Question 280
A registered model is later copied outside its workspace without its input transformation definitions. What should be included in the artifact packaging?
Explanation
Correct Answer
Inference portability requires the transformations used to produce the model's expected input features, not just trained coefficients.
Incorrect Answers
The alternative, Use the model's logged dependency specification to align production runtime libraries, belongs to a different technical condition: A scoring environment runs successfully in development, but the registered MLflow model has dependency versions incompatible with the production image. There the source-specific mechanism or proof is: MLflow packaging can capture required dependencies; incompatible framework versions can break an otherwise valid artifact. The condition asked about here is: A registered model is later copied outside its workspace without its input transformation definitions. A suitable correction is Package the required preprocessing code or model pipeline within the MLflow model artifact; the targeted justification or result is: Inference portability requires the transformations used to produce the model's expected input features, not just trained coefficients.
Retain the model version, producing run, environment, and related metadata under the retention policy. That operation would be relevant if: A model must be archived for governance but remain reproducible and inspectable. Its corresponding reason or verification is: Auditability depends on both artifacts and provenance; removing every associated trace makes past decisions unreproducible. However the problem in this question is: A registered model is later copied outside its workspace without its input transformation definitions. The requirement instead calls for Package the required preprocessing code or model pipeline within the MLflow model artifact; the question-specific check is: Inference portability requires the transformations used to produce the model's expected input features, not just trained coefficients.
Log the model in a supported MLflow flavor and register the resulting artifact. A separate source case supports that approach when: An MLflow training job logs loss and parameters but never logs the trained model artifact. The team wants to register it. The supporting rationale or observable result is: Registration needs an actual model artifact; scalar training metrics alone are not a deployable model package. This question describes a different issue: A registered model is later copied outside its workspace without its input transformation definitions. Address it through Package the required preprocessing code or model pipeline within the MLflow model artifact, with the required justification or evidence: Inference portability requires the transformations used to produce the model's expected input features, not just trained coefficients.