VanillaJSX.com(vanillajsx.com) |
VanillaJSX.com(vanillajsx.com) |
ultimately go to react.dev because: "Maturing is realizing React is best"
It started as a static site generator but added a bunch of support for client JavaScript too.
It also reminds me of that Douglas Adams line about flying: it's the trick of falling and completely missing the ground, so in order to do it, you can't think about it too hard.
I would guess there is more overhead in creating a dom element than a JS object (which JSX elements compile to).
First, you have to understand that at that time Firefox was about 500x faster at accessing the DOM than Chrome and about 250,000x faster accessing the DOM via the API methods than via querySelectors. Firefox and Chrome performed about equally in use of querySelectors with Chrome being a tiny bit faster. So, the DOM was already fast, but occupied a different memory space than JS.
At any rate the original motivation had nothing to do with performance. The goal was to introduce a template system that fit with React’s state/component system. JS modules weren’t a thing yet, so code organization was very different at that time and centered around concepts like AMD and Common.js, though it was mostly some form of AMD typically require.js.
The design of the template system in Vue was created to solve for the exact same conditions according to the internal organization of Vue.
- JS object appears to be at least 2x faster than document.createElement() (Chrome)
- Note: JS object only loosely represents JSX element so it is a bit unfair. But with actual JSX objects I would assume it is still somewhat faster than the DOM API.
https://youtu.be/DgVS-zXgMTk&t=1532
- Pete Hunt, one of the React devs, says "JSX is faster than the DOM because it is JS memory."
It also presents a wildly different use case. It’s common to rapidly create and destroy nodes only because that is what convenient for framework logic. Outside of frameworks DOM nodes are modified more frequently than created anew.
> - Pete Hunt, one of the React devs, says "JSX is faster than the DOM because it is JS memory."
That doesn’t make sense. Your only choices to display content in the browser using JS is via DOM modification or canvas. If not using canvas you are touching the DOM no matter what. It is true that DOM memory is different than JS memory, but the DOM API is fast enough that it doesn’t matter, especially since you are touching the DOM no matter what. If your approach to DOM interaction is too slow it’s because of string parsing on things like innerHTML and querySelectors. So if performance is important then don’t do things that parse strings and don’t abstract away the DOM interaction because you are touching it no matter what.
You want to return a description of the DOM, rather than the real DOM, because you want to be able to reevaluate your templates repeatedly with new state, and efficiently update the DOM where that template is rendered to.
All the examples here use imperative DOM APIs to do updates, like with this:
function TodoInput(attrs: { add: (v: string) => void }) {
const input = <input /> as HTMLInputElement;
input.placeholder = 'Add todo item...';
input.onkeydown = (e) => {
if (e.key === 'Enter') {
attrs.add(input.value);
input.value = '';
}
};
return input;
}
class TodoList {
ul = <ul class='todolist' /> as HTMLUListElement;
add(v: string) {
const item = <li>{v}</li> as HTMLLIElement;
item.onclick = () => item.remove();
this.ul.append(item);
}
}
(The problem with VDOMs is that diffing is slow, a problem solved by using templates that separate static from dynamic parts, like Lit - a library I work on).
In fact I now strongly believe it's counter productive, because most people come to it thinking "I can just trigger however many re-renders of this large piece of UI as I like, the vdom makes it ok" and it doesn't, the performance sucks, but now you have architected the app in a way that requires a rewrite to make the app perform well.
I have seen this exact sequence of events four times, by four different teams. The second, third and fourth, as a principal architect consulting for the team I tried to intervene and advocate for a vanilla architecture that is mindful about performance, citing the issues they would likely experience with react, but to no avail. There was a lot of "oh but there many ways to avoid those issues" followed by a list of things I was presumably ignorant about.
I guess most of us need to learn things the hard way.
Specifically, I'm hoping to show that vanilla architectures can be not only performant, but easy to maintain with well designed code that uses stable and known patterns. Using JSX just so happens to clean up the code nicely and make the relationship between React and vanilla very visible, but that's really all it did here.
Although to be fair, the hack required to get JSX-as-DOM to work is really unfortunately and I'm not very happy with it, and I would prefer JSX to just render as an object tree that anyone can render however they want. But when I tried that for a few months or a year, it was not nearly as performant as rendering them as strings as soon as they're evaluated, which can then be cached via standard module caching. At least, that's how I got immaculatalibrary to entirely render all HTML files in ~700ms initially and ~70ms on most file changes.
I'll try to do some experimentation next week to see if I can get more performance back out of having <foo bar={qux}>child</foo> to render as {foo:{bar:qux, children:[child]}} again though, because that would absolutely be the ideal, and would unfork JSX in the same way Typed Annotations proposes to unfork JavaScript types.
When people say that React is fast, what they mean is that React can dom-diff faster than a naive O(n) approach. It means that updating a component with a thousand nested divs won't crash out your browser, like it might if you were to write the code by hand. It doesn't mean it's an objectively high-performing framework.
What React is good at is forcing you to write code in a clear, comprehensible way. Having every engineer on your team obey F(props) = state is a strict improvement over virtually any other paradigm. (Yes, you can still make a tangle of components if you try hard enough, but the complexity of the tangle is capped significantly lower than the complexity of a tangle of JS without any framework attached.)
Separate source of truth from derived data. Separate possibly valid user intent from validated state. Use contexts to organize the data dependency graph of your application. Ensure all your widgets use a consistent value type in and out, don't let events contaminate it. Use something like cursors or optics to simplify mutations and derive setters automatically.
I've never had an easier time building very complex UI functionality than with React. But it requires you to actively start reasoning about change in your code (what doesn't change), and this is something most people are not used to.
