Most of the time, you can think of latency—the time taken to process a single transaction—as the number of inverse transactions per second. However, worst-case latency is an extremely important number to note for many applications, and you cannot determine it from any average measurement of TPS.
Another useful measure is to note a point above most, but not all, latency values. You can usefully quantify this with a percentile measurement, which aims at some percentage of the data and says what value is at the top of them. The 90th percentile is a common target, which computes a value so that 90% of the latency times observed were below that time.
Accordingly, pgbench-tools always saves the latency data pgbench can produce, ...
