Deferring Module Evaluation

previously known as "Lazy Module Initialization"

Status

Champion(s): Nicolò Ribaudo

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

Stage: 2.7

Slides:

• 2021-01 - Stage 1 (notes)
• 2022-11 - Take two: Defer Module Evaluation (notes)
• 2023-07 - Deferred import evaluation for Stage 2 (notes)
• 2024-04 - Deferred import evaluation for Stage 2.7
• 2024-06 - Deferred import evaluation for Stage 2.7 (2)

Background

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.

Motivation

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.

Proposal

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";

Semantics

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(() => { c.value }, 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(parsedModule.imports.map(loadModuleAndDependencies)); // load all dependencies
  return parsedModule;
}

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

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, {
        value,
        writable: true,
        configurable: true,
        enumerable: true,
      });
      return value;
    },
    configurable: true,
    enumerable: true,
  });

  return object;
}

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

etc.

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

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

Implementations

• engine262: https://github.com/nicolo-ribaudo/engine262/tree/defer-eval

Q&A

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, {
        value,
        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() {
  wrappedModule.MyModule.bar() // 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.

Why import defer * gives a different namespace object from import *?

If a deferred module throws while being evaluated, ns.foo will throw the evaluation error:

// module-that-throws1
export let a = 1;
throw new Error("oops");

// main1.js
import defer * as ns1 from 'module-that-throws';
try { ns1.a } catch (e) { console.log('caught', e) } // logs "oops"

Module namespace objects of modules that are already evaluated and threw during evaluation do now re-throw an error on property access:

// module-that-throws2
import * as ns2 from 'module-that-throws';
globalThis.ns2 = ns2;
export let a = 1;
throw new Error("oops");

// main2.js
import("module-that-throws").finally(() => {
  try { ns2.a } catch (e) { console.log('caught', e) } // Doesn't throw
});

This is not a problem today, because having access to the namespace object of a module that threw during evaluation is incredibly rare. However, it would be incredibly more common with import defer declarations. To guarantee that the behavior of main1.js is not affected by module previously loaded (and to avoid race conditions), ns2.foo must throw even if module-that-throws is already evaluated, and thus it cannot be the same namespace object as import *.

Another approach we considered (and discarded) was to always suppress evaluation errors on namespace property access, so that in the ecample above ns1.a would be guaranteed to never throw and thus not be affected by unrelated modules that might have already triggered evaluation of module-that-throws.

Why not re-use import attributes (import * as ns from "mod" with { defer: true })?

There are two reasons why we chose to use an "import modifier" rather than an attribute:

1. Import attributes affect what a module is, but cannot change basic semantics of how ECMAScript modules behave: they are similar to adding query parameters to the imported URL, except that attributes are handled by the running environment rather than by the server. For example, with { type: "json" } behaves as if the imported module was a JavaScript file wrapped in export default JSON.parse(` ... the file contents ... `);. import defer changes how namespace objects behave (by making them side-effectul, while before this proposal property access on namespace objects couldn't trigger any side effect): it cannot be expressed as a wrapped/modified "classic" ECMAScript module.
2. Together with the source phase imports proposal, we are exposing multiple "phases" of module loading. The phases we've identified are: resolving a module given a specifier, fetching the module (these two both happens in hosts and not in ECMA-262), attaching modules to their execution and resolution context, linking modules together, and finally executing them. We are using import modifiers to represent modules processed up to one of those phases, without going all the way to finishing execution. These modifiers give more guarantees than import attributes: while import "x" with { attr1: "val" } and import "x" with { attr2: "val2" } might be two completely different modules, import source s from "x", import defer * as ns from "x", and import "x" all are guaranteed to load the same module, and that module will be executed at most once regardless of which "phase" it gets temporarely paused at (and then continued from).