A. Incorrect.
The dashboard might have issues, but the next best step is to isolate the pipeline with controlled input.

B. Correct.
A fixed dataset helps determine whether parsing, transformation, windowing, or output logic drops records.

C. Incorrect.
Pub/Sub already shows successful publishing; the problem may be downstream.

D. Incorrect.
Dataflow reads from Pub/Sub subscriptions; changing push/pull semantics is not the right diagnostic action.

A. Incorrect.
SQL can transform data, but it is not the requested graphical analyst-friendly recipe tool.

B. Incorrect.
Python pipelines require developer skills and code management.

C. Correct.
Dataprep provides visual data preparation recipes that analysts can modify and schedule.

D. Incorrect.
Spark on Dataproc is powerful but more code-oriented and operationally heavier.

A. Incorrect.
Features correlated with the label are often useful and should not be removed solely for that reason.

B. Correct.
Combining redundant or highly co-dependent features can reduce dimensionality while preserving signal.

C. Incorrect.
Arbitrary averaging can destroy useful feature meaning.

D. Incorrect.
Some missingness can be handled with imputation or indicators; removing all such features may harm accuracy.

A. Correct.
Audit logs show who accessed what and help discover actual usage before redesigning access.

B. Incorrect.
IAM policies show permissions, not actual behavior.

C. Incorrect.
Slot metrics show utilization, not user data access patterns.

D. Incorrect.
Billing does not reveal how datasets are being used.

A. Incorrect.
Draining works for compatible updates and graceful completion, but the scenario says the new code is incompatible.

B. Incorrect.
Transform mapping helps update compatible pipeline graph changes, not incompatible redesigns.

C. Incorrect.
Sharing the same subscription risks messages being consumed by the wrong pipeline during migration.

D. Correct.
A new subscription lets the new incompatible pipeline process data independently without losing unacknowledged messages from the old pipeline.

A. Incorrect.
SLA misses are schedule-based and do not necessarily mean task failure.

B. Correct.
on_failure_callback directly runs custom notification logic when the task fails.

C. Incorrect.
sla_miss_callback relates to missed SLAs, not all unsuccessful task outcomes.

D. Incorrect.
on_retry_callback fires before retries and may notify even when the task later succeeds.

A. Incorrect.
Federated queries do not provide a robust cleansing and dead-letter pattern.

B. Incorrect.
Monitoring alerts do not process or preserve corrupt rows.

C. Incorrect.
Failing on the first bad row prevents loading good data and does not preserve bad rows for analysis.

D. Correct.
Dataflow can validate records, route valid output to BigQuery, and send malformed records to a dead-letter table.

A. Incorrect.
Threading affects execution, not generalization.

B. Incorrect.
Serialization stores models or data; it does not reduce overfitting.

C. Correct.
Dropout is a regularization method that helps neural networks generalize better.

D. Incorrect.
More complexity can worsen overfitting.

A. Incorrect.
Datastore is not the best ingestion and analytics path for real-time high-volume IoT streams.

B. Correct.
Pub/Sub ingests events, Dataflow processes streams, and BigQuery supports large-scale real-time analytics.

C. Incorrect.
Batch Dataproc does not meet the real-time processing requirement.

D. Incorrect.
Cloud SQL is not designed for this scale of IoT analytics.

A. Incorrect.
Fixed windows are based on wall-clock boundaries, not individual user inactivity.

B. Incorrect.
Sliding windows repeatedly overlap and do not naturally represent user sessions ending after inactivity.

C. Correct.
A session window with a 60-minute gap models per-user inactivity and can trigger abandonment logic.

D. Incorrect.
A global window lacks the natural per-session inactivity boundary.