[austincheney]

I love how big the numbers are in recent versions of Firefox. On my crappy work computer that can barely run Notepad++ I am pulling 1.8 billion ops for getElementById and getElementsByClassName. The performance gaps:

* getElementById vs equivalent querySelector is about 1000x.

* getElementsByClassName vs querySelectorAll is about 250,000x.

* On Chrome getElementsByClassName vs querySelectorAll is only about 426x.

This is a micro-benchmark. In the real world that disparity would magnify almost exponentially with the number of nodes in a given page and the frequency and depth of access requests to the DOM.

[UberMouse]

I decided to check back in this and this seems like it was entirely a wording problem in your first post. Your post implied that _parsing the query selector string_ was slowing down Javascript websites, which obviously made everyone do a hottake and jump in to say that can't be the case. What you seem to _actually_ be saying is that using query selectors instead of finding elements directly by class/id is slow which yes, that's certainly the case.

[austincheney]

Query selectors are slow, yes, but its because they have to parse a string. Other DOM methods don't have a string parsing step. You cannot optimize access to the DOM if there is a string parsing step that must occur first.

That barrier to efficiency is greatly magnified by the complexity of the query string, the size of the dynamically rendered page, and the number of query strings. If not for that string parsing step why would query selectors be any different from any other DOM access instruction computationally?

[filleduchaos]

You're quite frankly leaping to conclusions quite a bit, as well as starting from the (rather flawed) premise that parsing a, what, ten-character string is so slow that it can slow an operation down by three orders of magnitude.

There are any number of reasons why the querySelector API would be computationally different from the getElementX ones, especially considering that they don't even return the same thing.

[austincheney]

They return exactly the same thing: either null or a node or node list depending upon the method in question.

Any number of reasons like what? Could you provide an example of what would make those methods that much slower?

[filleduchaos]

> They return exactly the same thing: either null or a node or node list depending upon the method in question

...this is another thing that I'm surprised to find that people don't know. Do devs really just use these methods without ever looking at what they return or how they behave?

getElementsByClassName returns a (live) HTMLCollection, not a NodeList. querySelectorAll returns a (static) NodeList. That in itself is an obvious computational difference/potential bottleneck, because the simplest way to implement a live collection is to cache and return the same object on subsequent calls for it. And that's precisely what browsers do (getElementsByClassName('foo') === getElementsByClassName('foo')). In other words, getElementsByClassName called multiple times with the same class doesn't actually do any extra work. The real work for the browser engine, which doesn't actually happen at the point of function call, is watching any tracked class names and updating their associated HTMLCollections when an element matching that class name is added or removed from the DOM.

On the other hand, gathering the static NodeList for a querySelectorAll call requires actually iterating over the DOM to find element(s) matching the selector every time (in the naive implementation), with the trade-off of the engine not having to watch the collection internally.

As an aside, the querySelector method with an ID selector is slower but on the same order of magnitude (a difference of about a few hundred thousand ops/sec for me) as the getElementById method. So if one looks at all the data in the benchmark and not just the class ones in isolation, it becomes clear that merely parsing the selector string is not enough to drop millions/billions of operations per second down to single-digit thousands.