32 lines
1.3 KiB
Markdown
32 lines
1.3 KiB
Markdown
---
|
||
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
|