ODIn/spark-instrumented-optimizer
1
0
Fork 0
Code Issues Pull requests Projects Releases Wiki Activity

[SPARK-12898] Consider having dummyCallSite for HiveTableScan

Browse source 
Currently, HiveTableScan runs with getCallSite which is really expensive and shows up when scanning through large table with partitions (e.g TPC-DS) which slows down the overall runtime of the job. It would be good to consider having dummyCallSite in HiveTableScan.

Author: Rajesh Balamohan <rbalamohan@apache.org>

Closes #10825 from rajeshbalamohan/SPARK-12898.
This commit is contained in:
Rajesh Balamohan 2016-01-20 11:30:03 -08:00 committed by Reynold Xin
parent e75e340a40
commit ab4a6bfd11
1 changed files with 10 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

13
sql/hive/src/main/scala/org/apache/spark/sql/hive/execution/HiveTableScan.scala
View file

@ -32,6 +32,7 @@ import org.apache.spark.sql.catalyst.expressions._
import org.apache.spark.sql.execution._
import org.apache.spark.sql.hive._
import org.apache.spark.sql.types.{BooleanType, DataType}
import org.apache.spark.util.Utils
/**
* The Hive table scan operator. Column and partition pruning are both handled.
@ -133,11 +134,17 @@ case class HiveTableScan(
}
protected override def doExecute(): RDD[InternalRow] = {
// Using dummyCallSite, as getCallSite can turn out to be expensive with
// with multiple partitions.
val rdd = if (!relation.hiveQlTable.isPartitioned) {
hadoopReader.makeRDDForTable(relation.hiveQlTable)
Utils.withDummyCallSite(sqlContext.sparkContext) {
hadoopReader.makeRDDForTable(relation.hiveQlTable)
}
} else {
hadoopReader.makeRDDForPartitionedTable(
prunePartitions(relation.getHiveQlPartitions(partitionPruningPred)))
Utils.withDummyCallSite(sqlContext.sparkContext) {
hadoopReader.makeRDDForPartitionedTable(
prunePartitions(relation.getHiveQlPartitions(partitionPruningPred)))
}
}
rdd.mapPartitionsInternal { iter =>
val proj = UnsafeProjection.create(schema)