Intro
After the third try, we successfully migrated all existing Node.js code from CJS to native ESM, and the CI time of unit testing was significantly reduced.
Pull requests refactor: use ESM test: use native ESM
Before we start, I'd like to demonstrate the status quo ante for a better context. You may have different choices on the repo setting or toolchain, but the core steps and concepts should be the same:
- A TypeScript monorepo that includes both frontend and backend projects / packages.
- Total TypeScript code ~60k LOC (including frontend).
- Use
import
in TypeScript. - Use PNPM for workspace management.
- Use
tsc
to compile Node.js, and Parcel to bundle frontend projects. - Use Jest +
ts-jest
for unit testing. - Use package
module-alias
for internal path aliases.
BTW, our project Logto is an open-source solution for auth.
Why ESM?
When we noticed more and more NPM packages are "ESM-only", and we closed tons of PR because of it (except Parcel / Jest ones):
Disregard the war of ESM v.s. CJS, we found ESM does have several advantages:
No another-language-like code transpilation
Especially for TypeScript: Comparing the original code to the compiled version, ESM is much easier to read, edit, and debug.
Given a simple TypeScript snippet:
import path from 'path';
const replaceFile = (filePath: string, filename: string) =>
path.resolve(path.dirname(filePath), filename);
Results after tsc
:
// CJS
"use strict";
var __importDefault = (this && this.__importDefault) || function (mod) {
return (mod && mod.__esModule) ? mod : { "default": mod };
};
Object.defineProperty(exports, "__esModule", { value: true });
const path_1 = __importDefault(require("path"));
const replaceFile = (filePath, filename) => path_1.default.resolve(path_1.default.dirname(filePath), filename);
// ESM
import path from 'path';
const replaceFile = (filePath, filename) => path.resolve(path.dirname(filePath), filename);
Top-level await
This is one of our favorites. Finally no need for wrapping top-level async expressions into a function and using void
to execute it.
// CJS
(async () => {
await doSomething();
})();
// ESM
await doSomething();
Compatibility
- While ESM can easily load CJS modules, it'll be harder for CJS to load ESM. The primary reason is CJS
require()
is synchronous, while ESMimport
is asynchronous. You must useawait import()
in CJS which is painful w/o top-level await. - Plus, CJS is Node-only, which means a universal package needs to compile another version for browser users. (We know there are transpilers, but, huh)
Stick with the standard
ESM is the module standard of JavaScript. TypeScript also uses ESM syntax by default.
Immutable
This is more like a double-edged sword. ESM can significantly improve module security by its immutable design (see this article), but it also brings a little inconvenience for test mocking.
Dramatic testing time reduction
Yes, we're talking about Jest. While Jest is sticking with CJS and only has experimental ESM support, we've been using ts-jest
for a while, but obviously, it's a challenging task even for an M1 Pro MacBook. The fan spins with a noticeable sound, and the core temperature keeps going up while running unit tests.
After migrating all unit tests to ESM as well, my MacBook became silent and pro again! Here's the CI time comparison (with default machine spec. in GitHub Actions):
Execution time are not that stable in GitHub Actions, on average it shows 3x - 4x faster.
Code migration
For official docs, you may find Modules: ECMAScript modules and ECMAScript Modules in Node.js helpful.
Basic configuration
Let's start with tsconfig.json
. Two key points:
- Set
compilerOptions.moduleResolution
tonodenext
in order to tell TSC to use the "cutting-edge" Node.js module resolution strategy. - Set
compilerOptions.module
toesnext
to ensure the output also keeps ESM.
For Node.js, add "type": "module"
to your package.json
to let it treats the package as ESM.
Path aliases
We were mapping @/
to ./src/
using module-alias
alias but it doesn't work in ESM. The good news is Node.js provides a native support called Subpath imports by defining the imports
field in package.json
:
{
"imports": {
"#src/*": "./build/*" // Point to the build directory, not source
}
}
Note imports
can only starts with #
and must have a name. So we use #src/
to replace the original @/
. Also update your tsconfig.json
accordingly:
{
"compilerOptions": {
"paths": {
"#src/*": ["src/*"]
}
}
}
We also need to replace all @/
with #src/
in source code.
File extension
At this point, both TSC and Node.js start to work in ESM mode, but most likely some "Cannot find module..." errors float. Because the existing convention of importing is minimalistic and elegant:
import x from './foo';
// It can be
import x from './foo.js';
// Or
import x from './foo/index.js';
And the .js
extension can be replaced with .jsx
, .ts
, and .tsx
, etc.
However, this becomes unacceptable in native ESM. You must explicitly write the full path with the extension, e.g. import x from './foo/index.js';
.
So how it should be in TypeScript? Our first idea is changing the file extension to .ts
, which turns the path to './foo/index.ts'
, since that's the file we can find in the source directory, right?
Unfortunately, the TypeScript team has the principles like "TS is the superset of JS" and "TS doesn't rewrite paths". (You can see #13422 #16577 #42151 since 2017) So .ts
doesn't work here, and it led to the result: use .js
. :-)
It doe works, and I think I'm not qualified to judge the solution. So let's move to the actions we took to add extensions:
Since most packages in
node_modules
are not affected by this (at least for the main entry), we can omit them during the process.
