Free Snowflake SnowPro Advanced: Data Scientist (DSA-C03) Sample Questions

Correct answer: C — Multi-class classification, because each ticket receives exactly one label out of six mutually exclusive classes

• A. Incorrect. A single yes/no question for one queue ignores the other five queues, so it cannot route a ticket to its correct destination.
• B. Incorrect. Existing queue tags provide supervision, and discovered clusters would not map reliably onto the six business queues the team defined.
• C. Correct. One label from more than two exclusive categories, learned from tagged examples, is the definition of multi-class classification.
• D. Incorrect. The target is a category, not a number; scoring queues is how classifiers work internally rather than a separate regression framing.

Correct answer: D — Supervised image segmentation, since the model must predict a label for each pixel using the drawn masks as targets

• A. Incorrect. A whole-image label cannot give region boundaries or volume; the stated need is the exact outline, which requires pixel-level predictions.
• B. Incorrect. Grouping similar scans does not locate a tumour, and the masks already supply labels, so supervised learning is both possible and necessary.
• C. Incorrect. File size has no reliable link to tumour volume, and the first step is localising the region, which regression on a scalar cannot do.
• D. Correct. Predicting a mask that marks which pixels belong to the tumour is segmentation, supervised by the radiologists' annotations.

Correct answers: B, C — Subscribe to a Snowflake Marketplace listing for the demographics, which exposes the data through secure data sharing without copying files.; Configure Snowpipe with auto-ingest on the storage location so new clickstream files are loaded into a table shortly after they arrive.

• A. Incorrect. Manual downloads add latency and break reproducibility, and a daily upload cannot keep up with files arriving every minute.
• B. Correct. Marketplace and secure data sharing give live access to provider data in the account, with no ETL pipeline to maintain.
• C. Correct. Snowpipe provides continuous, event-driven loading of files as they land, which suits files arriving every minute.
• D. Incorrect. Scraping is brittle and unsupported by the provider, while a shared data listing is governed and kept current for you.
• E. Incorrect. A yearly snapshot would leave the model blind to recent behavior, which is exactly what the continuous feed is meant to supply.

Correct answers: C, E — A column storing the cancellation reason, which is filled in only after a customer has already churned.; Roughly 40% of the values in an important usage feature are null because the sensor was offline for some regions.

• A. Incorrect. A moderate expected correlation between two real features is normal and does not need correction before training.
• B. Incorrect. Unique identifiers should simply be excluded as features, which is a routine step rather than a discovered data defect.
• C. Correct. A field populated after the outcome leaks the label into the features and inflates offline metrics, so it must be removed.
• D. Incorrect. A 12% positive rate is moderate imbalance that standard classifiers and metrics can handle without blocking training.
• E. Correct. Heavy missingness has to be handled by imputation, an indicator feature or exclusion before the model can use that feature reliably.
• F. Incorrect. Skew matters for some linear models, but tree-based models split on thresholds and tolerate it, so it does not block training.

Correct answer: D — `xgboost.XGBRegressor` fitted on the features with the duration column as label, since the target is a continuous number.

• A. A classifier treats durations as unrelated categories, discards their ordering and cannot predict values missing from training, so it suits a continuous target poorly.
• B. KMeans is unsupervised and ignores the duration label during fitting. Its clusters are not optimized to predict the target, so it is not a proper regression approach.
• C. PCA is an unsupervised projection that maximizes variance, not agreement with the label. Its component is a rescaled mix of features and has no meaning as minutes.
• D. A continuous numeric label calls for supervised regression, and XGBRegressor learns that mapping and returns a predicted number of minutes for each delivery.

Correct answer: A — Define the problem, collect and explore data, engineer features, split data, train, evaluate, register the model, deploy, then monitor.

• A. This is the standard lifecycle: frame the problem, understand and prepare data, split before training, evaluate on held-out data, then deploy and monitor.
• B. Defining the problem after deployment and splitting after training are both out of order, and the test set would already have influenced the model.
• C. Training before exploration and feature engineering is premature, and splitting after evaluation means no held-out data was ever available to evaluate on.
• D. Deploying before the problem is defined or any model has been trained and evaluated puts an unvalidated artifact into production.

Correct answer: B — Two-proportion z-test on conversion rates, because large independent groups let a normal approximation describe the rate gap.

