Databricks Certified Machine Learning Associate - Blueprint Practice Exam 1

1. A classifier's validation data contains a category absent from training. What should the preprocessing design address?
2. A continuously arriving stream of sensor events must be scored as records are processed. What must the streaming design include?
3. A customer feature table should contain exactly one current row per customer_id. Before writing a new batch, what must the pipeline verify?
4. A data scientist computes Spark DataFrame summary statistics before training. What is the main purpose?
5. A dataset contains a few extremely large salary values. The analyst wants to describe the typical central value without those extremes dominating it. Which statistic is most suitable?
6. A dataset contains many measurements per machine. The intended deployment predicts failures for machines absent from training. Which validation split best fits?
7. A decision tree memorizes training identifiers, but those identifiers are new at deployment. Which change most directly addresses the modeling flaw?
8. A forecasting team evaluates a demand model by randomly mixing records from earlier and later weeks. What evaluation design better matches future prediction?
9. A fraud classifier is evaluated on 10,000 transactions containing 20 fraud cases. It predicts every transaction as legitimate. Why is accuracy misleading here?
10. A model must use the same feature definitions during training and scoring. Which practice reduces inconsistency?
11. A model predicts log-transformed duration, and an analyst reports its raw output as minutes. What is missing?
12. A positive transaction-amount feature has a long right tail. The team considers a log transformation, but some records contain zero. What should be checked before applying it?
13. A price model's evaluation should penalize occasional large numeric errors more strongly than MAE does. Which metric is a suitable candidate?
14. A request-time model needs rapid feature lookups for individual customers. How does an online feature store differ from a typical offline feature table?
15. A researcher changes the learning rate across several training runs and wants to compare that setting with validation performance. Which MLflow data should be logged?
16. A reviewer has the run ID recorded in a model release ticket. What does that identifier enable?
17. A scientist repeatedly installs common machine learning packages on a general-purpose runtime. Which benefit should be evaluated when choosing Databricks Runtime ML?
18. A search objective reports a loss value to a minimization algorithm. The desired business metric is higher-is-better accuracy. What mapping is appropriate?
19. A search tries two values for one parameter and four for another with three-fold cross-validation. Ignoring a final refit, how many training fits are required?
20. A serving request has the right number of values but places them under incorrect feature names. What should the client validate?
21. A team can afford only twelve trials in a large continuous hyperparameter space. Which approach avoids enumerating every point of a fixed grid?
22. A team can reproduce model code but cannot identify the package versions used by an earlier successful experiment. What should its MLOps process retain?
23. A team flags unusually distant feature values using an IQR rule. What should it do before deleting those records?
24. A team introduces a candidate model but wants only a small share of live requests to reach it while comparing results. Which deployment capability is relevant?
25. A team plans to run an unfamiliar AutoML result directly in production. What should it do first?
26. A team registers multiple model versions in Unity Catalog. What advantage does registration provide beyond keeping a loose model file?
27. A team repeatedly tunes against the same holdout until the score looks excellent. What is the main evaluation concern?
28. A team sees stale features in serving even though its offline table has fresh rows. What should it inspect?
29. A training feature represents one of four unordered device types. A linear model should not interpret those types as ordered distances. Which representation is suitable?
30. A training job predicts which customers will renew on a past decision date. Some feature records were produced after that date. What should the feature lookup ensure?
31. A training pipeline includes a transformer and an estimator. Which statement distinguishes their roles?
32. A validated model release is fixed, but an environment's application code must be tested with its own data access settings before release. Why might promoting code be preferred to copying a development model artifact?
33. A web application must return a model prediction within the current user request. Which inference pattern most directly fits?
34. A weekly maintenance report needs predictions for all machines after the week's measurements are finalized. It has no request-time latency requirement. Which design fits?
35. An alerting model should prioritize finding nearly every true failure, even if this raises false alarms. Which metric directly measures how many actual failures it detects?
36. An analyst investigates whether two continuous measurements move together across observations. Which initial analysis is most informative?
37. An experiment produces a confusion matrix image needed during model review. How should it be retained with the run?
38. An imputer learns a numeric replacement value before the dataset is divided into training and test partitions. Why is this problematic?
39. Several workspaces attached to the same Unity Catalog metastore need governed access to a shared customer feature table. What is a suitable design?
40. Two jobs trained models on different dataset snapshots but logged identical code versions. What additional information most directly helps explain their different results?

Comments — 0