Question

我对Scala Spark生态系统非常陌生，想知道什么是对链式数据框转换进行单元测试的最佳方法。所以这是我要测试的方法的代码示例

def writeToParquet(spark: SparkSession, dataFrame: DataFrame, col1: DataType1, col2:DataType2): Unit {
    dataFrame
        .withColumn("date", some_columnar_date_logic)
        .withColumn("hour", some_more_functional_logic)
        .... //couple more transformation logic
        .write
        .mode(SaveMode.Append)
        .partitionBy("col1", "col2", "col3")
        .parquet("some hdfs/s3/url")        
}

问题在于实木复合地板属于Unit返回类型，这使测试变得困难。问题是，转换本质上是不可变的，这使得模拟和监视变得有些困难

要创建数据框，我将测试数据集转储到了CSV

Answer 1

请找到用于数据帧单元测试的简单示例。您可以将其分为两部分。第一。测试转换，您可以执行简单的shell脚本来测试写入的文件

import com.holdenkarau.spark.testing._
import org.apache.spark.sql.{DataFrame, Row}
import org.apache.spark.sql.functions.lit
import org.apache.spark.sql.types.{IntegerType, StringType, StructField, StructType}
import org.scalatest.{FunSuite, Matchers}

class SomeDFTest extends FunSuite with Matchers with DataFrameSuiteBase    {
 import spark.implicits._

  test("Testing Input customer data date transformation") {


    val inputSchema = List(
      StructField("number", IntegerType, false),
      StructField("word", StringType, false)
    )
    val expectedSchema = List(
      StructField("number", IntegerType, false),
      StructField("word", StringType, false),
      StructField("dummyColumn", StringType, false)

    )
    val inputData = Seq(
      Row(8, "bat"),
      Row(64, "mouse"),
      Row(-27, "horse")
    )

    val expectedData = Seq(
      Row (8, "bat","test"),
      Row(64, "mouse","test"),
      Row(-27, "horse","test")
    )

    val inputDF = spark.createDataFrame(
      spark.sparkContext.parallelize(inputData),
      StructType(inputSchema)
    )

    val expectedDF = spark.createDataFrame(
      spark.sparkContext.parallelize(expectedData),
      StructType(expectedSchema)
    )


    val actual = transformSomeDf(inputDF)

    assertDataFrameEquals(actual, expectedDF) // equal



  }

  def transformSomeDf(df:DataFrame):DataFrame={
    df.withColumn("dummyColumn",lit("test"))
  }
}

Sbt.build配置

name := "SparkTest"

version := "0.1"

scalaVersion := "2.11.8"

val sparkVersion = "2.3.0"

libraryDependencies ++= Seq(
  "org.apache.spark" %% "spark-core" % sparkVersion,
  "org.apache.spark" %% "spark-sql" % sparkVersion,
  "org.apache.spark" %% "spark-hive" % sparkVersion % "provided",
"com.holdenkarau" %% "spark-testing-base" % "2.4.0_0.11.0" % Test

)

Answer 2

我在测试数据框时发现的第一件事就是将转换和IO分开

对于上述情况我们可以将上述链条分为三个部分

class Coordinator {
    def transformAndWrite(dataframe: Dataframe): Unit = {
transformedDf = dataFrame
        .withColumn("date", some_columnar_date_logic)
        .withColumn("hour", some_more_functional_logic)
        .... //couple more transformation logic

partitionedDfWriter = transformedDf.write
        .mode(SaveMode.Append)
        .partitionBy("col1", "col2", "col3")

和

partitionedDfWriter.parquet("some hdfs/s3/url")
}

现在我们可以将它们移到三个单独的类中，

DFTransformer，DFPartitioner 和 DataFrameParquetWriter extends ResourceWriter

所以代码将变成这样

class DFTransformer {
    def transform(dataframe:DataFrame): Dataframe = {
        return dataFrame
        .withColumn("date", some_columnar_date_logic)
        .withColumn("hour", some_more_functional_logic)
        .... //couple more transformation logic

}

class DfPartitioner {
    def partition(dataframe: DataFrame): DataFrameWriter = {
        return dataframe.write
        .mode(SaveMode.Append)
        .partitionBy("col1", "col2", "col3")
    }
}

和

class DataFrameParquetWriter extends ResourceWriter {
    overide def write(partitionedDfWriter: DataFrameWriter) = {
       partitionedDfWriter.parquet("some hdfs/s3/url") 

    }

class Coordinator(dfTransformer:DfTransformer, dfPartitioner: DFPartitioner, resourceWriter: ResourceWriter) {
    val transformedDf = dfTransformer.transform(dataframe)
    val partitionedDfWriter = dfPartitioner.partition(transformedDf)
    resourceWriter.write(partitionedDfWriter)
}

上述优点是，当您必须测试Coordinator类时，可以非常轻松地使用Mockito来模拟依赖项。
测试DFTransformer现在也很容易，您可以传递存根数据框并声明返回的数据框。（使用spark-testing-base）。我们还可以测试转换返回的列。我们也可以测试计数

单元测试Spark数据帧转换链

2 个答案: