Databricks Certified Machine Learning Professional Practice Test 37
Question 360
Review this production model incident: An endpoint's error rate spikes while data profiles look normal, and responders attempt to retrain the model instead of checking serving infrastructure. Which failure risk follows most closely from the unresolved behavior?
- Daily drift statistics reflect malformed time buckets rather than actual event periods.
- The monitoring pipeline either floods the SOC or ignores new material shifts.
- The live population drifts far from model training even though daily changes remain modest.
- Users cannot receive predictions because of infrastructure failures that retraining cannot fix.
Explanation
Correct Answer
D:This consequence follows from the specific unresolved condition. Users cannot receive predictions because of infrastructure failures that retraining cannot fix. The causal constraint is: Model health, performance and data drift are different failure categories that need separate investigation paths.
Incorrect Answers
A:That direct consequence 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. The more immediate risk here is users cannot receive predictions because of infrastructure failures that retraining cannot fix.
B:That direct consequence 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. The more immediate risk here is users cannot receive predictions because of infrastructure failures that retraining cannot fix.
C:That direct consequence is directed at a different drift detection and lakehouse monitoring situation: A model trained on last year's customer mix is monitored only against yesterday's already shifted cohort, so its original training-population mismatch is hidden. The more immediate risk here is users cannot receive predictions because of infrastructure failures that retraining cannot fix.
Question 361
In an enterprise ML workflow, the reported issue is: An ML operator promotes a new model version but its automated rollback points to a model that was deleted from the endpoint configuration. Which change addresses the real cause without undermining other requirements?
- Preserve the approved model artifact and rehearse supported traffic restoration before the release.
- Set a statistically and operationally adequate rollout observation period with strict safety abort thresholds.
- Use a supported routing design or keep the protected region on the approved model until scoped rollout is available.
- Send a small approved percentage of eligible live requests to the candidate and expand only after defined gates pass.
Explanation
Correct Answer
A:This corrects the relevant failure boundary instead of merely hiding the symptom. Preserve the approved model artifact and rehearse supported traffic restoration before the release. The issue matters because the team cannot recover from a bad promotion without an emergency rebuild.
Incorrect Answers
B:That corrective action is directed at a different deployment strategies situation: A candidate fraud model has near-identical average error to the champion, but an organization wants to compare rare fraudulent transactions during a brief canary. The relevant corrective action here is preserve the approved model artifact and rehearse supported traffic restoration before the release.
C:That corrective action is directed at a different deployment strategies situation: A model release requires all clients from a restricted region to keep the old version, but a simple percentage split sends some of those clients to the candidate. The relevant corrective action here is preserve the approved model artifact and rehearse supported traffic restoration before the release.
D:That corrective action is directed at a different deployment strategies situation: A travel site serves 30,000 bookings per minute and wants to evaluate a new fraud model while limiting customer impact during its first hour. The relevant corrective action here is preserve the approved model artifact and rehearse supported traffic restoration before the release.
Question 362
In an enterprise ML workflow, the reported issue is: A prediction endpoint returns HTTP 503 for intermittent bursts while the model's offline accuracy remains unchanged. What technical distinction is essential to this investigation?
- Alert windows and persistence conditions must align with the service objective to avoid transient false positives.
- Aggregate performance can conceal failure in important subpopulations and lifecycle stages.
- Serving availability errors are distinct from model predictive accuracy and require infrastructure telemetry.
- A normalized custom metric should account for the eligible denominator, not merely numerator volume.
Explanation
Correct Answer
C:The technical constraint is specific to the reported failure. Serving availability errors are distinct from model predictive accuracy and require infrastructure telemetry. To test it, correlate 5xx status bursts with request rate, resource use and deployment events.
Incorrect Answers
A:That underlying principle is directed at a different drift detection and lakehouse monitoring situation: A monitor alerts whenever a ratio exceeds 0.3 for a single minute, although normal traffic produces brief spikes and operations require ten-minute persistence. The current incident instead turns on serving availability errors are distinct from model predictive accuracy and require infrastructure telemetry.
