Spark-RDD高级算子

Spark课堂笔记

Spark生态圈:
Spark Core : RDD(弹性分布式数据集)
Spark SQL
Spark Streaming
Spark MLLib:协同过滤,ALS,逻辑回归等等 --> 机器学习
Spark Graphx : 图计算

重点在前三章

-----------------Spark Core------------------------
一、什么是Spark?特点?
https://spark.apache.org/
Apache Spark™ is a unified analytics engine for large-scale data processing.

特点:快、易用、通用性、兼容性(完全兼容Hadoop)

快:快100倍(Hadoop 3 之前)
易用:支持多种语言开发
通用性:生态系统全。
易用性:兼容Hadoop

spark 取代 Hadoop

二、安装和部署Spark、Spark 的 HA

1、spark体系结构
Spark的运行方式

Yarn

Standalone:本机调试(demo)

Worker:从节点。每个服务器上,资源和任务的管理者。只负责管理一个节点。

执行过程:
一个Worker 有多个 Executor。 Executor是任务的执行者,按阶段(stage)划分任务。————> RDD

客户端:Driver Program 提交任务到集群中。

1、spark-submit
2、spark-shell

2、spark的搭建
(1)准备工作:JDK 配置主机名 免密码登录
(2)伪分布式模式
在一台虚拟机上模拟分布式环境(Master和Worker在一个节点上)

export JAVA_HOME=/usr/java/jdk1.8.0_201
export SPARK_MASTER_HOST=node3
export SPARK_MASTER_PORT=7077

(3)全分布式环境
修改slave文件 拷贝到其他两台服务器 启动

3、Spark的 HA
回顾HA;
(*)HDFS Yarn Hbase Spark 主从结构
(*)单点故障

(1)基于文件目录的单点恢复
(*)本质:还是只有一个主节点Master,创建了一个恢复目录,保存集群状态和任务的信息。
当Master挂掉,重新启动时,会从恢复目录下读取状态信息,恢复出来原来的状态

用途:用于开发和测试,生产用zookeeper
export SPARK_DAEMON_JAVA_OPTS="-Dspark.deploy.recoveryMode=FILESYSTEM
-Dspark.deploy.recoveryDirectory=/usr/local/spark-2.1.0-bin-hadoop2.7/recovery"

(2)基于Zookeeper :和Hadoop类似

(*)复习一下zookeeper:
相当于一个数据库,把一些信息存放在zookeeper中,比如集群的信息。
数据同步功能,选举功能,分布式锁功能

数据同步:给一个节点中写入数据,可以同步到其他节点

选举:Zookeeper中存在不同的角色,Leader Follower。如果Leader挂掉,重新选举Leader

分布式锁:秒杀。以目录节点的方式来保存数据。

修改 spark-env.sh
export SPARK_DAEMON_JAVA_OPTS="-Dspark.deploy.recoveryMode=ZOOKEEPER
-Dspark.deploy.zookeeper.url=node3:2181,node4:2181,node5:2181
-Dspark.deploy.zookeeper.dir=/spark"

同步到其他两台服务器。

在node3 start-all node3 master node4 Worker node5 Worker
在node4 start-master node3 master node4 master(standby) node4 Worker node5 Worker

在node3上kill master
node4 master(Active) node4 Worker node5 Worker

在网页http://192.168.109.134:8080/ 可以看到相应信息

三、执行Spark的任务:两个工具

1、spark-submit:用于提交Spark的任务
任务:jar。

举例:蒙特卡洛求PI(圆周率)。

./spark-submit --master spark://node3:7077 --class

--class指明主程序的名字

/usr/local/spark-2.1.0-bin-hadoop2.7/bin/spark-submit --master spark://node3:7077
--class org.apache.spark.examples.SparkPi
/usr/local/spark-2.1.0-bin-hadoop2.7/examples/jars/spark-examples_2.11-2.1.0.jar 100

