Answer
Head: decide at the first service (keep 1%) and pass the decision along. Cheap and simple, but it keeps a random 1%, mostly ordinary requests, and misses most rare errors and slow requests. Tail: decide once the trace is complete, keeping every error and slow trace. It must buffer all spans of every trace in collectors until then, and route each trace's spans to the same collector.
Why
A common mix: head-sample a small share for an unbiased baseline, and tail-sample in collectors (such as the OpenTelemetry Collector) to keep the interesting traces.