Skip to content
BENIBIX Learn

Shadow Deployment

AI⚙️ MLOps & DeploymentModel Deployment

Also known as Shadow Mode, Shadow Testing, Dark Launch (ML)

A new model gets copies of real requests. Its answers are saved but not shown.

Think of it like

It is like a new driver who sits in the car and says what they would do. The real driver still steers.

Example

Lucy's bank has a new fraud model. It runs in shadow for two weeks, and the team checks its flags against the old model's before switching.

Why it matters

Test data never matches real life. Shadow mode tests a model on real traffic with no risk to users.

Definition

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.
The third panel is shadow deployment. Users get the old model's answer, and the new one's is only logged.

Common mistake

Letting the shadow model take actions. If it sends emails or blocks cards, it is not a shadow any more.

Key takeaways

  1. 1
    Real traffic. The new model sees live requests.
  2. 2
    Hidden answers. Saved, compared, never shown.
  3. 3
    Switch with proof. Swap in only after it does well.

Compare