The numbers

What latency numbers should I know for a system design interview?

A ladder of about a dozen, each a power of ten apart from its neighbours. A memory read is about 100 nanoseconds, a random SSD read about 100 microseconds, a round trip inside one region about half a millisecond, a spinning-disk seek about 10 milliseconds, and a round trip across an ocean about 150 milliseconds. Carry the ratios too: a seek costs 20 round trips inside a region, and an ocean crossing costs 300. Say each as a rule of thumb.

How many seconds are in a day for estimation?

Use 10^5. A day is 86,400 seconds, and rounding up to a hundred thousand turns dividing by a day into moving a decimal point five places. One request a second all day is 100 thousand a day, and a million a day is 10 a second. Then multiply by a peak factor of two to three and say which you picked.

How much traffic can one server handle?

As rules of thumb: a stateless app server about a thousand requests a second, a proxy or load balancer about 50 thousand, an in-memory cache node about 100 thousand operations, a relational node about 10 thousand indexed reads and 5 thousand synced writes. These are orders of magnitude, not measurements, so offer to change them if the interviewer has better figures.

How do I convert Gbps to MB/s?

Divide by eight, because a byte is eight bits. One gigabit a second is 125 megabytes a second, and a 10 gigabit card is 1.25 gigabytes a second. Sending 1 KB over a 1 Gbps link takes 8 microseconds.

How many servers do I need for a given peak?

Divide the peak by what one server does, then by the headroom you keep so that losing one box still fits. At 20 thousand requests a second, a thousand a server and 70 percent headroom, that is 29 servers. Sized exactly at the line, 20 servers lose to the peak the moment one dies: 19 thousand a second left against 20 thousand arriving.