2nd/10_Wiki/Topic_Programming/AI_and_ML/Science of Failure.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-science-of-failure	Science of Failure	10_Wiki/Topics	verified	self	Failure Science Postmortem Culture Learning from Failure	none	A	0.9	applied	reliability postmortem sre chaos-engineering learning-org		2026-05-10	pending	language framework english SRE

Science of Failure

매 한 줄

"매 failure 는 system 의 information signal — 매 blame 의 X, 매 learning 의 O". 매 origin 은 1979 Three Mile Island 와 NASA Challenger postmortem culture; 매 modern state 는 Google SRE blameless postmortem, Netflix Chaos Monkey, Honeycomb observability + AI-aided incident review (Claude Opus 4.7 transcript summarization).

매 핵심

매 failure 의 분류 (Westrum 1988 → 매 현대 적용)

• Pathological: 매 messenger shoot, 매 hide failure → 매 pre-mortem culture.
• Bureaucratic: 매 narrow responsibility, 매 novelty crush.
• Generative: 매 high cooperation, 매 inquiry, 매 messenger trained — 매 Google/Netflix 의 target.

매 blameless postmortem 의 5 components

• Timeline: UTC, 매 minute precision.
• Impact: user-facing metric (RPS, error budget burn).
• Root cause: 매 5 whys + contributing factors.
• Action items: owner + due date.
• Lessons: 매 process change, 매 not individual blame.

매 응용

1. SRE error budget — 매 SLO violation 시 launch freeze.
2. Chaos engineering — 매 prod fault injection 으로 latent failure surface.
3. Pre-mortem — 매 launch 전 "matrix this failed, why?".
4. Game days — 매 quarterly disaster sim.

💻 패턴

매 blameless postmortem template (Markdown)

# Incident: <name> (YYYY-MM-DD)

**Severity**: SEV-2
**Duration**: 47 min (14:03–14:50 UTC)
**Impact**: 12% of /api/v2 requests 5xx
**On-call**: @alice (commander), @bob (comms)

## Timeline (UTC)
- 14:03 — deploy v2.41.0 to prod
- 14:05 — error rate alarm fires (PagerDuty)
- 14:12 — rollback initiated
- 14:50 — error rate normal

## Root cause
DB migration added NOT NULL on `users.email` w/o backfill.
Old code paths (canary not yet drained) wrote NULL → constraint violation.

## Contributing factors
- Migration runner did not block on canary drain (process gap)
- Schema diff review missed NOT NULL implication (review gap)

## Action items
- [ ] @alice — migration runner: enforce canary-drain gate (P0, 2026-05-17)
- [ ] @bob — schema-diff bot: flag NOT NULL on existing column (P1, 2026-05-24)

## What went well
- Rollback under 10 min (rollback runbook v3 worked)
- On-call comms was fast

## What did not
- Canary drain assumption was tribal knowledge

## Lessons
Migration-runner gate is the structural fix.
Not "alice should have known" — process is the fix.

매 5-whys (chained, 매 not individual blame)

Why 5xx? → DB constraint violation
Why violation? → NULL written to NOT NULL col
Why NULL? → old canary still running old code
Why canary running? → migration ran w/o waiting for canary drain
Why no wait? → migration runner has no canary-state hook
→ FIX: migration runner must check canary state

매 chaos monkey (매 Litmus / Chaos Mesh, K8s native, 2026)

apiVersion: chaos-mesh.org/v1alpha1
kind: PodChaos
metadata:
  name: kill-payments-pod-randomly
spec:
  action: pod-kill
  mode: one
  selector:
    namespaces: [payments]
    labelSelectors:
      app: payments-api
  scheduler:
    cron: "@every 30m"  # 매 prod hour 동, 매 random pod kill

매 error budget burn alert (Google SRE, multi-window)

# 매 fast burn (1h window, 14.4x rate) + slow burn (6h, 6x) — 2-window
- alert: SLOFastBurn
  expr: |
    (1 - sum(rate(http_requests_success[1h])) / sum(rate(http_requests_total[1h])))
    > (1 - 0.999) * 14.4
  labels: { severity: page }
  annotations: { summary: "Burning SLO 14.4x — page on-call" }

- alert: SLOSlowBurn
  expr: |
    (1 - sum(rate(http_requests_success[6h])) / sum(rate(http_requests_total[6h])))
    > (1 - 0.999) * 6
  labels: { severity: ticket }

매 pre-mortem prompt (매 team session)

"매 6개월 후 — 매 launch 가 catastrophic failure.
매 NYTimes headline 이 'Company X loses $100M'.
매 어떻게 그 일이 일어났을지 — 매 5 most likely scenarios 작성."

→ 매 pre-mortem 은 cognitive bias (overconfidence) 회피, 매 risk 표면화.

매 incident summarizer (Claude Opus 4.7, transcript → postmortem draft)

import anthropic
client = anthropic.Anthropic()

slack_log = open("incident-2026-05-09.log").read()
msg = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=4096,
    system=(
        "You are an SRE writing a blameless postmortem. "
        "Extract: timeline (UTC), impact, root cause (5 whys), "
        "contributing factors, action items. Never name-blame; "
        "frame failures as process gaps."
    ),
    messages=[{"role": "user", "content": slack_log}],
)
print(msg.content[0].text)

매 결정 기준

상황	Approach
매 SEV-1 user-impacting	full blameless postmortem (24h SLA)
매 SEV-3 internal-only	lightweight 5-whys (1 page)
매 near-miss (no impact)	"near-miss log" — 매 still learn
매 individual error pattern	매 process gap 분석 (매 PIP X)

기본값: 매 SEV-2+ → blameless postmortem with action items + owners.

🔗 Graph

• 부모: SRE
• 변형: Chaos Engineering
• 응용: Postmortem

🤖 LLM 활용

언제: 매 Slack/PagerDuty transcript → postmortem first draft (Claude Opus 4.7 1M ctx 으로 매 long incident 통째로). 매 5-whys facilitation. 언제 X: 매 root cause 의 final attribution — 매 human judgment 필요. 매 LLM 의 "blame" hallucination 위험.

❌ 안티패턴

• Blame culture: 매 "who screwed up?" → 매 hide future failure.
• Action-item theater: 매 owner X, due date X → 매 never done.
• Single root cause: 매 real failure 는 multi-factor — 매 swiss-cheese model.
• Postmortem-as-punishment: 매 PIP 와 결합 → 매 honesty 죽음.

🧪 검증 / 중복

• Verified (Google SRE Book Ch.15, Westrum 1988, Sidney Dekker "Field Guide to Understanding Human Error").
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — blameless postmortem + chaos eng + LLM-aided draft