Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?


Failed to load latest commit information.
Latest commit message
Commit time
January 6, 2021 14:42
January 6, 2021 14:42
January 6, 2021 14:42
January 6, 2021 14:42
January 6, 2021 14:42
July 18, 2023 16:38
January 18, 2021 16:35
January 30, 2021 18:49
September 18, 2023 08:54

Deferring Module Evaluation

previously known as "Lazy Module Initialization"


Champion(s): Yulia Startsev and Nicolò Ribaudo

Author(s): Yulia Startsev, Nicolò Ribaudo and Guy Bedford

Stage: 2



JS applications can get very large, to the point that not only loading, but even executing their initialization scripts incurs a significant performance cost. Usually, this happens later in an application's life span - often requiring invasive changes to make it more performant.

Loading performance is a big and important area for improvement, and involves preloading techniques for avoiding waterfalls and dynamic import() for lazily loading modules.

But even with loading performance solved using these techniques, there is still overhead for execution performance - CPU bottlenecks during initialization due to the way that the code itself is written.


Avoiding unnecessary execution is a well-known optimization in the Node.js CommonJS module system, where there is a smaller gap between load contention and execution contention. The common pattern in Node.js applications is to refactor code to dynamically require as needed:

const operation = require('operation');

exports.doSomething = function (target) {
  return operation(target);

being rewritten as a performance optimization into:

exports.doSomething = function (target) {
  const operation = require('operation');
  return operation(target);

The consumer still is provided with the same API, but with a more efficient use of FS & CPU during initialization time.

For ES modules, we have a solution for the lazy loading component of this problem via dynamic import().

For the same example we can write:

export async function doSomething (target) {
  const { operation } = await import('operations');
  return operation(target);

This avoids bottlenecking the network and CPU during application initialization, but there are still a number of problems with this technique:

  1. It doesn't actually solve the deferral of execution problem, since sending a network request in such a scenario would usually be a performance regression and not an improvement. A separate network preloading step would therefore still be desirable to achieve efficient deferred execution while avoiding triggering a waterfall of requests.

  2. It forces all functions and their callers into an asynchronous programming model, without necessarily reflecting the real intention of the program. This leads to all call sites having to be updated into a new model, and cannot be made without a breaking API change to existing API consumers.

Problem Statement

Deferring the synchronous evaluation of a module may be desirable new primitive to avoid unnecessary CPU work during application initialization, without requiring any changes from a module API consumer perspective.

Dynamic import does not properly solve this problem, since it must often be coupled with a preload step, and enforces the unnecessary asyncification of all functions, without providing the ability to only defer the synchronous evaluation work.


The proposal is to have a new syntactical import form which will only ever return a namespace exotic object. When used, the module and its dependencies would not be executed, but would be fully loaded to the point of being execution-ready before the module graph is considered loaded.

Only when accessing a property of this module, would the execution operations be performed (if needed).

This way, the module namespace exotic object acts like a proxy to the evaluation of the module, effectively with [[Get]] behavior that triggers synchronous evaluation before returning the defined bindings.

The API will use the below syntax, following the phases model established by the source phase imports proposal:

// or with a custom keyword:
import defer * as yNamespace from "y";


The imports would still participate in deep graph loading so that they are fully populated into the module cache prior to execution, however it the imported module will not be evaluated yet.

When a property of the resulting module namespace object is accessed, if the execution has not already been performed, a new top-level execution would be initiated for that module.

In this way, a deferred module evaluation import acts as a new top-level execution node in the execution graph, just like a dynamic import does, except executing synchronously.

There are possible extensions under consideration, such as deferred re-exports, but they are not included in the current version of the proposal.

Top-level await

Property access on the namespace object of a deferred module must be synchronous, and it's thus impossible to defer evaluation of modules that use top-level await. When a module is imported using the import defer syntax, its asynchronous dependencies together with their own transitive dependencies are eagerly evaluated, and only the synchronous parts of the graph are deferred.

Consider the following example, where a is the top-level entry point:

// a
import "b";
import defer * as c from "c"

setTimeout(() => {
}, 1000);
// b
// c
import "d"
import "f"
export let value = 2;
// d
import "e"
await 0;
// e
// f

Since d uses top-level await, d and its dependencies cannot be deferred:

  • The initial evaluation will execute b, e, d and a.
  • Later, the c.value access will trigger the execution of f and c.

