In shadow deployment for a new ML model, what distinguishes it from both A/B testing and canary deployment?
Select an answer to reveal the explanation.
Short Explanation
Here's the deal — you're right to pick the option where all production traffic goes to both models, but only the incumbent model's predictions affect actual decisions. The new model runs invisibly for evaluation, so you can test it safely without customer impact.
Full Explanation
The correct choice sends all production traffic to both models, but only the incumbent model's predictions affect actual decisions; the new model runs invisibly for evaluation. The new model receives all production traffic and generates predictions, but those predictions are never used for actual decisions — only the production model's outputs affect customers. This allows safe evaluation with real traffic without customer impact. The option sending a subset of traffic to the new model and acting on its predictions describes split-testing or canary behavior. The option gradually increasing the new model's traffic share based on performance thresholds describes canary's progressive promotion. The option requiring separate infrastructure is an infrastructure consideration, not the defining characteristic.