MLflow Model Registry Quiz
Multiple Choice Questions
- What is a Registered Model?
- A) A GPU endpoint
- B) A collection of Model Versions under a logical name
- C) A training-data copy
- D) A record allowing only one run
Show Answer
Answer: B
For example, fraud-detector groups versions and aliases under one name.
- Which Model Version mutation statement is accurate?
- A) Every field and source byte is permanently immutable
- B) It receives a version number; descriptions/tags can change and artifact preservation is separate
- C) Every version expires after 30 days
- D) Every new model is merged into the previous version
Show Answer
Answer: B
Separate the practice of versioning new results from storage-level immutability.
- Does every Model Version have a training-run link?
- A) Always, and the link cannot be deleted
- B) model_id automatically preserves the dataset snapshot
- C) No; create_model_version run_id/model_id fields are optional
- D) The registry never supports origin links
Show Answer
Answer: C
Direct source URIs are allowed. Complete lineage requires explicit recording and retention.
- What is an alias?
- A) A built-in traffic percentage across several versions
- B) A mutable name pointing to one version
- C) A fixed database address
- D) An unchangeable content hash
Show Answer
Answer: B
Multiple aliases can reference one version, and an alias can move to another version.
- What is the legacy stage API status?
- A) Deprecated since 2.9.0, still present in 3.16.0
- B) Removed from every MLflow version
- C) An access-control policy equivalent to aliases
- D) Only Production is supported
Show Answer
Answer: A
Legacy stages are None/Staging/Production/Archived. Design new flows with aliases/tags and explicit permissions.
- How can registration happen when a model is logged?
- A) Set a tag only
- B) Pass registered_model_name to flavor log_model
- C) Delete an alias
- D) Rename a file to champion
Show Answer
Answer: B
Registering after logging with register_model is another path. Registration does not itself move an alias.
- What controls champion promotion?
- A) MLflow automatically approves the largest version
- B) A review_state tag alone completes permission separation
- C) Evaluation/approval evidence and authentication/authorization for the actor
- D) Every trainer always changes the alias
Show Answer
Answer: C
A tag string replaces neither approval workflow nor access control.
- What happens to an already-loaded model immediately after an alias changes?
- A) It is always replaced instantly
- B) It retrains automatically
- C) Shadow traffic appears automatically
- D) It may keep serving until reload/deployment/cache policy changes it
Show Answer
Answer: D
New resolution and the lifecycle of an existing loaded instance are separate.
Short Answer Questions
- Does READY alone prove inference compatibility and model quality?
Show Answer
No. It is a registration status; metadata fixtures can be READY. Validate real flavors, weights, dependency loading, and quality separately.
- Why can the registry not always reconstruct exact code and data lineage?
Show Answer
Run/model links are optional and source files, Runs, or artifacts can change or disappear. Preserve the serving version, hashes, commit, dataset snapshot, dependencies, and approvals.