Skip to content
AITroveRead. Build. Understand.
Reading list
Make this comfortable

Incident impact: window, denominator, and scope

Last updated: 7 Oct 202611 min read
tutorial
AdvancedBy AITrove Editorial

An impact statement needs a measurement window, population, unit, numerator, denominator, and source identity. A rate without a denominator invites readers to assume that every customer or request is affected. The prompt should ask whether retries, synthetic traffic, duplicate attempts, and delayed records are counted. Report measured attempts as attempts, not people. If a snapshot is incomplete, say it is preliminary and avoid a false precision. When several regions or payment paths differ, preserve their scope rather than averaging away a concentrated failure. The incident team should be able to reproduce the arithmetic from the frozen snapshot used for approval.

Operational case

Harbor's metrics lead freezes the 14:00-14:10 UTC checkout API window. It contains 18,600 attempts and 3,720 failed attempts, so the failure share is 3,720 divided by 18,600, or 20%. The numerator is failed checkout API attempts; the denominator is all checkout API attempts in that ten-minute window. This does not mean 20% of customers failed to buy, because one customer may retry and the snapshot says nothing about later completion. The public draft says 'about one in five checkout attempts failed during the measured window' only after the lead confirms the counter definitions and ingestion completeness.

Output
Window: 14:00-14:10 UTC.
Population: checkout API attempts in the measured service.
Failed attempts: 3,720; total attempts: 18,600.
Rate: 3,720 / 18,600 = 0.20 = 20%.
Do not translate attempts into affected customers.
Snapshot ID: HC-1410; counters and ingestion checked by metrics lead.

Performance and operating cost

For N event records, a direct count is O(N) time and O(1) aggregate memory; a precomputed metrics snapshot can reduce query time but adds ingestion and label-validation work. A single numerator and denominator are cheap to compute, while checking retries and missing telemetry takes longer. Freeze the window and snapshot ID in the draft so later recomputation does not silently change an already approved statement. If telemetry is incomplete, publish the bounded observation rather than a precise percentage that the team cannot defend.

Common Mistakes

  • Do not describe request attempts as distinct customers.
  • Do not combine a numerator from one window with a denominator from another.
  • Do not imply a measured rate covers an unmeasured region or path.

Connected lessons

prompt engineering
incident communication
Storage details