Databricks Machine Learning Associate Practice Exam

• A. Copy notebooks between workspaces by hand after each change
• B. Develop and test directly in the production workspace to avoid drift
• C. Keep code in Git and track runs and models with MLflow Correct answer
• D. Share one all-purpose cluster for training, testing and serving

Option C is correct because keeping code in Git and tracking runs and models with MLflow gives a reproducible, auditable workflow. Hand-copying notebooks, developing in production and sharing one all-purpose cluster for training, testing and serving all remove isolation and repeatability, so they are not sound MLOps practice.

• A. fe.create_table(name="ml.shop.customer_features", primary_keys=["customer_id"], df=features_df) Correct answer
• B. fe.create_table(name="ml.shop.customer_features", keys=["customer_id"], df=features_df)
• C. fe.write_table(name="ml.shop.customer_features", df=features_df, mode="merge")
• D. fe.create_feature_table(name="ml.shop.customer_features", keys=["customer_id"], features_df=features_df)

Option A is correct because FeatureEngineeringClient.create_table takes the full three-level Unity Catalog name, a primary_keys list and the source DataFrame in df. Option B uses a keys argument that create_table does not have, option C writes to a table that does not exist yet, and option D calls a method name that belongs to the older workspace feature store client.

• A. run_c3
• B. run_d4
• C. run_b2 Correct answer
• D. run_a1

Option C is correct because the filter metrics.rmse < 5 keeps run_a1 (3.9), run_b2 (4.2) and run_c3 (4.8) and drops run_d4 (6.1). Ordering by r2 descending puts run_b2 (0.91) first, ahead of run_c3 (0.88) and run_a1 (0.80), so row 0 is run_b2. Run_d4 has the highest r2 overall but is excluded by the rmse filter.

• A. client.set_registered_model_alias("ml.shop.churn", "champion", 5) Correct answer
• B. client.transition_model_version_stage("ml.shop.churn", 5, "Production")
• C. client.delete_registered_model_alias("ml.shop.churn", "challenger")
• D. client.set_model_version_tag("ml.shop.churn", 5, "alias", "champion")

Option A is correct because set_registered_model_alias(name, alias, version) points the alias at the given version, and an alias can reference only one version, so champion moves from version 3 to version 5. Stage transitions are not supported for Unity Catalog models, deleting challenger does not set champion, and a tag named alias does not create a real alias.

• A. Changes are versioned and reviewed, and each run can be traced to a commit Correct answer
• B. Commits to the repository retrain registered models without a job
• C. Notebook cells compile into a JAR, which shortens cluster start time
• D. Git folders store the trained model files, so no model registry is needed

Option A is correct because Git folders link the notebooks and code to a remote repository, so every change is versioned and reviewed, and a run records the commit it used. A commit does not retrain or register models by itself, notebooks do not compile into a JAR, and Git does not replace the model registry.

• A. Imputer(strategy="mean")
• B. Imputer(strategy="median") Correct answer
• C. Imputer(strategy="mode")
• D. df.fillna(0, subset=["monthly_spend"])

Option B is correct because the median is not pulled up by a few very large values, so it stays close to a typical customer in a right-skewed column. The mean is dragged upward by the large values, the mode is a poor summary for a continuous spend value, and filling with 0 would understate spend for every customer with a missing value.

• A. df.filter(~F.col("price").between(q1 - 1.5 * iqr, q3 + 1.5 * iqr))
• B. df.filter(F.col("price").between(q1 + 1.5 * iqr, q3 - 1.5 * iqr))
• C. df.filter(F.col("price").between(q1 - 1.5 * iqr, q3 + 1.5 * iqr)) Correct answer
• D. df.filter(F.col("price").between(q1 - iqr, q3 + iqr))

Option C is correct because the 1.5 x IQR rule treats values below q1 - 1.5 x iqr or above q3 + 1.5 x iqr as outliers, so keeping rows between those fences removes them. Option A keeps only the outliers, option B swaps the bounds so nothing can match, and option D uses a 1.0 multiplier instead of 1.5.

• A. OneHotEncoder takes the raw string column as input, then indexes it internally
• B. customer_age needs one-hot encoding before a linear model can use it
• C. A random forest can split on StringIndexer output, so one-hot encoding is optional Correct answer
• D. StringIndexer maps a payment_method value unseen in training to an extra index by default
• E. payment_method suits one-hot encoding before a logistic regression is fitted Correct answer

Options C and E are correct because tree-based models like a random forest can split on StringIndexer output, so one-hot encoding is optional there, while a logistic regression would wrongly read the indices of unordered categories as an order, so payment_method suits one-hot encoding. OneHotEncoder takes the indexed numeric column, not the raw string. A numeric column such as customer_age needs no encoding, and StringIndexer raises an error on unseen values by default.

• A. The 25%, 50% and 75% percentile rows Correct answer
• B. The mode and variance rows
• C. The null count and distinct count rows
• D. The skewness and kurtosis rows

Option A is correct because df.summary() adds the approximate 25%, 50% and 75% percentile rows to the count, mean, stddev, min and max rows that df.describe() returns. Mode, variance, null count, distinct count, skewness and kurtosis are not part of either output.

• A. display(df.describe())
• B. df.limit(100).printSchema()
• C. df.summary().explain()
• D. dbutils.data.summarize(df) Correct answer

Option D is correct because dbutils.data.summarize(df) renders a per-column profile in a Databricks notebook, with histograms for numeric columns. display(df.describe()) only shows a table of five statistics, printSchema() lists column names and types, and explain() prints a query plan.