• A. Incorrect. Shapiro-Wilk assesses whether data follow a normal distribution; it does not compare two group rates.
• B. Correct. Comparing two independent proportions with large counts is the standard use of a two-proportion z-test.
• C. Incorrect. The visitors in each variant are different people, so there is no natural pairing and a paired test would be invalid.
• D. Incorrect. Variant A's 4.0% is itself an estimate with sampling error, so treating it as a known constant understates uncertainty.

Correct answer: C — If the true mean difference were zero, results at least this extreme would appear about 3% of the time, so the null is rejected at 5%.

• A. Incorrect. One minus the p-value is not a reproducibility rate; replication depends on the true effect size and statistical power.
• B. Incorrect. P-values depend on sample size as well as effect size, so a small p-value can accompany a trivially small difference.
• C. Correct. A p-value is the probability of data at least as extreme as observed, assuming the null hypothesis is true; 0.03 is below 0.05, so the null is rejected.
• D. Incorrect. The p-value is not the probability that the null is true; that would need a prior and a Bayesian calculation.

Correct answer: A — Roughly 95% of values lie within two standard deviations of the mean, and the mean, median and mode all coincide at the centre.

• A. Correct. The 68-95-99.7 rule gives about 95% within two standard deviations, and a normal curve is symmetric so mean, median and mode are equal.
• B. Incorrect. About 99.7% falls within three standard deviations; one standard deviation covers about 68%, and mean, median and mode coincide.
• C. Incorrect. About 68% falls within one standard deviation, not two, and symmetry means the mean equals the median.
• D. Incorrect. About 68% falls within one standard deviation, and a normal distribution is symmetric rather than skewed.

Correct answer: B — It drops rows that have fewer than 3 non-null values across the four listed columns and keeps rows with at least 3.

• A. Incorrect. `thresh` counts required non-null values, not the number of nulls that triggers a drop; the two are not the same cut-off.
• B. Correct. `thresh` sets the minimum number of non-null values a row must have among the `subset` columns to be retained.
• C. Incorrect. `thresh` is honoured together with `subset`, and it takes precedence over the `how` setting.
• D. Incorrect. `dropna` works on rows, not columns, so no column is removed by this call.

Correct answers: A, D — Parse the date with `TRY_TO_DATE(payload:ts::STRING, 'DD/MM/YYYY')`, so the day-month order is explicit and malformed values return NULL.; Extract the amount with `payload:amount::NUMBER(10,2)` so the JSON string becomes an exact decimal column instead of remaining a VARIANT.

• A. Correct. Supplying the format removes the day-versus-month ambiguity, and the TRY variant avoids query failure on bad strings.
• B. Incorrect. FLOAT is binary floating point and cannot represent many decimal amounts exactly, so currency sums drift.
• C. Incorrect. `TRY_CAST` accepts no format argument; format models belong to `TRY_TO_DATE` and related conversion functions.
• D. Correct. The path expression plus a `::` cast produces a typed column, and NUMBER with scale preserves exact cents.
• E. Incorrect. `TO_DATE` raises an error on unparseable input, and without a format the day-first strings may be misread.

Correct answer: B — Compute the z-score with AVG and STDDEV as window functions OVER (PARTITION BY sensor_id) in a CTE, then filter on ABS(z) > 3 in the outer query.

• A. Window functions are not permitted in WHERE because it is evaluated before windows are computed; use a CTE or QUALIFY instead.
• B. Window functions keep every row while attaching per-sensor statistics, and the filter on the computed z-score goes in an outer query after the window is evaluated.
• C. Without PARTITION BY the mean and spread come from all sensors blended together, not from each sensor's own readings.
• D. Grouping collapses the rows to one per sensor, so the individual readings are no longer available to compare.

Correct answers: B, D — REGR_SLOPE(y, x) uses only those 750,000 complete pairs, so it can differ from a slope after imputing.; REGR_COUNT(y, x) returns 750,000, which is the number of rows in which both y and x are non-NULL values.

• A. The function skips incomplete pairs and returns a value from the remaining rows.
• B. Rows with a NULL in either column are excluded, so imputation would change the data the slope is fitted on.
• C. NULLs are not converted to zero; those rows are excluded entirely.
• D. The REGR functions use only pairs in which both values are present.
• E. The count is limited to complete pairs and would not include the 250,000 rows with a NULL.

