Distributed Tracing Data Simulation

Real trace data is hard to obtain for demos and tests, but naive random generators produce unusable output — uniform latency distributions, flat call trees, no error correlation. A realistic simulator must respect three invariants: (1) causality — child span startTime >= parent.startTime and child endTime <= parent.endTime; (2) service topology — calls flow along a predefined DAG (e.g., gateway → auth → orders → {inventory, payments} → db), not arbitrary service pairs; (3) latency realism — per-operation latencies sampled from log-normal or Pareto distributions, not uniform, so p99/p50 ratios resemble production (typically 5×–20×).

Structure the generator as a recursive tree walk. Start from a root span for the entry service, sample its total duration from the operation's distribution, then for each downstream call in the topology, recursively generate a child span whose startTime is offset by parent.startTime + sequentialWork and whose duration is bounded by the parent's remaining budget. Emit spans in a flat array keyed by spanId with parentSpanId pointers — this matches the OpenTelemetry wire format and lets visualizations consume production and synthetic data identically.

Inject faults deterministically via a seeded RNG so scenarios reproduce: a configurable per operation, a that multiplies duration by 10×–100× for a small fraction of spans, and optional that propagates errors up the tree when a child fails. Expose the seed and topology config in the UI so users can A/B compare "healthy" vs "degraded" traces side by side — this is what makes the demo teach the visualization, not just decorate it.