doc: weird scrolling behavior in Chrome

[mscdex]

📗 API Reference Docs Problem

Location

Affected URL(s):

• https://nodejs.org/docs/latest/api/modules.html

Description

At least on Linux and with Chrome, if you load a very long page, like the CommonJS modules API page, and you either press the End key or otherwise scroll to the bottom and then attempt to scroll up or hold down the Page Up key, the scroll bar makes a lot of strange jumps (backwards and forwards). Even starting at the top of the page and scrolling down causes the scroll bar to do some large, sudden, and unexpected jumps. However, using the Home and End keys seemingly work just fine.

My guess is there is either some javascript on the page listening for scrolling events in general or on particular elements (like code blocks) and that is somehow causing the weird issues.

[aduh95]

We're using the CSS property content-visibility (https://web.dev/content-visibility/) on the HTML version of the docs so it doesn't take forever to render pages on Chromium-based browsers. One of the downside of using this property is the scroll bar makes indeed some strange jumps. It was introduced in #37301.

[mscdex]

I wonder if the kind of solution described here or something similar would solve these problems?

It's very frustrating not being able to properly navigate the documentation a page at a time or by sliding the scroll bar. It seems to me like the browser should keep elements rendered after it's rendered the first time, which should solve the scrolling problem. It seems like the solution I linked to explicitly does this?

[aduh95]

Yep that looks promising indeed. Would you be interested in sending a PR? On the mean time (and I know this is not a satisfactory answer to that issue, more a temporary workaround), you may find Firefox more suitable to browse the docs as they haven't implemented content-visibility and they also able to render large pages in a reasonable time.

[himself65]

It works fine on macOS Chrome

[szmarczak]

I think this also causes an issue if you try to open any URL with a hash in a new tab. It doesn't scroll to the desired method / property.

[saltybuckets]

I am working on this rn

[noinkling]

I hate to pile on but this issue has made browsing the API docs an exercise in frustration for a while now. The scrollbar jumps all over the place when trying to drag it, as well as the yellow indicators when you do a ctrl+f. I never end up in the correct place when following an anchor link.

I'm surprised there haven't been more reports after all this time, surely it's more than just a handful of us affected? Chrome on Windows here. Firefox seems fine as mentioned.

[ovflowd]

@mscdex, for reading the current open Issue, for me, the takes are the following:

• The original content-visibility was introduced to reduce the re-paint on resizing of pages or other shenanigans
• It also had the benefit of a smaller initial rendering footprint

Then there's the discussion about if it should be removed:

• It would fix anchor link positioning behaviour, which is an important aspect of the experience
• It would allow scrolling behaviour to be respected

There's, of course, the approach of doing some JavaScript spaghetti, as I saw in some random blog post. Still, I'm severely against having arbitrary JavaScript code that tries to "fix" the web (by inherently trying to change how certain behaviours happen).

What I imagine as an "ideal" solution for now:

• Remove the content-visibility API.
  • 99th percentile of devices nowadays are efficient enough to render even the longest pages hassle-free. Of course, constant resizes of the page on desktop browsers will increase the CPU heap during the resize by a lot, but that's an OKAY outcome. People will only sometimes resize their pages; that is an edge case, imho.
• Add CSS grids for the two columns.
  • This page was probably done before the time of CSS grids.
  • It would allow us to remove a few margin-padding hacks
  • It would fix the width issue that happens due to content-visibility being removed
  • Enforces the maximum width of the second column and respects the page size

Regarding modern technologies:

Modern browsers and rendering engines are already smart enough to offload computing from complex parts.

I did some synthetic loads on the current API website with content-visibility disabled and requested Chrome to profile the page. I tried to do intensive scrolling behaviour to check for bottlenecks. I saw no significant bottlenecks, CPU-heap increase or significant delays/increase on the repainting/re-rendering of the page.

Moreover, disabling content-visibility shouldn't be a problem. I tried the same behaviour on an iPhone by using remote DevTools to disable those CSS rules. Didn't see any noticeable lag.

Of course, some real old phones or hardware would struggle a little. But given that technology is ever-evolving, I doubt someone would consistently open our API docs on a tiny smartphone from 7 years ago...

I feel inclined toward the approach/solution I mentioned above.

[aduh95]

• 99th percentile of devices nowadays are efficient enough to render even the longest pages hassle-free.

I wonder where you get that figure from, in my experience loading the all.html on a Chromium browser was a real struggle for my 2019 MBP, and I’m pretty sure that’s not the lowest spec you’ll find out there. But maybe Google or Microsoft fixed that since I made the measurements reported in #37301 (comment)

[ovflowd]

Well, the percentile is based on Browserslists on which browsers supporting supporting the 95th percentile. (https://browserslist.dev/?q=Y292ZXIgOTUl)

And sorry, I meant the 95th percentile not the 99th.

