macOS Web: Why I moved from React to Svelte
PuruVJ
Posted on August 29, 2021
macOS is an amazing operating system. It looks really good, works really fast(Thanks to Apple's hardware), is very consistent, and overall an amazing experience.
Personally, I have never been much infatuated with macOS. But that changed last year when macOS Big Sur came out with the super sleek and modern design with those beautiful blurs. I instantly fell in love with it π Just look how beautiful it is π
Unfortunately, I couldn't afford to buy the MacBook(You're overpriced, Apple π), but I couldn't not want it. I spent weeks just watching review videos and screenshots of it. I am not a designer, but I have a good taste in design, and Big Sur was it! Finally, when the craving became too much, I was like, "Screw it! I'll just make it myself"
So that's what I did π I recreated the interface of macOS Big Sur for Web. You can see it @ macos.vercel.app.
Early days (React)!
I began working on this project back in November 2020, with the stack being React with Snowpack as the bundler+dev server, JSS for styling, and restater for state management, and I created just the topbar(without the opening menus), the animated dock, and the background. Nothing much, really. Quite easy to do. You can visit it here macos-web-mmo5af3r6.vercel.app
The dock animation isn't as smooth as it is in the latest iteration if you look closely. Not to mention, this very simple app is using loads of dependencies. It loaded 146KB min+brotli of JS, and 0KB of CSS(It was CSS-in-JS, so no separate CSS file). For reference, the current version in Svelte loads only 28.5KB of JS, and 3.6KB of CSS, and it does a lot more than the November version.
Here's the package.json
dependencies
field of the November version π
{
"@material-ui/core": "^4.11.0",
"@mdi/js": "^5.8.55",
"@rooks/use-raf": "^4.0.2",
"@tippyjs/react": "^4.2.0",
"date-fns": "^2.16.1",
"framer-motion": "^2.9.4",
"react": "^17.0.1",
"react-dom": "^17.0.1",
"react-helmet-async": "^1.0.7",
"restater": "0.0.12",
"use-sound": "^2.0.1"
}
See?! So much stuff for so little!
So this was the state of the app. Then I didn't work on it. I was bored of it. Then I had an accident, broke my spine, was prescribed 3 months bed rest. I was so bored out of my mind, I thought, what the heck, I'll just continue working on it.
So I revived the project, made some heavy changes. Like I threw away restater
library to use Jotai, because it was much simpler.
Fast forward two months, I made some drastic changes. I replaced Snowpack(Bundler and dev server) with ViteJS. Why? No specific reason, I just felt it was more polished and had some superb features which I could use to make the app smaller in size.
I replaced the existing JSS style with styled-components
. Not perf-related really, just that writing CSS as JS objects wasn't fun.
I also got rid of some dependencies to reduce bundle size.
That brings us to 120KB min+brotli JS. You can see it for yourself: macos-web-git-feat-switch-to-vite-puruvj.vercel.app
I was still not satisfied with it. The app felt slow, the dock animation was janky, menus wouldn't open up instantly. Not satisfactory!
Then, I did something very drastic: I moved the app from React to Preact.
The era of Preact
Yepp!! I did that. I'm crazy that way π
Anyways, moving React to Preact was a very smooth experience. I just had to install the official @preactjs/preset-vite
Vite plugin by Preact team, and it took care of aliasing react
in import React from 'react'
to preact/compat
and everything. It was very seamless. It just worked. Huge kudos to the Preact team!!
And that, my friend, made all the difference in the world.
The site got fast. Animations lost their jank. Menus opened instantly. Everything just worked!.
And the best part: Bundle size went down by a lot!!. React + React DOM is 42KB. Preact Core + Compat is just 4KB, max!
You can check it out here yourself macos-web-git-feat-move-to-preact-puruvj.vercel.app.
You will find the bundle size is not drastically less here compared to the previous React iteration. That's because I added some dependencies in the period before transitioning.
This was great, but there was one problem: Too many dependencies.
