Databricks Certified Machine Learning Professional Practice Test 79
Question 780
Consider the following fictional Databricks production case: 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. What consequence most specifically follows from leaving the issue unchanged?
- Production traffic reaches a model not yet loaded or compatible with runtime dependencies.
- Older clients receive validation errors even though new client requests succeed.
- A full immediate cutover could expose every booking to an untested model defect.
- The team cannot recover from a bad promotion without an emergency rebuild.
Explanation
Correct Answer
C:This consequence follows from the specific unresolved condition. A full immediate cutover could expose every booking to an untested model defect. The causal constraint is: A canary rollout can expose a small controlled share of live traffic while retaining a proven model for most requests.
Incorrect Answers
A:That direct consequence is directed at a different deployment strategies situation: An endpoint update request succeeds but the newly configured model version is still loading; monitoring marks the release complete immediately. The more immediate risk here is a full immediate cutover could expose every booking to an untested model defect.
B:That direct consequence is directed at a different deployment strategies situation: A rollout updates the serving model but a renamed input column breaks requests from an older mobile client. The more immediate risk here is a full immediate cutover could expose every booking to an untested model defect.
D:That direct consequence is directed at a different deployment strategies situation: An ML operator promotes a new model version but its automated rollback points to a model that was deleted from the endpoint configuration. The more immediate risk here is a full immediate cutover could expose every booking to an untested model defect.
Question 781
In an enterprise ML workflow, the reported issue is: A fleet sensor table records measurement timestamps and analysts need daily drift trends; a snapshot-only profile erases the time dimension. What likely operational or model-quality harm does this issue create?
- Routine weekend traffic launches unnecessary retraining and distracts engineers.
- Temporal drift is hidden when a cumulative snapshot averages past and current behavior.
- The dashboard reports yesterday's metrics as current and misses a real shift.
- A deteriorating new model is concealed by still-good traffic on the old model.
Explanation
Correct Answer
B:This consequence follows from the specific unresolved condition. Temporal drift is hidden when a cumulative snapshot averages past and current behavior. The causal constraint is: Time-series data profiles use a timestamp and granularity to compare measurements across periods.
Incorrect Answers
A:That direct consequence 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 more immediate risk here is temporal drift is hidden when a cumulative snapshot averages past and current behavior.
C:That direct consequence is directed at a different drift detection and lakehouse monitoring situation: A team has accurate inference logs but its monitoring job never refreshes computed drift metrics after a table schema change. The more immediate risk here is temporal drift is hidden when a cumulative snapshot averages past and current behavior.
D:That direct consequence is directed at a different drift detection and lakehouse monitoring situation: An inference profile groups two active model versions into one aggregate and hides a regression introduced by the new release. The more immediate risk here is temporal drift is hidden when a cumulative snapshot averages past and current behavior.
Question 782
Consider the following fictional Databricks production case: A payment classifier's on-demand feature division returns infinity for zero denominators and fails serialization in the serving endpoint. What consequence most specifically follows from leaving the issue unchanged?
- Late sensor readings are assigned to the wrong time window and mislead incident models.
- Training consumes stale or differently computed features depending on who ran the notebook.
- Invalid numeric features produce serving errors or hidden downstream score distortions.
- Credential compromise exposes unrelated customers' private feature data.
Explanation
Correct Answer
C:This consequence follows from the specific unresolved condition. Invalid numeric features produce serving errors or hidden downstream score distortions. The causal constraint is: On-demand feature functions must explicitly handle invalid input and output type constraints.
Incorrect Answers
A:That direct consequence is directed at a different advanced feature store concepts situation: A sensor feature pipeline computes rolling 30-minute averages by arrival time, although delayed events carry trustworthy event timestamps. The more immediate risk here is invalid numeric features produce serving errors or hidden downstream score distortions.
B:That direct consequence is directed at a different advanced feature store concepts situation: A team manually reruns a feature notebook each week before training; feature definitions change without an execution history or stable destination. The more immediate risk here is invalid numeric features produce serving errors or hidden downstream score distortions.
D:That direct consequence is directed at a different advanced feature store concepts situation: A model serving service uses a broadly privileged shared token to read every online feature table when it needs only one approved feature set. The more immediate risk here is invalid numeric features produce serving errors or hidden downstream score distortions.
