The Trace Pipeline stitches together the spans a single request leaves behind as it crosses services, so you can see where latency or errors come from rather than guessing from per-service metrics.
Sampled, not exhaustive
Storing every span would be enormous and mostly redundant. A head-based sampling ratio of 0.1 keeps a representative slice; errors are tail-sampled so failing requests are always retained.
A log line in the Log Pipeline
carries a trace_id; clicking it opens the full trace.
A latency spike on a dashboard links to exemplar traces captured at that moment.
Span context is propagated on every internal call so the trace spans the entire request path.
Traces are retained 7d — long enough to debug a recent incident while keeping the ClickHouse index small and fast.