[01:48:11] dpifke: so.. https://phabricator.wikimedia.org/T259952 is afaik the way to trigger the "main()" notice [01:48:30] that's the only way to have tideways be consumed alreayd and thus be empty when our xhgui saver draws the short straw [01:48:38] in case you didn't have a repro :) [02:04:08] > 430702300.00% 4307.023 1 - wfIndexMain [02:04:16] oh right, we haven't fixed that silly bug yet :D [02:04:39] 100600.00% 1.006 81 - MediaWiki\MediaWikiServices::getGenderCache [02:10:13] gender is important [02:21:56] it seems somewhere in XhprofData.php or ProfilerXhprof.php, we're making %real effectively be just real * 100 [02:22:17] probably a /$total is broken somewhere where $total is 1 or something like that. [09:32:16] FOSDEM has finally made the final video links available, I've updated the Google doc with links to those. They seem to be higher quality transcodes than what Google Drive does [09:56:52] the whole video of welovespeed if you want to check it out: https://www.youtube.com/watch?v=6aupyVi_QMU&feature=youtu.be [12:45:34] how hard would it be to split the flamegraphs by stretch vs buster (on a temporary basis)? [12:45:43] the excimer flamegraphs (https://performance.wikimedia.org/php-profiling/) [16:56:24] legoktm: not very, we've done it before. Why? [16:58:29] We could emit from wmf config with extra virtual frame, and generate extra graphs from it like for some other factors we do already. [17:22:13] excimer + excimer-stretch + excimer-buster, as channels [17:22:35] Would be simplest probably especially with new headroom we have so don't have to squeeze it in a single log file [21:11:20] legoktm: ok, I understand now :) [22:24:28] Krinkle: yep, I was going to do a bit of CLI profling with parse.php but I think the excimer flamegraphs would make it pretty obvious where the difference [22:25:32] legoktm: do we have an mwdebug server of each flavour? [22:25:40] could do some xhgui profiles ad-hoc [22:25:43] yeah, let me check [22:25:45] might be quicker if it's obvious [22:25:59] otherwise yeah we can set up some new profiling but probably not by Tue/Wed next week [22:26:05] mwdebug1003.eqiad.mwnet is buster [22:26:10] 1001 and 1002 are still stretch [22:26:28] couple of warm ups first and then profile with xhgui wikimeidadebug :)