Grouping SageMaker Model Groups into Collections Without Modifying Underlying Model Artifacts

Why the Answer Is Correct

The requirement has two parts: organize the existing model groups into the three categories to improve discoverability at scale, and do so without affecting the integrity of the model artifacts or their existing groupings. A Model Registry collection sits above model groups as a logical grouping layer, and the documentation is explicit that any operation performed on a collection does not affect the integrity of the individual model groups it contains, because the underlying model group artifacts in S3 and ECR are not modified. That is precisely a non-destructive way to add a category level on top of the untouched registry structure. The vote was unanimous at 100 for D, with Linux_master, abrarjahin, GiorgioGss, and Saransundar all quoting the same documentation sentence about collections not affecting the integrity of contained model groups.

Why the Other Options Are Wrong

Creating a new model group for each category and moving the existing models into them (B) is the option that violates the constraint, because moving models out of their current groups restructures the registry and changes the existing groupings the question explicitly says must stay intact, and it also breaks the association between those models and the artifacts already recorded under their current group. Creating custom tags for each category and adding them to the model packages (A) is additive and non-destructive, and a4002bd and GiorgioGss both acknowledged it is a valid approach, but tags are metadata on individual packages rather than a grouping construct, so they do not deliver the improved discoverability at scale that organizational grouping provides. Using SageMaker ML Lineage Tracking to automatically identify and tag the model groups (C) misapplies a lineage feature, which tracks lineage relationships between models, datasets, and artifacts, and it cannot reorganize models into categories.

Community Comment Notes

The community was unanimous at 100 for D, and the discussion is notable because option A was recognized as genuinely defensible rather than merely wrong. a4002bd stated that creating custom tags is a valid approach but less effective for organizing models at scale compared to using collections, and GiorgioGss similarly said A could also be valid before citing the same collection documentation. abrarjahin, Saransundar, and GiorgioGss all quoted the decisive documentation line that operations on collections do not affect the integrity of the contained model groups and that the underlying artifacts in S3 and ECR are not modified, which is what distinguishes D from the destructive option B.