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

i'm curious about the acceptable level of precision for back-of-the-envelope calculations in system design. is an order-of-magnitude estimate usually sufficient, or do interviewers expect something more granular? i'm trying to build a personal cheat sheet of common numbers for things like tweet size, typical QPS for various services, network latencies (intra-datacenter, cross-datacenter), disk I/O speeds, etc. what are the essential numbers or heuristics you rely on for quick estimates? how do you quickly sanity-check your assumptions during an interview without getting stuck in minutiae?