文档章节

Spark Learning

王桥修道院副院长
 王桥修道院副院长
发布于 2017/09/06 15:41
字数 697
阅读 20
收藏 0

offcially manual : https://spark.apache.org/docs/latest/rdd-programming-guide.html

一: Spark versus Hadoop

Spark is faster than Hadoop cause hadoop execute disk io to retain failure tolerant function,whereas Spark through its functional programming .

二 :spark RDD(resilient distributed datasets)

TRANSFORMATION: LAZY to execute,like filter(),map(),flatMap() and so forth,spark could optimize chain operations,never execute intermediate process.

ACTION: EAGER to execute. like count(), foreach()  countByKey and so forth

三:spark job execution

四:COMMON USED API

transformation: groupBy   groupByKey  reduceBy reduceByKey  mapValues keys不会立即计算结果(lazy)

WikipediaRanking assignment:使用inverted index配合reduceByKey排序 比传统的遍历行查找内容aggregate速度快上一倍

package wikipedia

import java.util.stream.Collectors

import org.apache.spark.SparkConf
import org.apache.spark.SparkContext
import org.apache.spark.SparkContext._
import org.apache.spark.rdd.RDD
import org.apache.spark.storage.StorageLevel

case class WikipediaArticle(title: String, text: String) {
  /**
    * @return Whether the text of this article mentions `lang` or not
    * @param lang Language to look for (e.g. "Scala")
    */
  def mentionsLanguage(lang: String): Boolean = text.split(' ').contains(lang)
}

object WikipediaRanking {

  val langs = List(
    "JavaScript", "Java", "PHP", "Python", "C#", "C++", "Ruby", "CSS",
    "Objective-C", "Perl", "Scala", "Haskell", "MATLAB", "Clojure", "Groovy")

  val conf: SparkConf = new SparkConf().setAppName("Spark RDD").setMaster("local[*]").set("spark.executor.memory", "2g")
  val sc: SparkContext = new SparkContext(conf)

  // Hint: use a combination of `sc.textFile`, `WikipediaData.filePath` and `WikipediaData.parse`
  val wikiRdd: RDD[WikipediaArticle] = sc.textFile(WikipediaData.filePath).flatMap(lines => lines.split("\n")).map(x=>WikipediaData.parse(x))
  /** Returns the number of articles on which the language `lang` occurs.
   *  Hint1: consider using method `aggregate` on RDD[T].
   *  Hint2: consider using method `mentionsLanguage` on `WikipediaArticle`
   */
  def occurrencesOfLang(lang: String, rdd: RDD[WikipediaArticle]): Int = rdd.aggregate(0)((acc,article)=>
    if(article.mentionsLanguage(lang)) acc+1 else acc,(acc1, acc2) => (acc1 + acc2))

  /* (1) Use `occurrencesOfLang` to compute the ranking of the languages
   *     (`val langs`) by determining the number of Wikipedia articles that
   *     mention each language at least once. Don't forget to sort the
   *     languages by their occurrence, in decreasing order!
   *
   *   Note: this operation is long-running. It can potentially run for
   *   several seconds.
   */
  //Result��List(("Scala", 999999), ("JavaScript", 1278), ("LOLCODE", 982), ("Java", 42))
  def rankLangs(langs: List[String], rdd: RDD[WikipediaArticle]): List[(String, Int)] =
    langs.map((lang)=>(lang,occurrencesOfLang(lang,rdd))).sortWith((x,y)=>x._2>y._2)



  /* Compute an inverted index of the set of articles, mapping each language
   * to the Wikipedia pages in which it occurs.
   */
  def makeIndex(langs: List[String], rdd: RDD[WikipediaArticle]): RDD[(String, Iterable[WikipediaArticle])] = {
    rdd.map((w)=>(w,langs.filter((o)=>w.mentionsLanguage(o)).toList)).map(x=>x._2.map((ls)=>(ls,x._1))).flatMap(x=>x).groupByKey()
  }




  /* (2) Compute the language ranking again, but now using the inverted index. Can you notice
   *     a performance improvement?
   *
   *   Note: this operation is long-running. It can potentially run for
   *   several seconds.
   */
  def rankLangsUsingIndex(index: RDD[(String, Iterable[WikipediaArticle])]): List[(String, Int)] =
  index.map((o)=>(o._1,o._2.size)).sortBy(_._2,false).collect().toList