2、spark-shell 相当于REPL
作为一个独立的Application运行

两种模式:
(1)本地模式
spark-shell 后面不接任何参数,代表本地模式

Spark context available as 'sc' (master = local[*], app id = local-1554038459298).

sc 是 SparkContext 对象名。 local[*] 代表本地模式,不提交到集群中运行。

(2)集群模式
./spark-submit --master spark://node3:7077 提交到集群中运行

Spark context available as 'sc' (master = spark://node3:7077, app id = app-20190331212447-0000).

master = spark://node3:7077

Spark session available as 'spark'
Spark Session 是 2.0 以后提供的,利用 SparkSession 可以访问spark所有组件。

示例:WordCount程序

(*)处理本地文件,把结果打印到屏幕上
scala> sc.textFile("/usr/local/tmp_files/test_WordCount.txt")
.flatMap(_.split(" "))
.map((_,1))
.reduceByKey(_+_)
.collect

res0: Array[(String, Int)] = Array((is,1), (love,2), (capital,1), (Beijing,2), (China,2), (I,2), (of,1), (the,1))

(*)处理HDFS文件,结果保存在hdfs上
sc.textFile("hdfs://node1:8020/tmp_files/test_WordCount.txt")
.flatMap(_.split(" "))
.map((_,1))
.reduceByKey(_+_)
.saveAsTextFile("hdfs://node1:8020/output/0331/test_WordCount")

-rw-r--r-- 3 root supergroup 0 2019-03-31 21:43 /output/0331/test_WordCount/_SUCCESS
-rw-r--r-- 3 root supergroup 40 2019-03-31 21:43 /output/0331/test_WordCount/part-00000
-rw-r--r-- 3 root supergroup 31 2019-03-31 21:43 /output/0331/test_WordCount/part-00001

_SUCCESS 代表程序执行成功

part-00000 part-00001 结果文件,分区。里面内容不重复。

(*)单步运行WordCount ----> RDD

scala> val rdd1 = sc.textFile("/usr/local/tmp_files/test_WordCount.txt")
rdd1: org.apache.spark.rdd.RDD[String] = /usr/local/tmp_files/test_WordCount.txt MapPartitionsRDD[12] at textFile at <console>:24

scala> 1+1
res2: Int = 2

scala> rdd1.collect
res3: Array[String] = Array(I love Beijing, I love China, Beijing is the capital of China)

scala> val rdd2 = rdd1.flatMap(_.split(" "))
rdd2: org.apache.spark.rdd.RDD[String] = MapPartitionsRDD[13] at flatMap at <console>:26

scala> rdd2.collect
res4: Array[String] = Array(I, love, Beijing, I, love, China, Beijing, is, the, capital, of, China)

scala> val rdd3 = rdd2.map((_,1))
rdd3: org.apache.spark.rdd.RDD[(String, Int)] = MapPartitionsRDD[14] at map at <console>:28

scala> rdd3.collect
res5: Array[(String, Int)] = Array((I,1), (love,1), (Beijing,1), (I,1), (love,1), (China,1), (Beijing,1), (is,1), (the,1), (capital,1), (of,1), (China,1))

scala> val rdd4 = rdd3.reduceByKey(_+_)
rdd4: org.apache.spark.rdd.RDD[(String, Int)] = ShuffledRDD[15] at reduceByKey at <console>:30

scala> rdd4.collect
res6: Array[(String, Int)] = Array((is,1), (love,2), (capital,1), (Beijing,2), (China,2), (I,2), (of,1), (the,1))

RDD 弹性分布式数据集
(1)依赖关系 : 宽依赖和窄依赖
(2)算子:
函数:
Transformation : 延时计算 map flatMap textFile
Action : 立即触发计算 collect




说明:scala复习
(*)flatten:把嵌套的结果展开
scala> List(List(2,4,6,8,10),List(1,3,5,7,9)).flatten
res21: List[Int] = List(2, 4, 6, 8, 10, 1, 3, 5, 7, 9)


