10 free sample questions from a bank of 60, with the correct answers and explanations. No signup required — start practising right now.
Answer C — concept drift is a change in the relationship between inputs and the target, not just a shift in inputs or labels alone.
Concept drift vs data drift vs label drift, P(y|X) relationship change, model staleness over time.
C states the textbook definition: P(y|X) changes, so a model learned on old mapping becomes less accurate even if input ranges look similar. The vendor prose corroborates this relationship-change definition.
A describes covariate / data drift in an input, not concept drift. B describes prior / label drift in the target distribution. D describes prediction-distribution shift, a symptom rather than the definition. E is incorrect because C is correct.
Memorize the trio: input shift = data/covariate drift, target shift = label/prior drift, relationship shift = concept drift.
Answer D — Jensen-Shannon distance is preferred for large-scale numeric drift because it stays stable where KS becomes oversensitive.
Numeric feature drift detection, Jensen-Shannon distance properties, Kolmogorov-Smirnov sensitivity to sample size, threshold selection.
KS p-values flag trivial differences as significant when n is very large. JS distance is bounded, smoothed and symmetric, giving a comparable magnitude of divergence that remains interpretable at scale, hence more robust on large datasets.
A cannot hold because B, C and E are false. B reverses the fact: JS is normalized and smoothed. C is incorrect because D is valid. E is incorrect: JS still needs a business threshold for alerting; only the 0-1 scaling aids interpretation, it does not eliminate cutoffs.
If the exam mentions large datasets plus KS vs JS, choose JS robustness; remember KS = hypothesis-test p-value problem at scale.
Use spark.read.format("mlflow-experiment").load(exp_id) to get run-level results for an experiment as a Spark DataFrame.
mlflow-experiment Spark data source format documented at docs.databricks.com/en/query/formats/mlflow-experiment.htmlclient.list_run_infos, mlflow.search_runs, and Spark read.format.loadE matches the vendor-selected data source: spark.read.format("mlflow-experiment").load(exp_id) loads experiment exp_id run-level results directly as a Spark DataFrame, satisfying the requirement for programmatic Spark work with an active spark session.
client.list_run_infos(exp_id) returns RunInfo objects via the MLflow Client, not a Spark DataFrame.spark.read.format("delta").load(exp_id) treats exp_id as a Delta path, which is the wrong format and source for MLflow experiment data.mlflow-experiment format.mlflow.search_runs(exp_id) returns pandas results for fluent search, not a Spark DataFrame load via spark.read.When the stem asks for Spark DataFrame + experiment ID, choose format("mlflow-experiment").load(exp_id) over search_runs or list_run_infos.
Answer C — reload a scikit-learn flavor model with mlflow.sklearn.load_model to get the native object back, including feature_importances_.
MLflow flavors, sklearn native object vs pyfunc wrapper, model_uri format, cluster restart persistence.
The model was logged as a sklearn flavor, which preserves the pickle plus metadata. mlflow.sklearn.load_model restores the RandomForest object itself, so attributes like feature_importances_ are available again in the new session.
A mlflow.load_model is not the valid API name. B list_artifacts only lists files, it does not deserialize a model. D is false: the UI is convenient but not required for access. E pyfunc.load_model returns a generic PythonModel wrapper without sklearn attributes like feature_importances_.
Match load to log flavor: sklearn model needs mlflow.sklearn.load_model; pyfunc load loses native attributes.
For categorical feature drift, monitor mode, number of unique values, and percentage of missing values together.
C is the vendor-selected full set: Mode, number of unique values, and percentage of missing values. Together they capture central tendency for categories, cardinality expansion or collapse, and missing-value spikes, which single statistics alone miss.
If categorical drift offers a combo option with mode + unique count + missing, prefer the combo over any single statistic.
Answer B — the standard first response to confirmed drift is retrain and redeploy on more recent data.
Drift remediation, retraining, redeployment, monitoring loop, vs rebuild vs sunset.
Drift means P(X) or P(y|X) moved. Refreshing the training window re-aligns the model to current distribution and restores performance without redesigning the application.
A is incorrect because B is valid. C is incorrect because only B is broadly probable; rebuilding labels or sunsetting are extreme cases. D rebuilding with a new label is a re-scoping project, not routine drift response. E sunsetting is last-resort when the use case is no longer viable.
Default drift playbook for the exam: detect, retrain on recent data, redeploy, keep monitoring.
Answer E — mlflow.autolog() enables autologging for all supported libraries regardless of runtime defaults or admin settings.
MLflow Autologging scope, mlflow.autolog vs flavor-specific autolog, Databricks Runtime ML defaults, workspace admin overrides.
The generic mlflow.autolog() turns on autologging for every installed supported flavor upon import, overriding version-specific defaults and workspace-level toggles, which flavor-specific calls cannot guarantee across runtimes.
A only covers sklearn and misses other libraries. B only covers Spark ML. C spark.conf.set is not an MLflow API. D is false: automatic logging is possible via autolog.
All libraries no matter what = mlflow.autolog(); single library = mlflow.<flavor>.autolog().
Answer D — use DESCRIBE HISTORY on the Delta table to find when the column was dropped.
Delta Lake time travel, DESCRIBE HISTORY metadata, versioning, schema change audit.
DESCRIBE HISTORY returns operation, timestamp, user and parameters for every table version, including ALTER / DROP COLUMN events, so the team can pinpoint the version and time the column disappeared.
A VERSION alone is just a number, not a command that shows history. B DESCRIBE only shows current schema, not change log. C HISTORY alone is not valid Spark SQL syntax. E TIMESTAMP is used in time-travel reads, not to list the audit log itself.
Audit question = DESCRIBE HISTORY; read old data question = TIMESTAMP AS OF / VERSION AS OF.
Answer E — label drift is a change in the distribution of the target variable itself.
Label / prior drift vs concept drift vs covariate drift, P(y) shift.
Label drift means the class balance or target distribution in production differs from training, e.g. far more positives, which degrades thresholds and metrics even if inputs are stable.
A describes prediction drift, the model output shift. B is incorrect because E is correct. C describes input / covariate drift. D describes concept drift, the change in P(y|X) relationship.
Pair this with concept drift: label = P(y) moved, concept = P(y|X) moved.
Batch is the most common ML deployment paradigm.
D Batch matches the vendor-selected answer and thin prose that the most widely used deployment paradigm is typically batch, reflecting offline scheduled scoring that fits most ML projects.
Unless the stem stresses per-record immediacy or edge, default most-common-paradigm questions to batch.
Want the full bank of 60 questions for Certified Machine Learning Professional? See all practice exams.