COMPEL Glossary / shadow-deployment
Shadow Deployment
Shadow deployment (also called shadow mode) is a deployment pattern where a new AI model runs alongside the current production model, receiving the same real-world inputs but without its outputs being served to users or affecting business processes.
What this means in practice
This enables performance comparison under actual production conditions before committing to a model switch. Shadow deployment is particularly valuable for high-risk AI applications where the cost of a bad prediction is significant. In the COMPEL framework, shadow deployment is recommended as a standard practice during the Produce stage for high-risk AI systems, and its availability is assessed as part of MLOps maturity at Level 3.5 and above.
Why it matters
Shadow deployment enables performance comparison under actual production conditions before committing to a model switch, providing confidence that new models will perform well on real data without risking business impact. This is particularly valuable for high-risk applications where incorrect predictions carry significant costs. Organizations without shadow deployment capability make model updates based on test data alone, increasing production risk.
How COMPEL uses it
Shadow deployment is recommended as a standard practice during the Produce stage for high-risk AI systems. Its availability is assessed as part of MLOps maturity (Domain 7) at Level 3.5 and above. During the Model stage, shadow deployment requirements are specified in the deployment strategy. The Evaluate stage uses shadow deployment data to validate model improvements before production promotion.
Related Terms
Other glossary terms mentioned in this entry's definition and context.