Looking at [[https://performance.wikimedia.org/arclamp/svgs/daily/2021-09-24.excimer-wall.RunSingleJob.svgz|a recent flamegraph of the jobqueue]], it seems 23% of the wall time in the jobrunners is just being spent on queuing CirrusSearch jobs. @dcausse looked at these jobs and they don't contain anything large. In fact, the time it takes to write the actual content to Elastic is around 3% while pushing the jobs that contain only metadata is taking 23%.
An edit in wikidata is also similar. [[https://performance.wikimedia.org/xhgui/run/view?id=6153427113e6381f84c80315|XHGui says 2 seconds (out of 3.3 seconds) has been spent on queuing 10 jobs.
According to reverse flamegraphs. [[https://performance.wikimedia.org/arclamp/svgs/daily/2021-09-28.excimer-wall.index.reversed.svgz|4% of wall time in appservers]] and [[https://performance.wikimedia.org/arclamp/svgs/daily/2021-09-28.excimer-wall.RunSingleJob.reversed.svgz|46% of the wall time on appservers]] is being spent on just queuing the jobs.