Reference · Glossary
Traffic split
Sending only a **fraction of requests** to a new version while the rest stay on stable — core of canary and A/B deploys.
When to use
Model, prompt, or index rollouts where blast radius must stay small.
When not to
Zero-traffic staging only — you need real requests to compare metrics.
Example
90% → v1, 10% → v2 for 24h; promote if error rate and latency match baseline.