I'm trying to understand HyperHTML and how to get the best performance out of it.
Reading about how it works under the hood, it seems to imply that strong connections are made between a template and the DOM, which I take to mean that it requires a different mindset from VirtualDOM to optimise performance.
I wrote some code to show sorting of N elements in a table using hyperHtml vs normalHtml. The normalHtml version just flushes the table and rebuilds the elements. They both seem similar performance-wise. Am I comparing apples with oranges here? How can I make the hyperHtml version of the code perform better?
Code:
const numberOfElements = 10000
const items = Array.apply(null, Array(numberOfElements)).map((el, i) => i).sort(() => .5 - Math.random())
const sortMethods = [
() => 0,
(a, b) => a - b,
(a, b) => b - a
]
function hyperHtmlTest() {
const $node = document.createElement('div')
const $table = document.createElement('table')
const $button = document.createElement('button')
const tableRender = hyperHTML.bind($table)
let sortMethodIndex = 0
function render () {
const sortMethod = sortMethods[sortMethodIndex++ % sortMethods.length]
tableRender`${
items.concat().sort(sortMethod).map(item => {
return `<tr><td>${item}</td></tr>`
})
}`
}
$node.appendChild($button)
$node.appendChild($table)
$button.textContent = 'HyperHTml Sort'
$button.onclick = render
return $node
}
function normalHtmlTest() {
const $node = document.createElement('div')
const $table = document.createElement('table')
const $button = document.createElement('button')
let sortMethodIndex = 0
function render () {
const sortMethod = sortMethods[sortMethodIndex++ % sortMethods.length]
while($table.childNodes.length)
$table.removeChild($table.childNodes[0])
const frag = document.createDocumentFragment()
items.concat().sort(sortMethod).forEach(item => {
const tr = document.createElement('tr')
const td = document.createElement('td')
td.textContent = item
tr.appendChild(td)
frag.appendChild(tr)
})
$table.appendChild(frag)
}
$node.appendChild($button)
$node.appendChild($table)
$button.textContent = 'NormalHtml Sort'
$button.onclick = render
return $node
}
document.body.appendChild(hyperHtmlTest())
document.body.appendChild(normalHtmlTest())
Or on CodePen
To summarise the question: Why is HyperHTML as performant as plain ol' DOM manipulation in my code example, and how could I make HyperHTML more performant specifically when re-ordering DOM Nodes?
Update
hyperHTML 2.14 introduced the usage of domdiff V2 which brings in petit-dom like performance with the cost of an extra 0.6K to the library, hopefully worth the change.
The original demo also had huge problems on Safari for some weird reason, most likely related to the fact nodes are appended directly as TR to the TABLE, instead of into the table TBODY element.
Anyway, you can now compare performance of the wired demo through this CodePen.
Please note everything said about the old snabdom diffing also is not relevant anymore.
It seems like you could have read a bit further than that, reaching the wire part, since you got the bind one.
Basically, if you use an Array of strings as interpolated value, you are doing nothing different than an innerHTML like operation, with all the regular side effects:
To avoid replicating all these innerHTML gotchas and use hyperHTML properly, you need to wire items to the node these represent.
tableRender`${
items.concat().sort(sortMethod).map(item => {
return wire(item)`<tr><td>${item.text}</td></tr>`
})
}`
However, since memory is a concern, wire works through WeakMap so that numbers, unless used as :ids of the wire, aren't great.
The reality is that nobody has numbers or strings as items in the real world, the are 99% of the time represented by objects, and so let the object be for the demo sake.
Array.apply(null, Array(numberOfElements)).map(
(el, i) => new Number(i)
)
Once you have objects instead of primitives, which is beside the kind of object I've created for demo sake, a more realistic scenario, every time you'd invoke the render, or update, the rows won't be trashed and re-created each time, these will be simply re-ordered, as you can see in my CodePen fork of yours, basically summarized as such:
function hyperHtmlTest() {
const {bind, wire} = hyperHTML;
const render = bind(document.createElement('div'));
let sortMethodIndex = 0;
return update();
function update() {
const sortMethod = sortMethods[sortMethodIndex++ % sortMethods.length];
return render`
<button onclick=${update}>HyperHTml Sort</button>
<table>
${items.concat().sort(sortMethod).map(
item => wire(item)`<tr><td>${+item}</td></tr>`
)}
</table>`;
}
}
About Performance
Behind hyperHTML there is an engine called domdiff which purpose is to re-organize nodes.
The algorithm used in domdiff is, on average, pretty damn fast, but there are cases where it could be slower than a browser creating the same layout all at once.
You can easily spot in my pen that when you switch from ASC to DESC or vice versa, it's 5X faster than its vanilla DOM counterpart, but when it comes to ordered list to fully random one, the domdiff does a lot of checks that the DOM counterpart wouldn't care at all, so it can be slower.
In few words, while the vanilla DOM approach is linearly fast (or slow), the algorithm one has best cases and worst cases.
An algorithm that performs well in almost every case is the one used by petit-dom, but that whole logic a part has a weight that is IMO unnecessary for real-world scenarios, but surely impressive for non real-world benchmarks.
70000 Rows without a sweat
It's not a secret these days I'm working on a hyperHTML Custom Element which aim is to handle hundreds of thousands of rows, through a table that is both sortable, and searchable.
You can see an early prototype in this page, and realize it doesn't matter how a framework performs with 10000 table rows, because a good component should never put on the DOM so many nodes 'cause no user ever would be able to see all of them at once.
As you can see in that 70K table, sorting is a flash, and so is searching and scrolling, and that is HyperHTMLElement in all its glory.
These benchmarks? Not so interesting for daily, real world, tasks.
I hope this answered all your doubts.