A. Correct.
If historical fraud labels exist, supervised classification can learn patterns associated with fraudulent transactions.

B. Correct.
Unsupervised anomaly detection can flag unusual behavior when labeled fraud examples are limited or unavailable.

C. Correct.
Clustering groups similar transactions without needing labels and can reveal behavior segments.

D. Incorrect.
Reinforcement learning is for sequential decision-making with rewards, not direct transaction-location prediction.

E. Incorrect.
Predicting a known location label is a supervised learning problem, not unsupervised learning.

A. Correct.
Separate datasets provide a clean isolation boundary for access control.

B. Incorrect.
Partitions do not provide the same client isolation boundary as separate datasets with IAM controls.

C. Correct.
Dataset-level access should be limited to approved users for that client.

D. Incorrect.
A shared service account can become an overbroad access path and does not directly grant the right client users access.

E. Correct.
BigQuery IAM roles should be assigned according to least privilege for each client's users.

A. Incorrect.
Viewer can expose sensitive project data and resources unnecessarily.

B. Incorrect.
Dataflow Developer in the sensitive project still risks access to private data.

C. Incorrect.
Shared service account usage weakens accountability and can expose sensitive access.

D. Correct.
An anonymized sample in a separate project enables development while protecting private user data.

A. Incorrect.
This returns only one row for the entire query, not one row per unique event.

B. Incorrect.
Grouping by timestamp can preserve duplicates and can incorrectly aggregate facts.

C. Incorrect.
LAG can compare rows, but ROW_NUMBER is simpler for choosing one canonical record per ID.

D. Correct.
ROW_NUMBER over each unique ID lets you keep the latest or preferred record and filter duplicates.

A. Incorrect.
A pure timestamp key can create hotspotting because all new writes land near the same key range.

B. Incorrect.
A low-cardinality or skewed sensor ID can create hotspots for active sensors.

C. Correct.
Bigtable performance depends heavily on row-key design that distributes load across tablets.

D. Incorrect.
A single frequently updated row creates a severe hotspot.

A. Incorrect.
Applications should not try to restart database infrastructure as normal failure handling.

B. Correct.
Exponential backoff reduces load during failure and is a standard resilient client pattern.

C. Incorrect.
Retrying every second can amplify outages and overload the database.

D. Incorrect.
Reducing to hourly is a blunt operational change and does not provide responsive recovery behavior.

A. Incorrect.
The laptop is already the bottleneck, so local notebooks do not solve the compute constraint.

B. Incorrect.
Cloud Shell is useful for administration, but not ideal for heavy analysis and visualization workflows.

C. Incorrect.
Visualization alone does not support data preparation and ML dataset creation.

D. Correct.
A cloud-hosted notebook environment places compute near the data and supports analysis, labeling, and visualization.

A. Correct.
Anomaly detection assumes most examples are normal and unusual examples are rare.

B. Incorrect.
Balanced classes are better suited to supervised classification when labels exist.

C. Correct.
Anomaly detection can identify novel unusual patterns rather than only known mutation classes.

D. Incorrect.
If complete labels exist, supervised learning is often more direct.

E. Incorrect.
That describes supervised classification rather than unsupervised anomaly detection.

A. Correct.
Cloud Storage decouples storage from Dataproc compute and lowers persistent block storage needs while supporting Hadoop connectors.

B. Incorrect.
Preemptible VMs reduce compute cost, not the large persistent storage requirement.

C. Incorrect.
Right-sizing disks still keeps the expensive HDFS-like storage model.

D. Incorrect.
Cloud SQL is not a cold Hadoop data store.

A. Correct.
Role-based access is a core least-privilege control.

B. Incorrect.
Disabling writes does not prevent users from reading data they should not see.

C. Correct.
Only approved identities should be able to use BigQuery APIs against protected data.

D. Incorrect.
Encryption protects storage but does not define who can query which data.

E. Correct.
Segregating data by sensitivity makes access policies simpler and safer.