Perform a typed join in Scala with Spark Datasets

问题:

I like Spark Datasets as they give me analysis errors and syntax errors at compile time and also allow me to work with getters instead of hard-coded names/numbers. Most computations can be accomplished with Dataset’s high-level APIs. For example, it’s much simpler to perform agg, select, sum, avg, map, filter, or groupBy operations by accessing a Dataset typed object’s than using RDD rows’ data fields.

However the join operation is missing from this, I read that I can do a join like this

ds1.joinWith(ds2, ds1.toDF().col("key") === ds2.toDF().col("key"), "inner")

But that is not what I want as I would prefer to do it via the case class interface, so something more like this

ds1.joinWith(ds2, ds1.key === ds2.key, "inner")

The best alternative for now seems to create an object next to the case class and give this functions to provide me with the right column name as a String. So I would use the first line of code but put a function instead of a hard-coded column name. But that doesn't feel elegant enough..

Can someone advise me on other options here? The goal is to have an abstraction from the actual column names and work preferably via the getters of the case class.

I'm using Spark 1.6.1 and Scala 2.10

回答1:

Observation

Spark SQL can optimize join only if join condition is based on the equality operator. This means we can consider equijoins and non-equijoins separately.

Equijoin

Equijoin can be implemented in a type safe manner by mapping both Datasets to (key, value) tuples, performing join based on keys, and reshaping the result:

import org.apache.spark.sql.Encoder
import org.apache.spark.sql.Dataset

def safeEquiJoin[T, U, K](ds1: Dataset[T], ds2: Dataset[U])
    (f: T => K, g: U => K)
    (implicit e1: Encoder[(K, T)], e2: Encoder[(K, U)], e3: Encoder[(T, U)]) = {
  val ds1_ = ds1.map(x => (f(x), x))
  val ds2_ = ds2.map(x => (g(x), x))
  ds1_.joinWith(ds2_, ds1_("_1") === ds2_("_1")).map(x => (x._1._2, x._2._2))
}

Non-equijoin

Can be expressed using relational algebra operators as R ⋈θ S = σθ(R × S) and converted directly to code.

Spark 2.0

Enable crossJoin and use joinWith with trivially equal predicate:

spark.conf.set("spark.sql.crossJoin.enabled", true)

def safeNonEquiJoin[T, U](ds1: Dataset[T], ds2: Dataset[U])
                         (p: (T, U) => Boolean) = {
  ds1.joinWith(ds2, lit(true)).filter(p.tupled)
}

Spark 2.1

Use crossJoin method:

def safeNonEquiJoin[T, U](ds1: Dataset[T], ds2: Dataset[U])
    (p: (T, U) => Boolean)
    (implicit e1: Encoder[Tuple1[T]], e2: Encoder[Tuple1[U]], e3: Encoder[(T, U)]) = {
  ds1.map(Tuple1(_)).crossJoin(ds2.map(Tuple1(_))).as[(T, U)].filter(p.tupled)
}

Examples

case class LabeledPoint(label: String, x: Double, y: Double)
case class Category(id: Long, name: String)

val points1 = Seq(LabeledPoint("foo", 1.0, 2.0)).toDS
val points2 = Seq(
  LabeledPoint("bar", 3.0, 5.6), LabeledPoint("foo", -1.0, 3.0)
).toDS
val categories = Seq(Category(1, "foo"), Category(2, "bar")).toDS

safeEquiJoin(points1, categories)(_.label, _.name)
safeNonEquiJoin(points1, points2)(_.x > _.x)

Notes

• It should be noted that these methods are qualtiatively differnt from a direct joinWith application and require expensive DeserializeToObject / SerializeFromObject transformations (compared to that direct joinWith can use logical operations on the data).

  This is similar to the behavior described in Spark 2.0 Dataset vs DataFrame.

• If you're not limited to the Spark SQL API frameless provides interesting type safe extensions for Datasets (as of today its supports only Spark 2.0):

  import frameless.TypedDataset
  
  val typedPoints1 = TypedDataset.create(points1)
  val typedPoints2 = TypedDataset.create(points2)
  
  typedPoints1.join(typedPoints2, typedPoints1('x), typedPoints2('x))
  

• Dataset API is not stable in 1.6 so I don't think it makes sense to use it there.

• Of course this design and descriptive names are not necessary. You can easily use type class to add this methods implicitly to Dataset an there is no conflict with built in signatures so both can be called joinWith.

回答2:

Also, another bigger problem for not type safe Spark API is that when you join two Datasets, it will give you a DataFrame. And then you lose types from your original two datasets.

val a: Dataset[A]
val b: Dataset[B]

val joined: Dataframe = a.join(b)
// what would be great is 
val joined: Dataset[C] = a.join(b)(implicit func: (A, B) => C)