Self-join not working as expected with the DataFrame API

Question

I am trying to get the latest records from a table using self join. It works using spark-sql but not working using spark DataFrame API.

Can anyone help? Is it a bug?

I am using Spark 2.2.0 in local mode

Creating input DataFrame:

scala> val df3 = spark.sparkContext.parallelize(Array((1,"a",1),(1,"aa",2),(2,"b",2),(2,"bb",5))).toDF("id","value","time")
df3: org.apache.spark.sql.DataFrame = [id: int, value: string ... 1 more field]    

scala> val df33 = df3
df33: org.apache.spark.sql.DataFrame = [id: int, value: string ... 1 more field]

scala> df3.show
+---+-----+----+
| id|value|time|
+---+-----+----+
|  1|    a|   1|
|  1|   aa|   2|
|  2|    b|   2|
|  2|   bb|   5|
+---+-----+----+

scala> df33.show
+---+-----+----+
| id|value|time|
+---+-----+----+
|  1|    a|   1|
|  1|   aa|   2|
|  2|    b|   2|
|  2|   bb|   5|
+---+-----+----+

Now performing the join using SQL: works

scala> spark.sql("select df33.* from df3 join df33 on df3.id = df33.id and df3.time < df33.time").show
+---+-----+----+
| id|value|time|
+---+-----+----+
|  1|   aa|   2|
|  2|   bb|   5|
+---+-----+----+

Now performing the join using dataframe API: doesn't work

scala> df3.join(df33, (df3.col("id") === df33.col("id")) && (df3.col("time") < df33.col("time")) ).select(df33.col("id"),df33.col("value"),df33.col("time")).show
+---+-----+----+
| id|value|time|
+---+-----+----+
+---+-----+----+

The thing to notice is the explain plans: blank for the DataFrame API!!

scala> df3.join(df33, (df3.col("id") === df33.col("id")) && (df3.col("time") < df33.col("time")) ).select(df33.col("id"),df33.col("value"),df33.col("time")).explain
== Physical Plan ==
LocalTableScan <empty>, [id#150, value#151, time#152]

scala> spark.sql("select df33.* from df3 join df33 on df3.id = df33.id and df3.time < df33.time").explain
== Physical Plan ==
*Project [id#1241, value#1242, time#1243]
+- *SortMergeJoin [id#150], [id#1241], Inner, (time#152 < time#1243)
   :- *Sort [id#150 ASC NULLS FIRST], false, 0
   :  +- Exchange hashpartitioning(id#150, 200)
   :     +- *Project [_1#146 AS id#150, _3#148 AS time#152]
   :        +- *SerializeFromObject [assertnotnull(input[0, scala.Tuple3, true])._1 AS _1#146, staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString,
assertnotnull(input[0, scala.Tuple3, true])._2, true) AS _2#147, assertnotnull(input[0, scala.Tuple3, true])._3 AS _3#148]
   :           +- Scan ExternalRDDScan[obj#145]
   +- *Sort [id#1241 ASC NULLS FIRST], false, 0
      +- Exchange hashpartitioning(id#1241, 200)
         +- *Project [_1#146 AS id#1241, _2#147 AS value#1242, _3#148 AS time#1243]
            +- *SerializeFromObject [assertnotnull(input[0, scala.Tuple3, true])._1 AS _1#146, staticinvoke(class org.apache.spark.unsafe.types.UTF8String, StringType, fromString,
assertnotnull(input[0, scala.Tuple3, true])._2, true) AS _2#147, assertnotnull(input[0, scala.Tuple3, true])._3 AS _3#148]
               +- Scan ExternalRDDScan[obj#145]

Answer 1

No that's not a bug, but when you reassign the DataFrame to a new one like what you have done, it actually copies the lineage but it doesn't duplicate the data. Thus you'll be comparing on the same column.

Use spark.sql is slightly different because it's actually working on aliases of your DataFrames

So the correct way to perform a self-join using the API is actually aliasing your DataFrame as followed :

val df1 = Seq((1,"a",1),(1,"aa",2),(2,"b",2),(2,"bb",5)).toDF("id","value","time")

df1.as("df1").join(df1.as("df2"), $"df1.id" === $"df2.id" && $"df1.time" < $"df2.time").select($"df2.*").show
// +---+-----+----+
// | id|value|time|
// +---+-----+----+
// |  1|   aa|   2|
// |  2|   bb|   5|
// +---+-----+----+

For more information about self-joins, I recommend reading High Performance Spark by Rachel Warren, Holden Karau - Chapter 4.