Databricks Certified Machine Learning Professional Practice Test 42
Question 410
In an enterprise ML workflow, the reported issue is: A team runs independent GPU-heavy neural trials sequentially, despite having several isolated GPU workers that can fit each model. Before modifying the pipeline, which evidence should be collected first?
- Inspect worker weight hashes, gradient aggregation checkpoints and update barriers.
- Compare GPU inventory with the current trial task graph, resource reservations and wait times.
- Trace parameter ownership and update communication in the proposed training algorithm.
- Compare driver utilization, active executor CPU, task backlog and partition parallelism.
Explanation
Correct Answer
B:This is the most discriminating check for the observed symptoms. Compare GPU inventory with the current trial task graph, resource reservations and wait times. The underlying technical issue is that ray Tune can schedule independent hyperparameter trials with explicit per-trial resources rather than serial execution.
Incorrect Answers
A:That diagnostic evidence is directed at a different scaling and tuning situation: A training team averages gradients from multiple workers but lets each worker use a different untracked version of the model. For the stated incident, the decisive observation is compare GPU inventory with the current trial task graph, resource reservations and wait times.
C:That diagnostic evidence is directed at a different scaling and tuning situation: An ML engineer tries to parallelize a stateful neural-training loop by executing independent mapPartitions calls that share one mutable global model in the driver. For the stated incident, the decisive observation is compare GPU inventory with the current trial task graph, resource reservations and wait times.
D:That diagnostic evidence is directed at a different scaling and tuning situation: A feature-engineering job is CPU-bound on hundreds of independent daily partitions and a team upgrades only the driver to a much larger machine. For the stated incident, the decisive observation is compare GPU inventory with the current trial task graph, resource reservations and wait times.
Question 411
The engineering team encounters the following issue: A pipeline engineer fits a vectorizer after an estimator in the stage list and then tries to use the estimator's prediction as the vectorizer input. Which experiment most directly demonstrates a successful fix?
- The full fitted pipeline completes on a small test DataFrame and exposes the planned input and output columns.
- Every expected input route has exactly the intended output score after a distributed run with bounded driver memory.
- The tracked run records each parameter map, validation metric and final chosen model before untouched test evaluation.
- The assembled vector contains the ordered numeric fields and the estimator can consume it.
Explanation
Correct Answer
A:This verification measures the required result rather than just process liveness. The full fitted pipeline completes on a small test DataFrame and exposes the planned input and output columns. It confirms whether the targeted action succeeded: topologically order vectorizers, assemblers and estimators so required feature columns exist before model fitting.
Incorrect Answers
B:That acceptance test is directed at a different using spark ml situation: A logistics model must score 60 million daily routes already in Delta but the scoring script collects each route into Python before calling predict. Here the actual success criterion is the full fitted pipeline completes on a small test DataFrame and exposes the planned input and output columns.
C:That acceptance test is directed at a different using spark ml situation: An ML engineer manually compares two logistic-regression regularization values on a single holdout but has no reusable SparkML parameter-search plan. Here the actual success criterion is the full fitted pipeline completes on a small test DataFrame and exposes the planned input and output columns.
D:That acceptance test is directed at a different using spark ml situation: A team wants to convert several numeric inputs into one features vector but treats VectorAssembler as a clustering Estimator that should output predictions. Here the actual success criterion is the full fitted pipeline completes on a small test DataFrame and exposes the planned input and output columns.
Question 412
The engineering team encounters the following issue: An operations model uses a SparkML estimator that itself distributes data, yet the team launches dozens of concurrent full-cluster fits on the same fixed small cluster. What corrective action should the engineer prefer?
- Configure supported Ray Tune trials with declared GPU and CPU requirements and controlled concurrency.
- Tune outer concurrency based on actual cluster capacity and each trial's resource demand.
- Connect supported Optuna study execution to MLflow per-trial tracking and preserve the selected objective value.
- Adopt validated model-parallel or sharded training to fit the model and its optimizer state.
Explanation
Correct Answer
B:This corrects the relevant failure boundary instead of merely hiding the symptom. Tune outer concurrency based on actual cluster capacity and each trial's resource demand. The issue matters because oversubscribed trials spend time waiting and can destabilize shared model training.
Incorrect Answers
A:That corrective action is directed at a different scaling and tuning situation: A team runs independent GPU-heavy neural trials sequentially, despite having several isolated GPU workers that can fit each model. The relevant corrective action here is tune outer concurrency based on actual cluster capacity and each trial's resource demand.