(*)flatmap : 相当于一个 map + flatten

scala> var myList = List(List(2,4,6,8,10),List(1,3,5,7,9))
myList: List[List[Int]] = List(List(2, 4, 6, 8, 10), List(1, 3, 5, 7, 9))

scala> myList.flatMap(x=>x.map(_*2))
res22: List[Int] = List(4, 8, 12, 16, 20, 2, 6, 10, 14, 18)

myList.flatMap(x=>x.map(_*2))

执行过程:
1、将 List(2, 4, 6, 8, 10), List(1, 3, 5, 7, 9) 调用 map(_*2) 方法。x 代表一个List
2、flatten
3、在IDE中开发scala版本和Java版本的WorkCount。

(1)scala版本的WordCount

新建一个工程,把jar引入到工程中。

export jar 点击下一步下一步,不需要设置main class

把jar上传到服务器上。

spark-submit --master spark://node3:7077
--class day1025.MyWordCount
/usr/local/tmp_files/Demo1.jar
hdfs://node2:8020/tmp_files/test_WordCount.txt
hdfs://node2:8020/output/1025/demo1

(2)java版本的WordCount

./spark-submit --master spark://node3:7077 --class day0330.JavaWordCount /usr/local/tmp_files/Demo2.jar

四、分析Spark的任务流程

1、分析WordCount程序处理过程
见图片

2、Spark调度任务的过程

提交到及群众运行任务时,spark执行任务调度。

见图片

五、RDD和RDD特性、RDD的算子

1、RDD:弹性分布式数据集
(*)Spark中最基本的数据抽象。
(*)RDD的特性
* Internally, each RDD is characterized by five main properties:
*
* - A list of partitions

1、是一组分区。
RDD由分区组成,每个分区运行在不同的Worker上,通过这种方式来实现分布式计算。

* - A function for computing each split
在RDD中,提供算子处理每个分区中的数据

* - A list of dependencies on other RDDs

RDD存在依赖关系:宽依赖和窄依赖。

* - Optionally, a Partitioner for key-value RDDs (e.g. to say that the RDD is hash-partitioned)

可以自定义分区规则来创建RDD

* - Optionally, a list of preferred locations to compute each split on (e.g. block locations for
* an HDFS file)

优先选择离文件位置近的节点来执行


如何创建RDD?

(1)通过SparkContext.parallelize方法来创建
scala> val rdd1 = sc.parallelize(Array(1,2,3,4,5,6,7,8),3)
rdd1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[32] at parallelize at <console>:29

scala> rdd1.partitions.length
res35: Int = 3

scala> val rdd1 = sc.parallelize(Array(1,2,3,4,5,6,7,8),2)
rdd1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[33] at parallelize at <console>:29

scala> rdd1.partitions.length
res36: Int = 2

(2)通过外部数据源来创建
sc.textFile()

scala> val rdd2 = sc.textFile("/usr/local/tmp_files/test_WordCount.txt")
rdd2: org.apache.spark.rdd.RDD[String] = /usr/local/tmp_files/test_WordCount.txt MapPartitionsRDD[35] at textFile at <console>:29


2、 算子
(1)Transformation

map(func):相当于for循环,返回一个新的RDD

filter(func):过滤
flatMap(func):flat+map 压平


mapPartitions(func):对RDD中的每个分区进行操作
mapPartitionsWithIndex(func):对RDD中的每个分区进行操作,可以取到分区号。

sample(withReplacement, fraction, seed):采样

集合运算
union(otherDataset)
intersection(otherDataset)

distinct([numTasks])):去重

聚合操作:group by
groupByKey([numTasks])
reduceByKey(func, [numTasks])
aggregateByKey(zeroValue)(seqOp,combOp,[numTasks])