B:That underlying principle is directed at a different drift detection and lakehouse monitoring situation: A recommender reports stable overall accuracy while click-through prediction quality for first-time users deteriorates sharply. The current incident instead turns on serving availability errors are distinct from model predictive accuracy and require infrastructure telemetry.
D:That underlying principle is directed at a different drift detection and lakehouse monitoring situation: A monitoring team wants a business metric for the share of transactions exceeding a risk threshold but records only the raw count of high-risk transactions. The current incident instead turns on serving availability errors are distinct from model predictive accuracy and require infrastructure telemetry.
Question 363
A production ML review contains this 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. Which diagnostic check provides the strongest direct evidence?
- Measure repeated dataset exports and per-trial deserialization relative to model-fitting time.
- Trace parameter ownership and update communication in the proposed training algorithm.
- Compare dataset version, random seed, fold definition and metric contract attached to each MLflow run.
- Inspect runtime GPU memory per worker alongside Ray trial resource requests.
Explanation
Correct Answer
C:This is the most discriminating check for the observed symptoms. Compare dataset version, random seed, fold definition and metric contract attached to each MLflow run. The underlying technical issue is that tuning runs must use equivalent evaluation data and metric definitions before ranking configurations.
Incorrect Answers
A:That diagnostic evidence is directed at a different scaling and tuning situation: A hyperparameter search operates on large Spark-managed feature tables but each candidate model is small and independent, and workers repeatedly convert the entire table for each Ray task. For the stated incident, the decisive observation is compare dataset version, random seed, fold definition and metric contract attached to each MLflow run.
B: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 dataset version, random seed, fold definition and metric contract attached to each MLflow run.
D:That diagnostic evidence 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. For the stated incident, the decisive observation is compare dataset version, random seed, fold definition and metric contract attached to each MLflow run.
Question 364
Consider the following fictional Databricks production case: A company operates development and production models in one workspace under the same broad service account and frequently overwrites production tables during experiments. Which implementation is most appropriate for the stated incident?
- Define the supported model-serving and MLflow resources declaratively with validated environment targets.
- Separate or strictly scope dev/test/prod resources with explicit promotion and least-privilege identities.
- Migrate or retain the configuration only after verifying current supported fields and running an explicit bundle validation.
- Add bundle validation, permission checks and staging smoke tests before production promotion.
Explanation
Correct Answer
B:This corrects the relevant failure boundary instead of merely hiding the symptom. Separate or strictly scope dev/test/prod resources with explicit promotion and least-privilege identities. The issue matters because a developer's experiment can alter production feature tables or serving configuration.
Incorrect Answers
A:That corrective action is directed at a different environment architectures situation: A model serving endpoint is created manually in the UI and every staging deployment differs subtly from production because no deployment definition is saved. The relevant corrective action here is separate or strictly scope dev/test/prod resources with explicit promotion and least-privilege identities.
C:That corrective action is directed at a different environment architectures situation: An ML platform moves its configuration from older Databricks Asset Bundles terminology to newer Declarative Automation Bundles references and assumes every resource schema stayed identical. The relevant corrective action here is separate or strictly scope dev/test/prod resources with explicit promotion and least-privilege identities.
D:That corrective action is directed at a different environment architectures situation: A continuous-delivery job deploys a notebook change through a bundle but skips validation of referenced experiment paths and job permissions. The relevant corrective action here is separate or strictly scope dev/test/prod resources with explicit promotion and least-privilege identities.
Question 365
The engineering team encounters the following issue: A researcher distributes 500 tiny one-second trials across many Ray workers and sees coordination cost exceed compute time. What corrective action should the engineer prefer?
- Adopt validated model-parallel or sharded training to fit the model and its optimizer state.
- Use a supported distributed-training framework with explicit synchronization or adopt independent per-partition models only if the problem allows it.
- Assign each trial a unique tracked run with a shared parent experiment and stable trial metadata.
- Batch or redesign microtrials and choose a proportionate degree of Ray parallelism.
