MapReduce 概述
分布式计算框架
A MapReduce job consists of a number of
– Map tasks（Map task performs data transformation）
– Reduce tasks（Reduce task combines results of map tasks）
– (Internally) shuffle tasks （Shuffle task sends output of map tasks to right reduce tasks）
比较简单的流程（后续有详解）：
input data -> blocks in hdfs -> map function in parallel -> reduce -> output
MapReduce 1.x 流程（MapReduce最原始的处理流程，2.x之后使用YARN）
hadoop cluster.png
JobTracker： ===>集群资源管理与作业调度 + 与client进行通信
- Only one job tracker in a Hadoop cluster
- Takes requests from clients (MapReduce programs)
- Ask name node for location of data
- Assign tasks to task trackers near the data
- Reassign tasks if failed
TaskTracker ===>汇报心跳，一个是执行命令
- Accept (map, reduce, shuffle) tasks from job trackers
- Send heartbeats to job trackers: I am alive
- Monitor status of tasks and notify job tracker
JobTracker & TaskTracker.png
从 Map function和Recuce function 说起
python 中map和reduce的使用
https://www.runoob.com/python/python-func-map.html
map函数： map(function, iterable, ...)
>>> map(lambda x: x ** 2, [1, 2, 3, 4, 5])
[1, 4, 9, 16, 25]
# 提供了两个列表，对相同位置的列表数据进行相加
>>> map(lambda x, y: x + y, [1, 3, 5, 7, 9], [2, 4, 6, 8, 10])
[3, 7, 11, 15, 19]
reduce 函数 reduce(function, iterable[, initializer])
from functools import reduce
>>> reduce(lambda x, y: x+y, [1,2,3,4,5])
15
MapReduce 中map和reduce的使用
- map和reduce的输入输出均为<k,v>键值对
- map的输出作为reduce的输入（map输出其实还有一步group by的过程）
- System groups the intermediate key‐value pairs from map tasks by key
E.g., <hello, 1> <hello, 1> <hello, 1> <this, 1> => <hello, [1, 1, 1]>, <this, [1]>
MapReduce diagram.png
MapReduce in parallel.png
In hadoop
- A node may run multiple map/reduce tasks
- Typically, one map task per input split (chunk of data)
- One reduce task per partition of map output E.g., partition by key range or hashing
wordcount 例子：
word.txt
hello
hello world
Map function：
- Input: <offset of line, line> // line = a line of text in a document =>{0:"hello",6:"hello world"}
- Output: for each word in line, output <word, 1> => ["hello":1,"hello":1,"world":1]
Reduce function:
– Input: <word, list of 1’s> => {"hello":[1,1],"world":[1]} 所以其实reduce的输入在某种程度上来说不是map的输出
– Output: <word, count> where count is the number of 1's in the input list =>
{"hello":2,"world":1}
Map 和 Reduce 中的Java实现
对于MapReduce框架，其更像一套八股文，我们只需要编写Map和Reduce函数即可，其他的细节框架已经帮我们实现了，在Java中通过Mapper和Reducer类封装，所以我们可以编写子类继承父类然后overide相应的方法即可
Each map task runs an instance of Mapper
– Mapper has a map function
– Map task invokes the map function of the Mapper once for each input key‐value pair
Mapper.png
Each reduce task runs an instance of Reducer
– Reducer has a reduce function
– Reduce task invokes the reduce function of the Reducer once for every different intermediate key
– For the reduce function, values are NOT in any particular order
Reducer.png
Map 之后， Reduce之前的细节 --- Shuffling
Shuffle
- Process of distributing intermediate key‐values to the right reduce tasks
- It is the only communication among map and reduce tasks
- Individual map tasks do not exchange data directly with other map tasks
- They are not even aware of existence of their peer
Map可能在不同的机器上并行处理的，需要通过 shuffling 将相同 key 值的数据分发到同一个节点上去合并，这样才能统计出最终的结果，并且这一步也将一个由Map生成的<k,v> 变成<k,[v1,v2]>传给后续的Reduce来操作
Shuffle.png
图片来源：
https://xinze.fun/2020/01/10/Spark-Shuffle-%E5%92%8C-Spill-%E7%9A%84%E5%8C%BA%E5%88%AB/
Internal of shuffling
1.Map side：Partition, sort, spill & merge
- Partition data in the buffer into R parts (R = # of reduce tasks) # partition = # reduce task
- Sort data in each partition by key
- Spill/write data in the buffer to disk
- Merge the spills
- Notify job tracker: output complete
2.Reduce side： Fetch & merge
- Task tracker notified by job tracker: data ready
- Fetch/copy data from map side
- Merge data( Some data may sit on disk once fetched , this depends on the buffer size)
- Figure out groups from sorted data
details about Shuffle.png
回到开头 数据的输入输出格式
mapreduce with 2 nodes.png
输入
InputFormat
- Determine how input files are split and read
- Defined in the Java interface InputFormat
- Job:
- Split input file into chunks called InputSplits（logic concepts）
- Implement RecordReader to read data from splits
在Java中InputFormat是个抽象类，其有诸多的子类去读取特定的输入
• FileInputFormat (input from files in given dirs) --- 用的最多
• DBInputFormat (input data from a database)
• CombineFileInputFormat (input data by combining multiple files)
FileInputFormat
– Takes paths to files
– Read all files in the paths
– Divide each file into one or more InputSplits
FileInputFormat也是一个抽象类，其下有对应的子类来实现对应的内容输入
– TextInputFormat
– KeyValueTextInputFormat
– SequenceFileInputFormat
Subclasses of FileInputFormat.png
InputFormat 干的第一件事情：Split input file into InputSplits
聊一聊InputSplit，注意一点，InputSplit是一个逻辑划分的概念
• If a file is big, multiple splits may be created Typical split size = 128MB
• A map task is created for each split （a chunk of some input file）
InputFormat 干的第二件事情：Implement RecordReader to read data from InputSplits
- InputFormat defines an instanceof RR E.g., TextInputFormat provides LineRecordReader
- LineRecordReader
- Form a key‐value pair for every line of file
- Data type for key: LongWritable; value: Text
- Reader is repeatedly called until all data in the split are processed
输出
OutputFormat
- Define the format of output from Reducers（Output stored in a file）
- Defined in the Java interface OutputFormat
- Implemention: FileOutputFormat
-
Subclasses: TextOutputFormat, SequenceFileOutputFormat
Subclasses of OutputFormat.png
-
- All Reducers write to the same directory
- Set by FileOutputFormat.setOutputPath() method
- OutputFormat defines aRecordWrite which handles the write
可选操作：Combiner
个人理解：combiner 约等于 一个本地化的reduer
- Run on the node running the Mapper
- Perform local (or mini‐) reduction
- Combine Mapper results
- Before they are sent to the Reducers
- Reduce communication costs
- E.g., may use a combiner in WordCount – (cat, 1), (cat, 1), (cat, 1) => (cat, 3)
注意，使用combiner是为了节省空间成本、加快计算，但是不能对最终的结果有影响，例如说求个平均数什么的就不能用combiner
专业的说法就是：
May directly use the combiner
- If it is commutative and associative
- Meaning operations can be grouped & performed in any order
网友评论