Electrolysis/Meetings/2016-09-22

« previous week | index | next week »

mconley

• bug 1300411 - user reports excessive load spinning when switching tabs with e10s in 48.0.2 release
  • gps pointed me at UIforETW as a way of getting really nice, low-level cross-process profiles, using a nicer front-end for Microsoft's Event Tracing infrastructure.
  • Instructions to point users to.
  • billm's patch from bug 1279086 landed last week (and just made uplift!). Still waiting on Telemetry data to saturate and give us a few more days worth, but...
    • A nice improvement on FX_TAB_SWITCH_SPINNER_VISIBLE_LONG_MS shows up September 18th, according to Telemetry
    • FX_TAB_SWITCH_SPINNER_VISIBLE_MS saw an improvement that day too, unsurprisingly
    • FX_TAB_SWITCH_TOTAL_E10S_MS started to dip that day as well, which is nice
  • I think billm and I are going to split up bug 1303077 to try to get main thread stacks from the content process when the spinner is shown for a while.
  • More reports of spinner on Reddit. I'm engaging with the user.
• Could be some cases where spawning content processes takes longer on some systems? See bug 1304393 and and this video.
  • User had no add-ons. Looks like the content process is taking its sweet time getting started. Profile is useless because the problem is solved by the time the profiler backend starts working in the content process.

chutten

bug 1218576 landed and finally stuck. This brings content-process telemetry to the parent process so we can start reporting it on non-saved-session pings (primarily, on "main" pings). This also paves the way for things like bug 1304494 to add telemetry from non-content child processes (in this case, a GPU process).

Unfortunately, not all is amazing. First was an intermittent assert failure in bug 1304009 which means that, though the child is accumulating to valid histograms, somehow on the parent it is trying to accumulate to an invalid histogram id. Then was a leak bug 1304660 which I don't understand, as I'm doing the same thing as is done elsewhere without issue.

Finally, and most troubling, there is an infrequent crash in bug 1304519 which is pointing at memory corruption of a to-be-deallocated block as an potential culprit.

I'm at a loss of how to handle this. This is my first rodeo of this kind, so any pointers others can provide are greatly appreciated.

• In other news, I've been tracking release ER, which has shown a decent bump post-e10s-enabling: [1]. Hard to say for sure, as 48's release was a bit of a nightmare, and all releases have variable e10s over time... but there's no evidence that e10s is causing a decrease in release engagement.

bytesized

• bug 1301452 - (tab switcher telemetry probe test) finished and checked in.

felipe

• bug 1304113 - Change tab switching telemetry probes to be opt-out
• bug 1304164 - Disable e10s for russian users through the system add-on
• bug 1301131 - Disable e10s for users who see long tab spinner
• bug 1297755 - Beta 50 A/B test will all multiprocessCompatible add-ons, except TabMixPlus
• bug 1302846 - Add Test Pilot add-ons to e10s whitelist

aklotz

• bug 1297549 - Main showstopper for testing screenreaders - landed on 52, some stability fallout, more Talos regressions; BUT
• bug 1304449 - A better implementation of the above that completely eliminates all of the Talos a11yr regressions since a11y+e10s landed, and also actually boosts the score by 1 - 2%
  • This is the implementation that should be uplifted to 51. The first bug is a stopgap.
  • Once this lands, the perf work is shifting to focus on measuring and improving performance of out-of-process clients