incident-triage-runbook

Incident triage

If you change the order below, say why in #sre.

Order of operations

1. Pull deploys for the last 6h. Don't open the log first.
2. Line the deploy timestamps up against p99_latency_ms / error_rate for the paged service. State the gap ("deploy 14:31, p99 moves 14:33").
3. If a deploy lines up: pull the diff, read it. Check for the stuff in the next section.
4. Then grep the log to confirm. Don't grep to fish.
5. No deploy lines up → the service may be a victim. Check db_pool_utilization across checkout/cart/auth/inventory, then the upstream deps (db-primary, db-inventory, idp.internal). Find what moved first.

Ask get_metrics for a name that doesn't exist and it lists the ones that do.

Things that have burned us

In rough order of how often:

• per-row query where there used to be a batch
• cache decorator removed "temporarily"
• new query, no index
• blocking call in an async handler
• retry loop with no backoff

Write-up

One line at the bottom of your reply:

Root cause: <sha> — one sentence on the mechanism.

If it wasn't a deploy, put the component or upstream dep where the sha goes (db-primary, stripe-api, whatever). Still one sentence.

Everything above that line is evidence. Keep it short; the long version goes in the postmortem doc.

If you were handed an incident id, also save the structured version to /workspace/findings/<INCIDENT_ID>.json so the postmortem can be built from it:

{
  "incident_id": "INC-0000",
  "service": "checkout",
  "title": "one line, what users saw",
  "root_cause": {"ref": "<sha or component>", "summary": "one sentence on the mechanism"},
  "timeline": [{"time": "14:31:18", "event": "what happened"}],
  "evidence": ["one fact per line, with the number or timestamp that proves it"],
  "ruled_out": ["what you checked that turned out not to be the cause, and why"],
  "impact": {"metric": "p99_latency_ms", "baseline": 65, "peak": 3640, "unit": "ms"},
  "fix": "the immediate remediation",
  "caused_by_incident": null
}