Follow-up from https://wikitech.wikimedia.org/wiki/Incident_documentation/2021-03-30_Jobqueue_overload
Currently all of our jobrunners are also videoscalers. This is usually fine except sometimes in cases of high videoscaling demand can cause spikes of ffmpeg usage, overloading the host. There's also a difference in importance, the job queue is pretty important to be running smoothly, users will notice pretty quickly when it's delayed. But video scaling is usually less important as it only affects those people uploading new media and trying to use them in articles.
In this week's outage the job queue basically stalled while videoscalers sucked up all the CPU. I propose that we *always* have 2-4 jobrunners that are depooled from the videoscaling cluster that will just execute standard jobs so if there is another videoscaler overload it doesn't fully stop the job queue.