[16:50:58] <mszabo>	 I have noticed that in some speedscope profiles, Excimer timers constructed by request-timeout appear to be rather expensive. Not sure if this is just a profiler artifact / misattribution
[16:51:51] <mszabo>	 for instance, it supposedly takes 13% of request time here https://usercontent.irccloud-cdn.com/file/xogZ3yrR/https_sandbox_sus1.speedscope.json
[23:15:30] <TimStarling>	 mszabo: that's interesting, I would assume it's real
[23:16:11] <TimStarling>	 you could confirm by benchmarking with $wgCriticalSectionTimeLimit = INF or 0
[23:16:20] <TimStarling>	 then it won't start a timer for a critical section
[23:19:00] <TimStarling>	 If it's an artifact, it would mean some POSIX timer function like timer_create() is causing synchronous dispatch of queued events, that's possible I guess
[23:20:43] <TimStarling>	 performance of POSIX timers might change with glibc version or some details of the kernel