排序
sortByKey([ascending], [numTasks])
sortBy(func,[ascending], [numTasks])

join(otherDataset, [numTasks])
cogroup(otherDataset, [numTasks])
cartesian(otherDataset)
pipe(command, [envVars])
coalesce(numPartitions)

重分区:
repartition(numPartitions)
repartitionAndSortWithinPartitions(partitioner)

举例:
1、创建一个RDD,每个元素乘以2,再排序
scala> val rdd1 = sc.parallelize(Array(3,4,5,100,79,81,6,8))
rdd1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[36] at parallelize at <console>:29

scala> val rdd2 = rdd1.map(_*2)
rdd2: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[37] at map at <console>:31

scala> rdd2.collect
res37: Array[Int] = Array(6, 8, 10, 200, 158, 162, 12, 16)

scala> rdd2.sortBy(x=>x,true).collect
res39: Array[Int] = Array(6, 8, 10, 12, 16, 158, 162, 200)

scala> rdd2.sortBy(x=>x,false).collect
res40: Array[Int] = Array(200, 162, 158, 16, 12, 10, 8, 6)

def sortBy[K](f: (T) ⇒ K, ascending: Boolean = true)

过滤出大于20的元素:

scala> val rdd3 = rdd2.filter(_>20)
rdd3: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[53] at filter at <console>:33

scala> rdd3.collect
res41: Array[Int] = Array(200, 158, 162)

2、字符串(字符)类型的RDD

scala> val rdd4 = sc.parallelize(Array("a b c","d e f","g h i"))
rdd4: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[54] at parallelize at <console>:29

scala> rdd4.flatMap(_.split(" ")).collect
res42: Array[String] = Array(a, b, c, d, e, f, g, h, i)

3、RDD的集合运算:

scala> val rdd6 = sc.parallelize(List(1,2,3,6,7,8,9,100))
rdd6: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[56] at parallelize at <console>:29

scala> val rdd7 = sc.parallelize(List(1,2,3,4))
rdd7: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[57] at parallelize at <console>:29

scala> val rdd8 = rdd6.union(rdd7)
rdd8: org.apache.spark.rdd.RDD[Int] = UnionRDD[58] at union at <console>:33

scala> rdd8.collect
res43: Array[Int] = Array(1, 2, 3, 6, 7, 8, 9, 100, 1, 2, 3, 4)

scala> rdd8.distinct.collect
res44: Array[Int] = Array(100, 4, 8, 1, 9, 6, 2, 3, 7)


4、分组操作:reduceByKey

<key value>
scala> val rdd1 = sc.parallelize(List(("Tom",1000),("Andy",2000),("Lily",1500)))
rdd1: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[62] at parallelize at <console>:29

scala> val rdd2 = sc.parallelize(List(("Andy",1000),("Tom",2000),("Mike",500)))
rdd2: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[63] at parallelize at <console>:29

scala> val rdd3 = rdd1 union rdd2
rdd3: org.apache.spark.rdd.RDD[(String, Int)] = UnionRDD[64] at union at <console>:33

scala> rdd3.collect
res45: Array[(String, Int)] = Array((Tom,1000), (Andy,2000), (Lily,1500), (Andy,1000), (Tom,2000), (Mike,500))

scala> val rdd4= rdd3.groupByKey
rdd4: org.apache.spark.rdd.RDD[(String, Iterable[Int])] = ShuffledRDD[65] at groupByKey at <console>:35

scala> rdd4.collect
res46: Array[(String, Iterable[Int])] = Array(
(Tom,CompactBuffer(1000, 2000)),
(Andy,CompactBuffer(2000, 1000)),
(Mike,CompactBuffer(500)), (
Lily,CompactBuffer(1500)))

scala> rdd3.reduceByKey(_+_).collect
res47: Array[(String, Int)] = Array((Tom,3000), (Andy,3000), (Mike,500), (Lily,1500))

reduceByKey will provide much better performance.

