[SPARK-18642] Spark SQL: Catalyst is scanning undesired columns

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.6.2, 1.6.3
• Fix Version/s: 2.0.0
• Component/s: SQL
• Labels: performance

Environment

Ubuntu 14.04
Spark: Local Mode

• Assignee: Dongjoon Hyun
• Reporter: Mohit
• Created: 30/Nov/16 05:26
• Updated: 13/Dec/16 11:22
• Resolved: 05/Dec/16 06:37

Description

When doing a left-join between two tables, say A and B, Catalyst has information about the projection required for table B. Only the required columns should be scanned.

Code snippet below explains the scenario:

scala> val dfA = sqlContext.read.parquet("/home/mohit/ruleA")
dfA: org.apache.spark.sql.DataFrame = [aid: int, aVal: string]

scala> val dfB = sqlContext.read.parquet("/home/mohit/ruleB")
dfB: org.apache.spark.sql.DataFrame = [bid: int, bVal: string]

scala> dfA.registerTempTable("A")
scala> dfB.registerTempTable("B")

scala> sqlContext.sql("select A.aid, B.bid from A left join B on A.aid=B.bid where B.bid<2").explain

== Physical Plan ==
Project aid#15,bid#17
+- Filter (bid#17 < 2)
+- BroadcastHashOuterJoin aid#15, bid#17, LeftOuter, None
:- Scan ParquetRelationaid#15,aVal#16 InputPaths: file:/home/mohit/ruleA
+- Scan ParquetRelationbid#17,bVal#18 InputPaths: file:/home/mohit/ruleB

This is a watered-down example from a production issue which has a huge performance impact.
External reference: http://stackoverflow.com/questions/40783675/spark-sql-catalyst-is-scanning-undesired-columns