LLM 应用灰度发布与流量切换方案设计师

You are a senior ML platform engineer specializing in LLM deployment and release management. Design a complete canary/gradual release strategy for my LLM-powered application. Application context: - Application type: [chatbot/search/content_generation/code_assistant] - Current model: [MODEL_NAME] - New model to release: [NEW_MODEL] - Daily request volume: [NUMBER] - SLA requirements: [LATENCY_P99]ms, [UPTIME]% - Infrastructure: [cloud_provider/self-hosted] Design the following: 1. **Traffic Splitting Strategy** - Phased rollout plan (1% → 5% → 25% → 50% → 100%) - User cohort selection criteria (random, geo, feature flags, user tier) - Sticky session handling for consistent user experience - A/B test group isolation 2. **Quality Gates Between Phases** - Automated evaluation metrics: - Response quality score (LLM-as-judge pipeline) - Latency regression thresholds - Error rate ceilings - Token cost comparison - User satisfaction proxy metrics (thumbs up/down, retry rate, session length) - Statistical significance requirements before advancing - Automatic rollback triggers 3. **Monitoring Dashboard** - Real-time metrics to track (with Grafana/Datadog query examples) - Alerting rules for each rollout phase - Comparison views (old vs new model) 4. **Rollback Playbook** - Instant rollback procedure (< 30 seconds) - Partial rollback scenarios - Data handling for affected requests - Post-mortem template 5. **Implementation** - Architecture diagram (load balancer → router → model endpoints) - Feature flag configuration (LaunchDarkly/Unleash/custom) - Kubernetes manifests or serverless config for blue-green deployment - CI/CD pipeline stages 6. **Cost Analysis** - Parallel running cost estimate - Break-even analysis for model migration - Resource scaling recommendations Provide concrete, copy-pasteable configurations and code snippets.