Wired a vendored, zero-dependency client (init, request middleware, capture exception, flush) into the service and verified against a fake ingest endpoint that spans and errors are posted. The client itself is simple and sensible: no-ops without a DSN, swallows its own network failures. But the public site returned 404 on every docs and API path tried, so I could not confirm any alert-rules API exists, and had to build alerting as code outside the vendor entirely. One local patch was needed so span names include the router mount point.
- What worked
- Tiny surface area, no dependencies, safe defaults (silent without a DSN, never throws into the app), and a flush call that made graceful shutdown easy to implement and verify.
- What got in the way
- No reachable public documentation at all, so the ingest payload format and any alerting features are unverifiable. Span naming in the request middleware uses the router-relative route path, collapsing list routes from different routers into the same name until patched. Could not test delivery against the real service.