[
https://issues.apache.org/jira/browse/SPARK-28090?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Dongjoon Hyun updated SPARK-28090:
----------------------------------
Labels: (was: bulk-closed)
> Spark hangs when an execution plan has many projections on nested structs
> -------------------------------------------------------------------------
>
> Key: SPARK-28090
> URL: https://issues.apache.org/jira/browse/SPARK-28090
> Project: Spark
> Issue Type: Bug
> Components: Optimizer, SQL
> Affects Versions: 2.4.3
> Environment: Tried in
> * Spark 2.2.1, Spark 2.4.3 in local mode on Linux, MasOS and Windows
> * Spark 2.4.3 / Yarn on a Linux cluster
> Reporter: Ruslan Yushchenko
> Priority: Major
>
> This was already posted (#28016), but the provided example didn't always
> reproduce the error. This example consistently reproduces the issue.
> Spark applications freeze on execution plan optimization stage (Catalyst)
> when a logical execution plan contains a lot of projections that operate on
> nested struct fields.
> The code listed below demonstrates the issue.
> To reproduce the Spark App does the following:
> * A small dataframe is created from a JSON example.
> * Several nested transformations (negation of a number) are applied on
> struct fields and each time a new struct field is created.
> * Once more than 9 such transformations are applied the Catalyst optimizer
> freezes on optimizing the execution plan.
> * You can control the freezing by choosing different upper bound for the
> Range. E.g. it will work file if the upper bound is 5, but will hang is the
> bound is 10.
> {code:java}
> package com.example
> import org.apache.spark.sql._
> import org.apache.spark.sql.functions._
> import org.apache.spark.sql.types.{StructField, StructType}
> import scala.collection.mutable.ListBuffer
> object SparkApp1IssueSelfContained {
> // A sample data for a dataframe with nested structs
> val sample: List[String] =
> """ { "numerics": {"num1": 101, "num2": 102, "num3": 103, "num4": 104,
> "num5": 105, "num6": 106, "num7": 107, "num8": 108, "num9": 109, "num10":
> 110, "num11": 111, "num12": 112, "num13": 113, "num14": 114, "num15": 115} }
> """ ::
> """ { "numerics": {"num1": 201, "num2": 202, "num3": 203, "num4": 204,
> "num5": 205, "num6": 206, "num7": 207, "num8": 208, "num9": 209, "num10":
> 210, "num11": 211, "num12": 212, "num13": 213, "num14": 214, "num15": 215} }
> """ ::
> """ { "numerics": {"num1": 301, "num2": 302, "num3": 303, "num4": 304,
> "num5": 305, "num6": 306, "num7": 307, "num8": 308, "num9": 309, "num10":
> 310, "num11": 311, "num12": 312, "num13": 313, "num14": 314, "num15": 315} }
> """ ::
> Nil
> /**
> * Transforms a column inside a nested struct. The transformed value will
> be put into a new field of that nested struct
> *
> * The output column name can omit the full path as the field will be
> created at the same level of nesting as the input column.
> *
> * @param inputColumnName A column name for which to apply the
> transformation, e.g. `company.employee.firstName`.
> * @param outputColumnName The output column name. The path is optional,
> e.g. you can use `transformedName` instead of
> `company.employee.transformedName`.
> * @param expression A function that applies a transformation to a
> column as a Spark expression.
> * @return A dataframe with a new field that contains transformed values.
> */
> def transformInsideNestedStruct(df: DataFrame,
> inputColumnName: String,
> outputColumnName: String,
> expression: Column => Column): DataFrame = {
> def mapStruct(schema: StructType, path: Seq[String], parentColumn:
> Option[Column] = None): Seq[Column] = {
> val mappedFields = new ListBuffer[Column]()
> def handleMatchedLeaf(field: StructField, curColumn: Column):
> Seq[Column] = {
> val newColumn = expression(curColumn).as(outputColumnName)
> mappedFields += newColumn
> Seq(curColumn)
> }
> def handleMatchedNonLeaf(field: StructField, curColumn: Column):
> Seq[Column] = {
> // Non-leaf columns need to be further processed recursively
> field.dataType match {
> case dt: StructType => Seq(struct(mapStruct(dt, path.tail,
> Some(curColumn)): _*).as(field.name))
> case _ => throw new IllegalArgumentException(s"Field
> '${field.name}' is not a struct type.")
> }
> }
> val fieldName = path.head
> val isLeaf = path.lengthCompare(2) < 0
> val newColumns = schema.fields.flatMap(field => {
> // This is the original column (struct field) we want to process
> val curColumn = parentColumn match {
> case None => new Column(field.name)
> case Some(col) => col.getField(field.name).as(field.name)
> }
> if (field.name.compareToIgnoreCase(fieldName) != 0) {
> // Copy unrelated fields as they were
> Seq(curColumn)
> } else {
> // We have found a match
> if (isLeaf) {
> handleMatchedLeaf(field, curColumn)
> } else {
> handleMatchedNonLeaf(field, curColumn)
> }
> }
> })
> newColumns ++ mappedFields
> }
> val schema = df.schema
> val path = inputColumnName.split('.')
> df.select(mapStruct(schema, path): _*)
> }
> /**
> * This Spark Job demonstrates an issue of execution plan freezing when
> there are a lot of projections
> * involving nested structs in an execution plan.
> *
> * The example works as follows:
> * - A small dataframe is created from a JSON example above
> * - A nested withColumn map transformation is used to apply a
> transformation on a struct field and create
> * a new struct field.
> * - Once more than 9 such transformations are applied the Catalyst
> optimizer freezes on optimizing
> * the execution plan
> *
> */
> def main(args: Array[String]): Unit = {
> val sparkBuilder = SparkSession.builder().appName("Nested Projections
> Issue Example")
> val spark = sparkBuilder
> .master("local[4]")
> .getOrCreate()
> import spark.implicits._
> val dfInput = spark.read.json(sample.toDS)
> // Apply several negations. You can change the number of negations by
> changing the upper bound of the range up to 16
> val dfOutput = Range(1,12).foldLeft(dfInput)( (df, i) => {
> transformInsideNestedStruct(df, s"numerics.num$i", s"out_num$i", c =>
> -c)
> })
> dfOutput.printSchema()
> dfOutput.explain(true)
> dfOutput.show(false)
> }
> }
> {code}
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
