Work out trace sampling rate instantly with clear inputs, formula shown and shareable results.
Trace volume is traces times spans times span size, and with 24 spans per trace the multiplier is large enough that head-based sampling at one percent is common. The trade-off is that a failure occurring in one request per million is almost never captured, which is why tail-based sampling — deciding after the trace completes, keeping all errors and slow traces — is preferred where the backend supports it.
Trace volume
sampled traces = rate x sampling fraction; daily bytes = sampled traces x spans x bytes per span x 86400
The collector buffers complete traces and then decides, so it can keep every error and every slow request while discarding routine successes. It gives far better coverage at the same volume.
Yes, the decision must propagate in the trace context. Independent per-service decisions produce broken traces with missing spans.