Coordinator step-up spawns a thread to initialize for each of the coordinator stored on disk and when there are thousands of them, preliminary tests suggests that majority of the time is spent creating new threads.