官方不推荐使用 groupByKey 推荐使用 reduceByKey

5、cogroup

scala> val rdd1 = sc.parallelize(List(("tom", 1), ("tom", 2), ("jerry", 3), ("kitty", 2)))
rdd1: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[67] at parallelize at <console>:29

scala> val rdd2 = sc.parallelize(List(("jerry", 2), ("tom", 1), ("shuke", 2)))
rdd2: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[68] at parallelize at <console>:29

scala> val rdd3 = rdd1.cogroup(rdd2)
rdd3: org.apache.spark.rdd.RDD[(String, (Iterable[Int], Iterable[Int]))] = MapPartitionsRDD[70] at cogroup at <console>:33

scala> rdd3.collect
res48: Array[(String, (Iterable[Int], Iterable[Int]))] = Array(
(tom,(CompactBuffer(1, 2),CompactBuffer(1))),
(jerry,(CompactBuffer(3),CompactBuffer(2))),
(shuke,(CompactBuffer(),CompactBuffer(2))),
(kitty,(CompactBuffer(2),CompactBuffer())))

6、reduce操作(Action)

聚合操作
scala> val rdd1 = sc.parallelize(List(1,2,3,4,5))
rdd1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[71] at parallelize at <console>:29

scala> rdd1.reduce(_+_)
res49: Int = 15


7、需求:按照value排序。
做法:
1、交换,把key 和 value交换,然后调用sortByKey方法
2、再次交换

scala> val rdd1 = sc.parallelize(List(("tom",1),("jerry",3),("ketty",2),("shuke",2)))
rdd1: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[72] at parallelize at <console>:29

scala> val rdd2 = sc.parallelize(List(("jerry",1),("tom",3),("shuke",5),("ketty",1)))
rdd2: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[73] at parallelize at <console>:29

scala> val rdd3 = rdd1.union(rdd2)
rdd3: org.apache.spark.rdd.RDD[(String, Int)] = UnionRDD[74] at union at <console>:33

scala> val rdd4 = rdd3.reduceByKey(_+_)
rdd4: org.apache.spark.rdd.RDD[(String, Int)] = ShuffledRDD[75] at reduceByKey at <console>:35

scala> rdd4.collect
res50: Array[(String, Int)] = Array((tom,4), (jerry,4), (shuke,7), (ketty,3))

scala> val rdd5 = rdd4.map(t=>(t._2,t._1)).sortByKey(false).map(t=>(t._2,t._1))
rdd5: org.apache.spark.rdd.RDD[(String, Int)] = MapPartitionsRDD[80] at map at <console>:37

scala> rdd5.collect
res51: Array[(String, Int)] = Array((shuke,7), (tom,4), (jerry,4), (ketty,3))


(2)Action

reduce(func)

collect()
count()
first()
take(n)
takeSample(withReplacement,num, [seed])
takeOrdered(n, [ordering])
saveAsTextFile(path)
saveAsSequenceFile(path) 
saveAsObjectFile(path) 
countByKey()

foreach(func):与map类似,没有返回值。

3、特性:
(1)RDD的缓存机制
(*)作用:提高性能
(*)使用:标识RDD可以被缓存 persist cache

(*)可以缓存的位置:
val NONE = new StorageLevel(false, false, false, false)
val DISK_ONLY = new StorageLevel(true, false, false, false)
val DISK_ONLY_2 = new StorageLevel(true, false, false, false, 2)
val MEMORY_ONLY = new StorageLevel(false, true, false, true)
val MEMORY_ONLY_2 = new StorageLevel(false, true, false, true, 2)
val MEMORY_ONLY_SER = new StorageLevel(false, true, false, false)
val MEMORY_ONLY_SER_2 = new StorageLevel(false, true, false, false, 2)
val MEMORY_AND_DISK = new StorageLevel(true, true, false, true)
val MEMORY_AND_DISK_2 = new StorageLevel(true, true, false, true, 2)
val MEMORY_AND_DISK_SER = new StorageLevel(true, true, false, false)
val MEMORY_AND_DISK_SER_2 = new StorageLevel(true, true, false, false, 2)
val OFF_HEAP = new StorageLevel(true, true, true, false, 1)

