Reference · Glossary
Canary deployment
Last updated
Releasing a change to **a small slice of traffic first**, watching metrics, then promoting or rolling back.
When to use
New model versions, prompt changes, or RAG index updates in production.
When not to
Low-traffic demos where you can test manually before any deploy.
Example
10% of chat requests hit `gpt-4o-mini-v2` while 90% stay on v1 until error rate stays flat.