Talktodre-ops

latency-tracing

Talktodre-ops/claude-skills/latency-tracing/SKILL.md

How to find out where a slow API request actually spends its time, distilled from an investigation on a Django/DRF app whose dashboard was timing out. Covers the measurement ladder (the floor, the count, execution versus distance, behaviour under a burst), wiring OpenTelemetry and Jaeger v2 into a forking WSGI server without getting the two halves backwards, reading a waterfall for the four things worth looking for, and the fixes each finding implies: removing no-op round trips, caching a hot read without letting new data wait behind it, collapsing an N+1, and the worker model. Includes the evidence needed before reaching for an index or a search cluster, and the traps that make a developer misattribute the cost. Load when an endpoint is slow, when instrumenting a service for latency work, or when deciding whether a database change is justified.