Read replicas the application actually uses.

The application had been written when read replicas were "a problem for later." Later arrived. The team had read replicas provisioned but no clean way to route reads to them — the ORM didn’t differentiate, and surgically marking each query type was a months-long refactor nobody wanted to start.

RDS Proxy with a separate read-only endpoint solved the routing problem at the proxy layer. We added a thin SDK wrapper that the team adopted gradually — read-heavy endpoints first (reporting, dashboards), then the medium-traffic ones. The transactional writes stayed on the writer.

Within four weeks 41% of query volume was hitting replicas. The writer’s CPU dropped to a healthy 38% baseline. We scaled the writer down a class, which dropped the DB bill 32% net. The remaining read-heavy refactors continued in the background.