Personally I think React compiler is folly for this reason: they are taking the most interesting part of React, the part that lets you write apps that are incremental from top to bottom, and telling you it's too complicated for you to think about. Nonsense.
The issue is just that React makes pros feel like idiots unless they eat some humble pie and grok the principles and the reasons behind it. Which is that React is what you get when you try to come up with a UI architecture that can make entire classes of problems go away.
Without a VDOM, one way data flow, and diffing, your UI won't just be slow, it'll be full of secret O(n^2) cascades, random ifs to stop infinite cycles, random "let's update this state early so other code can immediately use it" ordering issues, and so on.
A rewrite on those occasions, I do not consider possible. Most businesses are way to deep into frameworks.
What I personally advocate is to choose frameworks with valid escape hatches, and then make sure to staff a team with JavaScript programmers rather than some-framework Andies.
In an even larger scale, I advocate for web-component based solutions. I try to stay close to vanilla as much as I can. Whether I like it or not, it's the highest value professional decision, imo.
As far as I understand, signals make it much easier to keep the DOM updates to a minimum.
More concisely: JSX is just an alternate function call syntax with some useful applications.
For example at my last company we used JSX to make test data factories that had an XML-like look but were using a builder-pattern in the element creation that was able to make contextual decisions about what was in the final data. Nothing to do with React, DOM, or inability to express the same thing declaratively without JSX.
Thats really interesting, can you elaborate more?
For example did you use a specific JSX compiler? Was that written in house or used a 3rd party library?
Depends who "you" are. I prefer to have my DOM nodes updated in place without all the reconciliation machinery. (no implication about what you want)
If all JSX does is return a DocumentFragment that you then need to imperatively add event listeners to and imperatively update, how is it much better than innerHTML?
I get the "no more imperative updates" dream. I've used these frameworks for probably a decade. I've mastered them.
Me personally, I prefer imperatively updating my DOM. I get completely fine-grained control over what's happening. I can architect it to be an extremely efficient machine. I can make it extremely easy to add/change/remove/fix features in my apps without forcing myself to think according to anyone else's opinionated methodology.
The farther you get into complex GUI territory, the more you want a declarative, functional approach, because it makes things simpler. The closer you are to a handful of controls with simple logic, the more you want to just imperatively tell them what to do, and leave the rest of the page alone, because it makes things simpler. We now just have better tools than jQuery for that.
I feel like this is one of the leakiest abstractions in all of computing. There's a reason there's an entire cottage industry around react; how to stop things rendering multiple times, refreshing needlessly, etc.
[0] https://blog.metaobject.com/2018/12/uis-are-not-pure-functio...
>you want to be able to reevaluate your templates repeatedly with new state
No you don't. It is inefficient and increases complexity. You then have to extract and keep track of state yourself where the platform/UI components could have done much of this themselves.
Calling the same method over and over again is wasteful.
So we don't do that.
First, we did not start with the obviously incorrect premise that the UI is a simple "pure" function of the model. Except for games, UIs are actually very stable, more stable than the model. You have chrome, viewers, tools etc. What is a (somewhat) pure mapping from the model is the data that is displayed in the UI, but not the entire UI.
So if we don't make the incorrect assumption that UIs are unstable (pure functions of model), then we don't have to expend additional and fragile effort to re-create that necessary stability.*List reconciliation still has to happen, but you don't need to pull out an entire vdom. You just have to have some mapping between list items and their resulting DOM nodes.
This mind numbing reliance upon layers of abstraction nonsense around state management is why I really don't like React. State management is ridiculously simple. State management, when done correctly, is the most primitive example of MVC with no abstractions needed.
You might be able to handle this while you're alone and you know everything about your codebase, but the moment you're working with someone else or in a team this will no longer be the case.
It's also not even an either-or. I've worked on a codebase that did both: React was loaded for some views & others were served with a lightweight JSX renderer.
I wanted to add my two pennies to the discussion. You are of correct that with that approach you lose the declarativeness but sometimes you don't need that, if the thing is mostly static.
I went this road many years ago for a project. The nice thing of this approach is getting rid of all the ugly DOM API but enjoying it's performance over innerHTML.
(1) I absolutely love lit-html but don't like the rest of the lit components framework. Luckily you can use it independently!
It works great! I was amazed at how you can do so much in templates, it's pretty much everything I could do in Vue templates, though a little more verbose.
I built my own `VanillaComponent` class, which has a mount() method which calls the render() function which I define on the child class.
My VanillaComponent class looks like this:
import { html, render } from "lit-html";
abstract class VanillaComponent {
abstract render(): ReturnType<typeof html>;
private _mountPoint?: HTMLElement;
mount(this: VanillaComponent, target: HTMLElement) {
target.textContent = '';
this._mountPoint = target;
this._update();
}
_update() {
let templateResult = this.render();
let rootPart = render(templateResult, this._mountPoint!);
}
}
class MyComponent extends VanillaComponent {
constructor(props: { label: string }) {
this._props = props;
}
render() {
return html`<button>${this._props.foo}</button>`;
}
}
let myComponent = new MyComponent({ label: "I am a button" });
let target = document.querySelector("#demo");
myComponent.mount(target);
myComponent.update()
However something I have not been able to solve with lit-html only, is when I compose parent and child components I have to do something like :
class MyDialog extends VanillaComponent {
render() {
let childComponent ... // another VanillaComponent previously instanced
return html`
<div>
${childComponent.render()}
</div>
But this means I can not do `childComponent.update()` because I don't know the root element of child component, since I did not mount it explicitly myself.
I mean technically because of the lit-html optimizations, I can do `.update()` on myDialog (the parent component) after any child component's props changes, and it will only re-render what is necessary... but let's say my child component has like 1000 cards... it seems very wasteful and it would be ideal if I could re-render only the child.
