DataFrames, like RDDs, are immutable. When you define a transformation on a DataFrame, this always creates a new DataFrame. The original DataFrame cannot be modified in place (this is notably different to pandas DataFrames, for instance).

Operations on DataFrames can be grouped into two: transformations, which result in the creation of a new DataFrame, and actions, which usually return a Scala type or have a side-effect. Methods like filter or withColumn are transformations, while methods like show or head are actions.

Transformations are lazy, much like transformations on RDDs. When you generate a new DataFrame by transforming an existing DataFrame, this results in the elaboration of an execution plan for creating the new DataFrame, but the data itself is not transformed immediately. You can access the execution plan with the queryExecution method.

When you call an action on a DataFrame, Spark processes the action as if it were a regular RDD: it implicitly builds a direct acyclic graph to resolve dependencies, processing the transformations needed to build the DataFrame on which the action was called.

Much like RDDs, we can persist DataFrames in memory or on disk:

scala> readingsDF.persist
readingsDF.type = [patientId: int, heightCm: int,...]

This works in the same way as persisting RDDs: next time the RDD is calculated, it will be kept in memory (provided there is enough space), rather than discarded. The level of persistence can also be set:

scala> import org.apache.spark.storage.StorageLevel
import org.apache.spark.storage.StorageLevel

scala> readingsDF.persist(StorageLevel.MEMORY_AND_DISK)
readingsDF.type = [patientId: int, heightCm: int, ...]