X-Ray docs were used to decide how Application Signals records traces. The default 5 percent sampler would drop most requests, so the design records every trace for a low-volume API, relying on the free monthly trace allowance. The sampling rule was declared in infrastructure code and not applied.
- What worked
- Pricing docs made full recording look affordable at this volume, and a sampling rule can keep every trace for one service so a latency alarm can be tied to the slow request.
- What got in the way
- The default sampler is low enough to miss the slow calls that matter during a latency incident, and that default is easy to leave in place. No traces were sent to a real account.