Explanation
Correct Answer
D:This corrects the relevant failure boundary instead of merely hiding the symptom. Batch or redesign microtrials and choose a proportionate degree of Ray parallelism. The issue matters because a larger cluster runs the hyperparameter search slower because each task is overhead-dominated.
Incorrect Answers
A: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 batch or redesign microtrials and choose a proportionate degree of Ray parallelism.
B:That corrective action 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. The relevant corrective action here is batch or redesign microtrials and choose a proportionate degree of Ray parallelism.
C:That corrective action is directed at a different scaling and tuning situation: An Optuna optimization reuses the same MLflow run ID for concurrent trials, causing overlapping parameter values and artifacts. The relevant corrective action here is batch or redesign microtrials and choose a proportionate degree of Ray parallelism.
Question 366
The engineering team encounters the following issue: A research group wants to register a custom classification model in Unity Catalog but uses a local path as its complete model identity and has no governed model name. What direct consequence should the owner prioritize if no fix is made?
- Automation falsely reports a completed rollout while clients still receive temporary failures.
- Unauthenticated calls are rejected or teams try to bypass endpoint controls.
- Stale candidate predictions persist after traffic has returned to the champion.
- Reviewers cannot audit or grant precise access to a reproducible registered version.
Explanation
Correct Answer
D:This consequence follows from the specific unresolved condition. Reviewers cannot audit or grant precise access to a reproducible registered version. The causal constraint is: Governed custom-model registration requires a model artifact plus catalog/schema/model identity and appropriate permissions.
Incorrect Answers
A:That direct consequence is directed at a different custom model serving situation: A Databricks SDK deployment script sends a model-version update but does not wait for the endpoint to reach a usable ready state before announcing success. The more immediate risk here is reviewers cannot audit or grant precise access to a reproducible registered version.
B:That direct consequence is directed at a different custom model serving situation: A client invokes a registered custom ML model using an unauthenticated HTTP POST to the serving endpoint and assumes network access implies authorization. The more immediate risk here is reviewers cannot audit or grant precise access to a reproducible registered version.
C:That direct consequence is directed at a different custom model serving situation: An integration caches predictions without recording the model version and returns yesterday's values after a controlled rollback. The more immediate risk here is reviewers cannot audit or grant precise access to a reproducible registered version.
Question 367
The engineering team encounters the following issue: A risk service measures feature drift but neglects observed-label performance and continues serving an accurate-looking yet ineffective model. What corrective action should the engineer prefer?
- Compare to a justified weekend reference or adjust baseline design with documented change control.
- Configure a documented threshold with the responsible on-call team and a tested delivery route.
- Add an approved later label join and maintain separate unlabeled input-drift metrics until labels arrive.
- Add delayed-ground-truth performance tracking alongside existing drift metrics and health signals.
Explanation
Correct Answer
D:This corrects the relevant failure boundary instead of merely hiding the symptom. Add delayed-ground-truth performance tracking alongside existing drift metrics and health signals. The issue matters because concept drift or changed label relationships degrade decisions without an input-drift alarm.
Incorrect Answers
A:That corrective action is directed at a different drift detection and lakehouse monitoring situation: An online feature monitor compares weekend shopping to weekday reference data and treats expected weekend behavior as a model failure. The relevant corrective action here is add delayed-ground-truth performance tracking alongside existing drift metrics and health signals.
B:That corrective action is directed at a different drift detection and lakehouse monitoring situation: A fraud-feature drift monitor crosses its threshold but the alert has no assigned recipient and no incident ownership. The relevant corrective action here is add delayed-ground-truth performance tracking alongside existing drift metrics and health signals.
C:That corrective action is directed at a different drift detection and lakehouse monitoring situation: A model's inference table records output probabilities but omits ground-truth labels, yet reviewers expect immediate precision and recall metrics. The relevant corrective action here is add delayed-ground-truth performance tracking alongside existing drift metrics and health signals.
Question 368
The engineering team encounters the following issue: A team switches a SparkML training stage to Ray without verifying that its chosen estimator supports Ray's distributed fit contracts. Which technical principle most directly explains the root of the issue?
