The freeze that started three layers away
The ticketing system froze every peak day. The load balancer logs led to the heaviest request paths, to dashboard code looping over an internal API on every login, to an API behind it that could not scale. The symptom was in one system, the cause in another, the evidence in a third. Rewrote the code path, scaled the API, season saved.
SYMPTOM → LOGS → CODE → CAUSE
When adding servers made nothing faster
A legacy application autoscaled onto a shared FSx cache, and latency never moved while AWS's own dashboards called the storage healthy. The saturation was in metadata operations, visible only after adding CloudWatch agent metrics and reading them against a load-test baseline. Moved the cache to NVMe instance volumes with stickiness. The system then held 170,000 requests per minute — 10,000 concurrent users in a 15-minute window — on architecture everyone had written off.
170,000 REQ/MIN · 10,000 CONCURRENT USERS