Makes reasoning about overall sequences hard if using server latency, produces 'late' messages otherwise: we just need guarantee of ordering, not simultaneity.