Building a Mission-Critical Integration Blind, Part II: 100,000 Requests Later

We loaded 100,000 real requests into production. Peak CPU: 0.047 cores per container. Only 6% completed in the first hour—the rest waited on the external system (expected given its 5‑day SLA). Our queue hit 150k messages, but adding worker replicas gave near‑linear throughput. Discovered a missing composite index on (status_code, created_at). Fixed a sleep(10) anti‑pattern that blocked workers on file‑heavy requests—replaced with DelayStamp. Found that DispatchAfterCurrentBusStamp from a control...