Comparing CSS in JS Libraries

A while back, I came across two awesome articles that introduced CSS in JS or the technique of styling your JS components using JS alone: 2017 frameworks and CSS in JS Battle.

Since that time, the popularity of CSS in JS has only grown! It's not hard to see why. With the advent of modern component-driven frontend frameworks, major improvements to stylesheet-driven styling (whether .scss, .css, or another styling language) have been to improve the modularity, reduce I/O, ease deduplication of classnames/styles, improve readability, and/or otherwise ease UI development:

I want to briefly introduce CSS in JS, why it matters, and then provide an update comparison using the latest specs.

In-Style Inline Styles

The advent of component-driven client architectures has shown the limitations of existing CSS archetypes (inline and external stylesheet).

    <style>
        //I'm so in
    </style>

Inline styling is advantageous since it allows all of our styling to be included in one file (a view or partial view of one kind or another) but becomes unreadable and cumbersome in larger apps. While inline styling reduces I/O by condensing the actual number of files, it's harder to manually deduplicate unnecessarily repeated styles or classes decreasing code reuse and magnifying the actual size of the styling.

In component-driven design, we might run into styling conflict scenarios where, due to complex component layout, the load order or dependency priority can cause some styles to override others in ways we did not want nor anticipate.

Moral of the Story: Can we put styles into the same file as our markup?

External Styling

External styling was originally intended to solve many of the issues raised above. Namely, external styling improves code reuse by allowing the same classes and styles to be efficiently reused throughout an application without multiple style references being used (avoiding duplication).

    <link rel="stylesheet" href="external.sheet.css" type="text/css">

Unfortunately, external styling likewise becomes cumbersome in a component-driven client architecture. Full .css files were originally intended to provided styling for an entire page of a multi-page application. With the advent of both the component-based along with single-page design patterns, such intentions (while well-intentioned) have become somewhat dated.

For example, importing .scss files in a single-page React app makes all of the imported styling global - e.g., the imported stylesheets are not scoped to the component but to the entire page. This can cause styling conflict scenarios and can become unmanageable in large applications.

Moral of the Story: Can we get both deduplication AND scoped styling?

Programmatic Styling to the Rescue?

    var el = document.getElementById('hey');
    el.style.height = 10000000%;

A third approach, programmatic-styling, has become more popular given the gradual enrichment of the native JavaScript Web and DOM API's. However, many common but advanced styling techniques are either not available or require complicated scripting to implement well. Furthermore, programmatic-styling moves styling computations out of the CSS interpreter into the runtime JavaScript interpreter which can massively reduce performance.

Moral of the Story: Maybe there's something to this JavaScript...

I CSS Your JS

The CSS in JS paradigm solves these problems by scoping styling to a specific JavaScript component using JavaScript but then injects said styles into the rendered and serve view!

Why CSS In My JS?

What all that means, in theory, is that we can get every awesome benefit of the preceding styling paradigms in a highly-performant way.

The aim of this article is to see what that theory amounts to in practice and how well different attempts to implement such approaches compare to each other!

Performance Testing

Let's set up a few tests for the most popular CSS in JS frameworks!

Aphrodite
CXS
Emotion
Glamorous
JSS
Radium
Styled-Components
Styletron

Specifically, we'll be adding these dependencies and versions into our test suite:

    "aphrodite": "=2.1.1",
    "cxs": "=6.2.0",
    "emotion": "=9.1.1",
    "glamor": "=2.20.40",
    "glamorous": "=4.12.3",
    "jss": "=9.8.1",
    "radium": "=0.24.0",
    "styled-components": "=3.2.5",
    "styletron": "=3.0.4"

If you know of any others, send me an email at adam.gerard@x-team.com, and I'll happily add them to the test suite :)

Performance Tests

We'll approach testing from five vantage points:

1. Phantom.js page load performance
2. Phantom.js I/0 performance
3. live-server 'http' request performance - time to complete 1000 requests
4. Webpack 3.6 compile time
5. Webpack 3.6 bundle size

These metrics will give us a better grasp of overall client-server interperformance. We'll need two dependencies that we can add to our package.json:

