How to join Datasets on multiple columns?

There's another way of joining by chaining where one after another. You first specify a join (and optionally its type) followed by where operator(s), i.e.

scala> case class A(id: Long, name: String)
defined class A

scala> case class B(id: Long, name: String)
defined class B

scala> val as = Seq(A(0, "zero"), A(1, "one")).toDS
as: org.apache.spark.sql.Dataset[A] = [id: bigint, name: string]

scala> val bs = Seq(B(0, "zero"), B(1, "jeden")).toDS
bs: org.apache.spark.sql.Dataset[B] = [id: bigint, name: string]

scala> as.join(bs).where(as("id") === bs("id")).show
+---+----+---+-----+
| id|name| id| name|
+---+----+---+-----+
|  0|zero|  0| zero|
|  1| one|  1|jeden|
+---+----+---+-----+


scala> as.join(bs).where(as("id") === bs("id")).where(as("name") === bs("name")).show
+---+----+---+----+
| id|name| id|name|
+---+----+---+----+
|  0|zero|  0|zero|
+---+----+---+----+

The reason for such a goodie is that the Spark optimizer will join (no pun intended) consecutive wheres into one with join. Use explain operator to see the underlying logical and physical plans.

scala> as.join(bs).where(as("id") === bs("id")).where(as("name") === bs("name")).explain(extended = true)
== Parsed Logical Plan ==
Filter (name#31 = name#36)
+- Filter (id#30L = id#35L)
   +- Join Inner
      :- LocalRelation [id#30L, name#31]
      +- LocalRelation [id#35L, name#36]

== Analyzed Logical Plan ==
id: bigint, name: string, id: bigint, name: string
Filter (name#31 = name#36)
+- Filter (id#30L = id#35L)
   +- Join Inner
      :- LocalRelation [id#30L, name#31]
      +- LocalRelation [id#35L, name#36]

== Optimized Logical Plan ==
Join Inner, ((name#31 = name#36) && (id#30L = id#35L))
:- Filter isnotnull(name#31)
:  +- LocalRelation [id#30L, name#31]
+- Filter isnotnull(name#36)
   +- LocalRelation [id#35L, name#36]

== Physical Plan ==
*BroadcastHashJoin [name#31, id#30L], [name#36, id#35L], Inner, BuildRight
:- *Filter isnotnull(name#31)
:  +- LocalTableScan [id#30L, name#31]
+- BroadcastExchange HashedRelationBroadcastMode(List(input[1, string, false], input[0, bigint, false]))
   +- *Filter isnotnull(name#36)
      +- LocalTableScan [id#35L, name#36]

You can do it exactly the same way as with Dataframe:

val xs = Seq(("a", "foo", 2.0), ("x", "bar", -1.0)).toDS
val ys = Seq(("a", "foo", 2.0), ("y", "bar", 1.0)).toDS

xs.joinWith(ys, xs("_1") === ys("_1") && xs("_2") === ys("_2"), "left").show
// +------------+-----------+
// |          _1|         _2|
// +------------+-----------+
// | [a,foo,2.0]|[a,foo,2.0]|
// |[x,bar,-1.0]|       null|
// +------------+-----------+

In Spark < 2.0.0 you can use something like this:

xs.as("xs").joinWith(
  ys.as("ys"), ($"xs._1" === $"ys._1") && ($"xs._2" === $"ys._2"), "left")