I can at least on all my devices open it perfectly without any issues. Using Chrome DevTools to mimic hardware constrained situations it also seems to work fine.

Of course to a certain degree I'm doing an opinionated guess.

[aduh95]

To clarify, do you mean that building the docs locally commenting out the content-visibility from the CSS has no effect on the rendering time? That’s surprising to say the least, but a very good news if true (I can’t confirm, I’m away from computer). Can you share the rendering time on both versions please?

[ovflowd]

To clarify, do you mean that building the docs locally commenting out the content-visibility from the CSS has no effect on the rendering time?

No need to rebuild the docs; just disable the property directly on the browser.

That’s surprising to say the least, but a very good news if true (I can’t confirm, I’m away from computer). Can you share the rendering time on both versions please?

I'm not sure I can give you a reliable rendering time or statistics as they're bound to my computer. But by even doing a 6x CPU slow down and checking GPU raster it sounds OK. Feel free to check by yourself.

[aduh95]

No need to rebuild the docs; just disable the property directly on the browser.

Yes you need to rebuild the docs, because content-visibility (used to?) affect the rendering time, so disabling it from DevTools won’t help there because it won’t let you disable it until after the first rendering is done (unless I’m missing something). I can try later, as I said I’m away from the computer at the moment.

I'm not sure I can give you a reliable rendering time or statistics as they're bound to my computer

That’s OK, what I’m interested in is the difference (on the same CPU) of rendering time, absolute values don’t matter in this case.

[ovflowd]

Yes you need to rebuild the docs, because content-visibility (used to?) affect the rendering time, so disabling it from DevTools won’t help there because it won’t let you disable it until after the first rendering is done (unless I’m missing something). I can try later, as I said I’m away from the computer at the moment.

Well, you know you can just download all the assets, change the style.css and then run a local server 😅

I'm not on my MacBook, hence why I didn't have the tools to build the docs. But afaik just downloading a copy of the built website (Ctrl + S) + Editing the CSS files + some random simple node server should be enough.

That’s OK, what I’m interested in is the difference (on the same CPU) of rendering time, absolute values don’t matter in this case.

Well, I saw a difference of just shy ~500ms (Mostly) on the Rendering Time. Note that I started the profiling before opening the Website.

[aduh95]

OK so I've got back to my computer, and tried on Chromium 112.0.5615.137 on my M2 MBP: using content-visibility: auto makes the rendering happens 6x faster, and it takes 70% of the time to load the same page.

With content-visibility: auto: rendering takes 269ms:

Without content-visibility: auto: rendering takes 1732ms:

(I promise I didn't cherry-picked it, I run the performance profiller once on all.html for each and took a screenshot; still using a sample of 1 is not a very scientific way of comparing the two, but IMO it's really not insignificant).

[ovflowd]

(I promise I didn't cherry-picked it, I run the performance profiller once on all.html for each and took a screenshot; still using a sample of 1 is not a very scientific way of comparing the two, but IMO it's really not insignificant).

Yeah agreed. For me, for example, the rendering time with content-visibility costs me around 870ms, without content-visibility, I get around 1540ms. So it is also dependent on your hardware, which is fair. Regardless, we see a consistent 120% uplift in rendering time, which is fine.

But honestly speaking, I feel that in the future (with the API Docs redesign), we should eliminate the "all.html" page. It genuinely doesn't make sense to have a gigantic page containing everything simultaneously. No other Docs Page I know does this. All other pages on our API Docs, including http.html, take around 33ms to load... So, all other pages have reasonable loading times.

[aduh95]

fs.html is another huge one that you can test to see how it affects the load. We can decide that it’s fine indeed, and ask folks to use a non-Chromium browser if they hardware has a hard time loading our docs, as long as that trade off that might work for most of our visitors (is it possible to have data on that?) and we are open about making that choice.

[ovflowd]

I genuinely feel right now the fact that we have content-visibility is actually breaking features of the docs pages such as scrolling and anchor links.

Not to mention the redesign docs will pretty much not have this, so my 2 cents:

• either we remove content-visibility and add the css grid
• we just ignore this until we switch to the new redesigned docs page that is owned by the Website Team

[dholbert]

Until this is hashed out, there's a trivial incremental improvement that we can make to mitigate some of the usability issues here (which are particularly bad/unusable in Firefox Nightly & Firefox-with-content-visibility-enabled-in-about-config). We can avoid much of the trouble by adding auto to the contain-intrinsic-size CSS declaration, which just lets the browser save the sizes that it lazily computes, once it's been forced to compute them (by scrolling an element on-screen).

See #48195 - if anyone with the ability to review/merge PRs here wouldn't mind taking a look over there, that would be much appreciated.

[KuthorX]

@mscdex this issue maybe resolve by pr above, could you please check it?