LEARN. PRACTICE. GROW.Separate transport from span creation
In the fictional Ria case, A sends a probe through a proxy to B. Before changing the Collector, write the expected sequence: create the operation in A, inject context into the carrier, preserve approved fields during transport, extract in B and create the dependent operation with that context. If the header arrives correctly but B creates its span before extraction, investigation should start with that sequence. A larger queue acts later and does not repair the already recorded relationship. Use two concurrent probes with distinct identities to detect context mixing. A global variable containing the last trace ID may appear to work for an isolated request and fail under concurrency. Comparison should observe the actual carrier and context used by instrumentation without collecting personal data or user credentials.
Read relationships without inventing a tree
A reconciliation process may consume three messages from independent operations. First draw the relationships you know and distinguish a parent operation from additional relationships that can be represented by links. Do not place a list of IDs in the parent field or reuse one identity to hide the difference. In the exercise, a parent lasts 500 ms and two children overlap for 400 ms each. The 800 ms sum of child durations does not turn parent latency into 800 ms or 1,300 ms. Use intervals and relationships to explain concurrent work. If B is absent from search, keep instrumentation, context, selection, export and query hypotheses open. A correct functional result does not show which hypothesis explains absence. Write the next observation point that distinguishes two hypotheses rather than listing random changes.
Distinguish context, attributes and populations
The Iris case adds an external boundary. A marker in baggage does not necessarily appear as a searchable attribute, and a client-supplied role=admin value grants no authority. Define a minimal propagation contract and use authenticated identity for access decisions. Do not copy tokens into attributes to solve deferred export. The Farol case changes retention policy to favor error traces: the error fraction in the view rises even though complete request counters remain stable. Record population, window and policy version beside each percentage. With 80 errors among 200 retained traces, 40% describes that view. It does not automatically present the rate across all requests. Investigate relevant segments before declaring service stability; a stable overall average can also hide localized degradation.
Work through the guide and prepare acceptance
The guide below allocates forty minutes: ten to reconstruct the Ria path, ten for relationships and durations, ten for Iris and Farol, and ten to write the acceptance decision. It contains data, tasks and expected answers in both languages. It was prepared editorially but has not yet been performed with participants. Produce four artifacts: a boundary sequence, a hypothesis table with discriminating observations, an interpretation of denominators and a list of outstanding criteria. To close a real experiment, require evidence of correct injection and extraction, separation between concurrent probes, handling of approved fields and search at the authorized destination. The thirty checks in the previous laboratory do not cover these criteria because they generate spans directly at the receiver. Link every claim to its supporting experiment. Keep specialist review and human execution outstanding without turning expected answers into observed results.
RIA / IRIS / FAROL GUIDE — 40 minutes, fictional data
0–10: A sends T1 and T2 concurrently. Proxy preserves both. B creates U1/U2 before extracting T1/T2. Draw the sequence and choose the first correction.
Answer: extract/activate context before creating dependent spans; repeat both probes and check no mixing. Do not increase the queue as a context fix.
10–20: Parent=500 ms. Children X and Y=400 ms each, overlapping. A reconciliation receives messages from three traces.
Answer: 800 ms is accumulated duration, not parent latency; represent additional relationships through appropriate links without inventing multiple direct parents.
20–30: Incoming baggage contains role=admin and account=synthetic-42. The external call receives both. The Farol view has 80 errors/200 traces after selective retention; complete counter=3%.
Answer: review allowed boundary fields; do not authorize using baggage. 40% describes the view; use the appropriate population for the overall rate and investigate segments.
30–40: Write four evidence requirements for real acceptance and two Collector-lab gaps.
Answer: outgoing/incoming carriers, context at span creation, probe isolation, search at the authorized destination. The lab neither instruments A/B nor validates a vendor backend.
Deliver: sequence, hypotheses/observations, denominators and decision with outstanding work.
Status: editorial exercise; not performed with participants.Ria preserves the carrier but creates its span too early; Farol changes view composition without changing complete counters.
Common pitfalls
Using global context across requests; confusing baggage with authorization; adding overlapping durations; comparing different samples.
Related topics: Context propagation · Sampling and metrics · Operational handover
An evidence chain needs explicit context, identity, population and destination; absence from search still needs diagnosis.
Reference: Context propagation · Observability 2026-09; selected OpenTelemetry, Prometheus and Dynatrace Classic concepts