Here's the package.json
dependencies
field π
{
"@mdi/js": "^5.9.55",
"@rooks/use-raf": "^4.9.2",
"@tippyjs/react": "^4.2.3",
"color2k": "^1.2.4",
"date-fns": "^2.18.0",
"framer-motion": "^3.10.0",
"idb-keyval": "^5.0.4",
"immer": "^8.0.1",
"jotai": "^0.15.0",
"preact": "^10.5.12",
"prop-types": "^15.7.2",
"react-helmet-async": "^1.0.9",
"react-is": "^17.0.1",
"react-lazily": "^0.9.0",
"react-rnd": "^10.2.4",
"react-slider": "^1.1.4",
"react-sound": "^1.2.0",
"styled-components": "^5.2.1",
"styled-reset": "^4.3.4",
"throttle-debounce": "^3.0.1",
"tippy.js": "^6.3.1",
"use-interval": "^1.3.0"
}
This is a lot of dependencies at this point. Doesn't look good, does it? π€
Goodbye styled-components π³οΈ
So in my quest to make the app smaller and faster, I found styled-components
to be one of the roadblocks.
What was the issue with it? Well, first thing, CSS-in-JS is always less performant than just raw CSS(Well written CSS. Bad, duplicated styled don't count πΌ). 2nd, styled-components
added 12KB bundle size to the app.
And the most important one: The output CSS wouldn't minify.
Generally, when you are writing in .css
or .scss
files, your CSS is minified completely, that is all spaces and line breaks are removed, making it smaller to send over the wire.
But with styled-components, because I'm writing the CSS inside JavaScript strings, the CSS tooling can't minify it by default, so that CSS remains unminified.
See? All the code is here as it is.
And I couldn't use the babel-plugin-styled-components, as it wouldn't work with Vite, for some reason. And leaving Vite wasn't an option.
So it was then that I decided to move to the basics, CSS Modules (SCSS Modules, to be exact π)
So I did that. Took a lot of effort, but the final result was worth it. It took the JS size from around 110KB to 85KB, and the CSS was 3.3KB. This was a huge amount of reduction and made the app feel really fast. The burden of supplying CSS was removed from the JS, and that alone made a lot of difference.
Here's an in-detail write-up of moving to CSS Modules: Why I moved from Styled Components to (S)CSS modules
And, because this process required me to touch the whole project at once, I got the chance to revisit some not-so-good code and improve it, resulting in me getting rid of a lot of dependencies. This is the package.json's dependencies
field after this change π
{
"@mdi/js": "^5.9.55",
"@reach/slider": "^0.14.0",
"@rooks/use-raf": "^4.10.0",
"@tippyjs/react": "^4.2.5",
"clsx": "^1.1.1",
"date-fns": "^2.19.0",
"framer-motion": "^4.0.3",
"idb-keyval": "^5.0.4",
"immer": "^9.0.1",
"jotai": "^0.16.0",
"preact": "^10.5.13",
"react-rnd": "^10.2.4",
"tippy.js": "^6.3.1",
"use-interval": "^1.3.0"
}
That's a lot of reduction compared to the previous iteration.
So all this brings us to the final situation of macOS: Built in preact, styled with CSS Modules, only 85KB loaded at first, and pretty smooth usage.
But me being me, I strived to make it even better. Because there wasn't much scope for radical techniques, I resorted to the ultimate technique: Code Splitting
Go Code splitting!
Implementing Code splitting was very easy. The components that aren't needed on the first load, I split them away into their own separate chunks.
One component where I do code splitting was the AppNexus.tsx
. This is basically a gateway to render the right app component when you open any app. This is as easy as π
import { AppID } from '__/stores/apps.store';
import { lazy } from 'react';
type AppNexusProps = {
appID: AppID,
isBeingDragged: boolean,
};
const Calculator = lazy(() => import('./Calculator/Calculator'));
const VSCode = lazy(() => import('./VSCode/VSCode'));
const Calendar = lazy(() => import('./Calendar/Calendar'));
const PlaceholderApp = lazy(() => import('./Placeholder/Placeholder'));
export const AppNexus = ({ appID, isBeingDragged }: AppNexusProps) => {
if (appID === 'calculator') return <Calculator />;
if (appID === 'vscode') return <VSCode isBeingDragged={isBeingDragged} />;
if (appID === 'calendar') return <Calendar />;
return <PlaceholderApp appID={appID} />;
};
See?? Super easy to do!!
