Should we parallelize a DataFrame like we parallelize a Seq before training

Question

Consider the code given here,

https://spark.apache.org/docs/1.2.0/ml-guide.html

import org.apache.spark.ml.classification.LogisticRegression
val training         


        

Answer 1

DataFrame is a distributed data structure. It is neither required nor possible to parallelize it. SparkConext.parallelize method is used only to distributed local data structures which reside in the driver memory. You shouldn't be used to distributed large datasets not to mention redistributing RDDs or higher level data structures (like you do in your previous question)

sc.parallelize(trainingData.collect()) 

If you want to convert between RDD / Dataframe (Dataset) use methods which are designed to do it:

1. from DataFrame to RDD:

   import org.apache.spark.sql.DataFrame
   import org.apache.spark.sql.Row
   import org.apache.spark.rdd.RDD
   
   val df: DataFrame  = Seq(("foo", 1), ("bar", 2)).toDF("k", "v")
   val rdd: RDD[Row] = df.rdd
   

2. form RDD to DataFrame:

   val rdd: RDD[(String, Int)] = sc.parallelize(Seq(("foo", 1), ("bar", 2)))
   val df1: DataFrame = rdd.toDF
   // or
   val df2: DataFrame = spark.createDataFrame(rdd) // From 1.x use sqlContext
   

Answer 2

You should maybe check out the difference between RDD and DataFrame and how to convert between the two: Difference between DataFrame and RDD in Spark

To answer your question directly: A DataFrame is already optimized for parallel execution. You do not need to do anything and you can pass it to any spark estimators fit() method directly. The parallel executions are handled in the background.