Update nexus wiki content
This commit is contained in:
39
wiki/concepts/Shadow-Traffic.md
Normal file
39
wiki/concepts/Shadow-Traffic.md
Normal file
@@ -0,0 +1,39 @@
|
||||
---
|
||||
title: "Shadow Traffic"
|
||||
type: concept
|
||||
tags: []
|
||||
sources: [engineering-autonomous-optimization-architect]
|
||||
last_updated: 2026-05-01
|
||||
---
|
||||
|
||||
# Shadow Traffic
|
||||
|
||||
## Definition
|
||||
|
||||
暗启动(Shadow Traffic)——将一小部分真实用户流量异步复制到实验模型,在不影响生产环境的前提下验证新模型的真实表现。
|
||||
|
||||
## Key Characteristics
|
||||
|
||||
- **真实数据**:使用实际用户请求,而非合成测试数据
|
||||
- **异步执行**:影子请求不影响主请求的响应时间
|
||||
- **隔离评估**:影子结果仅用于评分,不影响用户可见输出
|
||||
- **渐进放量**:初始 1-5% 流量,成功后逐步提升
|
||||
|
||||
## Workflow
|
||||
|
||||
1. **Phase 1: Baseline**:建立当前生产模型的基准评分
|
||||
2. **Phase 2: Shadow**:将 5% 流量同时路由到实验模型
|
||||
3. **Phase 3: Evaluate**:LLM-as-a-Judge 评估实验模型输出
|
||||
4. **Phase 4: Promote**:统计显著优于基准时,自动提升流量比例
|
||||
|
||||
## Safety Guarantees
|
||||
|
||||
- 用户永远只收到生产模型的输出
|
||||
- 实验模型超时/错误不影响用户请求
|
||||
- 熔断器随时保护异常端点
|
||||
|
||||
## Related
|
||||
|
||||
- [[Autonomous-Optimization-Architect]]:设计并执行影子测试的核心 Agent
|
||||
- [[LLM-as-a-Judge]]:对影子流量结果进行量化评分
|
||||
- [[Circuit-Breaker]]:影子测试期间的异常保护机制
|
||||
Reference in New Issue
Block a user