Sampling traces at 1% misses the request you care about

Head-based sampling decides before the request runs, so the eight-second outlier is in the sample only by luck.

1% head-based       cheap. and the slow request is
                    almost certainly not sampled.

tail-based          decide AFTER: keep every error and
                    everything over 2s. requires buffering
                    spans until the trace completes —
                    which is a collector, and memory.

what shipped:  1% baseline, 100% on two routes, and a
               forced trace on a debug header.

Tail-based sampling is the correct answer and needs infrastructure that a modest system will not have in 2021, since the OpenTelemetry collector is still young and the PHP instrumentation is pre-1.0. The compromise covers the investigations that actually happen: complete coverage on the two routes that matter, and a way to force a trace when reproducing something. The baseline exists for aggregate questions rather than for debugging.