Question 783
A production ML review contains this situation: A monitoring dashboard reports a statistically significant tiny difference after one million observations, and the team immediately treats it as operationally material drift. Which statement best explains the failure mode?
- Statistical significance and practical effect size differ; large samples can make negligible shifts detectable.
- Reference populations should match relevant seasonality and business context when interpreting drift.
- A snapshot profile suits a general current table when no prediction or event-time model-monitoring contract exists.
- Global thresholds can hide local critical slices; alerting should reflect operational segmentation where justified.
Explanation
Correct Answer
A:The technical constraint is specific to the reported failure. Statistical significance and practical effect size differ; large samples can make negligible shifts detectable. To test it, inspect sample size, numeric drift statistic and estimated business impact for the same comparison.
Incorrect Answers
B:That underlying principle 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 current incident instead turns on statistical significance and practical effect size differ; large samples can make negligible shifts detectable.
C:That underlying principle is directed at a different drift detection and lakehouse monitoring situation: An analyst wants to inspect today’s static customer table for missing values and profile changes but chooses an inference monitor requiring model IDs and predictions. The current incident instead turns on statistical significance and practical effect size differ; large samples can make negligible shifts detectable.
D:That underlying principle is directed at a different drift detection and lakehouse monitoring situation: An alert query uses one global transaction-failure count and masks a complete outage in a small high-value region. The current incident instead turns on statistical significance and practical effect size differ; large samples can make negligible shifts detectable.
Question 784
In an enterprise ML workflow, the reported issue is: A retraining pipeline ranks candidate models using metrics calculated on different dataset versions, then selects the numerically highest result. Which change addresses the real cause without undermining other requirements?
- Compute performance only on matured labeled cohorts and use separate permitted unlabeled drift signals where relevant.
- Run candidate training after the trigger but require explicit quality/security gates before any live promotion.
- Evaluate all eligible candidates against one controlled frozen benchmark and rank only comparable results.
- Gate retraining on validated drift or predictive performance signals with cooldown and approval rules.
Explanation
Correct Answer
C:This corrects the relevant failure boundary instead of merely hiding the symptom. Evaluate all eligible candidates against one controlled frozen benchmark and rank only comparable results. The issue matters because a model wins because it was tested on easier data rather than because it performs better.
Incorrect Answers
A:That corrective action is directed at a different automated retraining situation: A credit model's true-label feedback arrives 45 days late, yet the retraining scheduler assumes yesterday's unlabeled predictions provide yesterday's observed accuracy. The relevant corrective action here is evaluate all eligible candidates against one controlled frozen benchmark and rank only comparable results.
B:That corrective action is directed at a different automated retraining situation: An alert from a feature monitor automatically replaces the live model with a newly trained artifact without testing or approval. The relevant corrective action here is evaluate all eligible candidates against one controlled frozen benchmark and rank only comparable results.
D:That corrective action is directed at a different automated retraining situation: An automated retraining job triggers whenever a numerical feature mean shifts by 0.1, regardless of normal seasonal changes or sample size. The relevant corrective action here is evaluate all eligible candidates against one controlled frozen benchmark and rank only comparable results.
Question 785
The engineering team encounters the following issue: A sensor feature pipeline computes rolling 30-minute averages by arrival time, although delayed events carry trustworthy event timestamps. What investigation would most effectively isolate this behavior?
- Check entity join cardinality and the most recent eligible feature timestamp for each observation.
- Compare current feature-table history and prior Delta versions against the training-set timestamp requirements.
- Compare offline and serving results for synthetic zero-income, missing and normal customer fixtures.
- Compare event timestamps, ingestion timestamps and window assignment for late synthetic readings.
Explanation
Correct Answer
D:This is the most discriminating check for the observed symptoms. Compare event timestamps, ingestion timestamps and window assignment for late synthetic readings. The underlying technical issue is that streaming feature windows should use appropriate event-time semantics and a stated lateness policy.
Incorrect Answers
A:That diagnostic evidence is directed at a different advanced feature store concepts situation: A training dataset combines a feature table with multiple timestamped rows per device using an equality join on device ID only. For the stated incident, the decisive observation is compare event timestamps, ingestion timestamps and window assignment for late synthetic readings.