C:That corrective action is directed at a different scaling and tuning situation: An Optuna search runs 80 hyperparameter trials but records only the last winning score in MLflow, making comparisons impossible. The relevant corrective action here is tune outer concurrency based on actual cluster capacity and each trial's resource demand.
D:That corrective action is directed at a different scaling and tuning situation: A language model's weights cannot fit on one accelerator yet a team launches pure data-parallel replicas, each loading the entire model. The relevant corrective action here is tune outer concurrency based on actual cluster capacity and each trial's resource demand.
Question 413
Review this production model incident: An alert query uses one global transaction-failure count and masks a complete outage in a small high-value region. What end-to-end verification is most persuasive here?
- A labeled regression fixture raises the performance signal even when input features remain stable.
- All accepted test timestamps land in the correct expected monitoring interval.
- A synthetic v8-only error increase appears in its own model-version metric.
- A synthetic region-only failure triggers the scoped alert without misclassifying healthy regions.
Explanation
Correct Answer
D:This verification measures the required result rather than just process liveness. A synthetic region-only failure triggers the scoped alert without misclassifying healthy regions. It confirms whether the targeted action succeeded: create a dedicated approved critical-region alert with suitable minimum volume and failure-rate conditions.
Incorrect Answers
A:That acceptance test is directed at a different drift detection and lakehouse monitoring situation: A risk service measures feature drift but neglects observed-label performance and continues serving an accurate-looking yet ineffective model. Here the actual success criterion is a synthetic region-only failure triggers the scoped alert without misclassifying healthy regions.
B:That acceptance test is directed at a different drift detection and lakehouse monitoring situation: A time-series monitor uses a string-formatted date that cannot be reliably converted to timestamps and silently groups most rows into an unknown bucket. Here the actual success criterion is a synthetic region-only failure triggers the scoped alert without misclassifying healthy regions.
C:That acceptance test is directed at a different drift detection and lakehouse monitoring situation: A scoring endpoint quietly moves from model v7 to v8 during a rollout but an inference trend aggregates both versions in one error curve. Here the actual success criterion is a synthetic region-only failure triggers the scoped alert without misclassifying healthy regions.
Question 414
In an enterprise ML workflow, the reported issue is: A date feature function treats all incoming timestamp strings as UTC even though some include explicit offsets and a unit test uses only Z-suffixed inputs. What likely operational or model-quality harm does this issue create?
- Different releases disagree on whether a model passes the same quality gate.
- Sensitive events leak into lower-trust test assets while making debugging harder.
- A valid-looking endpoint reverses class interpretation for downstream automation.
- Feature dates shift, placing transactions in incorrect risk windows.
Explanation
Correct Answer
D:This consequence follows from the specific unresolved condition. Feature dates shift, placing transactions in incorrect risk windows. The causal constraint is: Date transformation tests need timezone and offset boundary fixtures aligned with documented semantics.
Incorrect Answers
A:That direct consequence is directed at a different validation testing situation: An engineering group duplicates a validation helper across six notebooks, and two copies enforce outdated threshold values. The more immediate risk here is feature dates shift, placing transactions in incorrect risk windows.
B:That direct consequence is directed at a different validation testing situation: An engineer uses only production replay data to validate a code change and exposes customer events in an unrestricted test workspace. The more immediate risk here is feature dates shift, placing transactions in incorrect risk windows.
C:That direct consequence is directed at a different validation testing situation: An end-to-end validation uses only a synthetic positive label and fails to detect a swapped positive/negative class mapping in inference. The more immediate risk here is feature dates shift, placing transactions in incorrect risk windows.
Question 415
Consider the following fictional Databricks production case: A mobile app requires thousands of concurrent low-latency scores with online features, but a team plans to query an offline historical feature Delta table for every request. Which implementation is most appropriate for the stated incident?
- Configure selection against the intended minimization metric and rerun model comparison.
- Order categorical indexing and encoding before assembling numeric and encoded features into a vector.
- Publish approved features to a supported low-latency online feature serving path and test endpoint integration.
- Add a justified tail-sensitive error measure such as RMSE and inspect failure slices in model selection.
Explanation
Correct Answer
C:This corrects the relevant failure boundary instead of merely hiding the symptom. Publish approved features to a supported low-latency online feature serving path and test endpoint integration. The issue matters because interactive scoring times out when each request launches heavy offline data scans.
Incorrect Answers
A:That corrective action is directed at a different using spark ml situation: A SparkML grid-search experiment selects the model with the largest error metric even though the evaluator reports RMSE to minimize. The relevant corrective action here is publish approved features to a supported low-latency online feature serving path and test endpoint integration.