Correct answer: C — `LabelEncoder` fitted on the single label column, which maps each distinct class to an integer between zero and the class count minus one.

• A. One-hot expansion produces several indicator columns, but a single-label classifier expects one target column of class codes, not a multi-column indicator matrix.
• B. KBinsDiscretizer divides continuous numeric ranges into intervals and does not accept arbitrary string categories as input.
• C. LabelEncoder is intended for target labels. It operates on one column and assigns integer codes to each distinct class value.
• D. Binarizer thresholds numeric values into two classes. It cannot map three or more string classes into distinct integer codes.

Correct answer: B — The second `fit` recomputes mean and standard deviation from test data, so test rows are scaled with different statistics.

• A. Transformers in Snowflake ML transform Snowpark DataFrames directly, with the work pushed down to the warehouse.
• B. The scaler should be fitted once on training data and reused through `transform` only, otherwise train and test features sit on different scales.
• C. Fitting on combined data pushes information about the test set into preprocessing, which leaks evaluation data into training.
• D. StandardScaler accepts distinct output column names, which is the normal way to keep original columns alongside the scaled ones.

Correct answer: C — The data likely mixes two user groups, such as quick browsers and long viewers, so analysis should segment them instead of using the mean.

• A. With two humps, the mean lies in the valley where very few sessions occur, so it misdescribes both groups.
• B. Multimodal data is common and legitimate, and unimodality is not a requirement for modelling.
• C. A bimodal distribution suggests distinct subpopulations, and the mean falls where few users actually sit, so segmentation is the useful next step.
• D. Two well-separated modes of this size are real structure; a single bin would just hide the pattern.

Correct answers: A, C — Treat them as seasonality rather than anomalies, and use a seasonal decomposition or per-month baseline before flagging deviations.; Present the line chart with the yearly overlay so the business sees the drop is recurring and planned inventory should follow it.

• A. A regular yearly pattern is expected behavior, so outlier rules should compare each month with its own seasonal baseline.
• B. Removing a recurring valid pattern would bias any forecast and rests on the false claim that all extreme values are errors.
• C. Overlaying years on one chart shows that the pattern repeats, which turns the flagged points into a planning insight.
• D. Imputing the global mean erases a real business signal and distorts the reported trend.
• E. Raising the threshold until nothing is flagged would hide true anomalies in other months as well.

Correct answer: C — Build a Snowpark `Session` from connection parameters and chain DataFrame transformations on `session.table`, so the SQL runs on a warehouse.

• A. `to_pandas()` materialises the whole table on the client, so the features would be computed locally rather than pushed down to the warehouse.
• B. Pulling all 2 billion rows to the client moves the data out of Snowflake and will exhaust laptop memory long before features are computed.
• C. Snowpark DataFrame operations are compiled to SQL and executed on the warehouse, so only small results ever reach the laptop.
• D. Unloading and downloading copies the data out of Snowflake, which is the opposite of connecting the tool directly to the data.

Correct answer: C — Import `modin.pandas` together with the Snowpark pandas plugin, so existing pandas-style calls are translated to SQL and run in Snowflake.

• A. The warehouse size does not change client memory, so 500 million rows would still have to fit on the machine running pandas.
• B. Those classes are ML transformers with their own API; they do not interpret arbitrary pandas calls.
• C. The Snowpark pandas API keeps the pandas syntax while pushing the computation down to Snowflake, so little code changes and no data is pulled locally.
• D. A UDF runs per row and cannot host a notebook that expects the whole DataFrame, so this does not translate pandas operations.

Correct answers: A, B — Embed each runbook chunk once and persist the vectors in a `VECTOR` column so they are not recomputed for every question.; Embed the user question with the same embedding model used for the chunks, then rank chunks by a vector similarity function.

• A. Precomputing and storing chunk embeddings avoids repeated token charges and makes similarity search fast.
• B. Query and document vectors must come from the same model to share one vector space; only then is similarity meaningful.
• C. Mixing models yields vectors in unrelated spaces, so similarity scores are meaningless, and dimensions may not even match.
• D. That replaces retrieval with fine-tuning, which is a different technique that cannot guarantee grounded, up-to-date answers.
• E. Whole-table prompts exceed context windows and multiply token cost, which is the problem retrieval is meant to avoid.

