Spark：写入CSV文件

几周前，我写了我是如何使用Spark探索芝加哥市犯罪数据集的 ，并得出了每起犯罪的数量，我想将其写入CSV文件。

Spark提供了一个saveAsTextFile函数，该函数允许我们保存RDD的代码，因此我将代码重构为以下格式，以允许我使用它：

import au.com.bytecode.opencsv.CSVParser
import org.apache.spark.rdd.RDD
import org.apache.spark.SparkContext._
 
def dropHeader(data: RDD[String]): RDD[String] = {
  data.mapPartitionsWithIndex((idx, lines) => {
    if (idx == 0) {
      lines.drop(1)
    }
    lines
  })
}
 
// https://data.cityofchicago.org/Public-Safety/Crimes-2001-to-present/ijzp-q8t2
val crimeFile = "/Users/markneedham/Downloads/Crimes_-_2001_to_present.csv"
 
val crimeData = sc.textFile(crimeFile).cache()
val withoutHeader: RDD[String] = dropHeader(crimeData)
 
val file = "/tmp/primaryTypes.csv"
FileUtil.fullyDelete(new File(file))
 
val partitions: RDD[(String, Int)] = withoutHeader.mapPartitions(lines => {
  val parser = new CSVParser(',')
  lines.map(line => {
    val columns = parser.parseLine(line)
    (columns(5), 1)
  })
})
 
val counts = partitions.
  reduceByKey {case (x,y) => x + y}.
  sortBy {case (key, value) => -value}.
  map { case (key, value) => Array(key, value).mkString(",") }
 
counts.saveAsTextFile(file)

如果从Spark shell运行该代码，则最终会得到一个名为/tmp/primaryTypes.csv的文件夹， 其中包含多个零件文件：

$ ls -lah /tmp/primaryTypes.csv/
total 496
drwxr-xr-x  66 markneedham  wheel   2.2K 30 Nov 07:17 .
drwxrwxrwt  80 root         wheel   2.7K 30 Nov 07:16 ..
-rw-r--r--   1 markneedham  wheel     8B 30 Nov 07:16 ._SUCCESS.crc
-rw-r--r--   1 markneedham  wheel    12B 30 Nov 07:16 .part-00000.crc
-rw-r--r--   1 markneedham  wheel    12B 30 Nov 07:16 .part-00001.crc
-rw-r--r--   1 markneedham  wheel    12B 30 Nov 07:16 .part-00002.crc
-rw-r--r--   1 markneedham  wheel    12B 30 Nov 07:16 .part-00003.crc
...
-rwxrwxrwx   1 markneedham  wheel     0B 30 Nov 07:16 _SUCCESS
-rwxrwxrwx   1 markneedham  wheel    28B 30 Nov 07:16 part-00000
-rwxrwxrwx   1 markneedham  wheel    17B 30 Nov 07:16 part-00001
-rwxrwxrwx   1 markneedham  wheel    23B 30 Nov 07:16 part-00002
-rwxrwxrwx   1 markneedham  wheel    16B 30 Nov 07:16 part-00003
...

如果我们查看其中的一些零件文件，我们可以看到它已记录了犯罪类型和预期数目：

$ cat /tmp/primaryTypes.csv/part-00000
THEFT,859197
BATTERY,757530
 
$ cat /tmp/primaryTypes.csv/part-00003
BURGLARY,257310

如果我们要将这些CSV文件传递到另一个基于Hadoop的作业中，这很好，但是我实际上只想要一个CSV文件，所以这不是我想要的。

实现此目的的一种方法是强制将所有内容都计算在一个分区上，这意味着我们只会生成一个零件文件：

val counts = partitions.repartition(1).
  reduceByKey {case (x,y) => x + y}.
  sortBy {case (key, value) => -value}.
  map { case (key, value) => Array(key, value).mkString(",") }
 
 
counts.saveAsTextFile(file)

