Recently we often get dispatch lag alerts for a few minutes each (takes 5 minutes until they recover, usually).
Currently we start one dispatcher instance, which runs for 10 minutes, every 3 minutes. This means we always have 4 dispatchers running once a new dispatcher starts (for a minute), after a minute one dispatchers stops, thus we have 3 dispatchers running then until the next one starts (for two minutes).
I suggest changing this so that we always have 4 dispatchers running in parallel. Very closely related to T193349: Ongoing dispatch problems (April 2018).