Free DSA-C03 practice test: 30 questions with answers

Predicting a yes or no outcome from labeled history is binary classification. Clustering ignores labels, revenue forecasting predicts a value over time, and anomaly detection does not use the churn label.

Clustering finds groups in unlabeled data. Logistic regression and classification need labels, and reinforcement learning uses rewards.

Regularization reduces overfitting by discouraging overly complex models. Adding label noise, fewer data points and removing the test set do not address generalization.

Early stopping on a validation set chooses the number of trees without using the test set. Tuning on the test set leaks information, training loss alone keeps decreasing, and maximum trees may overfit.

Metadata on micro-partitions enables pruning, and loading organizes data automatically. Users do not manage them manually, they apply to native tables, and they enable rather than prevent pruning.

Filter selectivity relative to clustering and warehouse resources drive query performance. Result colors, login counts and editor fonts do not.

Clustering keys organize micro-partitions so filters on the key scan fewer partitions. User counts, names and Time Travel settings do not improve pruning.

Virtual warehouses provide compute for queries and Snowpark code. Stages store files, shares expose data, and streams capture changes.

Cleaning, transforming and validating derived features form the core of feature engineering. Printing names, color changes and deleting source data are not pipeline steps.

Diagnosing drift by comparing recent inputs and labels with the baseline guides the right fix. Retraining on old data, disabling monitoring and replacing without analysis do not address the cause.

A conditional pipeline triggers retraining based on measured performance, which is efficient and responsive. Constant retraining, memory-based triggers and never retraining do not follow the requirement.

Logging with version and inputs to a governed table supports audits and debugging. Screenshots, memory-only storage and no logs do not support audit needs.

Secure data sharing exposes data to another account without copying it, and access can be limited. Email exports, public pages and owner roles create copies or excessive access.

Row access policies filter rows based on identity or role, which enforces the requirement. Column masking hides values rather than rows, while manual calls and separate tables do not enforce access.

Aggregation policies or secure views limit exposure to aggregated results. Full access, public stages and image storage expose or obscure data inappropriately.

Window functions with partitions and ordered ranges compute rolling averages per group. GROUP BY collapses rows, cross joins explode data, and DISTINCT does not compute averages.

LAG returns the prior row's value within an ordered partition. The other functions do not return the previous row's value.

Path notation and FLATTEN access nested fields in semi-structured data. Editors, images and ignoring keys do not produce features in SQL.

Computing the imputation statistic on the training split avoids leakage from the test split. Full-dataset statistics leak information, text placeholders break numeric types, and deleting all incomplete rows can discard most of the data.

Grouping rare categories and encoding the frequent ones controls dimensionality. Unlimited one-hot encoding creates huge inputs, raw names are not meaningful numbers, and dropping without analysis discards signal.

Windowed calculations limited to prior events prevent future information from entering features. Using future events, labels and only the final month introduce leakage or bias.

Snowpark ML trains models on Snowflake compute close to the data, which reduces movement. The other options describe unnecessary exports or incorrect limitations.

The model registry records versions, metadata and metrics for comparison and deployment. File names, warehouse comments and screenshots do not provide structured model lineage.

Registry-based batch inference on warehouse compute scales with the table and avoids data movement. Local row-by-row scoring, image conversion and scoring one row do not meet the requirement.

Snowpark pushes DataFrame operations to Snowflake compute so data does not need to leave the platform. Exporting to laptops, external copies and browser storage move or duplicate sensitive data unnecessarily.

A Snowpark DataFrame is lazily evaluated and executes transformations in Snowflake when an action such as collect runs. A stage stores files, a warehouse provides compute, and a task schedules work.

Snowpark defers execution until an action such as collect or show requests results, which lets Snowflake optimize the query. Replication, cloning and time travel are different features.

Null-rate checks with alerts catch quality issues early. Reactive checks, assumptions and deletion do not provide reliable early detection.

Comparing predicted class proportions with the baseline reveals distribution shifts. Column counts, file sizes and dashboard counts do not measure predicted classes.

Time Travel and query history reveal what changed and when. Deleting or ignoring the drop loses evidence, and comments do not show changes.