Model Registries Quiz

Answers and explanations appear as you take the quiz.

1. What is the main purpose of a model registry?
   • To label raw training data
   • To serve as a version-controlled catalog and single source of truth for trained models
   • To replace the GPU in a server
   • To write SQL queries
2. In current MLflow registry workflows, what can provide a named reference to a model version?
   • A reassignable model alias
   • A universal mandatory four-stage lifecycle in every registry
   • The training GPU’s display resolution
   • A random tensor without a model version
3. How can a registry support a model rollback?
   • By guaranteeing every server switches instantly
   • By deleting all previous versions
   • By retaining a previous artifact that the serving workflow can reload and validate
   • By changing the training labels automatically
4. What is a model signature in a registry?
   • The data scientist's handwritten autograph
   • A password for the database
   • The model's training duration
   • A schema describing the model's expected input and output types
5. Why is recorded lineage (dataset, code commit, hyperparameters) valuable?
   • It enables reproducibility and auditability of how each version was built
   • It makes the model run faster
   • It increases the model's accuracy automatically
   • It compresses the model file
6. What connects a registered model decision to an actual application rollout?
   • A guarantee that registering a version deploys it everywhere
   • A required fax to the GPU vendor
   • A configured deployment workflow that loads and verifies the chosen artifact
   • A model signature alone with no serving system
7. A champion alias is reassigned, but a long-running server still has the old artifact loaded. What should an operator verify?
   • Only that the alias has a new value
   • That all servers reload automatically in every integration
   • Only the number of files in the registry
   • The artifact and version actually serving requests
8. A model signature expects numerical columns but a request sends strings. What can signature-aware validation help detect?
   • Whether the training labels are morally correct
   • An input schema or type mismatch
   • Every possible accuracy regression
   • A guarantee that all numerical inputs are in distribution