Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.



ECMAScript async, generator and async generator functions with access to their state. The state can be clonable and serializable.

This enables:

  • universal apps, time traveling, hot reloading
  • custom jobs scheduling
  • long-running workflows (storing state in DB, e.g. while awaiting users e-mail confirmation)
  • various control manipulation operators


There is a simple counters demo - sources /transpiled. More details about it are in Async Generators as an alternative to State Management article.

It uses this library to implement:


$ npm install --save-dev @effectful/es
$ npm install --save @effectful/es-persist

In .babelrc:

  "presets": ["@effectful/es-persist/transform"]

If there are other babel presets, add "passPerPreset":true.

Import it in your module:

import * as R from "@effectful/es-persist"

// OR:
// const R = require("@effectful/es-persist")

async function main() {
  await say("hello world")


Like ECMAScript, transpiled async(generator) functions return an object with corresponding protocol support - Thenable and AsyncIterator resp. But unlike ECMAScript, these objects also store necessary local and closure captured variables values and execution control state (yield/await expression address where the function is suspended).

This object is serializable if the local variables are. If there are closures, the serialization library should be reference-aware. The transpiled code saves only variables with writes before any yield/await and reads after. It also resets them to undefined if its value isn't read after.

AsyncValue interface

interface AsyncValue {
  [R.awaitSymbol](r: Continuation)

The results of async functions and AsyncIterator protocol methods calls aren't Promises. There is another interface for them - AsyncValue. It is a replacement for Promises. The then method is replaced with [R.awaitSymbol]. It takes an object with Continuation interface instead of two callbacks. In the transpiled code AsyncValue interface is used instead of Thenable. All the AsyncValue objects in this library also implement Thenable, so everything (except a few control operators) works even if not transpiled.

The goal of this new interface is to make the continuation (the two callbacks methods of Thenable replacement) to be serializable. It is difficult to serialize a function if it has closure captured variables, or it is a result of Function#bind. But it is much simpler if instead of callbacks the program operates with objects where method are defined in its prototype.

The method name is a local symbol - R.awaitSymbol. This makes extending other classes (like Promise) safer.

Transpiled program expects await arguments to be AsyncValue. Passing a Promise there will work because the library extends its prototype. However, this will make the state not-serializable until it is settled.

Not-transpiled(native) version works only with Thenable. So if the code is supposed to be used in both modes (transpiled and not), the argument should implement both interfaces.

Unlike then method [R.awaitSymbol] returns nothing. So async operations can be composed only at the left (e.g., something like a.then(f).then(g) isn't possible, and its equivalent a.then(x => f(x).then(g)) to be used instead). This way we avoid allocations considering the API is designed to be used mostly by generated code, and the generated code composes computations only at the right.

Continuation interface

interface Continuation {

This is a protocol for [R.awaitSymbol] methods argument. Here resume/reject methods correspond to then arguments in Thenable interface.

These objects can be captured, saved (or cloned) and resumed after.

The ability to clone continuations makes it different to ECMAScript async (generator) functions. There the continuation can be executed at most once.

Some lightweight ES Proxy based copy on write can be used instead of a full clone for better performance. Or the object may be stored in some DB if it is suspended on some long-running remote operation.


JavaScript engines have a single global event loop and a queue to schedule async jobs. For simplicity, assume there is a global variable maintained by JavaScript runtime. I'm calling it Context here.

Unlike JS engines, this library supports multiple contexts. They may be assigned either globally, or per each async(generator) functions, or per each currently running async(generator) threads.

By threads here I mean cooperative multitasking threads - coroutines. The control can be switched to another thread only when the currently running thread suspends on yield/await. However, they are not really coroutines here. With the library, it is possible to resume them more than once from the same point. This feature makes coroutines different to continuations.

There are many use cases for having multiple contexts. Some custom scheduling strategies, say, animation jobs have bigger priority, etc. Or some (but not all) running instances may be shared between, say, front-end and back-end.

For running threads contexts are assigned to its [R.contextSymbol] property. It may be a property of the value itself or its prototype - the invoked function prototype property.

