Observability Reference/Alerting & Incidents

Incident Response & Postmortems

Incident roles, communication templates, timeline reconstruction with observability data, blameless postmortems, and action items that stick.

3/5Overview: 30m

Incident lifecycle

Detect → Triage → Mitigate → Resolve → Postmortem → Follow-up

Detect — alert, customer report, or synthetic monitor. Triage — confirm impact, assign severity, open incident channel. Mitigate — stop the bleeding (rollback, failover, scale, feature flag). Resolve — root cause fixed, SLO recovering. Postmortem — learn without blame.

Speed matters most in mitigate — root cause can wait.

Incident roles

RoleResponsibility
Incident Commander (IC)Coordinates, makes decisions, shields responders
Operations LeadExecutes technical mitigation
Communications LeadStatus page, stakeholder updates
Subject Matter ExpertsCalled in as needed

IC does not debug — they run the room. Critical at FAANG scale.

Communication

  • Dedicated channel (#inc-2026-07-10-checkout)
  • Regular updates: "Investigating", "Identified", "Monitoring"
  • Status page for external users — honest ETAs, not "we're looking into it" for 4 hours

Template: what we know, what we're doing, next update time.

Reconstructing timeline with observability

Build the timeline from data, not memory:

  1. Alert fired — burn-rate timestamp
  2. Deploy correlationdeployment.version attribute on spans
  3. Trace exemplars — slow/error traces from the window
  4. Logs filtered by trace_id — error details
  5. Metrics — when did error rate inflect?

Networking track adds wire evidence (dig for DNS, curl for endpoint health) when app telemetry points to network.

Blameless postmortems

Google SRE culture:

  • Focus on system failures, not individual mistakes
  • "What allowed this to happen?" not "who broke it?"
  • Psychological safety → honest reports → real fixes

Structure:

  1. Summary (customer impact, duration)
  2. Timeline (UTC, data-sourced)
  3. Root cause(s) — often multiple contributing factors
  4. What went well / what didn't
  5. Action items with owners and dates

Action items that stick

Bad: "Be more careful", "Add monitoring" (vague). Good: "Add SLO burn alert for payment dependency", "Integration test for retry storm scenario".

Track action items in the same system as sprint work — not a Google Doc graveyard.

Measuring incident health

  • MTTR — mean time to restore (mitigation, not root cause)
  • MTTD — mean time to detect (alert vs customer report)
  • Incident frequency by service
  • Repeat incidents — same root cause = process failure

Interview answer

"I'm IC for SEV2+. First action: confirm customer impact and assign roles. Mitigate via rollback before deep debugging. Postmortem within 5 business days, blameless, action items in Jira with owners."

Cross-reference

  • Production Debugging — triage workflow and exemplars
  • Distributed Systems — cascading failure patterns in root cause analysis

Senior signal

Distinguish mitigation (rollback stopped pain) from resolution (bug fixed). Incidents can close mitigated with follow-up tickets for root cause.

Further Reading