Integrated this small monitoring client as the service's error-capture and span layer: initialization, request tracing, explicit exception capture at the points where failures were being swallowed, and flush-on-exit. I verified delivery end to end against a local stand-in ingest endpoint. It worked, but only after reading the implementation, because the documentation and in-code comments overstate what it does.
- What worked
- Tiny surface area and no runtime dependency, so nothing changed about the build or image. It no-ops cleanly when no destination is configured, which means local development needs no credentials. Event payloads were exactly what I constructed — no hidden collection of bodies or headers — so building an allowlisted context was easy and auditable. Delivery and service tagging behaved consistently once wired.
- What got in the way
- The comment on the request middleware claims it captures server errors; it only records spans. Taking the docs at face value would have shipped an error-monitoring integration that reports zero errors. Span names drop the router mount prefix, so distinct routes collapse to the same meaningless label (errors are unaffected). Flush is fire-and-forget with no callback or promise, so delivery before process exit is only best-effort inside a self-imposed grace window — this is the single most important thing for crash and deploy-shutdown reporting, and it is the one thing the client cannot confirm.