GroupBy + custom aggregation on Dataset with Case class / Trait in the Key - scala

Question

I am trying to refactor some code and put the general logic into a trait. I basically want to process datasets, group them by some key and aggregate:
import org.apache.spark.sql.expressions.Aggregator
import org.apache.spark.sql.{ Dataset, Encoder, Encoders, TypedColumn }
case class SomeKey(a: String, b: Boolean)
case class InputRow(
SomeKey,
v: Double
)
trait MyTrait {
def processInputs: Dataset[InputRow]
def groupAndAggregate(
logs: Dataset[InputRow]
): Dataset[(SomeKey, Long)] = {
import logs.sparkSession.implicits._
logs
.groupByKey(i => i.key)
.agg(someAggFunc)
}
//Whatever agg function: here, it counts the number of v that are >= 0.5
def someAggFunc: TypedColumn[InputRow, Long] =
new Aggregator[
/*input type*/ InputRow,
/* "buffer" type */ Long,
/* output type */ Long
] with Serializable {
def zero = 0L
def reduce(b: Long, a: InputRow) = {
if (a.v >= 0.5)
b + 1
else
b
}
def merge(b1: Long, b2: Long) =
b1 + b2
// map buffer to output type
def finish(b: Long) = b
def bufferEncoder: Encoder[Long] = Encoders.scalaLong
def outputEncoder: Encoder[Long] = Encoders.scalaLong
}.toColumn
}
everything works fine: I can instantiate a class that inherits from MyTrait and override the way I process inputs:
import spark.implicits._
case class MyTraitTest(testDf: DataFrame) extends MyTrait {
override def processInputs: Dataset[InputRow] = {
val ds = testDf
.select(
$"a",
$"b",
$"v",
)
.rdd
.map(
r =>
InputRow(
SomeKey(r.getAs[String]("a"), r.getAs[Boolean]("b")),
r.getAs[Double]("v")
)
)
.toDS
ds
}
val df: DataFrame = Seq(
("1", false, 0.40),
("1", false, 0.54),
("0", true, 0.85),
("1", true, 0.39)
).toDF("a", "b", "v")
val myTraitTest = MyTraitTest(df)
val ds: Dataset[InputRow] = myTraitTest.processInputs
val res = myTraitTest.groupAndAggregate(ds)
res.show(false)
+----------+----------------------------------+
|key |InputRow |
+----------+----------------------------------+
|[1, false]|1 |
|[0, true] |1 |
|[1, true] |0 |
+----------+----------------------------------+
Now the problem: I want SomeKey to derive from a more generic trait Key, because the key will not always have only two fields, the fields won't have the same type etc. It will always be a simple tuple of some basic primitive types though.
So I tried to do the following:
trait Key extends Product
case class SomeKey(a: String, b: Boolean) extends Key
case class SomeOtherKey(x: Int, y: Boolean, z: String) extends Key
case class InputRow[T <: Key](
key: T,
v: Double
)
trait MyTrait[T <: Key] {
def processInputs: Dataset[InputRow[T]]
def groupAndAggregate(
logs: Dataset[InputRow[T]]
): Dataset[(T, Long)] = {
import logs.sparkSession.implicits._
logs
.groupByKey(i => i.key)
.agg(someAggFunc)
}
def someAggFunc: TypedColumn[InputRow[T], Long] = {...}
I now do:
case class MyTraitTest(testDf: DataFrame) extends MyTrait[SomeKey] {
override def processInputs: Dataset[InputRow[SomeKey]] = {
...
}
etc.
But now I get the error : Unable to find encoder for type T. An implicit Encoder[T] is needed to store T instances in a Dataset. Primitive types (Int, String, etc) and Product types (case classes) are supported by importing spark.implicits._ Support for serializing other types will be added in future releases.
.groupByKey(i => i.key)
I really don't know how to work around this issue, I tried lots of things without success. Sorry for this quite lengthy description but hopefully you have all the elements to help me understand... thanks!

Answer 1

Spark needs to be able to implicitly create the encoder for product type T so you'll need to help it work around the JVM type erasure and pass the TypeTag for T as an implicit parameter of your groupAndAggregate method.
A working example:
import org.apache.spark.sql.expressions.Aggregator
import org.apache.spark.sql.{ DataFrame, Dataset, Encoders, TypedColumn }
import scala.reflect.runtime.universe.TypeTag
trait Key extends Product
case class SomeKey(a: String, b: Boolean) extends Key
case class SomeOtherKey(x: Int, y: Boolean, z: String) extends Key
case class InputRow[T <: Key](key: T, v: Double)
trait MyTrait[T <: Key] {
def processInputs: Dataset[InputRow[T]]
def groupAndAggregate(
logs: Dataset[InputRow[T]]
)(implicit tTypeTag: TypeTag[T]): Dataset[(T, Long)] = {
import logs.sparkSession.implicits._
logs
.groupByKey(i => i.key)
.agg(someAggFunc)
}
def someAggFunc: TypedColumn[InputRow[T], Long] =
new Aggregator[InputRow[T], Long, Long] with Serializable {
def reduce(b: Long, a: InputRow[T]) = b + (a.v * 100).toLong
def merge(b1: Long, b2: Long) = b1 + b2
def zero = 0L
def finish(b: Long) = b
def bufferEncoder = Encoders.scalaLong
def outputEncoder = Encoders.scalaLong
}.toColumn
}
with a wrapping case class
case class MyTraitTest(testDf: DataFrame) extends MyTrait[SomeKey] {
import testDf.sparkSession.implicits._
import org.apache.spark.sql.functions.struct
override def processInputs = testDf
.select(struct($"a", $"b") as "key", $"v" )
.as[InputRow[SomeKey]]
}
and a test execution
val df = Seq(
("1", false, 0.40),
("1", false, 0.54),
("0", true, 0.85),
("1", true, 0.39)
).toDF("a", "b", "v")
val myTraitTest = MyTraitTest(df)
val ds = myTraitTest.processInputs
val res = myTraitTest.groupAndAggregate(ds)
res.show(false)
+----------+-----------------------------------------------+
|key |$anon$1($line5460910223.$read$$iw$$iw$InputRow)|
+----------+-----------------------------------------------+
|[1, false]|94 |
|[1, true] |39 |
|[0, true] |85 |
+----------+-----------------------------------------------+