B:That diagnostic evidence is directed at a different advanced feature store concepts situation: A model's feature generation step overwrites all feature-table history daily even though time-aware training requires earlier snapshots. For the stated incident, the decisive observation is compare event timestamps, ingestion timestamps and window assignment for late synthetic readings.
C:That diagnostic evidence is directed at a different advanced feature store concepts situation: A customer-risk feature needs a ratio of current debt to income, but online serving recomputes it differently from the training pipeline. For the stated incident, the decisive observation is compare event timestamps, ingestion timestamps and window assignment for late synthetic readings.
Question 786
In an enterprise ML workflow, the reported issue is: A moderately sized dataset fits on one worker but a team provisions a huge distributed cluster and observes overhead higher than model fit. How should the engineer validate the correction?
- A controlled A/B benchmark demonstrates the chosen configuration's cost-performance advantage.
- Each source record receives exactly one score from its assigned region-specific model version.
- Ray reports parallel trial schedules that respect assigned GPUs and complete the planned search.
- Tracing shows the intended supported distributed training path with correct outputs and resource use.
Explanation
Correct Answer
A:This verification measures the required result rather than just process liveness. A controlled A/B benchmark demonstrates the chosen configuration's cost-performance advantage. It confirms whether the targeted action succeeded: choose a measured right-sized execution mode instead of presuming more nodes means faster fit.
Incorrect Answers
B:That acceptance test 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. Here the actual success criterion is a controlled A/B benchmark demonstrates the chosen configuration's cost-performance advantage.
C:That acceptance test 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. Here the actual success criterion is a controlled A/B benchmark demonstrates the chosen configuration's cost-performance advantage.
D:That acceptance test is directed at a different scaling and tuning situation: A team switches a SparkML training stage to Ray without verifying that its chosen estimator supports Ray's distributed fit contracts. Here the actual success criterion is a controlled A/B benchmark demonstrates the chosen configuration's cost-performance advantage.
Question 787
In an enterprise ML workflow, the reported issue is: A sensor feature pipeline computes rolling 30-minute averages by arrival time, although delayed events carry trustworthy event timestamps. Which change addresses the real cause without undermining other requirements?
- Use an update strategy that preserves required historical feature states and governance metadata.
- Promote a versioned feature definition with backward-compatibility tests and controlled rollout.
- Assign a restricted serving identity and permit access only to required approved online feature assets.
- Apply supported event-time windows and watermark behavior to compute the intended rolling features.
Explanation
Correct Answer
D:This corrects the relevant failure boundary instead of merely hiding the symptom. Apply supported event-time windows and watermark behavior to compute the intended rolling features. The issue matters because late sensor readings are assigned to the wrong time window and mislead incident models.
Incorrect Answers
A:That corrective action is directed at a different advanced feature store concepts situation: A model's feature generation step overwrites all feature-table history daily even though time-aware training requires earlier snapshots. The relevant corrective action here is apply supported event-time windows and watermark behavior to compute the intended rolling features.
B:That corrective action is directed at a different advanced feature store concepts situation: A team updates a shared on-demand feature function without version control and unexpectedly changes both batch retraining and online serving behavior. The relevant corrective action here is apply supported event-time windows and watermark behavior to compute the intended rolling features.
C:That corrective action is directed at a different advanced feature store concepts situation: A model serving service uses a broadly privileged shared token to read every online feature table when it needs only one approved feature set. The relevant corrective action here is apply supported event-time windows and watermark behavior to compute the intended rolling features.
Question 788
In an enterprise ML workflow, the reported issue is: A sensor classifier silently reuses features from a year-old device after months without readings, despite a maximum allowed lookback of seven days. Before modifying the pipeline, which evidence should be collected first?
- Compare last available reading age with each prediction cutoff and its permitted freshness window.
- Inspect feature refresh run metadata, table versions and transformation code revisions.
- Compare event timestamps, ingestion timestamps and window assignment for late synthetic readings.
- Compare offline source update timestamps, online publication timestamps and stated staleness limits.
