インシデントに対応する

Every production system fails eventually. The difference between a minor blip and a disaster is how prepared and calm you are when it happens.

This lesson uses your AI assistant as an incident partner: helping you detect, diagnose, mitigate, and learn from outages without flailing.

Good alerts fire on what users feel: errors and slowness. Alerting on every internal hiccup trains you to ignore the dashboard.

Ask the assistant to define a small set of high-signal alerts tied to your golden signals, with thresholds that mean real trouble.

When an alert fires, the first job is to confirm it's real and decide if it's an incident. Declaring early gets the right people focused before things cascade.

Have the assistant draft a triage checklist so you assess severity quickly instead of staring at graphs.

During an active incident, stopping the bleeding comes before understanding the root cause. Users care that it works again, not why it broke.

Ask the assistant for fast mitigations: rolling back, disabling a feature flag, or shedding load, and the trade-offs of each.

Now diagnose. Feed the assistant the evidence you already collected: recent deploys, the spiking metric, and the error logs around the start time.

A model that sees logs, the metric shape, and what changed can correlate them far faster than scrolling alone.

Most incidents trace back to a recent change: a deploy, a config edit, a dependency update, or a traffic shift. "What changed?" is the highest-yield question in any outage.

Ask the assistant to line up the incident's start time against your change log.

During an incident, stakeholders need honest, jargon-free updates: what's affected, what you're doing, and when you'll update next.

Let the assistant draft a status update so you can keep people informed without losing focus on the fix.

An incident isn't over when the fix ships, it's over when metrics confirm users are healthy again. Closing too early invites a second outage.

Ask the assistant which signals must return to baseline and how long to watch before declaring all clear.

After recovery, capture what happened while it's fresh. A blameless postmortem focuses on the system and process, not on who to blame, so people stay honest.

Have the assistant turn your timeline and notes into a structured document with concrete follow-ups.

A postmortem that produces no change is just paperwork. The real value is converting each lesson into an automated guardrail that prevents a repeat.

Ask the assistant to translate action items into tests, alerts, or validation that would have caught this incident earlier.

Incident response is a skill you build before you need it. Alerts on symptoms, a triage checklist, fast mitigations, and a blameless culture turn chaos into routine.

Your AI assistant accelerates every phase, but the discipline of mitigate-then-diagnose and learn-then-guardrail is what keeps small failures small.

Effective incident response follows a clear arc: alert on user-facing symptoms, triage and declare, mitigate before diagnosing, then correlate logs, metrics, and recent changes to find the cause.

Communicate honestly, confirm recovery against baseline, and close the loop with a blameless postmortem that becomes automated guardrails.