
const c = Signal.scan(R.flip(R.append), [], b)
;[4, 1, -3, 8, 7].map(a)
c() //=> [0, -12, 18]

Nested signals, reads, writes

Although we advise against using functions with side-effects where possible, you can do so with Signal. Consider the following example. In link function of b, at label L0 flag in parent scope is updated to true. Updating a at L1 triggers the evaluation of a's link function. Current value of flag is true during this evaluation. Control is only returned to L2 after all of a's dependent links are evaluated and updated.

const acc = []
const push = x => acc.push(x)

const flag = Signal.of(false)
const a = Signal.of()
const b = Signal.of()

Signal.link(a => {
  push(`[2]:${a}`)
  push(`[3]:${flag()}`)
}, a)

Signal.link(b => {
    push(`[1]:${b}`)
L0: flag(true)
L1: a(b * 2)
L2: flag(false)
    push(`[4]${flag()}`)
}, b)

b(2); acc //=> ['[1]:2', '[2]:4', '[3]:true', '[4]:false']

Signals as well can be nested as one would expect.

const a = Signal.of(6)
const b = Signal.link(a => {
  const c = Signal.of(a * 2)
  const d = Signal.link(c => c / 3, c)
  return d()
}, a)
b() //=> 4

And finally, signals can be passed around like ordinary values (which they are). The following example is a 16-bit ripple-carry adder using full adders (2 x XOR, 2 x AND, 1 x OR). That's a grand total of 113 signals.

import * as R from 'ramda'

const xor = R.unapply(Signal.link((a, b) => a ^ b))
const and = R.unapply(Signal.link((a, b) => a & b))
const or = R.unapply(Signal.link((a, b) => a | b))
const toString = radix => s => s.toString(radix)
const padStart = (length, pad) => s => s.padStart(length, pad)
const split = separator => s => s.split(separator)
const decode = xs => parseInt(xs.reverse().join(''), 2)
const encode = R.compose(
  R.reverse, R.map(Number), split(''), padStart(16, '0'), toString(2)
)

const fullAdder = ([a, b, cin]) => {
  const x = xor(a, b)
  const s = xor(x, cin)
  const cout = or(and(x, cin), and(a, b))
  return [s, cout]
}

const parallelAdder = cin => R.range(0, 16).reduce(acc => {
  const ab = [Signal.of(), Signal.of()]
  const [s, cout] = fullAdder([...ab, acc.cout])
  acc.a.push(ab[0]); acc.b.push(ab[1]); acc.s.push(s)
  acc.cout = cout // c_in of next adder is c_out of previous adder
  return acc
}, { a: [], b: [], s: [], cout: cin }) // a, b, s: 16-bits each

const { a, b, s, cout } = parallelAdder(Signal.of(0))
encode(47813).forEach((v, i) => a[i](v))
encode(19987).forEach((v, i) => b[i](v))
decode([...s.map(s => s()), cout()]) //=> 67800

Fine-print: undefined

undefined (in contrast to null) is not considered a valid value for signals. We say a signal is undefined, when it holds undefined as its current value. Signals may start off undefined: Signal.of(), but once they have a valid value, there is no way back to undefined.

const a = Signal.of()
a() //=> undefined
a(1); a() //=> 1
a(undefined); a() //=> 1 (unchanged)
a(null); a(); //=> null (valid signal value)

The same holds for linked signals.

const a = Signal.of(3)
const b = Signal.link(a => a < 4 ? a * 2 : undefined, a)
b() //=> 6
a(4); b() //=> 6 (unchanged)
a(1); b() //=> 2

As a consequence, linked signals are only evaluated, if all input signals are defined.

const acc = []
const push = x => acc.push(x)

const a = Signal.of()
Signal.link(push, a)

acc //=> [] (not evaluated, yet)
a(1); a(2); acc //=> [1, 2]
a(3); acc //=> [1, 2, 3]

Fine-print: Glitch-free

Updates are efficient and glitch-free, i.e. there is no inconsistent intermediate state which can be observed. Consider the following diamond-shaped dependencies.

const acc = []
const push = x => acc.push(x)

const a = Signal.of(1)
const b = Signal.link(a => a + 3, a)
const c = Signal.link(a => a * 2, a)
const d = Signal.link((b, c) => b / c, [b, c])

Signal.link(push, d)
acc //=> [2] (d was evaluated once)
a(3); acc //=> [2, 1] (d was evaluated once again)

