Arun Allamsetty created SPARK-19954:
---------------------------------------
Summary: Joining to a unioned DataFrame does not produce expected
result.
Key: SPARK-19954
URL: https://issues.apache.org/jira/browse/SPARK-19954
Project: Spark
Issue Type: Bug
Components: SQL
Affects Versions: 2.1.0
Reporter: Arun Allamsetty
Priority: Blocker
I found this bug while trying to update from Spark 1.6.1 to 2.1.0. The bug is
that when we try to join two DataFrames, one of which is a result of a union
operation, the result of the join results in data as if the table was joined
only to the first table in the union. This issue is not present in Spark 2.0.0
or 2.0.1 or 2.0.2, only in 2.1.0. Here's how to reproduce it.
{{noformat}}
import spark.implicits._
import org.apache.spark.sql.functions.lit
case class A(id: Long, colA: Boolean)
case class B(id: Long, colB: Int)
case class C(id: Long, colC: Double)
case class X(id: Long, name: String)
val aData = A(1, true) :: Nil
val bData = B(2, 10) :: Nil
val cData = C(3, 9.73D) :: Nil
val xData = X(1, "a") :: X(2, "b") :: X(3, "c") :: Nil
val aDf = spark.createDataset(aData).toDF
val bDf = spark.createDataset(bData).toDF
val cDf = spark.createDataset(cData).toDF
val xDf = spark.createDataset(xData).toDF
val unionDf =
aDf.select($"id", lit("a").as("name"), $"colA", lit(null).as("colB"),
lit(null).as("colC")).union(
bDf.select($"id", lit("b").as("name"), lit(null).as("colA"), $"colB",
lit(null).as("colC"))).union(
cDf.select($"id", lit("c").as("name"), lit(null).as("colA"),
lit(null).as("colB"), $"colC"))
val result = xDf.join(unionDf, unionDf("name") === xDf("name") && unionDf("id")
=== xDf("id"))
result.show
{{noformat}}
The result being
{{noformat}}
+---+----+---+----+----+----+----+
| id|name| id|name|colA|colB|colC|
+---+----+---+----+----+----+----+
| 1| a| 1| a|true|null|null|
+---+----+---+----+----+----+----+
{{noformat}}
Force computing {{unionDf}} using {{count}} does not help change the result of
the join. However, writing the data to disk and reading it back does give the
correct result. But it is definitely not ideal. Interestingly caching the
{{unionDf}} also gives the correct result.
{{noformat}}
+---+----+---+----+----+----+----+
| id|name| id|name|colA|colB|colC|
+---+----+---+----+----+----+----+
| 1| a| 1| a|true|null|null|
| 2| b| 2| b|null| 10|null|
| 3| c| 3| c|null|null|9.73|
+---+----+---+----+----+----+----+
{{noformat}}
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