    "live-server": "^1.2.0",
    "request": "=2.85.0",
    "phantom": "=4.0.12"

live-server is, in my opinion, one of the best test-server libraries since it uses a simple configuration object:

const liveServer = require('live-server')

const opts = {
    host: '127.0.0.1',
    port: port,
    root: `./public/${subroot}`,
    file: 'index.html',
    wait: 0,
    open: true,
    logLevel: 1
}

liveServer.start(opts)

live-server supports hot-reloading, varying log-levels, manually-set ports and hostnames, etc. For our present purposes, it's one of the easiest libraries to use among many alternatives like Webdriver.

Since it supports easy startup and shutdown, we can partly simplify some of the architecture for our tests. We can also use the native Node http request object to realistically capture performance, load, and use. Specifically, we'll leverage live-server to be the workhorse of our test-suite.

Phantom.js is one of the best overall testing frameworks out there. We'll use it to primarily test page load and page rendering speed. For our testing purposes, we'll use the Node wrapper for Phantom.js (npm i phantom).

Project Architecture

In order to really dig in, we'll need to replicate a development and a production setup for each of the frameworks.

We will thus have eight (8) entirely separate React apps supported by distinct Webpack configuration files. Woot!

Our Webpack package.json dependencies will look like:

    "autoprefixer": "=7.2.2",
    "babel-core": "=6.26.0",
    "babel-loader": "=7.1.4",
    "babel-preset-es2015": "=6.24.1",
    "babel-preset-react": "=6.24.1",
    "babel-preset-stage-0": "=6.24.1",
    "babel-plugin-transform-decorators-legacy": "^1.3.4",
    "css-loader": "=0.28.11",
    "extract-text-webpack-plugin": "=3.0.2",
    "file-loader": "=1.1.5",
    "image-webpack-loader": "=3.4.2",
    "node-sass": "=4.7.2",
    "postcss-loader": "=2.0.9",
    "postcss-smart-import": "=0.7.6",
    "sass-loader": "=6.0.6",
    "style-loader": "=0.19.0",
    "webpack": "=3.6.0"

Bundle

We want our compiled React apps to be as production-esque as we can make them - thus, we will include in our package.json (as dependencies and not development dependencies):

    "react": "=16.3.1",
    "react-dom": "=16.3.1",
    "react-redux": "=5.0.7",
    "react-router-dom": "=4.2.2",
    "redux": "=3.7.2"

The relevant part of our webpack.config.js files is:

 //...
 rules: [
      {
        test: /\.js$/,
        exclude: /node_modules/,
        loader: 'babel-loader'
      },
      {
        test: /\.jsx$/,
        exclude: /node_modules/,
        loader: 'babel-loader'
      },
      //...
 ]

We'll also need to add the following .babelrc to our a project root:

{
  "presets": ["react", "es2015", "stage-0"],
  "plugins": ["transform-decorators-legacy"]
}

I've included both, a development and a production version here to make life a little easier and because, hey, I <3 you.

    $ npm run build-all

React Baseline

Our React baseline projects will include a Router, a Redux, and a CSS in JS component since, at the minimum, most React apps will have as much:

import React from 'react'
import { BrowserRouter, Route, Switch } from 'react-router-dom'
import CssInJsComponent from './CssInJsComponent'

export default () =>
  <BrowserRouter> <div> <nav> <ul> <li> <Link to={'/'}>Home</Link> </li> </ul> </nav> <Switch> <Route exact path={'/'} component={CssInJssComponent}/> </Switch> </div> </BrowserRouter>

import { dummy, DUMMY_ACTION } from './Actions'

export const DummyStorage = (state = {}, action) => {
  switch (action['cmd']) {
    case DUMMY_ACTION:
      state[action['key']] = action['value']
      return state
    default:
      return state
  }
}

import React from 'react'

export default () => <div>Hallo Planetoid!</div>

The variable we'll be altering in each test case is our CssInJsComponent. In certain cases (such as our Radium and Styletron test-cases), there will be some fairly extensive alterations to the baseline components specified above!

