Spark join produces wrong results

问题:

Presenting here before possibly filing a bug. I'm using Spark 1.6.0.

This is a simplified version of the problem I'm dealing with. I've filtered a table, and then I'm trying to do a left outer join with that subset and the main table, matching all the columns.

I've only got 2 rows in the main table and one in the filtered table. I'm expecting the resulting table to only have the single row from the subset.

scala> val b = Seq(("a", "b", 1), ("a", "b", 2)).toDF("a", "b", "c")
b: org.apache.spark.sql.DataFrame = [a: string, b: string, c: int]

scala> val a = b.where("c = 1").withColumnRenamed("a", "filta").withColumnRenamed("b", "filtb")
a: org.apache.spark.sql.DataFrame = [filta: string, filtb: string, c: int]

scala> a.join(b, $"filta" <=> $"a" and $"filtb" <=> $"b" and a("c") <=> b("c"), "left_outer").show
+-----+-----+---+---+---+---+
|filta|filtb|  c|  a|  b|  c|
+-----+-----+---+---+---+---+
|    a|    b|  1|  a|  b|  1|
|    a|    b|  1|  a|  b|  2|
+-----+-----+---+---+---+---+

I didn't expect that result at all. I expected the first row, but not the second. I suspected it's the null-safe equality, so I tried it without.

scala> a.join(b, $"filta" === $"a" and $"filtb" === $"b" and a("c") === b("c"), "left_outer").show
16/03/21 12:50:00 WARN Column: Constructing trivially true equals predicate, 'c#18232 = c#18232'. Perhaps you need to use aliases.
+-----+-----+---+---+---+---+
|filta|filtb|  c|  a|  b|  c|
+-----+-----+---+---+---+---+
|    a|    b|  1|  a|  b|  1|
+-----+-----+---+---+---+---+

OK, that's the result I expected, but then I got suspicious of the warning. There is a separate StackOverflow question to deal with that warning here: Spark SQL performing carthesian join instead of inner join

So I create a new column that avoids the warning.

scala> a.withColumn("newc", $"c").join(b, $"filta" === $"a" and $"filtb" === $"b" and $"newc" === b("c"), "left_outer").show
+-----+-----+---+----+---+---+---+
|filta|filtb|  c|newc|  a|  b|  c|
+-----+-----+---+----+---+---+---+
|    a|    b|  1|   1|  a|  b|  1|
|    a|    b|  1|   1|  a|  b|  2|
+-----+-----+---+----+---+---+---+

But now the result is wrong again! I have a lot of null-safe equality checks, and the warning isn't fatal, so I don't see a clear path to working with/around this.

Is the behaviour a bug, or is this expected behaviour? If expected, why?

回答1:

If you want an expected behavior use either join on names:

val b = Seq(("a", "b", 1), ("a", "b", 2)).toDF("a", "b", "c")
val a = b.where("c = 1")

a.join(b, Seq("a", "b", "c")).show
// +---+---+---+
// |  a|  b|  c|
// +---+---+---+
// |  a|  b|  1|
// +---+---+---+

or aliases:

val aa = a.alias("a")
val bb = b.alias("b")

aa.join(bb, $"a.a" === $"b.a" && $"a.b" === $"b.b" && $"a.c" === $"b.c")

You can use <=> as well:

aa.join(bb, $"a.a" <=> $"b.a" && $"a.b" <=> $"b.b" && $"a.c" <=> $"b.c")

As far as I remember there's been a special case for simple equality for a while. That's why you get correct results despite the warning.

The second behavior looks indeed like a bug related to the fact that you still have a.c in your data. It looks like it is picked downstream before b.c and the evaluated condition is actually a.newc = a.c.

val expr = $"filta" === $"a" and $"filtb" === $"b" and $"newc" === $"c"
a.withColumnRenamed("c", "newc").join(b, expr, "left_outer")