Troubleshoot Atlas safely

Follow one known observation through the signal path. Use sanitized synthetic data where possible, and never put a secret or real customer identifier into a diagnostic signal.

Triage sequence

1. Record the UTC observation window and expected source scope.
2. Confirm whether the source produced the observation and with which time.
3. Check source authentication, validation, rate, and rejection behavior.
4. Follow processing versions, filters, enrichment, aggregation, and queues.
5. Check retention and access boundaries, including tenant or purpose scope.
6. Compare end-to-end freshness and completeness with the decision threshold.
7. Verify the signal-path health observation itself.
8. Apply one reversible correction, then repeat the end-to-end check.

Symptom guide

Symptom	Likely boundaries	Disambiguating check
No observation	Source, collection, filtering, retention, or access	Trace one sanitized known event from source to consumer
Observation is late	Source clock, buffering, back pressure, processing, or query window	Compare source time, receive time, and access time
Duplicate observation	Source retry, transport retry, processing, or consumer query	Compare provenance and stable event identity where available
Value looks wrong	Source instrumentation, units, processing version, aggregation, or context	Inspect the unaggregated sanitized path and schema version
Only one consumer cannot see data	Access scope, query, visualization, or cached view	Compare authorization and equivalent raw result safely
Alert did not arrive	Source coverage, evaluation, suppression, routing, or recipient ownership	Test the full route with an approved synthetic event

Safe stop

Stop before deleting data, widening access, disabling validation, bypassing authentication, or changing retention to make a symptom disappear. Those actions can destroy evidence or create disclosure.

Escalate with product and release identities, redacted source class, UTC window, processing version, symptom, impact, and checks performed. Keep raw queries, labels, endpoints, customer data, and full exports in the approved private support system.

See the cross-product troubleshooting method.