  /* (3) Use `reduceByKey` so that the computation of the index and the ranking are combined.
   *     Can you notice an improvement in performance compared to measuring *both* the computation of the index
   *     and the computation of the ranking? If so, can you think of a reason?
   *
   *   Note: this operation is long-running. It can potentially run for
   *   several seconds.
   */
  def rankLangsReduceByKey(langs: List[String], rdd: RDD[WikipediaArticle]): List[(String, Int)] =
    rdd.map((w)=>(w,langs.filter((o)=>w.mentionsLanguage(o)).toList)).map(x=>x._2.map((ls)=>(ls,x._1))).flatMap(x=>x).map((m)=>(m._1,1)).reduceByKey(_+_).sortBy(_._2,false).collect().toList


  def main(args: Array[String]) {

    /* Languages ranked according to (1) */
    val langsRanked: List[(String, Int)] = timed("Part 1: naive ranking", rankLangs(langs, wikiRdd))

    /* An inverted index mapping languages to wikipedia pages on which they appear */
    def index: RDD[(String, Iterable[WikipediaArticle])] = makeIndex(langs, wikiRdd)

    /* Languages ranked according to (2), using the inverted index */
    val langsRanked2: List[(String, Int)] = timed("Part 2: ranking using inverted index", rankLangsUsingIndex(index))

    /* Languages ranked according to (3) */
    val langsRanked3: List[(String, Int)] = timed("Part 3: ranking using reduceByKey", rankLangsReduceByKey(langs, wikiRdd))

    /* Output the speed of each ranking */
    println(timing)
    sc.stop()
  }

  val timing = new StringBuffer
  def timed[T](label: String, code: => T): T = {
    val start = System.currentTimeMillis()
    val result = code
    val stop = System.currentTimeMillis()
    timing.append(s"Processing $label took ${stop - start} ms.\n")
    result
  }
}

pair RDDs

© 著作权归作者所有

共有 人打赏支持
王桥修道院副院长
粉丝 14
博文 46
码字总数 36605
作品 0
宁波
CTO(技术副总裁)
Spark的39个机器学习库-英文

Apache Spark itself 1. MLlib AMPLab Spark originally came out of Berkeley AMPLab and even today AMPLab projects, even though they are not in Apache Spark Foundation, enjoy a sta......

MoksMo
2015/11/04
0
1
你不能错过的 spark 学习资源

1. 书籍,在线文档 2. 网站 3. Databricks Blog 4. 文章,博客 5. 视频

u012608836
04/12
0
0
在 Databricks 可获得 Spark 1.5 预览版

我们兴奋地宣布,从今天开始,Apache Spark1.5.0的预览数据砖是可用的。我们的用户现在可以选择提供集群与Spark 1.5或先前的火花版本准备好几个点击。 正式,Spark 1.5预计将在数周内公布,和社区...

stark_summer
2015/08/25
0
0
A Vision for Making Deep Learning Simple

A Vision for Making Deep Learning Simple When MapReduce was introduced 15 years ago, it showed the world a glimpse into the future. For the first time, engineers at Silicon Vall......

openthings
05/17
0
0
利用Knime建立Spark Machine learning 模型 1:开发环境搭建

1、Knime Analytics 安装 从官方网站下载合适的版本 https://www.knime.com/downloads 将下载的安装包在安装路径解压 https://www.knime.com/installation-0 下图是knime启动后的欢迎页面...

forestwater
05/09
0
0

没有更多内容

加载失败,请刷新页面

加载更多

Redis开发规范

一、键值设计 1.1 key名设计 (1)【建议】: 可读性和可管理性 以业务名(或数据库名)为前缀(防止key冲突),用冒号分隔,比如业务名:表名:id ugc:video:1 (2)【建议】:简洁性 保证语义的前提下...

IT--小哥
15分钟前
0
0
crunch 練習

crunch 10 10 -t 09%%%%%%%% -o tw_mobile_number.lst

BaiyuanLab
30分钟前
0
0
Kafka技术资料总结(不断更新中)

1、Kafka实践:到底该不该把不同类型的消息放在同一个主题中 2、Kafka剖析系列: Kafka剖析(一):Kafka背景及架构介绍 Kafka设计解析(二):Kafka High Availability (上) Kafka设计解析...

九州暮云
今天
2
0
面向对象设计原则(OOP)

单一职责原则(Single responsibility principle)又称单一功能原则。它规定一个类应该只有一个发生变化的原因。 核心原则:低耦合,高内聚。 一个类,应该只有一个引起它变化的原因,也就是...

gackey
今天
5
0
C++ 锁

C++提供了两种常用的锁,std::lock_guard<Lockable &T>和std::unique_lock<Lockable &T>。通常使用场景下,这两个锁用法一致。即,在构造锁对象时上锁,在析构锁对象时解锁。使用户从上锁/解...

yepanl
今天
4
0

没有更多内容

加载失败,请刷新页面

加载更多

返回顶部
顶部