将带有struct-type键的映射传递给Spark UDF

Question

我想写一个Spark 1.6 UDF,它采用以下映射:

case class MyRow(mapping: Map[(Int, Int), Double])

val data = Seq(
  MyRow(Map((1, 1) -> 1.0))
)
val df = sc.parallelize(data).toDF()

df.printSchema()

root
 |-- mapping: map (nullable = true)
 |    |-- key: struct
 |    |-- value: double (valueContainsNull = false)
 |    |    |-- _1: integer (nullable = false)
 |    |    |-- _2: integer (nullable = false)

(作为旁注:我发现上面的输出很奇怪,因为键的类型打印在值的类型下面,为什么会这样？)

现在我将我的UDF定义为:

val myUDF = udf((inputMapping: Map[(Int,Int), Double]) =>
  inputMapping.map { case ((i1, i2), value) => ((i1 + i2), value) }
)

df
  .withColumn("udfResult", myUDF($"mapping"))
  .show()

但这给了我:

java.lang.ClassCastException: org.apache.spark.sql.catalyst.expressions.GenericRowWithSchema cannot be cast to scala.Tuple2

所以我尝试(Int,Int)用自定义替换case class,因为如果我想传递struct给UDF,我通常会这样做:

case class MyTuple2(i1: Int, i2: Int)
val myUDF = udf((inputMapping: Map[MyTuple2, Double]) => 
  inputMapping.map { case (MyTuple2(i1, i2), value) => ((i1 + i2), value) }
)

这奇怪地给出了:

org.apache.spark.sql.AnalysisException: cannot resolve 'UDF(mapping)' due to data type mismatch: argument 1 requires map<struct<i1:int,i2:int>,double> type, however, 'mapping' is of map<struct<_1:int,_2:int>,double> type.

由于类型匹配,我不理解上述异常.

我发现的唯一(丑陋)解决方案是传递a org.apache.spark.sql.Row然后"提取"结构的元素:

val myUDF = udf((inputMapping: Map[Row, Double]) => inputMapping
  .map { case (key, value) => ((key.getInt(0), key.getInt(1)), value) } // extract Row into Tuple2
  .map { case ((i1, i2), value) => ((i1 + i2), value) }
)

Answer 1

据我所知，Row在这种情况下没有逃避使用：映射（或另一个元组/案例类/数组...）中使用的元组（或案例类）是嵌套结构，因此它会表示为Row传递给 UDF 的时间。

我可以建议的唯一改进是使用Row.unapply稍微简化代码：

val myUDF = udf((inputMapping: Map[Row, Double]) => inputMapping
  .map { case (Row(i1: Int, i2: Int), value) => (i1 + i2, value) }
)