What is distributed tracing, and how does it work?

Question

Accepted Answer

Distributed tracing follows a single request end to end as it moves through every service that handles it. A trace is made of spans, where each span is one timed operation, an incoming request, a downstream call, a database query, and spans have parent-child relationships that form a tree showing the structure and timing. The mechanism that makes it work across services is context propagation. When one service calls another it passes the trace ID and the current span ID along, usually in the W3C traceparent header, so the receiving service creates its spans as children under the same trace. Stitch those spans together and you get a waterfall of where the time went. That is why tracing is the pillar for distributed latency. Metrics told us p99 checkout latency had gone from 400ms to 1.2 seconds, and logs told us what happened inside one service, but the trace showed the actual culprit: a single downstream call fanning out into 30 sequential database queries, about 700ms of the total. Because tracing every request is expensive, I use sampling, head-based or tail-based, often keeping something like 1 percent of normal traffic while tail-sampling holds on to every error and every slow request, so cost stays sane and the interesting traces are still there.

Answer

How does a trace span multiple services?

Context propagation: the trace ID and span ID are passed across service boundaries (typically via HTTP headers like traceparent), so each service attaches its spans to the same trace, reconstructing the full path.

Answer

Why and how do you sample traces?

Tracing every request is expensive at scale. Sampling (head-based at the start, or tail-based keeping slow/errored traces) reduces volume and cost while still capturing representative and interesting traces.

What is distributed tracing, and how does it work?

What is distributed tracing, and how does it work?

The follow-ups they ask next

What the interviewer is listening for

What sinks the answer

If you genuinely do not know

Keep going with observability

Knowing the answer is not the same as recalling it under pressure