What it is
A shadow deployment runs a new candidate model alongside the current production model. The candidate receives a copy of real production requests and produces predictions, but its outputs are never shown to users. The live model still serves every response.
Why do this
Offline tests use stale data, so a shadow run gives you real traffic without any risk.
- You see how the model behaves on the true production distribution
- You measure latency and resource use under real load
- You compare its predictions to the live model on identical inputs
What you watch
You log both models for the same requests and compare. You look at agreement rate, latency, and error rate. Because the candidate cannot affect users, even a crash only fails in the shadow path.
Limits
Shadow mode cannot measure the effect of the new model on user behavior, since users never see its output. For that you need a real experiment such as a canary or an A B test. Shadow mode answers does it run correctly and quickly, not does it make users happier.
Key idea
Shadow deployment validates a new model on real traffic without exposing its predictions to any user.