MLflow Tracking Quiz
Multiple Choice Questions
- How are Experiments and Runs related?
- A) An Experiment groups runs; a run can represent training or evaluation
- B) An Experiment is one GPU
- C) A Run is always a deployed model
- D) They are the same entity
Show Answer
Answer: A
Runs can also record preprocessing and comparison work.
- What is the default metadata backend in a new MLflow 3.16.0 environment?
- A) An S3 bucket
- B) sqlite:///mlflow.db
- C) Browser localStorage
- D) Mandatory PostgreSQL
Show Answer
Answer: B
SQLite is the default; inspect compatibility behavior if ./mlruns already exists. Explicit URIs remove ambiguity.
- What is the important MLflow 3 LoggedModel change?
- A) Independent model_id, status, and relationship tracking
- B) First-ever ability to call log_model without a start_run block
- C) Artifacts are no longer needed
- D) Immediate GPU deployment upon registration
Show Answer
Answer: A
Model.log in 2.22.0 already started a run implicitly when needed. Independent model identity is different from omitting an explicit run context.
- Which autologging statement is correct?
- A) It captures everything in arbitrary code
- B) It always removes PII
- C) Review each integration’s supported versions and collected inputs/outputs
- D) It automatically completes deployment
Show Answer
Answer: C
Behavior depends on the framework and options. Check input examples, raw data, and model-artifact collection.
- Which tracing and cost statement is accurate?
- A) Tracing first appeared in 3.x
- B) Every tool span has LLM cost
- C) Token-derived cost always equals the invoice
- D) Tracing arrived in 2.14.0; usage collection depends on integration
Show Answer
Answer: D
Separate later expansion from initial introduction. Cost cannot be guaranteed without model, usage, and pricing information.
- When can a client need S3 permissions despite using a remote tracking server?
- A) Non-proxy mode with a direct S3 artifact URI
- B) Only reading SQLite parameters
- C) It never needs permissions regardless of mode
- D) Every alias-name lookup requires S3 access
Show Answer
Answer: A
Metadata and artifact paths differ. Direct mode requires client storage permissions and network access.
- How do you inspect a metric logged at steps 0 and 1?
- A) The parameter changes automatically
- B) Only the second value is retained forever
- C) Use get_metric_history to inspect both observations
- D) Two models are registered automatically
Show Answer
Answer: C
Distinguish the current summary from timestamped, stepped metric history.
- What is an appropriate Tracking web UI query path?
- A) The browser connects directly to PostgreSQL
- B) The browser calls server HTTP APIs
- C) The UI always reads files directly from training Pods
- D) The browser needs a database administrator password
Show Answer
Answer: B
The server accesses its backend. Artifact proxy configuration is a separate concern.
Short Answer Questions
- Can a PENDING initialize_logged_model result immediately perform inference?
Show Answer
No. It can be metadata only; actual flavor, weights, artifact logging, and finalization are still required. READY is not quality or deployment approval.
- Why might an existing experiment still access S3 directly after server artifact flags change?
Show Answer
The recorded experiment/run artifact URI is not rewritten retroactively. Inspect that URI and the actual permissions and transfer path.