Starting around February 10th, 2021 we are seeing a significant vismet regression on Fenix cold pageload live site tests:

https://arewefastyet.com/androidPixel2/cold-page-load-live/overview?numDays=60

I've reproduced the regression in local tests.

https://docs.google.com/spreadsheets/d/18qCiz3SReDgDPwhbYfuDrbnBK1030FuVWGBHWwdgCFY/edit#gid=1681334699

The regression source isn't obvious yet .
When I add profiling, sometimes the performance difference is hidden.

This profile does capture many images that never complete
https://share.firefox.dev/38lfE2P

I noticed a similar regression for COLD VIEW FNPRMS around early February (there's an unexplained, prolonged blip in mid Jan. but the regression is indicated by the before time (early Jan) and the time after (early Feb)). The regression is on the order of 100ms for the S7 and 30ms for the S10. This regression isn't present in COLD MAIN FNPRMS. Note: FNPRMS also measures the full page load but does so from network.

To single things out, we could try to find the regression in the perftest data (which might be hard because it hasn't been stable) or run our COLD VIEW perftest locally – that could point to whether the regression is before or after navStart.

I have a better profile which is showing the socket thread being tied up by the SSL SQL lock:

https://share.firefox.dev/3epIYZR

It's very likely that this is caused by Bug 1689729 (backed out of beta)

I reproduced the regression (in geckoview_example) in the push that included the patch from Bug 1689729.

See the perf-alert for that one: Bug 1691898

Andrew, would you be so kind to add an appropriate wf whiteboard tag to this bug? :)

csadilek anecdotally noticed that cold MAIN tab restoration on cnn.com take longer on Nightly than Release – I wonder if these are the same issue. We're tracking in https://github.com/mozilla-mobile/fenix/issues/18741 (profiles in that bug)

(In reply to Michael Comella (:mcomella) [needinfo or I won't see it] from comment #7)

csadilek anecdotally noticed that cold MAIN tab restoration on cnn.com take longer on Nightly than Release – I wonder if these are the same issue. We're tracking in https://github.com/mozilla-mobile/fenix/issues/18741 (profiles in that bug)

That seems very possible because CNN.com connects to numerous hosts.

The patch that introduced the delay was backed out of beta 88
https://bugzilla.mozilla.org/show_bug.cgi?id=1691898#c22

This is caused by the same patch that's tracked in Bug 1691898.