I wonder if there is a trick to get around that with just lit-html?
I think in this case the ref() directive - i.e. https://lit.dev/docs/templates/directives/#ref - may be what you want. If it isn't exactly, reading how it's implemented would be my first step towards building something similar that is.
[edit] Oh also, this solution works really well for SEO. That's another problem I didn't find solved well in other JSX frameworks.
[1] https://engineering.fb.com/2010/02/09/developer-tools/xhp-a-...
https://github.com/tomtheisen/mutraction
It lets you do stuff like this.
const model = track({ clicks: 0});
const app = (
<button onclick={() => ++model.clicks }>
{ model.clicks } clicks
</button>
);
document.body.append(app);Another interesting thing is that other JSX libraries like Solid.JS also return DOM nodes, and I love that this idea is gaining traction
The closer we get to the platform we're using, the better. Being removed by layers of abstractions CAN be useful, but in practice, I haven't found a use for abstracting away the platform. (yet.)
Maybe huge projects like Facebook benefit from this tho (which I haven't worked on)
Technically this is the only code bundled with vanilla jsx:
Here’s an app written using Vanilla TSX: https://github.com/wisercoder/eureka/tree/master/webapp/Clie...
Uncaught (in promise) TypeError: Map.groupBy(...).entries().map is not a function h("div", {}, [
h("p", {}, "this is easy"),
...list.map((l) => h("li", {}, l),
])
No better article than this: https://blog.vjeux.com/2013/javascript/react-performance.htm...
Was super productive and easy to create a Cloudflare Worker Web App that’s free to host thanks to Cloudflare’s generous 100k daily worker request limit.
Generally don’t believe in serverless for larger Apps, but for small websites that you just want to create, deploy and ignore - it’s great!
el = <button>Click me</button> as HTMLButtonElement;
el = html.button('<button>Click me</button>');
That way no compilation step would be needed and debugging would be easier as the code executed in the browser is the same code the developer writes.
After “How would they handle large data?” it turns into an unreadable mess.
Communication between elements is not covered, global deps, DOM updates scheduling, content projection, and so on - you “just don't need it” in small demo examples, but you do need it in the real apps.
Clojure datastructures makes this so much more enjoyable. Everything is just basic lists and maps which makes it very flexible and powerful.
[:ul [:li "task 1"] [:li "task 2"]]
It's weird that it's not more common for making web apps.
h("ul", h("li", "task 1"), h("li", "task 2"))
There's also variants like h.ul(h.li("task 1"), h.li("task 2")) using Proxies now too.
Trouble is if you're used to HTML it can take a while to get used to it. It's like a learned helplessness or something.
Braces and semicolons exist so you can do that.
Sometimes it's just easier to read an array that's split over multiple lines.
Sometimes it's just easier to read a statement that's split over multiple lines.
Sometimes it's just easier to read two statements sharing the same line.
If your language's restrictions is making code harder to read, then it's making your job harder than it needs to be.
Not what you'd expect to see.
So I would like to explore the ability to use JSX in non-DOM environments. react-three-fiber does this with Threejs, but then it is still React oriented. I found this article about parsing JSX https://blog.bitsrc.io/demystifying-jsx-building-your-own-js.... And I know babel has something that parses JSX.
Does anyone have recommendations for doing this. Threejs to me a good candidate - a non React version, since it is a hierarchical system (scene, meshes, materials etc), but I suspect there are other applications.
I made an attempt to implement a Javascript version of Hickey's transducers - a sort of conveyor belt of functions and that is another instance of a series of processing steps that might be best represented in JSX
JSX doesn’t actually write anything, it’s just a templating language over React.createElement.
It’s the virtual DOM that actually syncs those structures created by createElement to the real DOM. So it’s the virtual DOM that allows you to write your code declaratively.
That’s evidenced by OP’s project, which is JSX without the declarative piece. You just get an Element and then you have to update it imperatively if you want to change anything.
<html lang="en">
<body>
<h1>Web Components Examples</h1>
<h2>Counter Component</h2>
<counter-component></counter-component>
<h2>Clickable Button Component</h2>
<clickable-button></clickable-button>
<h2>Toggler Component</h2>
<toggler-component></toggler-component>
<script>
class CounterComponent extends HTMLElement {
constructor() {
super();
this.count = 0;
this.button = document.createElement('button');
this.button.textContent = this.count;
this.button.addEventListener('click', () => {
this.count++;
this.button.textContent = this.count;
});
this.attachShadow({ mode: 'open' }).appendChild(this.button);
}
}
class ClickableButton extends HTMLElement {
constructor() {
super();
this.clicked = false;
this.button = document.createElement('button');
this.button.textContent = "Click me!";
this.button.addEventListener('click', () => {
this.clicked = !this.clicked;
this.button.textContent = this.clicked ? "Clicked!" : "Click me!";
});
this.attachShadow({ mode: 'open' }).appendChild(this.button);
}
}
class TogglerComponent extends HTMLElement {
constructor() {
super();
this.on = false;
this.button = document.createElement('button');
this.button.textContent = "OFF";
this.button.addEventListener('click', () => {
this.on = !this.on;
this.button.textContent = this.on ? "ON" : "OFF";
});
this.attachShadow({ mode: 'open' }).appendChild(this.button);
}
}
customElements.define('counter-component', CounterComponent);
customElements.define('clickable-button', ClickableButton);
customElements.define('toggler-component', TogglerComponent);
</script>
</body>
</html>Mostly I just do Vanilla.js, but the vanilla DOM creation functions turn really verbose, I got tired of that and created this to cut back on code size and increase readability.
