Why our p99 was exactly 100 ms
A latency floor that never moved turned out to be a retry timer in a client library, not the database. Notes on how to read a flat histogram.
A notebook about databases, networking and the unglamorous parts of keeping servers alive. Mostly written so I can find it again in two years.
A latency floor that never moved turned out to be a retry timer in a client library, not the database. Notes on how to read a flat histogram.
Sizing a pool by CPU count is a coin flip. Little's law gives a better first guess, and the arithmetic fits on a napkin.
Large responses hanging while small ones fly is almost always path MTU. Here is the five-minute version of the diagnosis.