B:That corrective action is directed at a different using spark ml situation: An IoT classifier passes raw text categories directly into VectorAssembler ahead of StringIndexer and OneHotEncoder, generating incompatible vector inputs. The relevant corrective action here is publish approved features to a supported low-latency online feature serving path and test endpoint integration.
D:That corrective action is directed at a different using spark ml situation: A forecasting model has a few catastrophic numeric errors hidden by a low median absolute error and operations care about large misses. The relevant corrective action here is publish approved features to a supported low-latency online feature serving path and test endpoint integration.
Question 416
A production ML review contains this situation: A team already has a vectorized Spark DataFrame pipeline and proposes converting everything to Ray datasets solely to perform a simple SQL aggregation. What production-safe change best matches the evidence?
- Keep SQL/DataFrame aggregation in Spark unless a distinct Ray capability justifies conversion.
- Declare correct per-trial GPU reservations and reduce concurrent trials until memory fits.
- Bind the selected model version to each group and join predictions on a unique provenance key.
- Use framework-managed synchronized updates and reproducible checkpoints for the parallel training loop.
Explanation
Correct Answer
A:This corrects the relevant failure boundary instead of merely hiding the symptom. Keep SQL/DataFrame aggregation in Spark unless a distinct Ray capability justifies conversion. The issue matters because unnecessary system transitions add serialization and operational maintenance without model benefit.
Incorrect Answers
B:That corrective action is directed at a different scaling and tuning situation: A Ray Tune training function requests zero GPU resources but initializes a large GPU model inside every trial. The relevant corrective action here is keep SQL/DataFrame aggregation in Spark unless a distinct Ray capability justifies conversion.
C:That corrective action is directed at a different scaling and tuning situation: A batch scoring pipeline trains one per-region model and then joins model predictions back using region alone, duplicating rows because each region contains multiple versioned models. The relevant corrective action here is keep SQL/DataFrame aggregation in Spark unless a distinct Ray capability justifies conversion.
D:That corrective action is directed at a different scaling and tuning situation: A training team averages gradients from multiple workers but lets each worker use a different untracked version of the model. The relevant corrective action here is keep SQL/DataFrame aggregation in Spark unless a distinct Ray capability justifies conversion.
Question 417
A production ML review contains this situation: A model's prediction error appears to spike at 2 a.m. local time but the monitoring table groups timestamps in UTC for a globally distributed region. Which diagnostic check provides the strongest direct evidence?
- Inspect source timestamps and selected monitoring timezone for affected geographic cohorts.
- Correlate 5xx status bursts with request rate, resource use and deployment events.
- Inspect inference table version IDs and label-matched errors before and after traffic shift.
- Compare historical baselines, current seasonal patterns and change approvals for the existing threshold.
Explanation
Correct Answer
A:This is the most discriminating check for the observed symptoms. Inspect source timestamps and selected monitoring timezone for affected geographic cohorts. The underlying technical issue is that time-of-day slices require explicit timezone interpretation and region-specific boundaries.
Incorrect Answers
B:That diagnostic evidence is directed at a different drift detection and lakehouse monitoring situation: A prediction endpoint returns HTTP 503 for intermittent bursts while the model's offline accuracy remains unchanged. For the stated incident, the decisive observation is inspect source timestamps and selected monitoring timezone for affected geographic cohorts.
C:That diagnostic evidence is directed at a different drift detection and lakehouse monitoring situation: A scoring endpoint quietly moves from model v7 to v8 during a rollout but an inference trend aggregates both versions in one error curve. For the stated incident, the decisive observation is inspect source timestamps and selected monitoring timezone for affected geographic cohorts.
D:That diagnostic evidence is directed at a different drift detection and lakehouse monitoring situation: A production alert threshold is hard-coded against last year's baseline, but seasonal traffic has shifted and the owner never reviewed the threshold. For the stated incident, the decisive observation is inspect source timestamps and selected monitoring timezone for affected geographic cohorts.
Question 418
The engineering team encounters the following issue: A text pipeline repeatedly fits a Tokenizer stage even though tokenization simply converts strings into token arrays without learning a model. What corrective action should the engineer prefer?
- Order categorical indexing and encoding before assembling numeric and encoded features into a vector.
- Use Tokenizer.transform in the correct pipeline position while reserving fit for actual Estimators.
- Evaluate the model with class-sensitive metrics and choose thresholds aligned with false-positive and missed-fraud tradeoffs.
