Shadow deployment, also known as a mirrored or dark launch, is a strategy where a new model runs in parallel with the current production model. It receives a copy of the live production traffic, and its predictions are logged for analysis but are not returned to the end-user. This allows the new model's performance, latency, and operational stability to be evaluated on real-world data without any risk or impact on the user experience. The results from the shadow model can be compared against the production model's results to make a data-driven decision about promotion.

A. Blue/green deployment involves switching all user traffic from an old environment (blue) to a new one (green), affecting all users at once.

B. Canary deployment rolls out the new model to a small subset of users, meaning those users are directly affected by the new model's predictions.

D. Batch deployment is a pattern for scoring data in batches on a schedule, not a real-time strategy for testing a new model with live traffic.

1. Microsoft Docs, "Safe rollout for online endpoints". Under the "Mirrored traffic (shadow mode)" section, it states, "Mirroring traffic... routes a copy of the live traffic to a deployment. Mirroring is useful for testing new features... without impacting existing clients."

2. Microsoft Learn, "Implement a controlled rollout strategy for a machine learning model". This module describes shadow deployment as a method where "the new model runs in parallel with the existing model... The predictions from the new model are not returned to the client application."

3. Azure Architecture Center, "MLOps for Python models using Azure Machine Learning". In the "Controlled rollout" section, it describes shadow mode as a technique to "test a challenger model against a production model by shadowing the HTTP requests that go to the production model."