A p95 latency chart tells you something is slow. It rarely tells you what. Distributed traces close that gap.
Averages Hide the Problem
Aggregate metrics smooth over the individual slow request. The endpoint looks fine on average while specific calls time out.
Follow the Span
A trace breaks a request into spans, so you see exactly which downstream call — a database query, a cache miss, a third-party API — consumed the time.
slow request
POST /api/v1/orders 812ms
billing-service charge → stripe 640ms ← bottleneck
db.query users 112ms
From Symptom to Cause
Latency is a symptom; the span is the cause. Tracing turns "the API is slow" into "the payment provider call is slow."