- Hold out a final test set before tuning and score the selected pipeline exactly once after selection.
Explanation
Correct Answer
B:This corrects the relevant failure boundary instead of merely hiding the symptom. Use Tokenizer.transform in the correct pipeline position while reserving fit for actual Estimators. The issue matters because the workflow fails or includes unnecessary invalid fit calls that cannot create a tokenizer model.
Incorrect Answers
A:That corrective action is directed at a different using spark ml situation: An IoT classifier passes raw text categories directly into VectorAssembler ahead of StringIndexer and OneHotEncoder, generating incompatible vector inputs. The relevant corrective action here is use Tokenizer.transform in the correct pipeline position while reserving fit for actual Estimators.
C:That corrective action is directed at a different using spark ml situation: A fraud detector predicts the majority no-fraud class for 99.4% of cases and advertises 99.4% accuracy while missing most positives. The relevant corrective action here is use Tokenizer.transform in the correct pipeline position while reserving fit for actual Estimators.
D:That corrective action is directed at a different using spark ml situation: After choosing the best parameter map, a data scientist reports the same cross-validation score as an independent final test result without any untouched data. The relevant corrective action here is use Tokenizer.transform in the correct pipeline position while reserving fit for actual Estimators.
Question 419
In an enterprise ML workflow, the reported issue is: An image model needs 90 GB accelerator memory for one training replica while each available GPU has 24 GB and engineers propose only adding more identical nodes. Which change addresses the real cause without undermining other requirements?
- Pin shared validation protocol and data snapshot for the compared Optuna trials.
- Use supported data parallelism with appropriately batched examples and convergence checks.
- Select a suitable larger device or validated model-parallel/sharded execution design before scaling replicas.
- Normalize per-group output columns and types before returning them to Spark.
Explanation
Correct Answer
C:This corrects the relevant failure boundary instead of merely hiding the symptom. Select a suitable larger device or validated model-parallel/sharded execution design before scaling replicas. The issue matters because every replica hits device OOM even though aggregate cluster memory appears ample.
Incorrect Answers
A:That corrective action is directed at a different scaling and tuning situation: A hyperparameter team compares two Optuna studies with different train/validation splits and chooses the lower loss as though both metrics were directly comparable. The relevant corrective action here is select a suitable larger device or validated model-parallel/sharded execution design before scaling replicas.
B:That corrective action is directed at a different scaling and tuning situation: A tabular classifier fits comfortably on one GPU and 100 million independent examples are available, but a team proposes splitting the small model's layers across hosts. The relevant corrective action here is select a suitable larger device or validated model-parallel/sharded execution design before scaling replicas.
D:That corrective action is directed at a different scaling and tuning situation: A per-city pandas UDF returns different prediction schemas for cities with absent categories, causing a distributed aggregation failure. The relevant corrective action here is select a suitable larger device or validated model-parallel/sharded execution design before scaling replicas.
Question 420
Review this production model incident: A team records experiments but has no controlled path from a selected registered model to scheduled scoring jobs. What evidence would best distinguish the actual fault from neighboring ML concerns?
- Inspect environment-specific data-source references and effective service-identity permissions.
- Compare release commit identifiers, approved artifacts and runtime dependency versions across dev, test and prod.
- Compare the logged training environment, lockfile and production model-load dependency versions.
- Inventory model run IDs, registered versions, job configuration and deployment references.
Explanation
Correct Answer
D:This is the most discriminating check for the observed symptoms. Inventory model run IDs, registered versions, job configuration and deployment references. The underlying technical issue is that tracking, registry, jobs and serving are distinct lifecycle capabilities that must be linked by artifact identity.
Incorrect Answers
A:That diagnostic evidence is directed at a different model lifecycle management situation: A team trains on curated development features but deploys the same job with credentials that accidentally read unapproved production source tables. For the stated incident, the decisive observation is inventory model run IDs, registered versions, job configuration and deployment references.
B:That diagnostic evidence is directed at a different model lifecycle management situation: A regulated company changes deployed Python notebook code directly in production to correct a model transform while its test environment runs a different untracked revision. For the stated incident, the decisive observation is inventory model run IDs, registered versions, job configuration and deployment references.
C:That diagnostic evidence is directed at a different model lifecycle management situation: A model is promoted to production with its code unchanged but a different unpinned training library version is installed, altering prediction preprocessing. For the stated incident, the decisive observation is inventory model run IDs, registered versions, job configuration and deployment references.