【大数据学习 | Spark-Core】Spark中的join原理

join是两个结果集之间的链接，需要进行数据的匹配。

演示一下join是否存在shuffle。

1. 如果两个rdd没有分区器，分区个数一致

，会发生shuffle。但分区数量不变。

scala> val arr = Array(("zhangsan",300),("lisi",400),("wangwu",350),("zhaosi",450))
arr: Array[(String, Int)] = Array((zhangsan,300), (lisi,400), (wangwu,350), (zhaosi,450))scala> val arr1 = Array(("zhangsan",22),("lisi",24),("wangwu",30),("guangkun",5))
arr1: Array[(String, Int)] = Array((zhangsan,22), (lisi,24), (wangwu,30), (guangkun,5))scala> sc.makeRDD(arr,3)
res116: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[108] at makeRDD at <console>:27scala> sc.makeRDD(arr1,3)
res117: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[109] at makeRDD at <console>:27scala> res116 join res117
res118: org.apache.spark.rdd.RDD[(String, (Int, Int))] = MapPartitionsRDD[112] at join at <console>:28scala> res118.collect
res119: Array[(String, (Int, Int))] = Array((zhangsan,(300,22)), (wangwu,(350,30)), (lisi,(400,24)))

2. 如果分区个数不一致，有shuffle，且产生的rdd的分区个数以多的为主。

3. 如果分区个数一样并且分区器一样，那么是没有shuffle的

scala> sc.makeRDD(arr,3)
res128: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[118] at makeRDD at <console>:27scala> sc.makeRDD(arr1,3)
res129: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[119] at makeRDD at <console>:27scala> res128.reduceByKey(_+_)
res130: org.apache.spark.rdd.RDD[(String, Int)] = ShuffledRDD[120] at reduceByKey at <console>:26scala> res129.reduceByKey(_+_)
res131: org.apache.spark.rdd.RDD[(String, Int)] = ShuffledRDD[121] at reduceByKey at <console>:26scala> res130 join res131
res132: org.apache.spark.rdd.RDD[(String, (Int, Int))] = MapPartitionsRDD[124] at join at <console>:28scala> res132.collect
res133: Array[(String, (Int, Int))] = Array((zhangsan,(300,22)), (wangwu,(350,30)), (lisi,(400,24)))scala> res132.partitions.size
res134: Int = 3

4. 都存在分区器但是分区个数不同，也会存在shuffle

scala> val arr = Array(("zhangsan",300),("lisi",400),("wangwu",350),("zhaosi",450))
arr: Array[(String, Int)] = Array((zhangsan,300), (lisi,400), (wangwu,350), (zhaosi,450))scala>  val arr1 = Array(("zhangsan",22),("lisi",24),("wangwu",30),("guangkun",5))
arr1: Array[(String, Int)] = Array((zhangsan,22), (lisi,24), (wangwu,30), (guangkun,5))scala> sc.makeRDD(arr,3)
res0: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[0] at makeRDD at <console>:27scala> sc.makeRDD(arr1,4)
res1: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[1] at makeRDD at <console>:27scala> res0.reduceByKey(_+_)
res2: org.apache.spark.rdd.RDD[(String, Int)] = ShuffledRDD[2] at reduceByKey at <console>:26scala> res1.reduceByKey(_+_)
res3: org.apache.spark.rdd.RDD[(String, Int)] = ShuffledRDD[3] at reduceByKey at <console>:26scala> res2 join res3
res4: org.apache.spark.rdd.RDD[(String, (Int, Int))] = MapPartitionsRDD[6] at join at <console>:28scala> res4.collect
res5: Array[(String, (Int, Int))] = Array((zhangsan,(300,22)), (wangwu,(350,30)), (lisi,(400,24)))scala> res4.partitions.size
res6: Int = 4

这里为啥stage3里reduceByKey和join过程是连在一起的，因为分区多的RDD是不需要进行shuffle的，数据该在哪个分区就在哪个分区，反而是分区少的RDD要进行join，要进行数据的打散。

分区以多的为主。

5. 一个带有分区器一个没有分区器，那么以带有分区器的rdd分区数量为主，并且存在shuffle

scala> arr
res7: Array[(String, Int)] = Array((zhangsan,300), (lisi,400), (wangwu,350), (zhaosi,450))scala> arr1
res8: Array[(String, Int)] = Array((zhangsan,22), (lisi,24), (wangwu,30), (guangkun,5))scala> sc.makeRDD(arr,3)
res9: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[7] at makeRDD at <console>:27scala> sc.makeRDD(arr,4)
res10: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[8] at makeRDD at <console>:27scala> res9.reduceByKey(_+_)
res11: org.apache.spark.rdd.RDD[(String, Int)] = ShuffledRDD[9] at reduceByKey at <console>:26scala> res10 join res11
res12: org.apache.spark.rdd.RDD[(String, (Int, Int))] = MapPartitionsRDD[12] at join at <console>:28scala> res12.partitions.size
res13: Int = 3scala> res12.collect
res14: Array[(String, (Int, Int))] = Array((zhangsan,(300,300)), (wangwu,(350,350)), (lisi,(400,400)), (zhaosi,(450,450)))

同理，stage6的reduceByKey过程和join过程是连在一起的，是因为有分区器的RDD并不需要进行shuffle操作，原来的数据该在哪在哪，而没有分区器的RDD要进行join要进行数据的打散，有shuffle过程，所以有stage4到stage6的连线。