Sync: add ai finops and deployment notes

This commit is contained in:
2026-04-26 12:35:45 +08:00
parent f09834b5a5
commit ecdf295ded
14 changed files with 7117 additions and 5832 deletions

View File

@@ -0,0 +1,31 @@
---
title: "DarkLaunching"
type: concept
tags: ["deployment", "release-management", "feature-rollout"]
sources: ["engineering-autonomous-optimization-architect"]
last_updated: 2026-04-26
---
## Aliases
- Dark Launch
- 暗启动
- 灰度发布
- Feature Flag Deployment
## Definition
暗启动是 [[AutonomousOptimizationArchitect]] 的模型引入策略——在不完全暴露给用户的前提下,将新模型部署到生产环境,通过 [[ShadowTraffic]] 验证其性能。分为三个阶段:影子测试(不返回用户)→ 灰度流量5% 用户)→ 全量切换。
## Mechanism
1. **Phase 1 - Shadow Deployment**:新模型接收影子流量,完全不影响用户
2. **Phase 2 - Canary**5% 真实流量切换到新模型,监控错误率和用户满意度
3. **Phase 3 - Full Rollout**:新模型通过所有检查后,全量替换旧模型
## Key Properties
- **风险可控**:任何阶段发现问题均可立即回滚
- **数据驱动**:每个阶段都有明确的量化指标门槛
- **与 CI/CD 集成**:暗启动可作为自动化发布流水线的组成部分
## Connections
- [[AutonomousOptimizationArchitect]] — 使用暗启动作为新模型引入框架
- [[ShadowTraffic]] — 暗启动 Phase 1 的核心实现方式
- [[CircuitBreaker]] — 提供暗启动失败时的自动保护机制