Sync: add ai finops and deployment notes
This commit is contained in:
31
wiki/concepts/CircuitBreaker.md
Normal file
31
wiki/concepts/CircuitBreaker.md
Normal file
@@ -0,0 +1,31 @@
|
||||
---
|
||||
title: "CircuitBreaker"
|
||||
type: concept
|
||||
tags: ["reliability", "fault-tolerance", "llm-ops"]
|
||||
sources: ["engineering-autonomous-optimization-architect"]
|
||||
last_updated: 2026-04-26
|
||||
---
|
||||
|
||||
## Aliases
|
||||
- Circuit Breaker
|
||||
- 熔断器
|
||||
- Circuit Breaker Pattern
|
||||
|
||||
## Definition
|
||||
熔断器模式是 [[AutonomousOptimizationArchitect]] 的核心安全机制——当某个 LLM Provider 的失败频率超过阈值(如 HTTP 402/429 错误、响应超时)时,自动切断该 Provider 并切换至廉价兜底方案,同时触发告警通知人工介入。
|
||||
|
||||
## Mechanism
|
||||
1. **监测**:追踪每个 Provider 的失败计数和失败率
|
||||
2. **触发**:当失败次数超过 `maxRetries` 阈值,或检测到 HTTP 402/429 错误流时,立即 trip 熔断器
|
||||
3. **降级**:所有请求切换到预配置的廉价兜底 Provider(如 Gemini Flash)
|
||||
4. **恢复**:人工确认问题解决后手动重置,或经过冷却期后自动尝试恢复
|
||||
|
||||
## Key Properties
|
||||
- **防止成本失控**:阻止 Token 消耗攻击(如恶意 bot 短时间内大量请求)
|
||||
- **防止无限重试**:每个 Provider 配置最大重试次数 `maxRetries`
|
||||
- **分级降级**:逐级切换到更便宜的备用 Provider,直到找到可用路径
|
||||
|
||||
## Connections
|
||||
- [[AutonomousOptimizationArchitect]] — 使用 CircuitBreaker 作为金融护栏的核心实现
|
||||
- [[LLMasJudge]] — 评估 Provider 降级后输出质量是否可接受
|
||||
- [[ShadowTraffic]] — 熔断触发后可异步在影子流量中测试备用 Provider
|
||||
Reference in New Issue
Block a user