Fine-print: Disposable

You might have noticed the absence of stream.end which flyd provides for ending streams and removing them from the dependency graph. That's because signals cannot be ended or closed. Also there is no way of removing signals or dependencies programmatically from the static dependency graph once they have been added. With one exception, which happens under the hood in private territory: chain. chain operates on a signal of signal of values Signal s => s s a. For each new signal of values chain receives, a linked signal (an effect actually) is created which update the outer output signal with values it receives from the current signal of values. Upon arrival of a new signal of values, this effect is unlinked and the dependency is removed from the previous signal of values. Now for the interesting part: When the current signals is replaced with a new one, chain checks if the signal incidentally has a dispose function. If so dispose is called during the process of unlinking the effect from the signal of values. dispose is completely optional, but can be useful to clean up resources of signals which are dynamically created and flattened through chain. fromListeners operator for example has a dispose function which removes registered listeners from the target.

const { chain, fromListeners } = Signal
const target = Signal.of()
const events = chain(fromListeners(['change']), target)

Whenever target is updated with a new target, change listener is registered on the new target and deregistered from the previous target. event is updated with (flattened) change events.

Super fine-print: When chain function does return a value other than a signal (null for example), the previous signal of values (if any) is cleaned up and no new effect is created until a signal of values is received again.

Fine-print: Error Handling

Easy! None. Using Maybe, Either or similar as signal values might be beneficial.

Why signals and not streams?

Streams don't have the notion of a current value, which can be queried at any given time. Hence we favor the term signal over stream. Signals (think of digital logic circuits), have discrete values which can and usually do vary over time. Also, we link one or more input signals to one output signal const AND = link((a, b) => a && b, [a, b]). In addition, signals have no buffering nor back-pressure handling.

Miscellaneous Operators

The following operators live under the Signal namespace and might be handy or they might not. There is no real reason to include them in this library, except that we use them for a different project. These operators can all be implemented with a few lines of code and only use the public API introduced so far.

fromListeners was already mentioned above. Note: fromListeners supports both addEventListener/removeEventListener and on/off protocols.

fromListeners :: [String] -> Target -> Signal Event

startWith injects a signal value where a signal would be undefined otherwise. This is mainly relevant for linked signals or signals created from listeners.

// startWith :: Signal s => a -> s a -> s a
// startWith :: Signal s => (() -> a) -> s a -> s a
const a = Signal.of()
const b = Signal.link(a => a + 1, a)
const c = Signal.startWith(0, b)
c() //=> 0

merge merges values of two signals into one resulting signal.

// merge :: Signal s => s a -> s b -> s (a | b)
const a = Signal.of()
const b = Signal.of()
const c = Signal.merge(a, b)
const d = scan(R.flip(R.append), [], c)
a(1); b('2'); b('3'); a(4); b('5')
d() //=> [1, '2', '3', 4, '5']

scan feeds back the calculated signal value as an accumulator for the next value.

// scan :: Signal s => (b -> a -> b) -> b -> s a -> s b
const a = Signal.of()
const b = Signal.scan((acc, a) => acc + a, 0, a)
R.range(0, 10).forEach(a)
b() //=> 45; sum 0..9

loop is similar to scan, but the value for the accumulator and the returned signal value can be different.

// loop :: Signal s => (b -> a -> [b, c]) -> b -> s a -> s c
const average = xs => xs.reduce((a, b) => a + b) / xs.length
const a = Signal.of()
const b = Signal.loop((xs, x) => {
  xs.push(x); xs = xs.slice(-10)
  return [xs, average(xs)]
}, [], a)
R.range(0, 20).forEach(a)
b() //=> 14.5; sum 10..19 / 10

lift applies the values of n signals to a n-ary function. It might not be obvious, but lift is pretty much the same as link, expect for the signal parameters, which are not given as an array of signals, but individually.

// lift :: Signal s => ((a -> b -> ...) -> x) -> s a -> s b -> ... -> s x
const a = Signal.of()
const b = Signal.of()
const c = Signal.lift((a, b) => a + b, a, b)
a(1); b(2); c() //=> 3

tap is used for side-effects while passing on the value (hopefully unchanged).

// tap :: Signal s => (a -> any) -> s a -> s a
const fn = R.compose(
  Signal.tap(x => console.log(x)), //=> 2
  R.map(x => x + 1)
)

const a = fn(Signal.of(1))
a() //=> 2