Spark dynamic DAG is much slower and different from hard-coded DAG

Question

I have an operation in a spark that needs to be performed for multiple columns in a data frame. There are usually two options for specifying such operations.

• hard

handleBias("bar", df)
  .join(handleBias("baz", df), df.columns)
  .drop(columnsToDrop: _*).show

• dynamically generate them from a list of code names

var isFirst = true
var res = df
for (col <- columnsToDrop ++ columnsToCode) {
  if (isFirst) {
    res = handleBias(col, res)
    isFirst = false
  } else {
    res = handleBias(col, res)
  }
}
res.drop(columnsToDrop: _*).show

The problem is that the dynamically generated dynamic DAG system is different and the execution time of the dynamic solution increases much more when more columns are used than for hard-coded operations.

I am curious how to combine the elegance of a dynamic design with fast lead times .

Here is a comparison for the DAG code example

80 , , DAG .

(2.0.2) DataFrames spark-sql

:

def handleBias(col: String, df: DataFrame, target: String = "FOO"): DataFrame = {
  val pre1_1 = df
    .filter(df(target) === 1)
    .groupBy(col, target)
    .agg((count("*") / df.filter(df(target) === 1).count).alias("pre_" + col))
    .drop(target)

  val pre2_1 = df
    .groupBy(col)
    .agg(mean(target).alias("pre2_" + col))

  df
    .join(pre1_1, Seq(col), "left")
    .join(pre2_1, Seq(col), "left")
    .na.fill(0)
}

foldleft DAG

, DAG, , , . , SQL , , . ?

Answer 1

1: handleBias .

2: .

, . -, "handleBias" , "withColumn", :

import org.apache.spark.sql.DataFrame
import org.apache.spark.sql.functions._
import org.apache.spark.sql.expressions.Window

def handleBias(df: DataFrame, colName: String, target: String = "foo") = {
  val w1 = Window.partitionBy(colName)
  val w2 = Window.partitionBy(colName, target)
  val result = 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")
  result
}

foldLeft, "" :

val df = Seq((1, "first", "A"), (1, "second", "A"),(2, "noValidFormat", "B"),(1, "lastAssumingSameDate", "C")).toDF("foo", "bar", "baz")

val columnsToDrop = Seq("baz")
val columnsToCode = Seq("bar", "baz")
val target = "foo"

val targetCounts = df.filter(df(target) === 1).groupBy(target)
  .agg(count(target).as("cnt_foo_eq_1"))
val newDF = df.join(broadcast(targetCounts), Seq(target), "left")

val result = (columnsToDrop ++ columnsToCode).toSet.foldLeft(df) {
  (currentDF, colName) => handleBias(currentDF, colName)
}

result.drop(columnsToDrop:_*).show()

+---+--------------------+------------------+--------+------------------+--------+
|foo|                 bar|           pre_baz|pre2_baz|           pre_bar|pre2_bar|
+---+--------------------+------------------+--------+------------------+--------+
|  2|       noValidFormat|               0.0|     2.0|               0.0|     2.0|
|  1|lastAssumingSameDate|0.3333333333333333|     1.0|0.3333333333333333|     1.0|
|  1|              second|0.6666666666666666|     1.0|0.3333333333333333|     1.0|
|  1|               first|0.6666666666666666|     1.0|0.3333333333333333|     1.0|
+---+--------------------+------------------+--------+------------------+--------+

, DAG, , , .

:

• Databricks : https://databricks.com/blog/2015/07/15/introducing-window-functions-in-spark-sql.html
• API- : http://spark.apache.org/docs/latest/api/scala/index.html#org.apache.spark.sql.functions $
• foldLeft: https://coderwall.com/p/4l73-a/scala-fold-foldleft-and-foldright