Correct answers: B, C — Training cost is input tokens multiplied by epochs trained, in addition to normal storage and warehouse charges.; Inference uses `AI_COMPLETE` (or `COMPLETE`) with the fine-tuned model's name in place of a base model name, and is billed on tokens.

• A. Cross-region inference is not supported for fine-tuned models.
• B. Fine-tuning is billed by trained tokens, which scale with both dataset size and epochs.
• C. The tuned model is invoked like any other model by name, with token-based billing.
• D. Cortex Fine-tuning uses parameter-efficient techniques and charges for trained tokens.
• E. They are called through the same SQL function interface as other Cortex models.

Correct answers: C, D — An XGBoost classifier logged from the xgboost Python package; A LightGBM regressor logged from the lightgbm Python package

• A. Incorrect: Keras models are not among the libraries supported for built-in explainability.
• B. Incorrect: deep learning models via PyTorch are not in the supported list for registry Shapley explanations.
• C. Correct: XGBoost models are supported for built-in explainability.
• D. Correct: LightGBM is among the supported libraries.
• E. Incorrect: Hugging Face pipelines are not covered by the tree and scikit-learn explainability support.

Correct answer: B — No, because the features use different scales; standardizing first makes coefficients comparable as per-standard-deviation effects.

• A. Incorrect: linear coefficients are directly interpretable when read in context.
• B. Correct: raw coefficients depend on units, so standardized coefficients are needed for comparing importance.
• C. Incorrect: linear coefficients are per-unit effects, not scale-free attributions.
• D. Incorrect: coefficient size depends on the feature's units.

Correct answer: A — Compute expected payout at every candidate threshold from its TPR, FPR, prevalence and the cost matrix, then take the maximum

• A. The best point on a ROC curve depends on prevalence and error costs, so payout must be computed per threshold. The cost-weighted maximum rarely sits at a default location.
• B. AUC summarizes ranking quality across all thresholds and says nothing about which cut-off is cheapest. The default 0.5 is not tied to the cost matrix.
• C. TPR equal to FPR is the line of a random classifier, so it carries no useful operating point. It ignores both costs and prevalence.
• D. Closest-to-corner treats false positives and false negatives as equally costly. With a ten-to-one cost ratio the optimum sits elsewhere on the curve.

Correct answers: C, D — Adjusted R² penalizes predictors that fail to improve fit enough, so it would likely drop and reveal the new features are not useful; Training R² never decreases when predictors are added to a least-squares model, so the 0.02 gain is not evidence of real signal

• A. Adjusted R² is less than or equal to R² for models with predictors, because of the penalty term.
• B. Any training-set metric rewards added complexity. Only held-out data or penalized metrics expose overfitting.
• C. Adjusted R² adds a penalty for each extra parameter. A tiny R² gain from 40 features usually produces a lower adjusted value.
• D. Extra predictors can only equal or raise in-sample R². Only held-out metrics show whether they generalize.
• E. A rise in training R² is guaranteed by adding terms, whatever their value. Worse validation error is the stronger evidence.

Correct answer: B — RMSE, because squaring each residual before averaging weights large forecast misses much more heavily than small ones.

• A. Log loss scores predicted class probabilities, not continuous values, and it is not a squared distance.
• B. RMSE squares errors, so a few large misses dominate the score. That matches a cost structure where large errors are disproportionately expensive.
• C. MAE weights errors linearly, so large misses are not penalized extra. It is more robust to outliers, which is the opposite need.
• D. AUC evaluates binary classifiers. It does not measure the size of numeric forecast errors.

Correct answer: A — Step 3, Step 5, Step 2, Step 4, Step 1

• A. The endpoint comes first, then the role that can invoke it, then the integration that references the role. The integration's values are used for the trust policy, and the function is created last.
• B. The trust policy update cannot come after creating the function, because calls through the function would fail until it is in place.
• C. The integration needs the ARN of the role, so the role must exist before the integration is created.
• D. The integration needs both the role ARN and the endpoint URL, so it cannot be created first.

