2nd/10_Wiki/Topics/Architecture/Fault-Tolerance.md

---
id: wiki-2026-0508-fault-tolerance
title: Fault Tolerance
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Fault Tolerance, 장애 내성, Resilience Engineering]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [architecture, distributed-systems, resilience, erlang]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: erlang
  framework: otp
---

# Fault Tolerance

## 매 한 줄
> **"매 system은 fail한다 — 매 question은 'when'이지 'if' 아님"**. 매 fault tolerance는 component failure에도 system이 계속 동작하도록 design — Erlang/OTP의 "let it crash" philosophy에서 modern Kubernetes self-healing까지 evolution. 2026 cloud-native에서는 chaos engineering, circuit breaker, bulkhead가 default.

## 매 핵심

### 매 Fault vs Error vs Failure
- **Fault**: 매 root cause (bug, hardware glitch, network partition)
- **Error**: 매 fault의 manifestation (incorrect state)
- **Failure**: 매 service가 contract 위반 (user-visible)
- 매 goal: fault → error containment, error → failure prevention

### 매 Erlang Philosophy
- **Let it crash**: 매 defensive coding 대신 supervisor가 restart
- **Process isolation**: 매 lightweight process per actor, shared-nothing
- **Hot code reload**: 매 zero-downtime upgrade
- 매 WhatsApp이 2 billion users를 50 engineers로 운영한 비결

### 매 응용
1. Erlang/OTP supervisor tree (telecom, WhatsApp, Discord).
2. Kubernetes pod restart + liveness probes.
3. Circuit breaker (Hystrix, resilience4j).
4. Distributed databases (Cassandra hinted handoff, Spanner).

## 💻 패턴

### Erlang Supervisor Tree
```erlang
-module(my_sup).
-behaviour(supervisor).
-export([start_link/0, init/1]).

start_link() ->
    supervisor:start_link({local, ?MODULE}, ?MODULE, []).

init([]) ->
    SupFlags = #{strategy => one_for_one,
                 intensity => 5,
                 period => 10},
    Children = [
        #{id => worker1,
          start => {worker, start_link, []},
          restart => permanent,
          shutdown => 5000,
          type => worker}
    ],
    {ok, {SupFlags, Children}}.
```

### Circuit Breaker (Python)
```python
from pybreaker import CircuitBreaker

db_breaker = CircuitBreaker(fail_max=5, reset_timeout=60)

@db_breaker
def query_db(sql: str):
    return db.execute(sql)

try:
    result = query_db("SELECT * FROM users")
except CircuitBreakerError:
    return cached_response()  # fallback
```

### Retry with Exponential Backoff
```python
import asyncio
import random

async def retry_with_backoff(fn, max_retries=5, base=1.0):
    for attempt in range(max_retries):
        try:
            return await fn()
        except Exception as e:
            if attempt == max_retries - 1:
                raise
            delay = base * (2 ** attempt) + random.uniform(0, 1)
            await asyncio.sleep(delay)
```

### Bulkhead Pattern (Go)
```go
import "golang.org/x/sync/semaphore"

type Service struct {
    dbSem    *semaphore.Weighted  // 10 concurrent DB calls
    apiSem   *semaphore.Weighted  // 50 concurrent API calls
}

func (s *Service) CallDB(ctx context.Context) error {
    if err := s.dbSem.Acquire(ctx, 1); err != nil {
        return err
    }
    defer s.dbSem.Release(1)
    return doDBWork()
}
```

### Kubernetes Liveness/Readiness
```yaml
apiVersion: v1
kind: Pod
spec:
  containers:
  - name: app
    livenessProbe:
      httpGet:
        path: /healthz
        port: 8080
      initialDelaySeconds: 15
      periodSeconds: 10
    readinessProbe:
      httpGet:
        path: /ready
        port: 8080
      periodSeconds: 5
```

### Chaos Engineering (Litmus)
```yaml
apiVersion: litmuschaos.io/v1alpha1
kind: ChaosEngine
spec:
  experiments:
  - name: pod-delete
    spec:
      components:
        env:
        - name: TOTAL_CHAOS_DURATION
          value: '60'
        - name: PODS_AFFECTED_PERC
          value: '50'
```

### Saga Pattern (Compensation)
```python
class OrderSaga:
    async def execute(self, order):
        steps = []
        try:
            payment = await charge_card(order)
            steps.append(("refund", payment.id))
            inventory = await reserve_stock(order)
            steps.append(("release", inventory.id))
            await ship_order(order)
        except Exception:
            for action, ref in reversed(steps):
                await compensate(action, ref)
            raise
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Telecom-grade uptime (5 nines) | Erlang/OTP supervisor tree |
| Microservices REST | Circuit breaker + retry + timeout |
| Stateful distributed DB | Quorum + hinted handoff |
| Container orchestration | K8s liveness/readiness + PodDisruptionBudget |
| Cross-service transactions | Saga + compensation |

**기본값**: 매 timeout + retry + circuit breaker 3종 세트 + chaos testing.

## 🔗 Graph
- 부모: [[Distributed Systems]]
- 변형: [[Circuit Breaker]]
- 응용: [[Kubernetes]] · [[Microservices]]
- Adjacent: [[Chaos Engineering]] · [[Eventual Consistency]] · [[CAP Theorem & PACELC]]

## 🤖 LLM 활용
**언제**: 매 distributed system design 시 failure mode enumeration, supervisor tree 설계, retry strategy 추천.
**언제 X**: 매 single-process script — fault tolerance overhead 가 value 보다 큼.

## ❌ 안티패턴
- **Catch-all exception swallow**: 매 error를 log만 하고 무시 → 매 silent corruption.
- **Infinite retry**: 매 backoff 없는 retry → 매 thundering herd, cascading failure.
- **Shared fate**: 매 단일 DB 의존 모든 service → 매 single point of failure.
- **No timeout**: 매 hang된 dependency가 매 caller exhaust.

## 🧪 검증 / 중복
- Verified (Joe Armstrong, "Making Reliable Distributed Systems in the Presence of Software Errors", 2003).
- Verified (Netflix Chaos Engineering principles, principlesofchaos.org).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — Erlang/OTP + modern resilience patterns |