A. Correct.
Delta Lake stores table actions and statistics in the transaction log, and Databricks can use that metadata for supported aggregate optimizations.

B. Incorrect.
Hive metastore row sampling is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Delta Lake stores table actions and statistics in the transaction log, and Databricks can use that metadata for supported aggregate optimizations.

C. Incorrect.
The driver cache always stores all previous COUNT results is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Delta Lake stores table actions and statistics in the transaction log, and Databricks can use that metadata for supported aggregate optimizations.

D. Incorrect.
Every COUNT query must scan all data files is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Delta Lake stores table actions and statistics in the transaction log, and Databricks can use that metadata for supported aggregate optimizations.

A. Incorrect.
Increase the number of shuffle partitions only is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Watermarks tell Structured Streaming how long to retain state for late events before finalizing windows.

B. Incorrect.
Disable checkpointing for the stream is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Watermarks tell Structured Streaming how long to retain state for late events before finalizing windows.

C. Incorrect.
Overwrite the target table on every trigger is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Watermarks tell Structured Streaming how long to retain state for late events before finalizing windows.

D. Correct.
Watermarks tell Structured Streaming how long to retain state for late events before finalizing windows.

A. Incorrect.
DROP and recreate the current table from full history every hour is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: MERGE performs efficient conditional updates and inserts without rebuilding the entire table.

B. Incorrect.
VACUUM the history table before each load is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: MERGE performs efficient conditional updates and inserts without rebuilding the entire table.

C. Correct.
MERGE performs efficient conditional updates and inserts without rebuilding the entire table.

D. Incorrect.
Append all hourly records directly to the current table is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: MERGE performs efficient conditional updates and inserts without rebuilding the entire table.

A. Incorrect.
databricks run daily_job --env prod is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Bundle commands understand target-specific settings from databricks.yml.

B. Correct.
Bundle commands understand target-specific settings from databricks.yml.

C. Incorrect.
databricks execute daily_job -e prod is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Bundle commands understand target-specific settings from databricks.yml.

D. Incorrect.
databricks jobs run-now daily_job --target prod is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Bundle commands understand target-specific settings from databricks.yml.

A. Correct.
Delta schema evolution can add compatible fields while existing files are interpreted with NULLs for missing columns.

B. Incorrect.
The update is always rejected because nested fields cannot evolve is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Delta schema evolution can add compatible fields while existing files are interpreted with NULLs for missing columns.

C. Incorrect.
The new field is written to a separate rescue column is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Delta schema evolution can add compatible fields while existing files are interpreted with NULLs for missing columns.

D. Incorrect.
All old Parquet files must be rewritten immediately is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Delta schema evolution can add compatible fields while existing files are interpreted with NULLs for missing columns.

A. Incorrect.
Only inside one notebook with spark.conf.set is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Cluster Spark configuration applies to workloads attached to that cluster.

B. Incorrect.
In a SQL warehouse parameter is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Cluster Spark configuration applies to workloads attached to that cluster.

C. Incorrect.
In a job task parameter called spark_conf is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Cluster Spark configuration applies to workloads attached to that cluster.

D. Correct.
Cluster Spark configuration applies to workloads attached to that cluster.

A. Incorrect.
Partition by ingestion cluster ID is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Topic partitioning aligns the physical layout with the policy boundary.

B. Incorrect.
Do not partition because value is binary is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Topic partitioning aligns the physical layout with the policy boundary.

C. Correct.
Topic partitioning aligns the physical layout with the policy boundary.

D. Incorrect.
Partition by Kafka offset is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Topic partitioning aligns the physical layout with the policy boundary.

A. Incorrect.
A SQL dashboard refresh schedule is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Asset Bundles provide declarative project deployment across environments.

B. Correct.
Asset Bundles provide declarative project deployment across environments.

C. Incorrect.
Notebook-scoped widgets is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Asset Bundles provide declarative project deployment across environments.

D. Incorrect.
Cluster libraries only is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Asset Bundles provide declarative project deployment across environments.

A. Correct.
Structured Streaming stores offsets and state in checkpoints for fault-tolerant restart.

B. Incorrect.
A display() call in the notebook is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Structured Streaming stores offsets and state in checkpoints for fault-tolerant restart.

C. Incorrect.
A larger driver node is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Structured Streaming stores offsets and state in checkpoints for fault-tolerant restart.

D. Incorrect.
An OPTIMIZE command after every micro-batch is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Structured Streaming stores offsets and state in checkpoints for fault-tolerant restart.

A. Incorrect.
Sorting descending returns the largest region_id values first, while the requirement asks for the smallest three.

B. Incorrect.
RESTORE the table is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: OPTIMIZE bin-packs small files into larger files to reduce file listing and scan overhead.

C. Incorrect.
GRANT SELECT on the table is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: OPTIMIZE bin-packs small files into larger files to reduce file listing and scan overhead.

D. Correct.
OPTIMIZE bin-packs small files into larger files to reduce file listing and scan overhead.