The Tracer

Explain the difference between metrics, logs, and traces. When would you reach for each?

How does distributed trace context propagation work? What are the W3C Trace Context headers?

What is tail-based vs head-based trace sampling? What is the tradeoff?

How would you design a time-series database for storing metrics? What are the key design decisions?

Design an alerting system that avoids alert fatigue. What mechanisms prevent flapping?

Not propagating trace context across async boundaries

Symptom: Traces break at async task boundaries; child spans from thread pool or async queue workers appear as detached root spans with no parent.

Sampling at a fixed rate without tail-based sampling

Symptom: Traces from slow or failed requests — the ones you most need — are missing from your tracing backend because they were sampled out.

Using high-cardinality values as span attribute keys

Symptom: Your tracing backend runs out of memory or slows to a crawl; index sizes grow unboundedly.

Not setting span status on error

Symptom: Failed requests appear as successful spans in your tracing dashboard; error rates are undercounted.

Creating too many spans with no aggregation strategy

Symptom: Traces contain hundreds or thousands of spans per request; the tracing backend is overwhelmed; trace rendering in the UI is unusably slow.

Distributed Observability Signals: Traces vs Metrics vs Logs

Traces, metrics, and logs are complementary observability signals. They answer different questions and have different cost and cardinality trade-offs.