如何在 Spark SQL 中定义和使用用户定义的聚合函数？

Question

我知道如何在SparkSQL中编写UDF：

def belowThreshold(power: Int): Boolean = {
        return power < -40
      }

sqlContext.udf.register("belowThreshold", belowThreshold _)

我可以做类似的事情来定义聚合函数吗？这是怎么做到的？

对于上下文，我想运行以下SQL查询:

val aggDF = sqlContext.sql("""SELECT span, belowThreshold(opticalReceivePower), timestamp
                                    FROM ifDF
                                    WHERE opticalReceivePower IS NOT null
                                    GROUP BY span, timestamp
                                    ORDER BY span""")

它应该会返回类似于

行（span1，false，T0）

我希望聚合函数告诉我，在由< code>span和< code>timestamp定义的组中，是否有任何< code>opticalReceivePower的值低于阈值。我需要把我的UDAF写得和我上面粘贴的UDF不同吗？

Answer 1

要在火花（3.0） Java 中定义和使用 UDF，请执行以下操作：

private static UDF1<Integer, Boolean> belowThreshold = (power) -> power < -40;

        

注册虚拟硬盘录像机：

SparkSession.builder()
.appName(appName)
.master(master)
.getOrCreate().udf().register("belowThreshold", belowThreshold, BooleanType);

使用UDF by SparkSQL：

spark.sql("SELECT belowThreshold('50')");

Answer 2

火花

Scala < code > UserDefinedAggregateFunction 已被弃用(SPARK-30423 Deprecate UserDefinedAggregateFunction)，取而代之的是注册的< code >聚合器。

火花

矢量化udf(仅限Python):

from pyspark.sql.functions import pandas_udf
from pyspark.sql.functions import PandasUDFType

from pyspark.sql.types import *
import pandas as pd

df = sc.parallelize([
    ("a", 0), ("a", 1), ("b", 30), ("b", -50)
]).toDF(["group", "power"])

def below_threshold(threshold, group="group", power="power"):
    @pandas_udf("struct<group: string, below_threshold: boolean>", PandasUDFType.GROUPED_MAP)
    def below_threshold_(df):
        df = pd.DataFrame(
           df.groupby(group).apply(lambda x: (x[power] < threshold).any()))
        df.reset_index(inplace=True, drop=False)
        return df

    return below_threshold_

用法示例：

df.groupBy("group").apply(below_threshold(-40)).show()

## +-----+---------------+
## |group|below_threshold|
## +-----+---------------+
## |    b|           true|
## |    a|          false|
## +-----+---------------+

另请参阅在PySpark中对GroupeData应用UDF（具有功能的python示例）

火花

可以在类型化的数据集上使用聚合器：

import org.apache.spark.sql.expressions.Aggregator
import org.apache.spark.sql.{Encoder, Encoders}

class BelowThreshold[I](f: I => Boolean)  extends Aggregator[I, Boolean, Boolean]
    with Serializable {
  def zero = false
  def reduce(acc: Boolean, x: I) = acc | f(x)
  def merge(acc1: Boolean, acc2: Boolean) = acc1 | acc2
  def finish(acc: Boolean) = acc

  def bufferEncoder: Encoder[Boolean] = Encoders.scalaBoolean
  def outputEncoder: Encoder[Boolean] = Encoders.scalaBoolean
}

val belowThreshold = new BelowThreshold[(String, Int)](_._2 < - 40).toColumn
df.as[(String, Int)].groupByKey(_._1).agg(belowThreshold)

火花

在 Spark 1.5 中，您可以像这样创建 UDAF，尽管它很可能是矫枉过正：

import org.apache.spark.sql.expressions._
import org.apache.spark.sql.types._
import org.apache.spark.sql.Row

object belowThreshold extends UserDefinedAggregateFunction {
    // Schema you get as an input
    def inputSchema = new StructType().add("power", IntegerType)
    // Schema of the row which is used for aggregation
    def bufferSchema = new StructType().add("ind", BooleanType)
    // Returned type
    def dataType = BooleanType
    // Self-explaining 
    def deterministic = true
    // zero value
    def initialize(buffer: MutableAggregationBuffer) = buffer.update(0, false)
    // Similar to seqOp in aggregate
    def update(buffer: MutableAggregationBuffer, input: Row) = {
        if (!input.isNullAt(0))
          buffer.update(0, buffer.getBoolean(0) | input.getInt(0) < -40)
    }
    // Similar to combOp in aggregate
    def merge(buffer1: MutableAggregationBuffer, buffer2: Row) = {
      buffer1.update(0, buffer1.getBoolean(0) | buffer2.getBoolean(0))    
    }
    // Called on exit to get return value
    def evaluate(buffer: Row) = buffer.getBoolean(0)
}

用法示例：

df
  .groupBy($"group")
  .agg(belowThreshold($"power").alias("belowThreshold"))
  .show

// +-----+--------------+
// |group|belowThreshold|
// +-----+--------------+
// |    a|         false|
// |    b|          true|
// +-----+--------------+

火花 1.4 解决方法：

我不确定我是否正确理解了您的需求，但据我所知，这里简单的旧聚合应该足够了:

val df = sc.parallelize(Seq(
    ("a", 0), ("a", 1), ("b", 30), ("b", -50))).toDF("group", "power")

df
  .withColumn("belowThreshold", ($"power".lt(-40)).cast(IntegerType))
  .groupBy($"group")
  .agg(sum($"belowThreshold").notEqual(0).alias("belowThreshold"))
  .show

// +-----+--------------+
// |group|belowThreshold|
// +-----+--------------+
// |    a|         false|
// |    b|          true|
// +-----+--------------+

火花

据我所知，目前（Spark 1.4.1），除了Hive之外，没有对UDAF的支持。Spark 1.5应该可以（参见SPARK-3947）。

在内部，Spark 使用许多类，包括命令性聚合和声明性聚合。

有用于内部使用，可能会更改，恕不另行通知，所以它可能不是您想在生产代码中使用的东西，但只是为了完整性BelowThresholdwithDeclarativeAggregate可以这样实现（使用Spark 2.2-SNAPSHOT测试）：

import org.apache.spark.sql.catalyst.expressions.aggregate.DeclarativeAggregate
import org.apache.spark.sql.catalyst.expressions._
import org.apache.spark.sql.types._

case class BelowThreshold(child: Expression, threshold: Expression) 
    extends  DeclarativeAggregate  {
  override def children: Seq[Expression] = Seq(child, threshold)

  override def nullable: Boolean = false
  override def dataType: DataType = BooleanType

  private lazy val belowThreshold = AttributeReference(
    "belowThreshold", BooleanType, nullable = false
  )()

  // Used to derive schema
  override lazy val aggBufferAttributes = belowThreshold :: Nil

  override lazy val initialValues = Seq(
    Literal(false)
  )

  override lazy val updateExpressions = Seq(Or(
    belowThreshold,
    If(IsNull(child), Literal(false), LessThan(child, threshold))
  ))

  override lazy val mergeExpressions = Seq(
    Or(belowThreshold.left, belowThreshold.right)
  )

  override lazy val evaluateExpression = belowThreshold
  override def defaultResult: Option[Literal] = Option(Literal(false))
} 

应该用等效的< code > withAggregateFunction 进一步包装它。