Latency numbers every engineer should know, made human
The famous list of latency figures is easy to memorise and hard to feel. The last column rescales it: pretend one nanosecond lasts one second, and a cross-region network round trip takes about five years.
| Operation | Time | If 1 ns were 1 second | Scale (log) |
|---|
What to take from it: memory is roughly a thousand times faster than SSD, SSD is roughly a hundred times faster than a spinning-disk seek, and a trip to another continent costs about as much as 1.5 million memory reads. That is why caching, batching, keeping hot data in memory and keeping chatty services in one region show up in almost every design.
Read these numbers with care
- They are the classic figures, popularised by Jeff Dean around 2010. Treat them as orders of magnitude, not specifications. Modern NVMe SSDs and faster networks are better than the list, but the ratios between the rows are what interviews rely on.
- Same-datacenter round trips, disk and SSD figures vary a lot by hardware and load. Quote a range and say it is an estimate.
- Always say the unit. A number without "microseconds" or "milliseconds" is a common slip.
Use them in a design
- Turn these into load with the capacity calculator.
- Why a cache helps and what it costs: Caching.
- Why spreading data across machines costs a network hop: Sharding and Replication.
- Browse all system design lessons, free, with animated walkthroughs.