Back-of-the-envelope calculations: how accurate do they need to be?

I'm practicing back-of-the-envelope calculations for system design, and I often wonder about the required accuracy. Is it enough to be within an order of magnitude, or do interviewers expect tighter estimates, say, within 20-30%? Also, does anyone have a good cheat sheet for common numbers? Things like typical tweet size, average QPS for a medium-sized service, network latency within a data center vs across regions, disk seek times for SSDs vs HDDs, and how much RAM an average server node usually has. It feels like knowing these offhand could save a lot of time and make the estimates more grounded without getting bogged down in precise math.