There are other libraries that do something similar, but in my own very biased opinion this is one of the better.
It looks simpler at first glance/ to a untrained eye; but it's just adding complexity without really solving any problems.
I like approaches like Kotlinx.html, scalatags, Elm's HTML package or HtmlFlow. They are also abstractions, but they add typesafety that html-as-a-string does not offer. On top of that you get breakpoints, code completion, and you can keep working in one language.
That said, JSX can be used easily with Astro as long as you get used to at least a bit of The `.astro` syntax for wrapper components and pages/layouts.
Another user below said
> We've recently moved one service from next to Astro and it was just removing a ton of boilerplate and 'dance around' code.
And I get why it happens. When you first try out a new framework, you allow yourself to learn and add its inherent complexity, knowingly and intentionally. You say to yourself, "it's part of the dream, it's going to work out; there's a vision, just trust the process." This is true with literally all frameworks.
But they never deliver. The complexity is never worth it, and in the end, the intentionally added complexity is always intentionally and gladly removed when it becomes clear that it was unnecessary complexity. This is what I am glad to have learned so thoroughly that I no longer try to learn new frameworks when I initially see its complexity, imagine adopting it in view of my experience, and recognize that its almost always not worth it.
Look at the code on vanillajsx.com. Besides JSX and types, it's plain JavaScript and DOM manipulation. Translating it to document.createElement would add almost no lines of code. There's no unnecessary complexity. That's the whole point of the site. The simplicity of discovering and removing unnecessary complexity is wonderful and refreshing, and I think a lot of people agree.
Coming from Vue I was really surprised it does a lot of what Vue templating does, including attaching events, with just vanilla JS templates. And when you use VSCode lit extension, you get syntax highlighting and full type checking inside the templates.
I learned about lit-html after a tweet from Marc Grabanski, where he said he used lit-html with vanillajs, not Lit.
After some experimenting I found it works great and it seems like you are trying to solve something very similar.
When you use the lit-html template package you can do basically evetything that is described in the Templates chapter
https://lit.dev/docs/templates/overview/
... without all the other abstraction of components that are part of lit-element.
https://lit.dev/docs/libraries/standalone-templates/#renderi...
Var button = html(’<button>im a button</button>’);
The html() function is trivial, but it just doesn’t feel like real programming to do this, even though there’s nothing else to it in the end.
But maybe we can revive the general idea with a modern take: https://github.com/WICG/webcomponents/issues/1069
Isn't this what we have in TFA?
But even considering the single implementation problem, it also was just not a good language model, nor was it well specified or defined and it brought with it a pile of weird baggage and complexity.
Then because it was The Future there was no real thought into proper interop with JS (it was essentially a completely independent spec so adopted general syntax but specified in a way that meant JS could not simply adopt that syntax).
Firefox never shipped the optional E4X DOM APIs. I wrote a polyfill for them at the time.[1]
VanillaJSX seems to suffer from the same problem though.
Honestly, the real interesting part about my framework is literally everything else. Returning strings from JSX on the ssg-side; being able to import raw source directories and manipulate string|Buffer at ssg-time; the extremely efficient and lightning fast module system I wrote on top of chokidar and swc; probably more I'm forgetting, but basically the JSX-as-DOM is only the most visually interesting part. But really just a party trick.
[edit] Case in point: the source code to vanillajsx.com is extremely concise and clear and short, I literally wrote the whole thing today with zero deps (besides imlib), and the JSX-as-DOM demos are the least innovative part of it: https://github.com/sdegutis/vanillajsx.com/tree/main/site
For those that appreciate this approach of JSX returning concrete DOM elements, Solid works exactly like this, with the addition of a proper reactivity layer.
* https://www.immaculatalibrary.com/books.html (src = https://github.com/sdegutis/immaculatalibrary.com/blob/main/...)
* https://www.immaculatalibrary.com/prayers/ (src = https://github.com/sdegutis/immaculatalibrary.com/blob/main/...)
Your best bet would be to queue up specific UI changes that need to be made as diff's rather than checking the entire UI state. At that point, though, you might as well run them immediately as the change is needed.
If that was still a perf problem you would end up chasing a very complex solution like react fiber to partially update the UI on a loop while periodically pausing for user events.
Also, it sounds like the only browser to ever support it was Firefox? That was probably much more of a limiting factor for adoption.
However, the XML selector syntax was a godsend. Recursively parsing an XML tree is really a pain. E4X would allow you to do things like:
var foo = someXml..childNodes.@attribute;
<someXml>
<intermediateNodeYouWantToSkip>
<childNode attribute="1" />
<childNode attribute="2" />
<childNode attribute="3" />
<unrelatedNode />
</intermediateNodeYouWantToSkip>
</someXml>
Basically JSON as a format won out over XML and ES4/XHTML were ditched. Frankly, a world that revolved around XML/SOAP would have been a nightmare, so I guess killing off the easy processing of XML in JavaScript helped to stave off that potential future. XSS, XSLT and E4X were all casualties.
<div>
<p>this is easy</p>
{list.map((l) => <li>{l}</li>)}
</div>
To be fair to JSX, you use regular loops, interpolation, etc without any different syntax (`{}` accepts a vanilla JS expression), you just obviously need the compiler step to de-sugar the element tags to `createElement` calls.
And that is what a lot of the pushback against React seems to come down to as well -- it is overkill for simple things. And it is. But different people doing different apps with different levels of complexity... all have different needs.
So it is not as simple as one having appeal over the other. They are different tools for different problems.
But to your point, JSX doesn’t really do much. Your h function is basically what React.creatElement does. Google “React without JSX” and you’ll see how it looks.