part-00000现在看起来像这样：

$ cat !$
cat /tmp/primaryTypes.csv/part-00000
THEFT,859197
BATTERY,757530
NARCOTICS,489528
CRIMINAL DAMAGE,488209
BURGLARY,257310
OTHER OFFENSE,253964
ASSAULT,247386
MOTOR VEHICLE THEFT,197404
ROBBERY,157706
DECEPTIVE PRACTICE,137538
CRIMINAL TRESPASS,124974
PROSTITUTION,47245
WEAPONS VIOLATION,40361
PUBLIC PEACE VIOLATION,31585
OFFENSE INVOLVING CHILDREN,26524
CRIM SEXUAL ASSAULT,14788
SEX OFFENSE,14283
GAMBLING,10632
LIQUOR LAW VIOLATION,8847
ARSON,6443
INTERFERE WITH PUBLIC OFFICER,5178
HOMICIDE,4846
KIDNAPPING,3585
INTERFERENCE WITH PUBLIC OFFICER,3147
INTIMIDATION,2471
STALKING,1985
OFFENSES INVOLVING CHILDREN,355
OBSCENITY,219
PUBLIC INDECENCY,86
OTHER NARCOTIC VIOLATION,80
NON-CRIMINAL,12
RITUALISM,12
OTHER OFFENSE ,6
NON - CRIMINAL,2
NON-CRIMINAL (SUBJECT SPECIFIED),2

这可以工作，但是比我们跨分区进行聚合时要慢很多，因此并不理想。

相反，我们能做的是利用Hadoop的合并功能之一， 将部分文件压缩到一个文件中。

首先，我们将Hadoop导入到我们的SBT文件中：

libraryDependencies += "org.apache.hadoop" % "hadoop-hdfs" % "2.5.2"

现在，让我们将合并功能引入Spark shell：

import org.apache.hadoop.conf.Configuration
import org.apache.hadoop.fs._
 
def merge(srcPath: String, dstPath: String): Unit =  {
  val hadoopConfig = new Configuration()
  val hdfs = FileSystem.get(hadoopConfig)
  FileUtil.copyMerge(hdfs, new Path(srcPath), hdfs, new Path(dstPath), false, hadoopConfig, null)
}

现在让我们利用它：

val file = "/tmp/primaryTypes.csv"
FileUtil.fullyDelete(new File(file))
 
val destinationFile= "/tmp/singlePrimaryTypes.csv"
FileUtil.fullyDelete(new File(destinationFile))
 
val counts = partitions.
reduceByKey {case (x,y) => x + y}.
sortBy {case (key, value) => -value}.
map { case (key, value) => Array(key, value).mkString(",") }
 
counts.saveAsTextFile(file)
 
merge(file, destinationFile)

现在我们两全其美：

$ cat /tmp/singlePrimaryTypes.csv
THEFT,859197
BATTERY,757530
NARCOTICS,489528
CRIMINAL DAMAGE,488209
BURGLARY,257310
OTHER OFFENSE,253964
ASSAULT,247386
MOTOR VEHICLE THEFT,197404
ROBBERY,157706
DECEPTIVE PRACTICE,137538
CRIMINAL TRESPASS,124974
PROSTITUTION,47245
WEAPONS VIOLATION,40361
PUBLIC PEACE VIOLATION,31585
OFFENSE INVOLVING CHILDREN,26524
CRIM SEXUAL ASSAULT,14788
SEX OFFENSE,14283
GAMBLING,10632
LIQUOR LAW VIOLATION,8847
ARSON,6443
INTERFERE WITH PUBLIC OFFICER,5178
HOMICIDE,4846
KIDNAPPING,3585
INTERFERENCE WITH PUBLIC OFFICER,3147
INTIMIDATION,2471
STALKING,1985
OFFENSES INVOLVING CHILDREN,355
OBSCENITY,219
PUBLIC INDECENCY,86
OTHER NARCOTIC VIOLATION,80
RITUALISM,12
NON-CRIMINAL,12
OTHER OFFENSE ,6
NON - CRIMINAL,2
NON-CRIMINAL (SUBJECT SPECIFIED),2

• 如果您想使用完整的代码，则可以根据其要点进行操作。

翻译自: https://www.javacodegeeks.com/2014/12/spark-write-to-csv-file.html