Explanation
Correct Answer
A:This is the most discriminating check for the observed symptoms. Compare last available reading age with each prediction cutoff and its permitted freshness window. The underlying technical issue is that temporal feature joins may need an explicit bounded lookback so stale records are not mistaken for current observations.
Incorrect Answers
B:That diagnostic evidence is directed at a different advanced feature store concepts situation: A team manually reruns a feature notebook each week before training; feature definitions change without an execution history or stable destination. For the stated incident, the decisive observation is compare last available reading age with each prediction cutoff and its permitted freshness window.
C:That diagnostic evidence is directed at a different advanced feature store concepts situation: A sensor feature pipeline computes rolling 30-minute averages by arrival time, although delayed events carry trustworthy event timestamps. For the stated incident, the decisive observation is compare last available reading age with each prediction cutoff and its permitted freshness window.
D:That diagnostic evidence is directed at a different advanced feature store concepts situation: A team publishes an online feature table but neglects its refresh schedule while the predictive model assumes values no older than 15 minutes. For the stated incident, the decisive observation is compare last available reading age with each prediction cutoff and its permitted freshness window.
Question 789
A production ML review contains this situation: A retailer's features arrive two hours late, and offline training uses event time alone even though the online system could not have seen those late arrivals at prediction time. Which diagnostic check provides the strongest direct evidence?
- Inspect state store durability, output table persistence and recovery behavior under restart.
- Measure endpoint latency attributable to offline table reads versus approved online lookup latency.
- Compare feature event time, publish time and prediction decision time for a delayed fixture.
- Compare offline and serving results for synthetic zero-income, missing and normal customer fixtures.
Explanation
Correct Answer
C:This is the most discriminating check for the observed symptoms. Compare feature event time, publish time and prediction decision time for a delayed fixture. The underlying technical issue is that point-in-time correctness may require both event time and actual feature availability when ingestion latency matters.
Incorrect Answers
A:That diagnostic evidence is directed at a different advanced feature store concepts situation: A business deploys a model whose features are computed from a stream in memory only and disappear when its compute cluster restarts. For the stated incident, the decisive observation is compare feature event time, publish time and prediction decision time for a delayed fixture.
B:That diagnostic evidence is directed at a different advanced feature store concepts situation: A pricing endpoint queries a large offline Delta feature table for every user request, and P95 response time reaches several seconds. For the stated incident, the decisive observation is compare feature event time, publish time and prediction decision time for a delayed fixture.
D:That diagnostic evidence is directed at a different advanced feature store concepts situation: A customer-risk feature needs a ratio of current debt to income, but online serving recomputes it differently from the training pipeline. For the stated incident, the decisive observation is compare feature event time, publish time and prediction decision time for a delayed fixture.
Question 790
Review this production model incident: A model serving service uses a broadly privileged shared token to read every online feature table when it needs only one approved feature set. What end-to-end verification is most persuasive here?
- The model input is missing or flagged stale when no eligible feature exists inside the authorized window.
- A restart restores feature values consistent with the pre-restart control dataset.
- Authorized prediction requests work while a test access to an unrelated feature table is denied.
- A past labeled event can still retrieve the correct historical feature state.
Explanation
Correct Answer
C:This verification measures the required result rather than just process liveness. Authorized prediction requests work while a test access to an unrelated feature table is denied. It confirms whether the targeted action succeeded: assign a restricted serving identity and permit access only to required approved online feature assets.
Incorrect Answers
A:That acceptance test is directed at a different advanced feature store concepts situation: A sensor classifier silently reuses features from a year-old device after months without readings, despite a maximum allowed lookback of seven days. Here the actual success criterion is authorized prediction requests work while a test access to an unrelated feature table is denied.
B:That acceptance test is directed at a different advanced feature store concepts situation: A business deploys a model whose features are computed from a stream in memory only and disappear when its compute cluster restarts. Here the actual success criterion is authorized prediction requests work while a test access to an unrelated feature table is denied.
D:That acceptance test is directed at a different advanced feature store concepts situation: A model's feature generation step overwrites all feature-table history daily even though time-aware training requires earlier snapshots. Here the actual success criterion is authorized prediction requests work while a test access to an unrelated feature table is denied.