JSX is just syntactic sugar over React.creatElement. And that is what makes it so nice… there _are_ no special constructs for loops, or variables, or components. They are actual JavaScript loops, JavaScript variables, and JavaScript function.
It makes JSX easier to reason about than most templating languages.
https://svelte.dev/blog/virtual-dom-is-pure-overhead
React is also at the point where re-rendering the whole app is a fiction the library maintains for you while being smarter and doing less, why not go the whole way?
I'm asking because, many of React (or friends) introductory material naturally focus on building things like TODO lists or Tic Tac Toe; while those offer insights into how to work with React (& cie), they're not showcasing cases where the performance gains are perceptible, and IMO not even cases where the "organizational" benefits of such libraries are salient.
While there's no doubt that React and virtual DOM offer advantages, it's essential to clearly demonstrate where and how these benefits manifest in real-world applications.
> they're not showcasing cases where the performance gains are perceptible
According to this commenter, it's not even about the performance gains:
https://news.ycombinator.com/item?id=41271272
> and IMO not even cases where the "organizational" benefits of such libraries are salient
Apparently, that is what it ultimately boils down to:
If you assume programmers who don't know what they are doing it is a very different question. Some people will manage to make a train wreck both with and without a framework. But if we assume that there is a skill level where people will manage to make something useful with a framework, but not without it, or vice versa, then I really do not know which way it swings.
One project I worked on was a BI tool with a fair amount of complex state. Before we introduced react we were using backbone, event listeners, and managing dom state ourselves. With react we made it declarative, which simplified things.
Another project was an email client with a lot of complex state and caching. That started with and continued to use react so I don’t have a direct performance comparison. But again, managing the state manual would have been a nightmare.
You still end up with the bloated bundle size but with one of the more modern react alternatives you can eliminate that too. So at least for me, I don’t mind the build complexity for the power I get; especially now that node itself is supporting typescript, the build side is getting simpler to set up as well.
I don't understand why people take such offense to calling document.createElement() or document.getElementById() or kind of document. or window. function. It's consistent and native.
class Item extends EventTarget {
#checkbox = <input type='checkbox' /> as HTMLInputElement;
li;
constructor(private list: List, text: string) {
super();
this.li = (
<li class='item'>
{this.#checkbox}
<span onclick={() => this.toggle()}>{text}</span>
<button class='close' onclick={() => this.remove()}></button>
</li> as HTMLLIElement
);
this.#checkbox.onclick = () => this.toggle();
}
}
class Item extends EventTarget {
#checkbox;
li;
constructor(private list: List, text: string) {
this.#checkbox = createElement('input');
this.#checkbox.setAttribute('type', 'checkbox');
this.li = document.createElement('li');
this.li.className = 'item';
this.li.appendChild(this.#checkbox);
const span = document.createElement('span');
span.onclick = (() => this.toggle());
span.textContent = text;
this.li.appendChild(span);
const button = document.createElement('button');
button.className = 'close';
button.onclick = (() => this.remove());
this.li.appendChild(button);
}
}
With the second of you have stringa.
If you can control your app’s structure, it primarily adds significant increases in the RAM and CPU required for your app and slows load time because you are using a huge amount of JavaScript to emulate the carefully tuned C++ code built in to the browser. If you notice, most of the benchmarks from when React launched claiming performance wins were compared to heavyweight frameworks or complex jQuery plug-in combinations where a single user interaction might trigger cascading updates forcing the browser to rerender things which didn’t change along or to reflow multiple times in cascading update-measure-update chains. Pure DOM implementations were always faster, often by multiple orders of magnitude and once you could drop IE6, and then IE11, the DOM APIs and CSS were rich enough that much of the library code is now a net negative as well (e.g. people used to use complex code trying to build layouts which CSS grids solved).
It can also simplify some stuff surrounding event handling (but that's not it's main goal I think)
So people wrote various ways to defer/batch/denounce updates.
Virtual DOM is a general solution/implementation. It's not the only one, but I think you always need at least a tiny runtime to avoid too much DOM access (ie Svelte, Solid JS are fairly minimal)
Unless you use lit-html, which has a very efficient diffing algorithm that only updates the nodes that have changed.
People often mistakingly describe the vdom as faster than the dom, this is incorrect. It would be faster than throwing away the whole components dom and rebuilding, so the same templating code building a new dom, rather than a vdom that's then diffed. Hand crafter mutations will be faster than a vdom diff, simply because the computer is doing les work, however much more error prone.
Garbage collection is less efficient, but it is sometimes very difficult to figure out exactly when a piece of memory stops being used, which leads to use-after-free, double-free and memory leak bugs.
Same goes for classical UI approaches. In classical UI, most pieces of state are kept in at least two places, once in code and at least once in the DOM.
For example, in a shopping cart, the total might appear three times, implicitly in the code (as a function that sums the prices of all the items), once as the label of the "open cart" button in the navbar, and once as text in the "your cart" modal, which that button shows or hides. THe cart may be modifiable from different places, the cart modal itself, product pages, product collection pages, order history (when re-ordering recently purchased items) etc.
In the classical approach, you need to make sure that all modifications to the cart accurately change the state in all three places. You also need to ensure that if you remove a product from the cart using the modal and you're currently on a page that lets you order the product in any way, the "remove from cart" button on that page needs to turn back into "add to cart", and there may be hundreds of different such buttons, which the cart modal needs to handle somehow. It is very easy to make mistakes here and have the state in the code (array of products) fall out of sync with what the user sees on the page.
In React, there's just one array of products, one function to calculate the total, and a lot of places in the code that use this array. WHenever the array changes, the pieces of the page that rely on it automatically re-render, while everything else stays the same. There's no way for the UI and the array to fall out of sync, and there's no need to track where the array is being used and where it's being modified.
