Currently, the OTel integration (#189) uses the default ParentBasedAlwaysOn sampler — 100% of traces are captured. This is fine for initial rollout but we should add configurable sampling for production use.
Options to consider
- Collector-side tail-based sampling — keep errors and slow requests, drop routine 200s. No code changes needed, just collector config. This should be the first approach, and implies no code changes.
- SDK-side head sampling — e.g.
TraceIdRatioBased(0.1) for 10% sampling. Reduces overhead on the API process. Useful at very high request volumes.
- Per-route sampling — different rates for different endpoints (e.g. 100% for webhooks, 1% for feature flags evaluation).
Context
- The
BatchSpanProcessor is lightweight, so SDK-side sampling is unlikely to be needed until high request volumes.
- Health/readiness endpoints are already excluded via
OTEL_TRACING_EXCLUDED_URL_PATHS.
- This is also relevant for eventually replacing Sentry SDK tracing with OTel-native tracing.
Related: #182
Currently, the OTel integration (#189) uses the default
ParentBasedAlwaysOnsampler — 100% of traces are captured. This is fine for initial rollout but we should add configurable sampling for production use.Options to consider
TraceIdRatioBased(0.1)for 10% sampling. Reduces overhead on the API process. Useful at very high request volumes.Context
BatchSpanProcessoris lightweight, so SDK-side sampling is unlikely to be needed until high request volumes.OTEL_TRACING_EXCLUDED_URL_PATHS.Related: #182