A. Correct.
groupBy followed by aggregation is a wide transformation that moves records with the same key across partitions, creating a shuffle boundary and a new stage.

B. Incorrect.
filter() is a narrow transformation. It removes rows within each partition and does not require records to move across partitions, so it does not create a shuffle stage.

C. Incorrect.
withColumn() computes or replaces a column row by row. Unless the expression itself performs a wide operation, it does not redistribute data across the cluster.

D. Incorrect.
select() only projects columns or expressions from existing rows. It is a narrow transformation and does not group records by key or shuffle data.

A. Incorrect.
na.drop("all") removes only rows where all checked columns are null. A row with just one missing field would remain, which violates the requirement.

B. Incorrect.
dropna(how="all") is the opposite threshold from what is needed here: it keeps rows that have at least one non-null value, instead of dropping rows with any null.

C. Incorrect.
raw_users.na.drop("any") is equivalent in Spark, but this option is less explicit than the requested dropna(how="any") form used by the answer set. The key behavior is dropping rows with any null value.

D. Correct.
dropna(how="any") removes rows that have at least one NULL value in any column, which matches the requirement to remove records with any missing field.

A. Incorrect.
broadcast() does not increase partition sizes. It replicates the small side of the join so Spark can avoid shuffling both sides.

B. Incorrect.
The join condition already controls matching id values. broadcast() affects performance, not the logical equality condition.

C. Correct.
broadcast(df1) sends the small DataFrame to each worker so Spark can join locally with the larger DataFrame partitions and avoid a large shuffle.

D. Incorrect.
broadcast() does not filter join keys. It changes the physical join strategy by sending the smaller DataFrame to workers.

A. Incorrect.
partitionBy() controls output layout. Appending requires an explicit write mode such as mode("append").

B. Correct.
partitionBy writes data into directory paths based on the specified partition columns, such as color=red/fruit=apple, enabling partition pruning.

C. Incorrect.
partitionBy() does not coalesce output into one file. Spark still writes part files, organized under partition directories.

D. Incorrect.
Spark can write null partition values. Nulls do not cause partitionBy() to fail by themselves.

A. Correct.
Actions such as collect() trigger Spark jobs. The scheduler splits jobs into stages at shuffle boundaries and runs tasks per partition.

B. Incorrect.
A Spark application is the submitted program. The application is not split into multiple applications; actions create jobs inside the application.

C. Incorrect.
Spark transformations are lazy. A job is triggered by an action such as collect(), not by the whole script before an action runs.

D. Incorrect.
Tasks are created per stage and partition. Spark does not create exactly one task for each transformation or action.

A. Incorrect.
A large size difference is exactly the kind of case where broadcasting the small side can help, provided the small table is under the broadcast threshold.

B. Incorrect.
Adaptive Query Execution can improve join planning, but Spark can choose a broadcast join statically when one side is below the broadcast threshold.

C. Incorrect.
A broadcast hint is optional. Spark can automatically broadcast a small relation when it is below spark.sql.autoBroadcastJoinThreshold.

D. Correct.
Spark can automatically choose a broadcast hash join when one side is below spark.sql.autoBroadcastJoinThreshold; an 8 MB table is below the common 10 MB threshold.

A. Incorrect.
Spark transformations are lazy. They build a logical plan, but execution waits until an action is called.

B. Incorrect.
Spark does not require manual intervention to run transformations. The pipeline starts automatically when an action triggers a job.

C. Correct.
Spark transformations are lazy and build a logical lineage plan. Execution begins only when an action such as count(), collect(), save(), or show() is called.

D. Correct.
Spark transformations are lazy and build a logical lineage plan. Execution begins only when an action such as count(), collect(), save(), or show() is called.

A. Incorrect.
A positive days value moves the date backward. A negative days value subtracts a negative number, which moves the date forward.

B. Correct.
date_sub subtracts the days value from the start date. Subtracting a negative number effectively adds days to the date.

C. Incorrect.
Spark supports negative values for the days argument. Passing a negative value does not produce an invalid-parameter error.

D. Incorrect.
The date only stays the same when the days value is 0. A negative value changes the result by adding days.

A. Correct.
Spark's distributed, in-memory processing model is the core capability that lets it process high-volume data across a cluster quickly.

B. Incorrect.
Spark can process small data, but its main advantage in this scenario is scalable distributed processing for large, high-velocity data.

C. Incorrect.
SQL support is useful, but it does not by itself solve the volume and velocity problem described in the scenario.

D. Incorrect.
MLlib can help with recommendations or modeling, but the core feature that handles 2 TB daily data is distributed in-memory processing.

A. Incorrect.
Adding executors increases task capacity, but it does not move or strengthen the driver when the bottleneck is the client-side driver.

B. Incorrect.
Increasing client-machine driver memory may be a temporary patch, but cluster mode is the stronger solution when the client driver is the limiting factor.

C. Incorrect.
Local mode runs the driver and executors on one machine, which makes the resource bottleneck worse for large workloads.

D. Correct.
In client mode the driver runs on the submitting machine. Cluster mode moves the driver into the cluster, giving it cluster-managed resources and avoiding a weak client machine bottleneck.