Details
-
Bug
-
Status: Resolved
-
Minor
-
Resolution: Incomplete
-
2.0.2, 2.2.0
-
None
Description
hiveContext.emptyDataFrame.except(hiveContext.emptyDataFrame).count() yields 1 against expected 0.
This was not the case with spark 1.5.2. This is an api change from usage point of view and hence I consider this as a bug. May be a boundary case, not sure.
Work around - For now I check the counts != 0 before this operation. Not good for performance. Hence creating a jira to track it.
As Young Zhang explained in reply to my mail -
Starting from Spark 2, these kind of operation are implemented in left anti join, instead of using RDD operation directly.
Same issue also on sqlContext.
scala> spark.version
res25: String = 2.0.2
spark.sqlContext.emptyDataFrame.except(spark.sqlContext.emptyDataFrame).explain(true)
== Physical Plan ==
*HashAggregate(keys=[], functions=[], output=[])
+- Exchange SinglePartition
+- *HashAggregate(keys=[], functions=[], output=[])
+- BroadcastNestedLoopJoin BuildRight, LeftAnti, false
:- Scan ExistingRDD[]
+- BroadcastExchange IdentityBroadcastMode
+- Scan ExistingRDD[]
This arguably means a bug. But my guess is liking the logic of comparing NULL = NULL, should it return true or false, causing this kind of confusion.
