We have several concurrency suites that initial sync nodes in the background while fsm workloads run (like concurrency_sharded_initial_sync). If we injected random failures into the initial syncs that required restarting them, we could get more coverage for starting servers with partially synced data.