You'll get better performance with _carefully crafted_ DOM access, but that's easier said than done, especially on a larger applications.
vDOM takes care of the "carefully crafted" part with some trade offs, especially if it also defers rendering and doesn't wccess the DOM on every update.
So yes, it's easier to write declarative UIs with it, but it's also there to address common performance issues with unchecked/eager DOM access. Even if you don't throw away the whole tree and insert a new one, it can be very slow. Just _reading_ from the DOM is slow _and_ everything stops while that's being done too.
I was painfully reminded of that while visiting Europe last month on a data plan which was clearly deprioritized on partner networks - the old sites with .php in the URLs loaded in a few seconds and worked perfectly, but every time something failed to load in less than 5 minutes or partially loaded but didn’t work a quick trip over to webpagetest.org showed a lot of NextJS, React, et al. scripts trickling in because clearly a form with half a dozen fields needs 25MB of JavaScript.
The root cause is obvious: you get what you measure. If businesses prioritize Jira tickets closed per day, they’re going to get this soup of things promising to be easy to use for high developer velocity and they’re never going to get around to the optimizing it. If they’re trying to be able to hire as cheaply as possible, they’re going to look for the current tool boot camps are pushing and hire based on that, not looking for deeper knowledge of web standards or experience which costs more and shrink the candidate pool. If they’re looking for a safe choice, Facebook’s marketing means all of the big consulting companies will push React and few people will pause long enough to ask whether they’re building the same kind of apps it was designed to build (long session times, tons of local state being mutated, etc.) or whether they’re willing to invest the time needed to get it to perform reliably and well.
In all these frameworks, mutations happen inside effects, and effects are scheduled such that they all happen at the end of the tick, avoiding reflows and thrashing.
As the user, not you the author of some code, make changes to the UI it will be via events. In the respective event handlers just update a big state object with the identity of the node and how it is changed. When the page loads simply read from that state object to recreate the respective node exactly how they were when the user modified them. That's it. Nothing more.
Its how I wrote a full OS gui for the browser. It does far more than SPA framework, does it substantially faster, and requires only a tiny fraction of the code.
Yes, yes, here come all the excuses, like: I can't code or it wont work in a team, because other people cant code. Stop with the foolishness. Provide some governance, preferably fully automated, around event handling and you have complete obsoleted your giant nightmare framework.
I describe it here: https://github.com/prettydiff/wisdom/blob/master/state_manag...
There is a persistent ‘learned helplessness’ tendency among some developers to assume that the frameworks they are using have access to magical mystical powers above and beyond those that their own code can make use of.
Framework code might well be better optimized or more tuned than the code you would write - but if you cared to employ similar techniques you could achieve those same gains; on the other hand, since by definition it’s more flexible than single-purpose code, it might not be optimal for your usecase.
Because now everybody know that it's not the case. But if you check the first video announcement from Facebook a decade ago or so, performance was definitely at the front, above everything else.
How many frameworks did you have experience with?
> Without a VDOM, one way data flow, and diffing
you wanted to write "or" not "and", didn't you?
you can adhere to the same principles (one way data flow) without vdom. Not saying it's easy at large scale but it's possible. I don't appreciate people invoking fud towards anyone opting out of their tech choice.
But there was no real progress after years of them pushing this syntax, but failing to actually define a type system that was coherent, or a model that would allow it.
As a result I proposed `for (of)` largely to prevent sane enumeration from being blocked on the intransigence of two people.
It's also worth noting that for(:) enumeration would not even preclude their syntax - it's certainly not grammatically ambiguous - and most real world code in languages that support enumeration directly and support inference doesn't explicitly specify the types , so the ugliness of `for(let a:type:expression)` would have be rare anyway.
shrug
Given that ECMA literally killed E4X a few years later the blanket ban on "for each" or "foreach" (because it would be "confusing" in E4X) is arguably worth than for(:), but again shrug
Its a shame, E4X was really nice
Definitely; it's a struggle to find precise, concrete arguments in this direction. And there are many good reasons to be conservative: e.g. inheritance-based OO was sold with "VW inherits Car"; looks great on paper, but not as much in front of real-world issues.
> Apparently, that is what it ultimately boils down to:
If so, I'd be left wondering how much of this is actually caused by a lack of discipline, as seems to be for example indicated by the "dumb reflows" issues.
Anyway I prefer the html looking syntax if anything because it looks like the output on the page. That’s dependent on a project coding style that doesn’t unnecessarily wrap things in components, which for my company’s product I’ve set as a standard.
var data:XML = <foo><bar>hello</bar></foo>
and then data was an object instance like you’d expect
> managing the state manually would have been a nightmare
Now, of course, the annoying follow-up question would be "why?". For sure, there are poor ways to handle complex states without React, but that doesn't mean it's impossible to do it fairly well either.
For example, a divide and conquer strategy might make the situation less miserable already: associate "components" (DOM nodes) to relevant portions of the state only. But it's still too difficult to get a clear idea without actual code (hence my original question: most of the discussions about this are either objectively too hypothetical because they deal with complex/private codebases, or unconvincing toys)
There were two groups I was hoping vanillajsx would resonate with. The first is people who still buy into the React dream but are beginning to be disillusioned with its inability to deliver on its promises, and the second is people who already are fully disillusioned.
I don't know if you've seen it, but Alex Russell just did a blog series where he directly talks about this disillusion and proposes a move away from React for most web apps: https://infrequently.org/series/reckoning/
I am not as anti-React as that myself, but I do agree it is hard to scale up and have it perform well, not at all like the promise. As always, there are no silver bullets and you have to pick a stack that you can understand.