/**
* Persist this RDD with the default storage level (`MEMORY_ONLY`).
*/
def persist(): this.type = persist(StorageLevel.MEMORY_ONLY)

/**
* Persist this RDD with the default storage level (`MEMORY_ONLY`).
*/
def cache(): this.type = persist()


举例:测试数据,92万条

scala> val rdd1 = sc.textFile("hdfs://192.168.109.131:8020/tmp_files/test_Cache.txt")
rdd1: org.apache.spark.rdd.RDD[String] = hdfs://192.168.109.131:8020/tmp_files/test_Cache.txt MapPartitionsRDD[82] at textFile at <console>:29

scala> rdd1.count --> 直接出发计算
res52: Long = 923452

scala> rdd1.cache --> 标识RDD可以被缓存,不会触发计算
res53: rdd1.type = hdfs://192.168.109.131:8020/tmp_files/test_Cache.txt MapPartitionsRDD[82] at textFile at <console>:29

scala> rdd1.count --> 和第一步一样,触发计算,但是,把结果进行缓存
res54: Long = 923452

scala> rdd1.count --> 从缓存中直接读出结果
res55: Long = 923452

(2)RDD的容错机制:通过检查点来实现。

/**
* Mark this RDD for checkpointing. It will be saved to a file inside the checkpoint
* directory set with `SparkContext#setCheckpointDir` and all references to its parent
* RDDs will be removed. This function must be called before any job has been
* executed on this RDD. It is strongly recommended that this RDD is persisted in
* memory, otherwise saving it on a file will require recomputation.
*/
(*)复习检查点:HDFS中的检查点:有SecondaryNamenode来实现日志的合并。

(*)RDD的检查点:容错
概念:血统 Lineage
理解:表示任务执行的生命周期。
WordCount textFile ---> redceByKey

如果血统越长,越容易出错。

假如有检查点,可以从最近的一个检查点开始,往后面计算。不用重头计算。

(*)RDD检查点的类型:

(1)基于本地目录:需要将Spark shell 或者任务运行在本地模式上(setMaster("local"))

开发和测试

(2)HDFS目录:用于生产。

sc.setCheckPointDir(目录)

举例:设置检查点
scala> var rdd1 = sc.textFile("hdfs://192.168.109.131:8020/tmp_files/test_Cache.txt")
rdd1: org.apache.spark.rdd.RDD[String] = hdfs://192.168.109.131:8020/tmp_files/test_Cache.txt MapPartitionsRDD[1] at textFile at <console>:24

设置检查点目录:
scala> sc.setCheckpointDir("hdfs://192.168.109.131:8020/sparkckpt")

标识rdd1可以执行检查点操作
scala> rdd1.checkpoint

scala> rdd1.count
res2: Long = 923452

(3)依赖关系:宽依赖,窄依赖。
划分任务执行的stage

见讲义。

六、RDD的高级算子
1、mapPartitionsWithIndex:对RDD中的每个分区(带有下标)进行操作,下标用index表示
通过这个算子,我们可以获取分区号。

def mapPartitionsWithIndex[U](
f: (Int, Iterator[T]) ⇒ Iterator[U],
preservesPartitioning: Boolean = false)(
implicit arg0: ClassTag[U]): RDD[U]