Correct answers: B, D, F — The task is resumed with `ALTER TASK ... RESUME` after creation, since scheduled tasks start in a suspended state.; The role owning the task has the account-level `EXECUTE TASK` privilege, along with usage on the warehouse the task uses.; The stored procedure lists `snowflake-ml-python` in its `PACKAGES` so the Snowflake ML library is available at runtime.

• A. Incorrect: aliases are not a prerequisite for logging a version.
• B. Correct: without resuming, the schedule never fires.
• C. Incorrect: tasks can reference objects in other schemas when privileges allow.
• D. Correct: tasks need the execute privilege and compute to run.
• E. Incorrect: Python procedures run on a warehouse and do not require a compute pool.
• F. Correct: the procedure must declare the packages it imports, including the Snowflake ML package.

Correct answers: C, F — A source table or view holding logged inferences, including a timestamp column and the prediction columns.; The specific model version and the model method, such as predict, whose inputs and outputs are being tracked.

• A. Incorrect: tags are not a prerequisite for computing metrics.
• B. Incorrect: monitor refreshes use a warehouse rather than a compute pool.
• C. Correct: the source holds the inference records, and the timestamp column enables time-based aggregation.
• D. Incorrect: a Dataset version is not required to define a monitor, and it never supplies compute.
• E. Incorrect: monitors read their source table directly and do not need a stream.
• F. Correct: a monitor is created for a particular model version and function.

Correct answers: C, E — Warehouse deployment resolves dependencies from the Snowflake Anaconda channel, so the unavailable library blocks that target platform; Snowpark Container Services deployment can install the library with `pip_requirements`, because the service image is built with those

• A. Snowflake restricts conda resolution for warehouses to its own channel, and arbitrary channels are not consulted.
• B. `relax_version` loosens the version pinning of dependencies. It does not widen the set of channels.
• C. Warehouse inference installs packages from the Snowflake channel, so an unpublished library cannot be satisfied there.
• D. Registered models can use pip packages when served from containers, so a rewrite is unnecessary.
• E. Container-based serving builds an image, which allows packages from PyPI through `pip_requirements`.

Correct answer: D — Log the model with a signature > validate the new version > set it as default > score with `mv.run` > save results to a table

• A. Saving results before scoring or validating is impossible because no predictions exist yet.
• B. Inference with `mv.run` needs a logged model version, so it cannot come first.
• C. A default version cannot be set before the version exists, and promotion should follow validation.
• D. The model must be logged before it can be validated, promoted and run, and predictions are persisted at the end.

Correct answer: A — Subclass `CustomModel`, expose a method with `@custom_model.inference_api`, and log the instance with `log_model`

• A. A custom model class wraps arbitrary Python logic and its inference methods become callable model methods.
• B. The registry stores model objects it logs. A pickle on a stage is not a logged model and has no versions or methods.
• C. Unsupported frameworks can be wrapped with a custom model class, so no rewrite is needed.
• D. No vendor change is needed because the custom model class exists for this case.

Correct answer: A — It provides reference distributions that production rows in the source table are compared against when drift metrics are computed

• A. The optional baseline table exists for comparative operations such as drift. Without it, drift metrics have no reference distribution to compare live data with.
• B. Computed metrics are managed by the monitor itself and read through the monitor functions. The baseline is user-supplied input, not output storage.
• C. Model monitors observe and compute metrics; they never retrain a model. The baseline is a reference dataset, not a retraining queue.
• D. The baseline does not filter the source table. Time windows come from the timestamp column and the aggregation settings.

Correct answer: B — `F1_SCORE`, the harmonic mean of precision and recall, which drops sharply if either type of error grows too large

• A. MAPE is a regression metric and is not offered for binary classification monitors.
• B. F1 balances precision and recall and suits imbalanced problems.
• C. Accuracy is dominated by the majority class and hides both error types.
• D. MSE is a regression metric and does not balance precision and recall.

Correct answer: B — `MODEL_MONITOR_STAT_METRIC`, which returns counts and null counts for the columns logged in the source table

• A. Performance functions evaluate predictions against actuals, not feature nulls.
• B. The statistic metric function exists for counts and null values on monitored columns.
• C. Drift names are distribution distances, and RMSE is a performance metric.
• D. No such function exists in Snowflake.