Skip to content

Triage and Investigation

Triage establishes ownership, impact, and urgency. Investigation then gathers the smallest evidence set that can explain the incident without turning a read operation into a hidden mutation path.

An investigation request names the incident, target resources, time range, and latency budget. Resource analyzers read provider evidence and return structured detected issues. The coordinator builds a timeline, correlations, an optional root-cause hypothesis, and prioritized recommendations.

The report is read-only. A recommendation naming a fix is still only a proposal and must re-enter the typed action pipeline.

A short lookup and a wide sweep across a subscription shouldn’t feel the same. Investigations run in one of three modes, and the choice is about your attention, not about what FDAI is allowed to see.

ModeYou getUse it when
DirectThe full result in the same responseThe question is small and you’re waiting on it
StreamedProgress and partial results as analyzers finishIt takes a while and you want to watch it land
DetachedA task id immediately, and the result laterIt’s long, or you want several running while you do something else

For a detached investigation you can check progress, subscribe to its updates, cancel it, or fetch the finished result by its task id. Cancelling is best-effort: in-flight work stops and no result is written.

Two behaviors are worth knowing before you rely on this during an incident:

  • Cancelling undoes nothing in Azure. Investigations only read, so there is nothing to roll back. Cancelling stops the observation, and any recommendation in a partial result is still a proposal that has to go through the normal approval path.
  • Closing the tab doesn’t kill the work. A detached investigation keeps running and you can pick it back up by its task id. If the process running it is lost, the attempt is recorded as unknown rather than silently retried, and you can ask again with the same key to reuse or redo it.
  • Resource scope limits which resources an analyzer may inspect.
  • Time range prevents an unbounded history query.
  • Latency budget records whether the investigation completed in time.
  • Provider failures become unavailable evidence, not invented facts.
  • Priorities rank recommendations as P1, P2, or P3 without granting execution authority.

Evidence availability is explicit rather than inferred from a missing field. Priority is local ordering inside the report; it is not severity, confidence, or an autonomy decision unless a separate policy says so.

Evidence stateMeaningDownstream behavior
AvailableProvider returned bounded, fresh dataMay support detected issues and hypotheses
EmptyQuery succeeded with no matching recordsReport absence with query scope
UnavailableProvider failed or dependency is unhealthyMark the gap and suppress dependent claims
StaleData exists but exceeds its freshness policyHold dependent conclusions for review
  1. Confirm incident severity, owner, affected resources, and user impact.
  2. Check whether telemetry and inventory are fresh enough to investigate.
  3. Run analyzers only for the declared resource types.
  4. Build the ordered timeline before asserting causality.
  5. Separate correlated observations from grounded root-cause hypotheses.
  6. Route actionable recommendations to an incident response plan or normal action proposal.
SectionQuestion it answers
Detected issuesWhat did each resource analyzer observe?
TimelineIn what order did changes and symptoms occur?
CorrelationsWhich observations move together?
RCA hypothesisWhat cause is supported by cited evidence?
RecommendationsWhat should be inspected, simulated, or proposed next?
Budget resultDid evidence gathering finish within its declared limit?

A wedged analyzer is bounded and produces a no-action result. An exception is recorded as unavailable evidence rather than crashing the response and losing the audit trail. Cancellation still aborts the investigation cleanly.

Analyzers fail independently. Completed analyzer results remain in a partial report while failed or timed-out analyzers contribute explicit gaps. When the overall latency budget expires, the coordinator stops gathering new evidence, records whether the budget was met, and returns only supported observations. It does not fill missing sections with model prose or turn a partial report into an action.

Before using a recommendation, verify that its supporting analyzer completed, the cited evidence is fresh, and the recommendation remains within the declared resource and time scope. A high report priority can accelerate review, but it cannot bypass RCA evidence check, risk classification, or approval.

To learn aboutRead
How the incident record changesIncident management
How cited hypotheses are gatedRoot-cause analysis
How a recommendation becomes a proposalResponse plans and mitigation
How to inspect supporting recordsRead the audit log