- Concurrent study workers need a supported shared trial-storage design and correct coordination, not a fragile local-only backend.
- Selecting Ray over Spark must be based on the actual training library and distributed API, not the existence of a cluster.
- Horizontal replica count alone does not combine GPU memory for one model without a supported parallelization strategy.
- Ray resource declarations must match actual device use to avoid GPU oversubscription.
Explanation
Correct Answer
B:The technical constraint is specific to the reported failure. Selecting Ray over Spark must be based on the actual training library and distributed API, not the existence of a cluster. To test it, inspect the estimator's documented execution support and task resource requirements.
Incorrect Answers
A:That underlying principle is directed at a different scaling and tuning situation: A researcher parallelizes Optuna trials while sharing one writable SQLite file on a fragile network mount and observes duplicate reservations and storage locks. The current incident instead turns on selecting Ray over Spark must be based on the actual training library and distributed API, not the existence of a cluster.
C:That underlying principle is directed at a different scaling and tuning situation: 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. The current incident instead turns on selecting Ray over Spark must be based on the actual training library and distributed API, not the existence of a cluster.
D:That underlying principle 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 current incident instead turns on selecting Ray over Spark must be based on the actual training library and distributed API, not the existence of a cluster.
Question 369
Consider the following fictional Databricks production case: An inference pipeline sends new categorical labels unseen during training to StringIndexerModel with error handling configured to reject unknown categories. Which implementation is most appropriate for the stated incident?
- Use model.transform on new feature vectors and retrain only in an approved training workflow.
- Choose a justified supported unknown-category handling strategy and test it without retraining on confidential evaluation labels.
- Evaluate the model with class-sensitive metrics and choose thresholds aligned with false-positive and missed-fraud tradeoffs.
- Configure selection against the intended minimization metric and rerun model comparison.
Explanation
Correct Answer
B:This corrects the relevant failure boundary instead of merely hiding the symptom. Choose a justified supported unknown-category handling strategy and test it without retraining on confidential evaluation labels. The issue matters because a rare new category can fail an entire scoring batch rather than produce a controlled result.
Incorrect Answers
A:That corrective action is directed at a different using spark ml situation: An analyst loads a fitted LogisticRegressionModel but calls fit on it for every scored batch rather than transform. The relevant corrective action here is choose a justified supported unknown-category handling strategy and test it without retraining on confidential evaluation labels.
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 choose a justified supported unknown-category handling strategy and test it without retraining on confidential evaluation labels.
D: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 choose a justified supported unknown-category handling strategy and test it without retraining on confidential evaluation labels.
Question 370
Consider the following fictional Databricks production case: 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. What consequence most specifically follows from leaving the issue unchanged?
- A costly driver upgrade leaves worker-side backlog and deadline misses unchanged.
- Multiple trials contend for the same accelerator and fail with memory exhaustion.
- Seasonal behavior is blurred and no store receives its required independent fit.
- Trials are lost or repeated even when compute nodes appear healthy.
Explanation
Correct Answer
A:This consequence follows from the specific unresolved condition. A costly driver upgrade leaves worker-side backlog and deadline misses unchanged. The causal constraint is: Data-parallel workloads may benefit from more executor cores and appropriately partitioned tasks rather than vertical driver scaling.
Incorrect Answers
B:That direct consequence 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 more immediate risk here is a costly driver upgrade leaves worker-side backlog and deadline misses unchanged.
C:That direct consequence is directed at a different scaling and tuning situation: A retailer needs one independent forecasting model per store but concatenates all store histories before fitting one global model. The more immediate risk here is a costly driver upgrade leaves worker-side backlog and deadline misses unchanged.
D:That direct consequence is directed at a different scaling and tuning situation: A researcher parallelizes Optuna trials while sharing one writable SQLite file on a fragile network mount and observes duplicate reservations and storage locks. The more immediate risk here is a costly driver upgrade leaves worker-side backlog and deadline misses unchanged.