参数:f是个函数参数 f 中第一个参数是Int,代表分区号,第二个Iterator[T]代表分区中的元素

举例:把分区中的元素,包括分区号,都打印出来。

scala> val rdd1 = sc.parallelize(List(1,2,3,4,5,6,7,8),3)
rdd1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[3] at parallelize at <console>:24

scala> def fun1(index:Int, iter:Iterator[Int]) : Iterator[String] = {
| iter.toList.map(x => "[partId : " + index + " , value = " + x + " ]").iterator
| }
fun1: (index: Int, iter: Iterator[Int])Iterator[String]

scala> rdd1.mapPartitions
mapPartitions mapPartitionsWithIndex

scala> rdd1.mapPartitionsWithIndex(fun1).collect
res3: Array[String] = Array(
[partId : 0 , value = 1 ], [partId : 0 , value = 2 ],
[partId : 1 , value = 3 ], [partId : 1 , value = 4 ], [partId : 1 , value = 5 ],
[partId : 2 , value = 6 ], [partId : 2 , value = 7 ], [partId : 2 , value = 8 ])

2、aggregate:聚合操作。类似于分组。

(*)先对局部进行聚合操作,再对全局进行聚合操作。

调用聚合操作
scala> val rdd2 = sc.parallelize(List(1,2,3,4,5),2)
rdd2: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[5] at parallelize at <console>:24

scala> rdd2.mapPartitionsWithIndex(fun1).collect
res4: Array[String] = Array(
[partId : 0 , value = 1 ], [partId : 0 , value = 2 ],
[partId : 1 , value = 3 ], [partId : 1 , value = 4 ], [partId : 1 , value = 5 ])

scala> import scala.math._
import scala.math._

scala> rdd2.aggregate(0)(max(_,_),_+_)
res6: Int = 7

说明:aggregate
(0) 初始值是 0
(max(_,_) 局部操作的函数
, _+_ 全局操作的函数
)

scala> rdd2.aggregate(100)(max(_,_),_+_)
res8: Int = 300

分析结果:初始值是100,代表每个分区多了一个100
全局操作,也多了一个100
100+100+100 = 300


对RDD中的元素进行求和
1、RDD.map
2、聚合操作

scala> rdd2.aggregate(0)(_+_,_+_)
res9: Int = 15

MapReduce Combiner

scala> rdd2.aggregate(10)(_+_,_+_)
res10: Int = 45

(*)对字符串操作
scala> val rdd2 = sc.parallelize(List("a","b","c","d","e","f"),2)
rdd2: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[7] at parallelize at <console>:27

scala> rdd2.aggregate("")(_+_,_+_)
res11: String = abcdef

scala> rdd2.aggregate("*")(_+_,_+_)
res12: String = **def*abc

结果分析:
1、*abc *def
2、**def*abc

(*)复杂的例子:
1、
scala> val rdd3 = sc.parallelize(List("12","23","345","4567"),2)
rdd3: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[8] at parallelize at <console>:27

scala> def fun1(index:Int, iter:Iterator[String]) : Iterator[String] = {
| iter.toList.map(x => "[partId : " + index + " , value = " + x + " ]").iterator
| }

scala> rdd3.mapPartitionsWithIndex(fun1).collect
res17: Array[String] = Array(
[partId : 0 , value = 12 ], [partId : 0 , value = 23 ],
[partId : 1 , value = 345 ], [partId : 1 , value = 4567 ])

scala> rdd3.aggregate("")((x,y)=> math.max(x.length,y.length).toString,(x,y)=>x+y)
res13: String = 42

执行过程:
第一个分区:
第一次比较: "" "12" 长度最大值 2 2-->"2"
第二次比较: “2” “23” 长度最大值 2 2-->"2"

第二个分区:
第一次比较: "" "345" 长度最大值 3 3-->"3"
第二次比较: “3” “4567” 长度最大值 4 4-->"4"


2、
rdd3.aggregate("")((x,y)=> math.min(x.length,y.length).toString,(x,y)=>x+y)

scala> rdd3.aggregate("")((x,y)=> math.min(x.length,y.length).toString,(x,y)=>x+y)
res18: String = 11

执行过程:
第一个分区:
第一次比较: "" "12" 长度最小值 0 0-->"0"
第二次比较: “0” “23” 长度最小值 1 1-->"1"

