Full page screenshot crashes Chrome on certain URLs

[patrickhulce]

Provide the steps to reproduce

1. Run LH on https://www.theverge.com

What is the current behavior?

Chrome crashes when taking the full page screenshot.

What is the expected behavior?

Chrome doesn't crash.

Other Notes

I'm not sure how widespread this sort of failure is (all my other typical test URLs seem fine), but probably worth kicking off a GCP run on a larger list without failure retries to see how many are affected if we need a workaround before v7.

I spent a bit of time investigating and here's what I found.

I originally suspected the unlimited size change so I tried reducing the max height param.
4000px tall Chrome crashes
2000px tall Chrome crashes
1000px tall renderer crashes
800px tall works

Clearly not a file size issue because the normal screenshot is 640px tall at 2.5x the DPR. The fact that there's a difference in size between Chrome entirely crashing and just the renderer crashing is interesting though. I reported all the crashes through chrome://crashes so hopefully Chromium can work on the issue from their side as well 🤷

Environment Information

• Affected Channels: CLI
• Lighthouse version: master
• Chrome version: canary and stable :/
• Node.js version: 12.18.1
• Operating System: macOS Catalina

[paulirish]

oof.

are these GCP runs headless or w/ xvfb?

potentially related issue around crashes with fullPage screenshot() in pptr: puppeteer/puppeteer#5300

[connorjclark]

xvfb-run

[patrickhulce]

are these GCP runs headless or w/ xvfb?

This was just on my local mac, but I was suggesting we try to do a GCP run to see if we can repro on more sites too. I imagine we'd want to prioritize xvfb-run as that's how more less technical users knowledgable of workarounds will experience things.

[paulirish]

Might be relevant; #11536 (comment)

[patrickhulce]

On a much broader list of URLs using our GCP scripts I could not reproduce this full page screenshot fatal error on any of them (though 4% of runs did fail with PROTOCOL_TIMEOUT on other gatherer phases, alarming result compared to my previous inability to repro PROTOCOL_TIMEOUT on GCP, but maybe encouraging for resuming our potential investigation there?)

I suspect whatever was triggering this locally for me is similar to the mac GPU quirkiness @paulirish and @connorjclark experienced in #11536. Either way I don't think it's worth blocking the release over as I don't have much evidence to say this is a widespread issue that justifies a workaround.

Any differing opinions feel free to comment and reopen :)

[ikifar2012]

Still an issue Chrome 90

[patrickhulce]

@ikifar2012 on https://theverge.com? Do you have reproduction details? I'm no longer able to repro.

[ikifar2012]

I was just trying to get a full page screenshot of a news article on TVO and my browser crashed

[ikifar2012]

Interestingly I cannot recreate it with the verge