production-incident-responder

Production Incident Responder

Source mapping: Tier 2 high-value skill derived from Kotlin_Spring_Developer_Pipeline.md (SK-24).

Mission

Restore service safely before chasing perfect explanations. Keep mitigation, diagnosis, communication, and evidence preservation disciplined and explicit.

First Principles

• Mitigate first.
• Prefer reversible actions over heroic code changes.
• Preserve evidence while the system is still exhibiting the problem.
• Separate confirmed facts from working hypotheses.

Inputs To Gather

• Current alert state, user impact, and blast radius.
• Recent deploys, config changes, feature-flag changes, and dependency incidents.
• Key dashboards: latency, error rate, saturation, dependency health, queue depth, pool usage.
• Correlated traces and logs for the failing path.
• Known runbooks, rollback mechanisms, and feature flags.

Response Sequence

1. State impact and likely severity.
2. Stop unsafe changes and identify the fastest reversible mitigation:
   • rollback
   • disable feature
   • reduce concurrency
   • shed load
   • rate-limit callers
   • isolate or degrade a dependency
3. Preserve high-signal evidence while the symptom still exists.
4. Compare timeline of incident onset with recent changes.
5. Localize the failing layer: application, database, downstream dependency, queue, infrastructure, or configuration.
6. Propose long-term corrective actions only after the service is stable.

Advanced Incident Heuristics

• Restarting everything can destroy the best evidence and amplify a connection storm. Use restarts deliberately, not reflexively.
• Scaling the app tier does not help when the database or a downstream service is the bottleneck.
• Rate-limiting or queue pausing may protect core flows better than full rollback when only one feature path is toxic.
• A config-only incident can look like a code regression; compare effective runtime values before patching application code.
• A healthy dependency at low volume can still fail under retry storms from your own fleet.
• If the system uses caches, verify whether bad cache fill, stampede, or stale data amplified the incident.
• If the incident is intermittent, preserve timing and hypothesis logs. Races and saturation patterns are easy to lose after mitigation.
• Post-incident work must include detection and prevention, not only the code fix.

Incident Command Nuances

• One person should own technical command during a serious incident. Parallel debugging without a decision owner often slows mitigation.
• Communication cadence matters. Operators and stakeholders need regular updates even when the technical picture is incomplete.
• Rollback is not always safe if data shape or side effects have already changed. Assess rollback safety before pressing the button.
• Canary comparison, feature-flag cohort analysis, and effective-config diffing often localize incidents faster than code inspection.
• Preserve version, commit, config, and infrastructure fingerprints in the incident notes while they are still recoverable.

Expert Heuristics

• Choose the first mitigation that reduces blast radius and buys time, not the one that feels most technically satisfying.
• Prefer mitigations that also test a hypothesis when that can be done safely.
• If the incident spans several layers, identify the current bottlenecked layer first. Solving secondary symptoms wastes the window of action.
• A good postmortem action item changes detection, defaults, rollout strategy, or operational safety nets, not just one line of code.

Output Contract

Return these sections:

• Impact: who or what is affected and how badly.
• Immediate mitigation: the safest reversible action to reduce pain now.
• Evidence: the strongest signals collected so far.
• Working hypothesis: the leading explanation plus uncertainty.
• Next diagnostic step: the most informative next action once stable.
• Follow-up: long-term fix, monitoring change, and postmortem actions.

Guardrails

• Do not recommend code changes as the very first incident action when a reversible mitigation exists.
• Do not claim root cause certainty without evidence.
• Do not optimize for elegance over containment during an outage.
• Do not forget operator communication and blast-radius tracking while debugging.

Quality Bar

A good run of this skill reduces user pain quickly and leaves the team with a cleaner path to root cause. A bad run jumps to speculative code fixes while the service remains unstable and evidence disappears.