This brought the initial bundle size down to just 62KB min+brotli with 3.9KB of CSS. 23KB smaller
by just dynamically importing stuff π€―π€―
And this is when I took a break from this project(again!)...
Until 12th August 2021, when I was so bored I decided to just rebuild the whole thing again in Svelte. So that's what I did.
Svelte: The golden age
I began migrating on the morning of 12th August, kept at it the whole day, and got the Dock, the menubar, and the action center working, as it was a matter of copy-pasting only.
As for the dock animation, I had already done an article about it, so all I had to do was just copy-paste stuff, again π.
I continued on the 13th too, working the whole day, implementing the different apps. Finally completed by the end of the day.
Results
Unsurprisingly, the bundle size went down and runtime performance went up. If you have been following Web Dev news on Twitter and Reddit, you probably saw this coming from a mile away.
But the extent of these improvements was very surprising. I'll just let the cat out of the bag: Initial bundle size went down to 28.5KB of JS, min+brotliπ€―π€―π€―π€―π€―. That's less than half of what it was before.
And runtime performance went up by a lot. With the previous version, the dock animation could get janky when my computer was overloaded. Not by a lot, but it was visible. After I moved to Svelte, even that much jank was gone. Fully gone!
The process
The process of migrating to Svelte from a React app was gruelling and hard and rough and mentally taxing. It was exhausting.
Ignore the line above, it was total BS π€£
Moving to Svelte was an extremely simple process. It was literally copying the JSX code, pasting it as is, removing the CSS Modules stuff, and adding regular old classes.
Because Svelte's state is very close to React's useState, moving it was a matter of copy-pasting all the state-related code and modifying it to be a regular Variable.
So this π
const [xPos, setXPos] = useState(0);
const [yPos, setYPos] = useState(0);
const [isMenuVisible, setIsMenuVisible] = useState(false);
became this in svelte π
let xPos = 0;
let yPos = 0;
let isMenuVisible = false;
This is it. Literally got rid of a lot of noise, but the skeleton structure remained the same.
But what about Global state?
Yes, global state would've been a problem, but the thing is: It Wasn't.
Moving React global state would be a huge pain if I were using the React Context or Redux, where person A tells person B to tell person C to change the state. I personally never liked this Reducer pattern. It made sense to use it on super big codebases, but I never encountered the problem myself.
So I searched rigorously for a global management system as simple as possible. I already had a taste of the simplicity that StencilJS store and Svelte Stores provided.
After searching long, I found this amazing library called Jotai which provided a global state very very similar to Svelte Stores, and I freaking loved it!!! ππ
It was exactly what I needed.
Here's a little snippet of how it works π
// Create global store
const themeAtom = atom('light');
// Inside the component
const [theme, setTheme] = useAtom(themeAtom);
It makes the state management very linear, just like useState
. Which made migrating it to Svelte extremely easy.
I copied the atom
s as is, renamed their function to writable
, remove the whole const [val, setVal] = useAtom(someAtom);
syntax entirely and put a $
before the store where I was using it. And that's it! This is literally how I migrated away.
What about Styles?
The same thing literally. Because Styles were just plain SCSS, independent of the framework, there was literally nothing to do here. Just copy-pasted styles from the .scss file to the Svelte SFCs, and we're done.
This whole process was so simple, that it took me just 2 days to completely migrate away.
Observations
Here are my takeaways from this whole process π
Svelte Code is much simpler
This comes as no surprise, Svelte code is much shorter, more readable than JSX.
I tweeted a little part of it π
Every single component in the new Svelte Codebase is around 20-30% smaller. The number of files is literally half as before, cuz I don't need to separate .scss
files anymore. Styles exists in the one .svelte
file only.
And some components that are now longer than before, are longer because I got rid of some abstracted small components entirely, which I had refactored away earlier. Because Svelte is so simple in terms of readability, I could just inline these components and the code still looked very clean. π― points for Svelte!!
Svelte Motion = π₯
macOS is an Operating system, and no OS is complete without animations. This is where Svelte's built in Motion stores and transitions come in.