Rough sketch

If we split out the components of Module loading and initialization, we could roughly sketch out the intended semantics:

⚠️ The following example does not take cycles into account

// LazyModuleLoader.js
async function loadModuleAndDependencies(name) {
  const loadedModule = await import.load(`./${name}.js`); // load is async, and needs to be awaited
  const parsedModule = loadedModule.parse();
  await Promise.all(; // load all dependencies
  return parsedModule;

async function executeAsyncSubgraphs(module) {
  if (module.hasTLA) return module.evaluate();
  return Promise.all(;

export default async function lazyModule(object, name) {
  const module = await loadModuleAndDependencies(name);
  await executeAsyncSubgraphs(module);
  Object.defineProperty(object, name, {
    get: function() {
      delete object[name];
      const value = module.evaluateSync();
      Object.defineProperty(object, name, {
        writable: true,
        configurable: true,
        enumerable: true,
      return value;
    configurable: true,
    enumerable: true,

  return object;

// myModule.js
import foo from "./bar";


// module.js
import LazyModule from "./LazyModuleLoader";
await LazyModule(globalThis, "myModule");

function Foo() {
  myModule.doWork() // first use



What happened to the direct lazy bindings?

The initial version of this proposal included direct binding access for deferred evaluation via named exports:

import { feature } from './lib' with { lazyInit: true }

export function doSomething (param) {
  return feature(param);

where the deferred evaluation would only happen on access of the feature binding.

There are a number of complexities to this approach, as it introduces a novel type of execution point in the language, which would need to be worked through.

This approach may still be investigated in various ways within this proposal or an extension of it, but by focusing on the module namespace exotic object approach first, it keeps the semantics simple and in-line with standard JS techniques.

Is there really a benefit to optimizing execution, when surely loading is the bottleneck?

While it is true that loading time is the most dominant factor on the web, it is important to consider that many large applications can block the CPU for of the range of 100ms while initializing the main application graph.

Loading times of the order of multiple seconds often take the focus for performance optimization work, and this is certainly an important problem space, but the problem of freeing up the main event loop during initialization remains a critical one when the network problem is solved, that doesn't currently have any easy solutions today for large applications.

Is there prior art for this in other languages?

The standard libraries of these programming languages includes related functionality:

  • Ruby's autoload, in contrast with require which works in the same way as JS import
  • Clojure import
  • Most LISP environments

Our approach is pretty similar to the Emacs Lisp approach, and it's clear from a manual analysis of billions of Stack Overflow posts that this is the most straightforward to ordinary developers.

Why not support a synchronous evaluation API on ModuleInstance

A synchronous evaluation API on the module expression and compartments ModuleInstance object could offer an API for synchronous evaluation of modules, which could be compatible with this approach of deferred evaluation, but it is only in having a clear syntactical solution for this use case, that it can be supported across dependency boundaries and in bundlers to bring the full benefits of avoiding unnecessary initialization work to the wider JS ecosystem.

What can we do in current JS to approximate this behavior?

The closest we can get is the following:

// moduleWrapper.js
export default function ModuleWrapper(object, name, lambda) {
  Object.defineProperty(object, name, {
    get: function() {
      // Redefine this accessor property as a data property.
      // Delete it first, to rule out "too much recursion" in case object is
      // a proxy whose defineProperty handler might unwittingly trigger this
      // getter again.
      delete object[name];
      const value = lambda.apply(object);
      Object.defineProperty(object, name, {
        writable: true,
        configurable: true,
        enumerable: true,
      return value;
    configurable: true,
    enumerable: true,
  return object;

// module.js
import ModuleWrapper from "./ModuleWrapper";
// any imports would need to be wrapped as well

function MyModule() {
 // ... all of the work of the module

export default ModuleWrapper({}, "MyModule", MyModule);

// parent.js
import wrappedModule from "./module";

function Foo() { // first use

However, this solution doesn't cover deferring the loading of submodules of a lazy graph, and would not acheive the characteristics we are looking for.