Reconstruct how one failure can become many symptoms

Trace identifiers and parent-child relationships help connect recorded work across services, provided context propagation and instrumentation are correct. A missing child span can reflect sampling, export loss, an uninstrumented hop, or work that never began. A parent span can include waiting, retries, and parallel children, so summing every child duration can double-count elapsed time. Establish whether the identifiers denote one logical request, one attempt, or a batch before using the trace to count work or locate the critical path.

W3C Trace Context standardizes propagation fields; it does not certify that a recorded relationship is an experimentally established cause of the incident. Compare traces with request counters, queue occupancy, version evidence, and resource metrics. Account for clock uncertainty before interpreting cross-host timestamps. Also include shared dependencies outside the apparent request tree, such as a host-level memory limit, model-loading storage, or a collector that competes with application work. The instrumented graph can omit the resource that actually couples the symptoms.

A user request may fan out to several downstream calls, each with its own retry policy. A retry at an outer layer can repeat an entire fan-out, while a retry within one child repeats only that child. These are different load multipliers. Read the actual client and proxy policies, including maximum attempts, deadlines, and which failures are retryable. Reporting two retries as two total attempts is another common off-by-one error: an initial attempt plus two retries is three attempts.

A simple expectation can estimate work when assumptions are declared, but overload often creates dependence and feedback. More attempts increase queueing, which causes more timeouts, which induce more attempts. A fixed independent failure probability then stops describing the system. Measure attempts per logical operation and useful success separately. A mitigation that reduces retries may improve capacity while changing user-visible completion behavior, so verify those outcomes rather than presenting a lower downstream request rate as proof that the underlying dependency recovered.

If feature-service delay is alleged to exhaust inference workers, identify the retained resource and predict how its occupancy changes with delay and concurrency. Then inspect a control path that does not use that dependency, along with local worker and downstream evidence. A dashboard that turns red later is consistent with propagation but also with smoothing delay or a shared host failure. Record contradictory observations, such as unaffected requests sharing the same supposed bottleneck, and refine the model rather than removing them from the narrative.

The queue lab shows a normalized resource imbalance only. Reset, set both λ=3 and μ=2, predict the growth from four initial work units, then Play or Step. Reset again and set both λ=2,μ=3 for a separate drain comparison. Edits retain elapsed time and recompute from the initial four, rather than switching live rates. The model has no dependency graph, retries, finite worker pool, or causality test. Use its visible imbalance to motivate measuring actual retained work along the proposed propagation path.