A production bug you cannot reproduce: what information should you collect?
An incident report template to correlate an intermittent issue with a version and prepare a reproduction without exposing secrets.
'The export does not work' is not enough
In this fictional example, someone reports that an export stays empty, then succeeds on the second attempt. The developer cannot reproduce the problem locally. A correction chosen at random may change behaviour without explaining the incident. The first task is to preserve an actionable report.
Record the first failure's time and time zone, the displayed version and the exact action. Also specify the observed frequency: one failure in two attempts is a different observation from a report without a count. Keep 'frequency unknown' when no one recorded the attempts. Collect this information within the limits of access rights and planned data retention.
A report template to copy
Attendu : exporter les commandes du jour, sans doublon
Observé : fichier vide au premier essai, second essai correct
Date : 2026-10-20T08:12:00+02:00
Version : demo-3 ; navigateur : version à relever
Contexte : filtre « validées », deux onglets ouverts
Identifiant technique de corrélation : demo-request-42
Impact : clôture retardée, étendue encore inconnue
Preuve : statut HTTP et extrait expurgé de journal
Hypothèse : concurrence entre deux traitements, non démontrée
Prochain contrôle : rejouer avec deux requêtes sur données fictivesNone of these details describes a real client incident. The correlation identifier connects logs from the same journey; it must be propagated between cooperating components. Avoid using an email address as a technical diagnostic identifier.
Connect the facts before concluding
- Record the sequence: click, request sent, response, file generation and download. Retain clock readings and known offsets.
- Connect the request to the application version and available logs. A recent interface may call an older server.
- Compare a successful case and a failing case with the same filters and fictional data.
- Separate observed facts, reported information and hypotheses. Look for a check that could disprove each hypothesis.
An error in the logs may be a consequence of the defect or concern another request. Temporal proximity alone does not prove causation. Error-tracking tools help group signals, but their collection may be incomplete.
Prepare a reproduction without copying production
Remove tokens, cookies, passwords and personal data before attaching a trace. Reconstruct a minimal dataset that preserves the property relevant to the defect: for example, two simultaneous orders without reusing customer names. A complete network capture may contain secrets even if the screen looks harmless.
Once a reproduction exists, preserve the scenario before the fix and replay it afterwards. If it is still missing, request specific access or an observation and limit instrumentation to the established need. The next delivery may be that targeted observation, with a criterion for removing temporary logs.