By the way, I made my own pitch for fully vanilla web development here: https://plainvanillaweb.com/
- If the UK GSD is anything like USDS, using them for comparison is like comparing a pro sports team to your local high school’s. They are an outlier specifically created to be better than the average, so tautologically their stuff will be better. Code For America is a similarly odd comparison.
- The US has a massive gap in pay and prestige between public and private sector developer jobs. It’s not that this means “worse” people work at public jobs, but in general they start less experienced and can wind up in a non-learning cycle as they don’t get mentorship/guidance from more expert folks, and if they do get good independently they leave. It’s really hard to convince people to take a pay cut to work these jobs, and many of the few willing to do so instead go to CFA, USDS, etc because they want prestige and avoid all the other inefficiencies in public jobs.
I could go on about the structural problems leading to this, but suffice it to say that blaming React and other JS frameworks is a miss. For some services it’s lucky they are online at all, and a slow web page is still orders of magnitude faster than physical mail or god forbid going to a physical office. The sites could definitely be better but this is not fundamentally a problem of technology choice.
Frameworks have evolved over time, as we've identified better ways of doing things and as the browsers have implemented native solutions to problems that frameworks were invented to address.
For example, we don't use dojo anymore. Or ember. Or backbone. The list of once-popular frameworks that have faded is long.
The point that I (and and increasing number of other developers) have been making for a while now is that React has hit this stage. Many of the problems it was invented to solve are mostly not currently relevant as native solutions have become more widely adopted.
This has caused React to evolve into bloatware in an effort to maintain mind share. I think that evolution has had the opposite effect. I think it has driven many developers to seek simpler solutions.
I think it's completely valid to recognize framework's strengths and (especially) weaknesses. You can call this "blame" or just a justified critique.
I agree with your point that lower-skilled developers can make a mess out of any technology, but one of the supposed benefits to adopting a framework is to provide guard rails against this. If that's not working, it makes you question the fundamental value of using a framework at all.
It was attempting to do something along the same lines as you first suggest
Write it in your own vanilla framework, and you will effectively re-invent all the complexity of React, but in a way that no one has ever done before. It's easy at small application scales, but once your app gets large, good luck debugging the thing that exists primarily in your principal engineer's head.
React is cool because it's widely used (hence more tooling and learning material) and because JSX is very convenient (it composes very well), not because the vdom works well to create an immediate mode API.
This in theory could do something to copy that local state over, or diff the two DOMs directly without a VDOM (though from the sound of it, it probably doesn't).
You can have imperative event listeners and updates.
These are two independent dimensions. I made UI framework called mutraction that produces real DOM elements from JSX expressions. It also updates any contents or attributes of those DOM nodes based on their dependencies without requiring imperative DOM interaction from application code.
https://github.com/tomtheisen/mutraction
Here's a click counter. `track()`, as you might guess creates a proxy so that reads and writes can be converted into dependencies.
const model = track({ clicks: 0});
const app = (
<button onclick={() => ++model.clicks }>
{ model.clicks } clicks
</button>
);
document.body.append(app);2) Autocomplete for element props.
3) IDE support such as refactors and jump to definition/jump to usages.
4) Proper syntax highlighting out of the box instead of the editor just saying "there's a string here".
5) A uniform pattern for defining custom components that work the same as primitives, rather than defining custom components as helper functions returning string fragments or something like that.
And so on. JSX has a lot going for it regardless of the semantics chosen. It's just a syntax that is very convenient for lots of kinds of tooling, and it's completely unopinated about the semantic context in which it is used.
> We've recently moved one service from next to Astro and it was just removing a ton of boilerplate and 'dance around' code.
---
> And I get why it happens. When you first try out a new framework, you allow yourself to learn and add its inherent complexity, knowingly and intentionally. You say to yourself, "it's part of the dream, it's going to work out; there's a vision, just trust the process." This is true with literally all frameworks.
I am quite picky and have strong opinions. I've used Astro for more than a year and still love it. There is complexity (especially if you use SSR), but for the use case of "I just want a static site" it is wonderful.
> Look at the code on vanillajsx.com. Besides JSX and types, it's plain JavaScript and DOM manipulation. Translating it to document.createElement would add almost no lines of code. There's no unnecessary complexity. That's the whole point of the site. The simplicity of discovering and removing unnecessary complexity is wonderful and refreshing, and I think a lot of people agree.
I don't disagree, but this doesn't replace what you might want for SSG. For one, this requires JS on the client. Astro compiles to static HTML despite using JSX-like syntax.
As an example, here's my Astro site and source code:
It just runs your code in a node.js process, and translates your JSX expressions into jsx() calls. On the node.js side[2], jsx() returns a string from its tag/attrs/children. On the browser side[3], jsx() returns DOM elements.
Combined with a little bit of architecture, it becomes something extremely well suited to creating static sites. I guess SSG is an outdated term now, maybe it's a framework? Or a platform?
In any case, it seems to do something similar to Astro, but in a significantly simpler way. The only "bundle" it needs in the browser is /@imlib/jsx-browser.js [4] which in itself is just jsx-dom.ts (its impl is overridable by the "framework" user). And on the node.js side, it's implemented as a very small "vm" of sorts [5].
I'm not against Astro, I just get all the same benefit people here are saying Astro has, but with orders of magnitude more simplicity imo.
I've used imlib to make a relatively large website [6], in fact imlib was developed as this website and extracted from it over the past year. I have absolutely no difficulty breaking down my site into various reusable and encapsulated JSX components, both on the ssg-side and the browser-side. Development time is lightning fast. IDE support is essentially automatic. The site loads instantaneously in the static parts, and as quickly as reasonable in the dynamic parts.
[1] https://github.com/sdegutis/imlib
[2] https://github.com/sdegutis/imlib/blob/main/src/jsx-strings....