There is also a global context used to assign the prototype's property when function's objects are constructed. It can be changed/accessed with R.context(newContext?:Context):Context function.

interface Context {
  running?: Set<Async|AsyncGenerator>
  scheduler: Scheduler

Here reg/unreg callbacks are called when the function's invoked and exited, running field is optional, not used by the library, but maintained by default Context implementation for use in applications.


It is an interface for scheduler field in Context objects. It defines an order of Jobs executing in this context. The default implementation uses JavaScript engines event loop.

interface Scheduler {
  enqueue(job: Job)
  onIdle(job: Job)

interface Job {

The library schedules jobs to be executed using enqueue or onIdle methods. The first one asks scheduler to run the job in an unspecified order and the second only if no other jobs are enqueued on this schedule. The idle queue only needed by R.idle. If it is not used the method may be omitted too.


The library itself doesn't contain any serialization or cloning routine. Any third party can be applied.

As an option there is @effectful/serialization library and there are another library binding these two - @effectful/es-persist-serialization

Accessing execution state

The library may be used without transpiling sources, and it will work the same way ECMAScript effects work for most of the programs. There is a set of constants to access the extras. And they will work only if the sources are transpiled. Their type extends AsyncValue, so to access the value use either await or its [R.awaitSymbol] method.

R.current: AsyncValue<Continuation>

Awaiting this value returns current continuation object to the calling thread. This object may be used to resume the computation from the same await R.current expression, but next times it returns values passed to the continuation resume function, or it throws an exception if reject function is called.

The returned continuation object is a reference. It is not a clone. If some variable is changed or the calling thread suspends on next await/yield the object will be updated too. It is the responsibility of the caller to clone it.

This can be used to fork threads (like POSIX fork function). The forked thread can even survive process restart if stored in DB.

let cont

async function forkable() {
    // ....
    const current = await R.current
    if (current) {
       cont = clone(current)
       // continue the same thread
    } else {
       // child thread

async function startChild() {

The resume can receive any argument here. It will be a result of await R.current in forkable. The test for a child thread in the sample may need a change though.

Cloning and re-running the same continuation several times will execute finally blocks several times as well. This may be not desired behavior if the blocks are used to clean some resources. Your library will need a special combinator dynamic-wind to handle this, and it requires delimited continuations implemented in @effectful/cc.

It works only in transpiled code.

R.idle: AsyncValue<Continuation>

It is the same as R.current but the await expression will be settled only if the current scheduler has nothing enqueued.

Using it instead of R.current makes programs more deterministic. Also no needs to serialize scheduler state. The default scheduler isn't even serializable. It uses JS engine's event loop.

It works only in transpiled code.

R.abort: AsyncValue

Awaiting R.abort stops execution without returning any result. It can be used to cancel further execution if the corresponding continuation is cloned and will be executed after.

It is similar to awaiting never settling Promise but without holding a reference to its local variables. They can be garbage collected. However finally blocks aren't executed. If a program needs cleaning some resources in finally block, better throw some exception and catch it somewhere in top-level.

It works only in transpiled code.

R.managed: AsyncValue<boolean>

This returns true if the currently running thread is transpiled to support the other control combinators (R.current/R.idle/R.abort/R.lock).


R.all(arg: Iterable<AsyncValue>): AsyncValue<any[]>

An alternative for ECMAScript Promise.all. The result is serializable if all elements of arg are.

R.any(arg: Iterable<AsyncValue>): AsyncValue

An alternative for ECMAScript Promise.race. The result is serializable if all elements of arg are.

R.producer(): Producer

Returns an AsyncIterator object with functions send/stop to provide some next value and to finish it resp.

R.lock(): AsyncValue & Continuation

A helper to implement different threads combinators. Returns an object which is both an AsyncValue and a Continuation. Awaiting the object will block the thread until resume or reject method of the same object is called.

It works only in transpiled code.


At the time of writing this, @effectful generators implementation is faster than natives for all mainstream engines. But anyway, this is usually a very small part of the whole program to care about.

Implementing persistence requires replacing local variable references with object's member access. It is an overhead, but modern engines handle it well.


Copyright © 2018 Vitaliy Akimov

Distributed under the terms of the The MIT License (MIT).