Svelte's Motion stores are extremely lightweight, and extremely powerful. They power the dock animation when you hover over it, and using them is extremely easy.
Previously I was using Framer Motion. Framer Motion is a great library and my go-to thing for animations in P/React. But there's just one little issue with it: It's not exactly small π₯²
This is the bundlephobia size of framer-motion
This isn't very big, but this ain't small either. 32KB Gzip is a lot π€
Framer Motion's size is indicative of its capabilities. Framer Motion can do a lot. Really, it can do anything animation related under the sun.
But ofc, there's no such thing as a good tool and a bad tool. There's only the right tool for the JOB, and Framer Motion was overqualified for my use case.
But Svelte fit right in. It had exactly what I needed(Except for one small piece, which I got from popmotion
. Jut 3Kb big π), and most importantly, it allowed me to programmatically handle it, which is easier than going the declarative route.
Here's a little example. Go to macos.vercel.app, click on any icon on the dock. It jumps up and down once.
Animating that with Framer Motion looked somewhat like this π
const [animateObj, setAnimateObj] = useState({ translateY: ['0%', '0%', '0%'] });
<motion.span
onTap={() => setAnimateObj({ translateY: ['0%', '-39.2%', '0%'] })}
initial={false}
animate={animateObj}
transition={{ type: 'spring', duration: 0.7 }}
transformTemplate={({ translateY }) => `translateY(${translateY})`}
>
{/* Markup */}
</motion.span>;
This is pretty simple, but there are a few issues with it: Have to tell it to not run in the beginning, how to translate it. All these can be condensed into 1 big issue: It's declarative.
I'm a JS guy first. Declarative programming doesn't come naturally to me. I see the world as cause and effect. You click, something happens.
Doing the same thing in Svelte was much easier and programmatic π
<script>
// Spring animation for the click animation
const appOpenIconBounceTransform = tweened(0, {
duration: 400,
easing: sineInOut,
});
async function openApp(e: MouseEvent) {
/* State related stuff */
// Animate the icon
await appOpenIconBounceTransform.set(-39.2);
// Now animate it back to its place
appOpenIconBounceTransform.set(0);
}
</script>
<span style="transform: translate3d(0, {$appOpenIconBounceTransform}%, 0)">
<!-- Stuff -->
</span>
It's very clean and simple now. Plus it's not declarative anymore. I tell the program explicitly when to go up, and when to come back down. The program literally reads like this: Go up -39.2%, and when you're done with that, go down to base position
. As simple as that Β―\_(γ)_/Β―
Those transitions π
Transitions in Svelte are basically animations to be run when some DOM is created or destroyed. This is nearly impossible to do in P/React without relying on external libraries like Framer Motion. But in Svelte, this is in-built and extremelyyyyyyy easy to do.
So, ofc, I put in transitions everywhere I could. Earlier, when you opened/closed any app, there was no transition. Now, when you close a window, it collapses into itself while closing, giving an awesome transition.
Adding this window close transition was literally just this π
<script>
function windowCloseTransition(el, { duration = 300 }) {
const existingTransform = getComputedStyle(el).transform;
return {
duration,
easing: sineInOut,
css: (t) => `opacity: ${t}; transform: ${existingTransform} scale(${t})`,
};
}
</script>
<section out:windowCloseTransition></section>
And that is it!! Now there's a really nice transition, with around 10 lines of code only.
I see this as an absolute win ~~ Professor Hulk
I added in transitions in many other places which I couldn't add before with Framer Motion. Not because they weren't possible, but because the syntax felt unnatural there, but here in Svelte, it just fits right in!
Final Words
Really, moving to Svelte made me realize how extremely powerful it is. A lot of React developers treat Svelte as a toy(I was one of them), but really, it isn't a toy. In fact I'd say it's more powerful than React. Yes, less control over everything than React, but it is a better choice for 95% of developers than React anyway.
Hope you liked this half-historical half-technical article, and it gave you some enthusiasm to try Svelte, if you haven't already.
Peace out βοΈ
Posted on August 29, 2021
Join Our Newsletter. No Spam, Only the good stuff.
Sign up to receive the latest update from our blog.
Related
November 30, 2024
November 30, 2024