NCA-ADS - Introductory MLOps Practices (10% of the exam) - Section 6.5
Manage artefacts.
Manage pipeline artefacts - datasets, preprocessors, models, and evaluation reports - by storing them with unique identifiers and tracing their provenance back to a specific experiment run. Recognise how MLflow artefact logging or equivalent tooling supports this lineage.
A data scientist wraps a supported-framework training step in an MLflow run and enables MLflow autologging instead of writing manual logging calls. She expects autologging to remove the need for explicit logging code for the standard items it covers. For a typical supported estimator, which set of items does MLflow autologging capture automatically without explicit per-item logging calls?
AOnly the final evaluation metric, leaving the hyperparameters and the model artefact to be logged by hand.
BThe raw training dataset rows and the full GPU memory profile of the training process.
CThe model hyperparameters, the training metrics, and the trained model artefact for the run.check_circle Correct
DOnly descriptive run tags such as the user name and the source file, with no parameters or metrics recorded.
MLflow autologging on a supported framework records hyperparameters, training metrics, and the model artefact without explicit per-item logging calls. Autologging hooks a supported framework's training routine and emits the estimator's parameters, the metrics generated during fitting, and the fitted model as a logged artefact, replacing the explicit parameter, metric, and model logging a developer would otherwise write by hand.
Why A is wrong: It is plausible to assume autologging is minimal, but it also captures hyperparameters and the model artefact, so restricting it to one metric understates what the feature does.
Why B is wrong: Capturing inputs and resource profiles sounds thorough, but autologging does not snapshot raw data rows or detailed GPU memory profiles as part of its standard behaviour.
Why C is correct: Autologging for supported frameworks records the estimator's hyperparameters, the metrics produced during fitting, and the resulting model artefact, which is precisely the standard logging it removes from manual code.
Why D is wrong: Run context tags are recorded by tracking, which makes this tempting, but autologging's purpose is the parameters, metrics, and model, not tags alone.
Answers that look right on this material and are not. Each one is a distractor from a different question in the NCA-ADS bank for this domain.
Promote each run to the model registry and read the production stage, because only registered models retain their logged metrics.
errorWhy it is wrong: The registry is a real MLflow component and sounds authoritative, but it manages model versions and stages and does not rank runs by an arbitrary metric, and unregistered runs keep their metrics.
Compression of the model binaries, which shrinks each stored artefact so that the production and evaluation copies occupy less disk space
errorWhy it is wrong: Tempting because efficient storage sounds like a registry benefit, but compression only reduces footprint and says nothing about which model is live versus under evaluation.
Only the final validation accuracy and the wall-clock training time, since these two figures uniquely identify which run produced the best model
errorWhy it is wrong: Tempting because these summarise a run at a glance, but two numbers describe outcomes; they cannot reconstruct the inputs, code, or environment needed to rebuild the model.