Reference · Cheatsheet
Canary ship card
Safe rollout checklist for models, prompts, and RAG indexes.
Safe rollout checklist for models, prompts, and RAG indexes.
Before traffic
- v1 still deployable for instant rollback
- Smoke tests pass on v2 in staging
- Traces include `model_version` / `index_version`
Canary (5–10%)
- Split traffic with sticky session or random %
- Compare p95 latency, 5xx, groundedness vs baseline
- Hold 24h if traffic is low
Promote or rollback
| Signal | Action |
|--------|--------|
| Quality −30% DoD | Rollback |
| 2× p95 latency | Page + pause promote |
| Metrics stable | Ramp 25% → 50% → 100% |
**Related:** glossary `traffic-split`, `ab-testing`, `rollback`