In each, we'll be testing the equivalent of:

div {
    background-color: black,
    color: white,
    opacity: .9
}

Hardware

The increasing popularity of Dell's XPS laptops and the advent of Microsoft's Linux Subsystem on Windows 10 (which I just set up) gives developers a lot of flexibility when choosing hardware and budgeting.

This particular test suite was performed on a Dell 15 XPS from 2016 equipped with:

 Intel Core i7-6700HQ Quad-Core (8 Logical Cores) at 2.60 GHz 32 GB RAM Windows 10 Pro + Linux Subsystem and Cygwin NVIDIA 960M GPU 256GB m2 PCIe SSD 15.6" FHD Screen 

I'm also a big fan of Apple products (I use all the gadgets when I can) since they are required to use Swift and develop for iOS. The comparable version from Apple would be one of the new MacBook Pro 15's!

As of April 2018, the underlying hardware between the two makers is derived primarily from Intel, NVIDIA, or AMD (Radeon) so, despite some lessening differences between operating systems, laptop models are roughly the same under-the-hood. As a result, your findings should mirror the results of the article despite the particular model or make of the testing equipment.

Results

Let's go over the results!

Build Time and Bundle Size!

To run the tests yourself:

    $ npm run build-all

To toggle between development and production modes alter:

module.exports = {
  production: false
}

Unminified (Development):

Testing build time for development-like scenarios (time to rebuild/recompile on change):

Build Time Winner: Aphrodite (4879.66 avg ms)
Vendor Dependencies Size Winner: CXS (858 KB)
App Size Winner: CXS (3.94 KB)
Overall Size Winner: CXS (861.94 KB)

Minified (Production):

Testing build time and size for production-like scenarios:

Build Time Winner: CXS (8366.66 avg ms)
Vendor Dependencies Size Winner: CXS (165 KB)
App Size Winner: Glamorous (1.72 KB)
Overall Size Winner: CXS (166.74 KB)

live-server - Time To Complete 1000 Requests

In this test, we see how well the different frameworks hold-up against each other in high-use scenarios. We want to see how much throughput our server can handle given the choice of a particular CSS in JS framework (Please note that this is a test of server-side performance rather than client page-loading performance):

    $ npm run number-request

The Winner: Styled-Components (2491.33 avg ms)

Phantom.js - Page Load Performance

In this test, we'll be testing the time to execute a request and receive a response back from our test server. We want to see how fast we can request, render, and receive a production page when equipped with a particular CSS in JS framework:

    $ npm run phantom-page

The Winner: Radium (2936.33 avg ms)

Phantom.js - I/O Performance

To further quantify overall I/O performance, we'll also test the time to load each supplied index.html with all production assets this time using Phantom.js and combining the results with our previous build size:

    $ npm run phantom-io

The Winner: Glamorous (77.7719 avg KB/S)

Conclusion

Let's wrap this up!

Highlights

CXS is an overall strong contender taking almost all of the development build first-place build spots and coming in third for the overall time at a respectable build time average of 5124.66 ms (versus Aphrodite's 4879.66 avg ms).

CXS also places first for almost every production build spot. In systems where multiple builds are required daily, CXS is a great choice. CXS is also very fast in terms of clientside page loading.

From a server-side standpoint, Glamorous represents a strong choice - it comes first in our KB/S test (testing clientside load time against buildsize), third in our 1000 request test, and fourth in raw clientside page load time.

Radium reveals a similar profile coming in first for raw clientside page load performance and running middle of the pack for the other test cases.

Aphrodite underperformed and came in last or near last in all three non-build-related test scenarios.

Comparison to Previous Tests

HelloFresh is a great company both because it's tasty and for experimenting with CSS in JS tests previously. At that time, HelloFresh concluded there was no clear winner.

Like before, CXS remains a strong build and clientside page loading library capable of efficiently injecting styles at render.

This time, we've also tested for raw serverside performance and the metrics we've obtained for Glamorous and Radium mirrors their continued rise in popularity.

Check out the code and tests used in this article over on GitHub!