The Real Cost of Read Replicas (and When They Stop Paying Off)

There is a reflex every engineering team develops once read traffic starts climbing: the primary is sweating, so you spin up another read replica.

It works, but it does not actually reduce the work being done. It spreads the same expensive queries across more copies of the same database, each one independently recomputing identical results. Costs keep climbing, replication large becomes another operational concern, and every replica adds more infrastructure to manage.

We wrote about where replicas earn their keep, where they stop paying off, and what it looked like when Lemit hit a hard ceiling at 13,000 reads per second on fully maxed out hardware before pushing past 109,000 QPS by caching repetitive reads instead of adding more replicas.