第二个分区:
第一次比较: "" "345" 长度最小值 0 0-->"0"
第二次比较: “0” “4567” 长度最小值 1 1-->"1"

val rdd3 = sc.parallelize(List("12","23","345",""),2)
rdd3.aggregate("")((x,y)=> math.min(x.length,y.length).toString,(x,y)=>x+y)

scala> val rdd3 = sc.parallelize(List("12","23","345",""),2)
rdd3: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[10] at parallelize at <console>:27

scala> rdd3.aggregate("")((x,y)=> math.min(x.length,y.length).toString,(x,y)=>x+y)
res19: String = 10

scala> rdd3.aggregate("")((x,y)=> math.min(x.length,y.length).toString,(x,y)=>x+y)
res20: String = 01

3、aggregateByKey:类似于aggregate,区别:操作的是 key value 的数据类型。

scala> def fun3(index:Int, iter:Iterator[(String,Int)]) : Iterator[String] = {
| iter.toList.map(x => "[partId : " + index + " , value = " + x + " ]").iterator
| }

scala> pairRDD.mapPartitionsWithIndex(fun3).collect
res22: Array[String] = Array(

[partId : 0 , value = (cat,2) ], [partId : 0 , value = (cat,5) ], [partId : 0 , value = (mouse,4) ],


[partId : 1 , value = (cat,12) ], [partId : 1 , value = (dog,12) ], [partId : 1 , value = (mouse,2) ])

(1)将每个动物园(分区)中,动物数最多的动物,进行求和

动物园0
[partId : 0 , value = (cat,2) ], [partId : 0 , value = (cat,5) ], [partId : 0 , value = (mouse,4) ],

动物园1
[partId : 1 , value = (cat,12) ], [partId : 1 , value = (dog,12) ], [partId : 1 , value = (mouse,2) ])

pairRDD.aggregateByKey(0)(math.max(_,_),_+_)

scala> pairRDD.aggregateByKey(0)(math.max(_,_),_+_).collect
res24: Array[(String, Int)] = Array((dog,12), (cat,17), (mouse,6))

(2)将所有动物求和
pairRDD.aggregateByKey(0)(_+_,_+_).collect

scala> pairRDD.reduceByKey(_+_).collect
res27: Array[(String, Int)] = Array((dog,12), (cat,19), (mouse,6))

aggregateByKey效率更高。

4、 coalesce与repartition
与分区有关
都是对RDD进行重分区。

区别:
coalesce 默认不会进行Shuffle 默认 false 如需修改分区,需置为true
repartition 会进行Shuffle

scala> val rdd1 = sc.parallelize(List(1,2,3,4,5,6,7,8,9),2)
rdd1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[17] at parallelize at <console>:27

scala> val rdd2 = rdd1.repartition(3)
rdd2: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[21] at repartition at <console>:29

scala> rdd2.partitions.length
res28: Int = 3

scala> val rdd3 = rdd1.coalescse(3,true)
<console>:29: error: value coalescse is not a member of org.apache.spark.rdd.RDD[Int]
val rdd3 = rdd1.coalescse(3,true)
^

scala> val rdd3 = rdd1.coalesce(3,true)
rdd3: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[25] at coalesce at <console>:29

scala> rdd3.partitions.length
res29: Int = 3

scala> val rdd4 = rdd1.coalesce(4)
rdd4: org.apache.spark.rdd.RDD[Int] = CoalescedRDD[26] at coalesce at <console>:29

scala> rdd4.partitions.length
res30: Int = 2

5、其他高级算子
http://homepage.cs.latrobe.edu.au/zhe/ZhenHeSparkRDDAPIExamples.html


七、编程案例

(1)分析日志
需求:找到访问量最高的两个网页
(*)第一步:对网页的访问量求和
(*)第二步:排序,降序

(2)创建自定义分区

(3)使用JDBCRDD 操作数据库


posted @ 2019-05-01 14:03  jareny  阅读(235)  评论(0编辑  收藏  举报