- Replace all
from '\.'
(RegExp) withfrom './index'
. - Replace all
from '\./(.*)'
(RegExp) withfrom './$1.js'
. - If you have path alias, use the similar technique in step 2 to add extensions to them.
- Try to compile the project. It may show some errors for those paths with omitted
/index
, e.g. a./foo
which actually points to./foo/index.js
. They are updated to wrong paths like./foo.js
in step 2. - Try to compile again and it should show no error this time.
Misc.
It's exciting to see the output files are almost the same as the input, but when you run the project, Node.js may complain about some special variables, for example:
ReferenceError: __dirname is not defined in ES module scope
Don't give up, we're almost there! Read the Node.js official doc Differences between ES modules and CommonJS to handle them, then you're good to go.
Recap
We successfully migrated our Node.js packages from CJS to ESM. Now the entry-point file is runnable after tsc
.
However, some issues still remain in unit tests:
- As of today (12/26/22), Jest only has experimental support for ESM.
- ESM is immutable, thus
jest.mock()
will not work andjest.spyOn()
also doesn't work on first-level variables (export const ...
). This also applies to other test libraries like Sinon. - You may find some libraries for mocking ESM, but almost all of them are creating "a new copy" of the original module, which means if you want to import module A that depends on module B, you must import A AFTER B is mocked to get it to work.
The solution is the key to boosting the CI time. Since this article already pulled out a lot of things to digest, we'll cover them in the next chapter:
Migrate a 60k LOC TypeScript (NodeJS) repo to ESM and testing become 4x faster (2/2)
You can find our ts-with-node-esm
repo for the key result of this series:
https://github.com/logto-io/ts-with-node-esm
Thank you for reading, feel free to comment if you have any questions!
This series is based on our experience with Logto, an open-source solution for auth.
Top comments (13)
Yes, I am on board with this too with wj-config. I transpiled it from TypeScript to ES Modules. That's it. Look at some code snippet comparison when used in Common JS and ES Modules:
Set Up
Common JS
ES Modules
Consumption
Common JS
ES Modules
Just having top-level await is enough to sway like 90% of people to ES Modules.
We are going through a similar experiment right now, but for migrating to ESM from RequireJS. Ironically, it seems to be much easier, largely because there are nearly zero adaptations/hotfixes required for front-end compatibility, where all the transpilation was destined for anyway. If anything, the weird edge cases come from situations (like automated unit testing) where some degree of NodeJS compatibility is still required.
It was a problem with tooling, not with CJS. Migration of tests from requires to imports shouldn't cause such drastic speed up, it's definitely a bug in the
ts-jest
or in a related tooling. I'm using @swc/jest for tests and it takes couple of seconds to run a thousand tests written in TS.The point of this article is that ESM brings a compilation speedup, so what was the change of the build step of your project?
Thank you! Totally agree there's something that happened in
ts-jest
orts-node
magnificently slows down the testing speed. Will try@swc/jest
.The faster testing speed is something out of our expectations. The point of this article is to discuss why we decided to transform to native ESM and our know-how of the migration. No big changes for the build steps.
I read more carefully and I see that main reason was a non-compatible packages that enforces ESM to be enabled. That's sad to write extensions in import pathes, have experimental Jest, to loose
jest.mock()
which I'm using regularly, and all because some libs cannot setup rollup properly to have both commonjs and esm outputs. In your project it's probably important to support all packages, but luckily in other node.js projects it's fine to use an alternative lib.Non-compatible packages is a trigger, not the key reason. We kept those PR closed for months.
Extensions and experimental flag are not blockers to us. We don't lose `jest.mock(), as you can see the practice in chapter 2. Actually the code becomes more readable and predictable afterwards.
And it's all about choices and balances - since in most cases our project serves as a service/solution which is not for importing, and our users are using container to run it, so single transpilation strategy is good to us.
We are using rollup for dual-package publish as well, but it's a transitive plan and soon the repo will be pure-ESM.
And now time to swap to Vitest!
Switching to Vitest fixed many issues that we've had with Jest. Like React hooks didn't work as expected within components when running within Jest which had led to cumbersome hook mocking. Especially useContext was very troublesome. I don't know if this is issue always with Jest but the issue did replicate on two different repos that we have.
Vitest is also faster.
Thanks guys. Will take a look on Vitest!
We have a code base that I have tried to move to ESM but the problem that I have run into is how to mock imports in a unit test. We use Jasmine and Karma for our specs and if I am testing a module that imports another module (e.g. utils.js), I am unable to stub or mock the exports of utils.js. How are you handling this? You mentioned the issue in the Recap, but I'm wondering if anyone has a solution. Thanks
Mocking is explained in chapter 2. The principle will be the same.
Thanks!
Thank you for the post, this has been really helpful.
But I have a question about the subpath imports. You have the imports defined as:
Why are you pointing to the build dir?
In your github example you don't have any subpath imports defined. So what is the thought process behind the subpath import pointing to build and further, does it also imply that you need to be running a watcher on the code to continuously update the build output?
Thanks