[3] https://github.com/sdegutis/imlib/blob/main/src/jsx-dom.ts
[4] https://vanillajsx.com/@imlib/jsx-browser.js
[5] https://github.com/sdegutis/imlib/blob/main/src/runtime.ts
I'll look into imlib a little more.
We’ve recently moved one service from next to Astro and it was just removing a ton of boilerplate and “dance around” code.
Typescript and similar can do it because they don't have to specify the type system, and can't change it in meaningful ways over time. Things in the language standard cannot be easily changed, if they can be changed at all.
It's clearly not strictly necessary though. Python has shown that.
I mean I agree it is pretty mad to just say "you can write types but they mean whatever" but surprisingly in practice it seems to work ok.
The Python implementation can do whatever it wants because “Python” does not mean “The Python Language Specification”. It means the one specific implementation, and whatever that impl does is definitionally correct.
The ability for a language specification does to hand wave behaviour is very limited, and for JS is non existent (the only places where there is divergence between implementations is some squirrely edge cases of property and prototype chain mutation during for(in) enumeration).
So you can’t say “types mean whatever”, you have to specify what the implementation is required to do when it encounters those annotations. Even if they are not meant to have any semantic impact the lack of semantic impact must be specified: e.g the language specification would be required to state “here is the valid grammar for these annotations”, and specify that they are explicitly ignored and must not be evaluated or examined in any way.
https://lit.dev/docs/libraries/standalone-templates/
"lit-html lets you write HTML templates in JavaScript using template literals with embedded JavaScript expressions. lit-html identifies the static and dynamic parts of your templates so it can efficiently update just the changed portions."
> efficiently update just the changed portions
Since actually applying each change to the real DOM is too slow, the only way to efficiently update is to batch changes and then apply the delta to the actual DOM.
That means we need to keep track of some state, namely the previously applied state and the current goal state, which you then compare.
Now, you may have noticed that we've just independently invented the concept of diffing. And the extra state that needed to be tracked can be given a spiffy name, like "virtual DOM", since it's like the DOM, but not the real thing.
So, I'm quite unconvinced by Lit-html's claim that they are able to efficiently mutate the DOM without using a vDOM anywhere.
Either their method is not efficient (for example it falls over for rapid updates), or there is a data structure under the hood that is analogous to a vDOM, even if they prefer to give that data structure a different name.
No you're misunderstanding how it works in Python. This isn't like Rust where "the implementation is the specification".
The Python type checking standards explicitly don't define semantics (though I think they give guidelines). The standard Python implementation - CPython - does not include a static type checker. There is no "official" implementation.
In fact there are at least 4 Python static type checkers and they are all third party projects and they do differ in interpretation of types sometimes. The most popular ones by far are Mypy and Pyright (and Pyright is the far superior option).
So it is exactly the same as what is proposed for JavaScript. It definitely sounds mad and I do agree that it would be better if they just actually specified semantics, but not bothering isn't the complete disaster you might imagine.
OptionalType := (‘:’ <Expression>)?
Does the expression get evaluated? E.g
let x : Foo.Bar = …;
Is the expression unrestricted? E.g could it be (a=>a)()
If you want something to be invoked at runtime you have to specify how and when that occurs (and you’re now going to have to specify what is being passed).
You have to specify when evaluation or calls happen, etc.
The problem for an environment like JS is you cannot add a language feature and not specify the exact behaviour.
E.g it’s not “you must define a type system” (though for the parties involved in pushing this when I was involved it would have been), it’s that even if you aren’t actually interested in defining a type system you have to do a lot of design and specification work because there cannot be ambiguity or gaps where different engines will disagree on what is valid, or will disagree on what portions result in any evaluation, or what semantic effects occur. The specification also needs to handle other things that don’t matter in the Python use cases: what happens if I do have a library that does type checking, but then my code is included in an environment that also does type checking but does it differently.
In Python it’s acceptable to say “don’t do that”, but in JS that’s not sufficient, the implementations need to agree on the result, so the result needs to be specified, and ideally the specification would need to provide semantics that simply support that.
Note that none of this is unsolvable, it’s just a lot of work, and not defining the type system doesn’t remove that specification work.
https://krausest.github.io/js-framework-benchmark/current.ht...
1. It leans on support for the <template> element from browsers to hold fragments of HTML
2. For most use cases, it has a more restricted programming model compared to React or other vdom libraries, because templates are not allowed to change the shape (the tree structure of nodes) of the templated DOM.
3. For some cases where you want it to act more like React (for example, dynamically picking an HTML tag) you must use other mechanisms such as unsafeStatic. The docs explicitly tell you this immediately triggers a re-render = slow.
So I guess that answered my own curiosity: the vDOM is mostly replaced by a collection of static templates that don't need to be diffed, because the onus is on the dev to write DOM fragments where the tree does not change.
This is a more restrictive model that what React gives you, where you can generate any fragment tree, including ones with different shapes, entirely programatically.
If you do want the tree's shape to change, then lit-html isn't promising you good performance. You'll need to use methods like unsafeStatic which are slow. All in all, this is pushing more work off onto the developer.
Is this a good tradeoff? I think for most websites you can probably work within Lit's programming model. But the benchmarks you yourself linked points to many, many vDOM libraries that are about as performant as Lit (including React, whose main downside somewhat more memory usage) and has a more convenient React-like programming model.
I disagree about the trade-off in convenience tho, React programming model (especially related to managing state) is quite confusing, verbose and error-prone.
Multiple useless re-renders are the norm.
useEffects inside the body of functions a very poor way to manage the lifecycle of a component and this is way simpler with Lit.
All in all, just for it to be more like React pre-15 I would choose Lit.