• A. fmin(fn=objective, space=space, algo=rand.suggest, max_evals=32, trials=Trials())
• B. fmin(fn=objective, space=space, algo=tpe.suggest, max_evals=32, trials=Trials()) Correct answer
• C. fmin(fn=objective, space=space, algo=tpe.suggest, max_evals=32)
• D. fmin(fn=objective, space=space, algo=tpe.suggest, n_iter=32, trials=Trials())

Option B is correct because algo=tpe.suggest selects the TPE algorithm, max_evals=32 sets the number of evaluations, and passing a Trials() object keeps the loss and other details of every trial. Option A uses random search, option C has no Trials object to inspect, and option D uses n_iter, which is not an fmin argument.

• A. 18
• B. 24
• C. 72
• D. 73 Correct answer

Option D is correct because the grid has 3 x 2 x 3 = 18 settings, and with 4 folds each setting is trained 4 times, giving 72 models. The final refit of the best setting on the full training data adds 1 more, so 73. The parallelism of 2 changes how fast the models train, not how many are trained.

• A. About 59,874 dollars
• B. About 1.22 dollars
• C. About 0.20 dollars
• D. About 13,256 dollars Correct answer

Option D is correct because the model works in log space, so dollar values come from exponentiating: e^11.2 is about 73,130 and e^11.0 is about 59,874, a difference of about 13,256 dollars. Option C is just the log-space error of 0.20, option B is a wrong conversion, and option A is the predicted price, not the error.

• A. Model B, with an F1 of 0.60 against about 0.33 for Model A Correct answer
• B. Model A, because F1 weights precision more than recall
• C. Model A, because its accuracy is higher on imbalanced data
• D. Model A, because its precision of 0.90 is the higher value

Option A is correct because F1 is the harmonic mean of precision and recall. Model B gives 0.60, while Model A gives 2 x 0.90 x 0.20 / (0.90 + 0.20), which is about 0.33. F1 weights precision and recall equally, accuracy is misleading with only 1 percent positives, and a high precision alone ignores that Model A misses most fraud.

• A. Estimating next month's revenue for each store Correct answer
• B. Grouping stores by their weekly sales pattern
• C. Forecasting the minutes a parcel needs to reach a customer Correct answer
• D. Deciding whether a card payment is fraudulent
• E. Assigning each support ticket to one of six queues

Options A and C are correct because they predict a continuous numeric value (revenue, minutes), which is regression. Option B groups stores without a label, so it is clustering. Options D and E assign a category (fraud or not, one of six queues), so they are classification.

• A. Predictions are returned to callers over a REST endpoint within milliseconds
• B. The model is reloaded and retrained each time a user opens a session
• C. Records are scored continuously as they arrive from a source such as Kafka
• D. A large set of records is scored on a schedule and the predictions are written to a table Correct answer

Option D is correct because batch inference scores a large set of records on a schedule, often as a job, and writes the predictions to a table. Millisecond responses over a REST endpoint describe real-time serving, continuous scoring of arriving Kafka records describes streaming inference, and retraining the model on every session is not inference at all.

• A. A second endpoint, with the version 3 endpoint workload size set to 90 percent
• B. Both versions as served entities, with traffic_percentage 90 and 10 in the traffic config Correct answer
• C. Two endpoints, with a 90 and 10 weight set on the champion and challenger aliases
• D. Provisioned concurrency of 90 and 10 on the two versions' served entities

Option B is correct because a single Model Serving endpoint can host several served entities, and the traffic configuration assigns each a traffic_percentage, here 90 for version 3 and 10 for version 4. Workload size sets compute capacity, not traffic share, aliases carry no endpoint weights, and provisioned concurrency controls scaling, not request routing.

• A. A notebook loads the model once and predicts for a pandas DataFrame held in driver memory
• B. An application sends one row over HTTPS and waits for the prediction in the response
• C. A running query scores records as they arrive in a stream and appends predictions to a table Correct answer
• D. A scheduled job reads a stored table in full and writes one prediction per row to a results table

Option C is correct because streaming inference runs as a continuous query that scores records as they arrive and appends the predictions to a table. Option A describes single-node pandas inference, option B describes a real-time endpoint request, and option D describes batch inference over a stored table.

• A. A Structured Streaming source that reads the endpoint's output table
• B. A scheduled job that writes predictions to a file in cloud storage
• C. A REST API request sent to the endpoint over HTTPS Correct answer
• D. A JDBC query run against a SQL warehouse that holds the scores

Option C is correct because a Model Serving endpoint exposes a REST API, and the application sends an HTTPS request and receives the prediction in the response. The other options score data through tables, files or SQL, which is batch or streaming scoring and not how an endpoint is queried.

• A. Flagging a sensor reading within seconds of it reaching the ingestion topic
• B. Ranking search suggestions after each keystroke in a mobile app
• C. Approving or declining a card payment while the cashier waits
• D. Scoring every subscriber overnight to refresh a retention-risk table Correct answer
• E. Producing next week's call list for the sales team from a stored table Correct answer

Options D and E are correct because both score stored data on a schedule with no one waiting for the answer, which is what batch inference is for. Options A, B and C each need a score within seconds or while a user waits, so they need streaming or a real-time endpoint.