Josh Goldberg

Switching a Jest Project from Babel to SWC

Feb 21, 202215 minute read

Speeding up and simplifying the transpiler configuration for a Jest project with Speedy Web Compiler (SWC).

Thereā€™s been a lot of buzz on frontend dev Twitter over the last year about the next wave of JavaScript tooling. Tools such as Speedy Web Compiler (SWC) use lower-level languages such as Rust to transpile JavaScript much more quickly than older tools such as Babel. Iā€™ve been wanting to try them out for a while, so this past weekend I migrated the tslint-to-eslint-config project from babel-jest to @swc/jest.

Spoilers: if you just want to see code, check the final pull requestā€™s file changes.

Existing Babel Config

tslint-to-eslint-configā€™s existing testing configuration was pretty typical of many modern TypeScript Node apps. It used the babel-jest package for Jest to transpile source files. The repositoryā€™s babel.config.js looked like:

module.exports = {
	presets: [
		["@babel/preset-env", { targets: { node: "current" } }],
		"@babel/preset-typescript",
	],
};

I donā€™t use ts-jest because Iā€™ve found it simpler to use Jestā€™s Babel support instead. I also donā€™t want tests to be slowed down by type checking, though we should note ts-jest can disable type checking.

Moving to SWC

SWC boasts significant performance improvements over Babel for being written in Rust rather than JavaScript. SWCā€™s docs on usage with Jest indicate there are two quick steps to onboard to SWC:

  1. Install @swc/jest as a dev dependency along with jest

    npm i -D jest @swc/core @swc/jest
  2. Add a transform line to your jest.config.js file indicating to use @swc/jest:

    module.exports = {
    	// ...
    	transform: {
    		"^.+\\.(t|j)sx?$": ["@swc/jest"],
    	},
    };

    (I omitted the x? because my project is a Node app that doesnā€™t use .jsx/.tsx files)

Those steps are enough to direct Jest to use SWC instead of Babel to transpile files. Some tests were already passing, but any test that included async/await needed a bit more work.

Runtime Generators and Regenerator Runtime

My first test run saw many tests fail with Cannot find module 'regenerator-runtime':

 FAIL  src/cli/runCli.test.ts
  ā— Test suite failed to run

    Cannot find module 'regenerator-runtime' from 'src/cli/runCli.test.ts'

      1 | import { EOL } from "os";
    > 2 |
        | ^

regenerator-runtime is the runtime package injected by Regenerator, the source code transform used by many transpilersā€™ transforms for generator functions. Fun fact: async/await in JavaScript is built on generators, so if your app has any async or await keywords, the default @swc/jest compiler settings add a runtime regenerator-runtime dependency.

I found two good ways to fix the tests crashing:

Generators and async/await have been fully supported in Node for years, including all active and maintenance LTS versions of Node. Thereā€™s really no good reason to feel a need to keep transpiling away generators for apps made to run in Node.

I changed the SWC configuration in jest.config.js to specify a target of ES2021:

module.exports = {
	// ...
	transform: {
		"^.+\\.(t|j)s$": [
			"@swc/jest",
			{
				jsc: {
					target: "es2021",
				},
			},
		],
	},
};

ā€¦and filed Consider defaulting jsc target to the current Node versionā€™s supported level as an issue on the swc-project/jest GitHub repo.

Edit 2/24/2022: Versions 0.2.18 of @swc/jest will auto-detect the highest supported jsc.target on your system now. Hooray!

At this point, all my tests passed (yay!), but the code coverage calculation was slightly lower than expected. šŸ˜•

Code Coverage Differences

SWC has an open issue on incorrect Jest coverage. It was much worse in my project before setting jsc.target to "es2021". After that change there was only one file with less coverage than before: src/index.ts, which exclusively contains export statements for the packageā€™s Node API.

export { convertFileCommentsStandalone as convertFileComments } from "./api/convertFileCommentsStandalone";
export { convertTSLintConfigStandalone as convertTSLintConfig } from "./api/convertTSLintConfigStandalone";
// ...

src/index.ts had gone from 0/0 = 100% (code coverage math is funny sometimes) to marking every one of its lines as uncovered. It isnā€™t included in any unit tests so that change felt like an improvement in reporting accuracy.

Solution: I added "!./src/index.ts" to the collectCoverageFrom array in jest.config.js.

At this point, all tests were passing and code coverage was the same shiny 100%-except-for-excluded-files that it was before! Hooray! šŸŽ‰

Performance Comparison

One of the most important rules of making decisions based on performance is to always measure results scientifically. Itā€™s not enough to read a docs site that claims blazing fast speed. You need to measure how your specific scenario matches up with and without a change, in as similar an environment as possible.

I ran the same test commands in both the Babel and SWC app builds with the following methodology:

CommandBabel TimeSWC Time(old - new) / old x 100%
jest18-20 seconds11-14 seconds30-40% faster
jest --coverage --maxWorkers=223-28 seconds20-25 seconds10-13% faster

A true scientific experiment should calculate confidence intervals or similar data science metrics. I would encourage you to repeat this experiment on your own systems.

Look at that speed improvement! Itā€™s much more pronounced in the faster dev-time version, but even a 10% improvement in test time is a success if you ask me.

One interesting note is that even with the significantly faster SWC transpiler, the standalone jest command still took more than 10 seconds for its remaining 60-70% time. Much of that time is spent by Jest setting up and tearing down Node environments (which it does for each test file to keep tests isolated) and running tests.

Dependencies Size

Runtime performance is not the only important metric for application development. The size a packageā€™s dependencies take up is important too. The larger and more numerous a projectā€™s dependencies, the longer it will take to install and more disk space its node_modules will take up.

Yarn v2+ and other fancy package/project managers can partially mitigate size issues by sharing dependencies across packages locally but donā€™t yet improve CI caching as much.

Iā€™d been optimistic about the node_modules folder size changes from switching to SWC because the project package-lock.json had shrunk by about 2,000 lines. Running du -sh node_modules to check the size of my package dependencies crushed my heart:

Babel SizeSWC Size(old - new) / old x 100%
186M330M77% larger

šŸ˜¢.

The 30-40% local test running time improvement is still worth an extra 144M for me, but thatā€™s still a little painful.

Digging Into Dependencies

Looking a little deeper into where SWCā€™s size comes from:

du -sh node_modules/@swc/*
444K    node_modules/@swc/core
74M     node_modules/@swc/core-linux-x64-gnu
78M     node_modules/@swc/core-linux-x64-musl
24K     node_modules/@swc/jest

Interestingly, running rm -rf node_modules/@swc/core-linux-x64-musl didnā€™t prevent subsequent tests from passing. 78M of 152.5MB -more than half!- of the @swc/core package install is totally unused on my system.

I filed Both gnu and musl core distributions installed for Ubuntu Linux as an issue on the SWC project about the duplicate architecture package.

As of February 2022, an npm RFC for package distributions is in review now that would allow for solving this issue.

Furthermore, a bunch of Babel package still exist in my projectā€™s package-lock.json because Jest packages have them as dependencies. Running rm -rf node_modules/*babel* reduced node_modules by another 6M. A little improvement, but still a nice one.

In Conclusion

SWC comes with significant transpiler performance improvements over Babel. Those runtime performance improvements come at a cost of a much larger dependency size if youā€™re not already using SWC. Still, the improved performance makes the Babel-to-SWC migration very much worth it for at least my Jest projects. āš”