Skip to content

Alert Tuning Runbook

Use this runbook when false positives, false negatives, duplicate incidents, or stale routing reduce the usefulness of an alert. It keeps tuning measurable by comparing one proposed change against a frozen baseline in observation mode, where FDAI observes and logs but does not act.

Keep environment-specific detector names, dashboard queries, notification destinations, and promotion commands in the downstream fork. Do not put customer values in this upstream template.

Start this runbook when one or more of these signals persist beyond the normal observation window:

  • False positives: alerts fire but labeled evidence shows no actionable condition.
  • False negatives: a confirmed incident has no matching detector or route event.
  • Duplicates: one condition creates multiple incidents or repeated notifications.
  • Late delivery: the detector fires on time but correlation or routing misses its deadline.
  • Stale routing: delivery targets, ownership, or escalation policy no longer match the service.

Do not use tuning to hide a currently active incident. Complete incident triage first, then return with a labeled case.

ItemRequired input
OwnerPerson accountable for the detector or route under review
ReviewerA person other than the owner who approves promotion or rollback
Scenario setFrozen positive, negative, duplicate, and delivery-failure cases
BaselineDetector, correlation, routing, catalog, and configuration versions
Guard metricsMissed incidents, notification latency, duplicate ratio, and policy escapes
Observation windowA fixed duration or event count used for baseline and treatment

The owner can prepare and run the comparison. Promotion needs an independent review so the person optimizing alert volume does not approve a weaker safety signal alone.

Before changing configuration:

  1. Confirm the scenario labels have an evidence source and review date.
  2. Record the current detector, correlation, routing, and catalog versions.
  3. Confirm baseline and treatment runs use the same scenarios and observation window.
  4. Verify observation mode results write audit records without opening, closing, or mutating incidents.
  5. Save the current configuration as the rollback reference.
  6. Pause tuning if an active incident depends on the detector being changed.

Classify the observed defect before selecting a change. One defect can have multiple symptoms, but each treatment run changes only one axis.

SymptomInspect firstTypical treatment axis
Stable noise at all hoursBaseline and thresholdBaseline window or threshold
Noise at predictable timesSeasonality modelSeasonal window or schedule
Repeated alerts for one conditionDeduplication and debounceCorrelation key or debounce interval
Related signals split into incidentsCorrelation evidenceCorrelation rule or time window
Confirmed incident had no alertCoverage and missing-data handlingDetector condition or data-quality route
Correct alert reached the wrong responderOwnership and channel policyRoute mapping or escalation policy
Correct route arrived late or failedDelivery outcome and retry auditDelivery retry or fallback route
  1. Measure the baseline. Run the frozen scenario set and capture fire rate, precision, recall, duplicate ratio, cold-start holds, delivery latency, and terminal delivery outcomes.
  2. Choose one treatment. State the defect classification, the single configuration axis to change, and the expected metric movement.
  3. Run in observation mode. Apply the treatment only to the observation mode evaluator and rerun the same scenario set under the same observation window.
  4. Compare results. Calculate baseline and treatment values for every primary and guard metric. Explain missing samples instead of treating them as zero.
  5. Review failures. Inspect every newly missed incident, policy escape, uncorrelated duplicate, and failed delivery before considering promotion.
  6. Decide. Promote only when the target metric improves, guard metrics remain within their declared bounds, and the independent reviewer accepts the evidence.
  7. Observe the rollout. Keep the prior configuration available and monitor the promoted version for the declared observation window.
ResultAction
Target metric improves and all guard metrics passApprove the configured promotion path
Target metric improves but a guard metric regressesReject the treatment and restore the baseline
Results are inconclusiveExtend or relabel the scenario set, then restart from baseline
A policy-violation escape appearsBlock promotion and route the case for safety review
Live observation differs materially from observation modeRoll back and preserve both result sets for review

Stop the run when any of these conditions occurs:

  • Unreliable labels: evidence is missing, stale, or disputed.
  • Invalid comparison: baseline and treatment use different scenarios or windows.
  • Safety regression: missed incidents or policy-violation escapes increase.
  • Hidden scope change: more than one configuration axis changed.
  • Active dependency: an ongoing incident needs the detector in its current form.

Do not suppress an alert solely to reduce volume. A lower alert count is not an improvement when detection or delivery quality declines.

Restore the recorded baseline version when promotion guard metrics fail or live observation diverges from observation mode. Rerun a small canary subset after rollback and confirm that detector, correlation, and delivery outcomes match the prior baseline. Keep the failed treatment as evidence; do not overwrite it with the rollback run.

Attach these records to the tuning decision:

  • Identity: owner, reviewer, detector or route ID, and change reference.
  • Versions: detector, correlation, routing, catalog, baseline, and treatment versions.
  • Dataset: scenario-set hash, label provenance, and observation window.
  • Measurements: baseline and treatment values for every primary and guard metric.
  • Exceptions: missed cases, duplicates, failed deliveries, and policy escapes.
  • Decision: promote, reject, extend, or roll back, with the approving principal.
  • Outcome: rollout window, rollback reference, and final configuration version.

Close the tuning work only when the decision is audited, the active configuration version is known, and either the promoted treatment or restored baseline has passed its observation window. Send newly discovered response gaps to the postmortem workflow rather than expanding this change.

To continue withRead
Scope and classify an active incidentIncident triage
Validate a service-level objective burnSLO burn response
Turn a response gap into owned follow-upPostmortem workflow