spark convert spark-SQL to RDD API

Question

Spark SQL is pretty clear to me. However, I am just getting started with spark's RDD API. As spark apply function to columns in parallel points out this should allow me to get rid of slow shuffles for

def handleBias(df: DataFrame, colName: String, target: String = this.target) = {
    val w1 = Window.partitionBy(colName)
    val w2 = Window.partitionBy(colName, target)

    df.withColumn("cnt_group", count("*").over(w2))
      .withColumn("pre2_" + colName, mean(target).over(w1))
      .withColumn("pre_" + colName, coalesce(min(col("cnt_group") / col("cnt_foo_eq_1")).over(w1), lit(0D)))
      .drop("cnt_group")
  }
}

In pseudo code: df foreach column (handleBias(column) So a minimal data frame is loaded up

val input = Seq(
    (0, "A", "B", "C", "D"),
    (1, "A", "B", "C", "D"),
    (0, "d", "a", "jkl", "d"),
    (0, "d", "g", "C", "D"),
    (1, "A", "d", "t", "k"),
    (1, "d", "c", "C", "D"),
    (1, "c", "B", "C", "D")
  )
  val inputDf = input.toDF("TARGET", "col1", "col2", "col3TooMany", "col4")

but fails to map correctly

val rdd1_inputDf = inputDf.rdd.flatMap { x => {(0 until x.size).map(idx => (idx, x(idx)))}}
      rdd1_inputDf.toDF.show

It fails with

java.lang.ClassNotFoundException: scala.Any
java.lang.ClassNotFoundException: scala.Any

An example can be found https://github.com/geoHeil/sparkContrastCoding respectively https://github.com/geoHeil/sparkContrastCoding/blob/master/src/main/scala/ColumnParallel.scala for the problem outlined in this question.

evan.oman · Accepted Answer · 2017-01-03T16:42:12.010

When you call .rdd on a DataFrame you get an RDD[Row] which is not strongly typed. If you want to be able to map over the elements you will need to pattern match over Row:

scala> val input = Seq(
     |     (0, "A", "B", "C", "D"),
     |     (1, "A", "B", "C", "D"),
     |     (0, "d", "a", "jkl", "d"),
     |     (0, "d", "g", "C", "D"),
     |     (1, "A", "d", "t", "k"),
     |     (1, "d", "c", "C", "D"),
     |     (1, "c", "B", "C", "D")
     |   )
input: Seq[(Int, String, String, String, String)] = List((0,A,B,C,D), (1,A,B,C,D), (0,d,a,jkl,d), (0,d,g,C,D), (1,A,d,t,k), (1,d,c,C,D), (1,c,B,C,D))

scala> val inputDf = input.toDF("TARGET", "col1", "col2", "col3TooMany", "col4") 
inputDf: org.apache.spark.sql.DataFrame = [TARGET: int, col1: string ... 3 more fields]

scala> import org.apache.spark.sql.Row
import org.apache.spark.sql.Row

scala> val rowRDD = inputDf.rdd
rowRDD: org.apache.spark.rdd.RDD[org.apache.spark.sql.Row] = MapPartitionsRDD[3] at rdd at <console>:27

scala> val typedRDD = rowRDD.map{case Row(a: Int, b: String, c: String, d: String, e: String) => (a,b,c,d,e)}
typedRDD: org.apache.spark.rdd.RDD[(Int, String, String, String, String)] = MapPartitionsRDD[20] at map at <console>:29

scala> typedRDD.keyBy(_._1).groupByKey.foreach{println}
[Stage 7:>                                                          (0 + 0) / 4]
(0,CompactBuffer((A,B,C,D), (d,a,jkl,d), (d,g,C,D)))
(1,CompactBuffer((A,B,C,D), (A,d,t,k), (d,c,C,D), (c,B,C,D)))

Otherwise you can use a typed Dataset:

scala> val ds = input.toDS
ds: org.apache.spark.sql.Dataset[(Int, String, String, String, String)] = [_1: int, _2: string ... 3 more fields]

scala> ds.rdd
res2: org.apache.spark.rdd.RDD[(Int, String, String, String, String)] = MapPartitionsRDD[8] at rdd at <console>:30

scala> ds.rdd.keyBy(_._1).groupByKey.foreach{println}
[Stage 0:>                                                          (0 + 0) / 4]
(0,CompactBuffer((0,A,B,C,D), (0,d,a,jkl,d), (0,d,g,C,D)))
(1,CompactBuffer((1,A,B,C,D), (1,A,d,t,k), (1,d,c,C,D), (1,c,B,C,D)))

As I want to use this in a ml.Pipeline and the output step is DataFrame the the "schema is lost" e.g. I will need to use pattern matching? is this correct? But there are quite a lot of columns is there a way to "infer" them somewhat (partial shcema? — Georg Heiler, Jan 03 '17 at 06:19
Yeah the `DF => RDD` conversion doesn't use the schema at all unfortunately (and I don't think there is a good way to force its use). However take a look at my new `Dataset` example: there is no need to use an intermediary `Dataframe` and it looks like the `DataSet` infers the types nicely (in Spark 2.0 I think anything you could do with a DF can also be done with a DS) — evan.oman, Jan 03 '17 at 16:44
thanks. Indeed you are right. However a spark transformer in a ml pipeline will output only dataframes ;) even with dataset as the input. So I think the schema will be lost for subsequent transformer steps — Georg Heiler, Jan 03 '17 at 17:04
I posted a follow-up question here http://stackoverflow.com/questions/41445571/spark-migrate-sql-window-function-to-rdd-for-better-performance maybe you have a suggestion as well. — Georg Heiler, Jan 04 '17 at 08:19

spark convert spark-SQL to RDD API

1 Answers1