[SPARK-26572] Join on distinct column with monotonically_increasing_id produces wrong output

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.0.2, 2.1.3, 2.2.2, 2.3.2, 2.4.0
• Fix Version/s: 2.3.4, 2.4.1, 3.0.0
• Component/s: SQL
• Labels: correctness

Environment

Running on Ubuntu 18.04LTS and Intellij 2018.2.5

• Assignee: Peter Toth
• Reporter: Sören Reichardt
• Created: 08/Jan/19 13:16
• Updated: 02/Mar/20 19:36
• Resolved: 15/Feb/19 01:05

Description

When joining a table with projected monotonically_increasing_id column after calling distinct with another table the operators do not get executed in the right order. 

Here is a minimal example:

import org.apache.spark.sql.{DataFrame, SparkSession, functions}

object JoinBug extends App {

  // Spark session setup
  val session =  SparkSession.builder().master("local[*]").getOrCreate()
  import session.sqlContext.implicits._
  session.sparkContext.setLogLevel("error")

  // Bug in Spark: "monotonically_increasing_id" is pushed down when it shouldn't be. Push down only happens when the
  // DF containing the "monotonically_increasing_id" expression is on the left side of the join.
  val baseTable = Seq((1), (1)).toDF("idx")
  val distinctWithId = baseTable.distinct.withColumn("id", functions.monotonically_increasing_id())
  val monotonicallyOnRight: DataFrame = baseTable.join(distinctWithId, "idx")
  val monotonicallyOnLeft: DataFrame = distinctWithId.join(baseTable, "idx")

  monotonicallyOnLeft.show // Wrong
  monotonicallyOnRight.show // Ok in Spark 2.2.2 - also wrong in Spark 2.4.0

}

It produces the following output:

Wrong:
+---+------------+
|idx| id         |
+---+------------+
| 1|369367187456 |
| 1|369367187457 |
+---+------------+

Right:
+---+------------+
|idx| id         |
+---+------------+
| 1|369367187456 |
| 1|369367187456 |
+---+------------+

We assume that the join operator triggers a pushdown of expressions (monotonically_increasing_id in this case) which gets pushed down to be executed before distinct. This produces non-distinct rows with unique id's. However it seems like this behavior only appears if the table with the projected expression is on the left side of the join in Spark 2.2.2 (for version 2.4.0 it fails on both joins).

Issue Links

• links to: GitHub Pull Request #23731