.tip

Head-based sampling at 10 percent drops 90 percent of the failures, because the coin flips before the request knows it will fail. Decide at the end: keep everything with an error or above the p99, and 2 percent of the rest by trace id.

tail_sampling:
  decision_wait: 30s
  policies: [status_code: ERROR, latency: 800ms, probabilistic: 2]

Set decision_wait longer than your slowest request, or the slow ones get dropped as "not slow yet".