Shadow Deployment vs Canary Release
People think a shadow deployment and a canary release are the same careful way to launch a model, and they are not. In a canary, a few real users get the new version and see what it does. In a shadow deployment, no user sees the new model at all. It gets a copy of real requests, and its answers are only saved. A canary tests on people. A shadow tests in secret.
Shadow Deployment
A new model gets copies of real requests. Its answers are saved but not shown.
Shadow deployment is a way to test a new model on real traffic without users seeing its answers. Each live request still goes to the old model, which answers as usual. A copy of the request also goes to the new model. Its answers are saved and compared, but never shown. If the new model does well, it can replace the old one with less risk.
Canary Release
Give a new version to a few users first, then to more and more.
A canary release sends the new version to a small share of users first, like 5 percent. The team watches errors and speed for that group. If all looks fine, the share grows step by step until everyone has it. If errors rise, the canary is pulled and only a few users were hurt. The name comes from canaries that miners once took underground to warn of bad air.
In one line
- Shadow Deployment
- A new model gets copies of real requests. Its answers are saved but not shown.
- Canary Release
- Give a new version to a few users first, then to more and more.
These live in different corners of the glossary (AI and Development), which is exactly why they get confused.