Right now, taskmasters on empty queues sleep for 30 seconds. With a default setup (4 taskmasters), this averages out to 7.5 seconds between the time you do anything that queues something and the time that the taskmasters start work on it.
On instances, which currently launch a smaller number of taskmasters, this wait is even longer.
Instead, sleep for the number of seconds that there are taskmasters, with a random offset. This makes the average wait to start a task from an empty queue 1 second, and the average maximum load of an empty queue also one query per second.
On loaded instances this doesn't matter, but this should dramatically improve behavior for less-loaded instances without any real tradeoffs.