What is a shadow deployment of a model?
Options
- A model trained on data from other companies
- A copy of the model in another region
- A model that only runs at night
- The new model receives copies of real requests and its predictions are logged and compared, but users never see them (Correct option)
Answer
The new model receives copies of real requests and its predictions are logged and compared, but users never see them
Why
Shadow mode tests a model on real inputs with zero user impact. You can compare its predictions and latency with the live model.
Its limit: it gives no evidence on how users react, because they never see its output. For that you need a canary or an A/B test.
It does double the compute for the shadowed traffic.
Remember
Shadow = real traffic, no user impact.