维基百科对 MapReduce 的部分介绍

【维基中文】

MapReduce是 Google 提出的一个软件架构，用于大规模数据集（大于1TB）的并行运算。概念“Map（映射）”和“Reduce（化简）”，及他们的主要思想，都是从函数式编程语言借来的，还有从矢量编程语言借来的特性。

当前的软件实现是指定一个 Map（映射）函数，用来把一组键值对映射成一组新的键值对，指定并发的 Reduce（化简）函数，用来保证所有映射的键值对中的每一个共享相同的键组。

映射和化简

简单来说，一个映射函数就是对一些独立元素组成的概念上的列表（例如，一个测试成绩的列表）的每一个元素进行指定的操作（比如，有人发现所有学生的成绩都被高估了一分，他可以定义一个“减一”的映射函数，用来修正这个错误。）。事实上，每个元素都是被独立操作的，而原始列表没有被更改，因为这里创建了一个新的列表来保存新的答案。这就是说，Map 操作是可以高度并行的，这对高性能要求的应用以及并行计算领域的需求非常有用。

而化简操作指的是对一个列表的元素进行适当的合并（继续看前面的例子，如果有人想知道班级的平均分该怎么做？他可以定义一个化简函数，通过让列表中的奇数（odd）或偶数（even）元素跟自己的相邻的元素相加的方式把列表减半，如此递归运算直到列表只剩下一个元素，然后用这个元素除以人数，就得到了平均分）。虽然他不如映射函数那么并行，但是因为化简总是有一个简单的答案，大规模的运算相对独立，所以化简函数在高度并行环境下也很有用。

分布和可靠性

MapReduce 通过把对数据集的大规模操作分发给网络上的每个节点实现可靠性；每个节点会周期性的把完成的工作和状态的更新报告回来。如果一个节点保持沉默超过一个预设的时间间隔，主节点（类同 Google 档案系统中的主服务器）记录下这个节点状态为死亡，并把分配给这个节点的数据发到别的节点。每个操作使用命名文件的不可分割操作以确保不会发生并行线程间的冲突；当文件被改名的时候，系统可能会把他们复制到任务名以外的另一个名字上去。（避免副作用）。

化简操作工作方式很类似，但是由于化简操作在并行能力较差，主节点会尽量把化简操作调度在一个节点上，或者离需要操作的数据尽可能近的节点上了；这个特性可以满足 Google的需求，因为他们有足够的带宽，他们的内部网络没有那么多的机器。

【维基英文】

MapReduce is a programming model for processing large data sets with a parallel, distributed algorithm on a cluster.^[1]

A MapReduce program is composed of a Map() procedure that performs filtering and sorting (such as sorting students by first name into queues, one queue for each name) and a Reduce() procedure that performs a summary operation (such as counting the number of students in each queue, yielding name frequencies). The "MapReduce System" (also called "infrastructure" or "framework") orchestrates by marshalling the distributed servers, running the various tasks in parallel, managing all communications and data transfers between the various parts of the system, and providing for redundancy and fault tolerance.

The model is inspired by the map and reduce functions commonly used in functional programming,^[2] although their purpose in the MapReduce framework is not the same as in their original forms.^[3] Furthermore, the key contributions of the MapReduce framework are not the actual map and reduce functions, but the scalability and fault-tolerance achieved for a variety of applications by optimizing the execution engine once.

MapReduce libraries have been written in many programming languages, with different levels of optimization. A popular open-source implementation is Apache Hadoop. The name MapReduce originally referred to the proprietary Google technology but has since been genericized.

Overview

'MapReduce' is a framework for processing parallelizable problems across huge datasets using a large number of computers (nodes), collectively referred to as a cluster (if all nodes are on the same local network and use similar hardware) or a grid (if the nodes are shared across geographically and administratively distributed systems, and use more heterogenous hardware). Computational processing can occur on data stored either in a filesystem (unstructured) or in a database (structured). MapReduce can take advantage of locality of data, processing data on or near the storage assets to decrease transmission of data.

"Map" step: The master node takes the input, divides it into smaller sub-problems, and distributes them to worker nodes. A worker node may do this again in turn, leading to a multi-level tree structure. The worker node processes the smaller problem, and passes the answer back to its master node.

"Reduce" step: The master node then collects the answers to all the sub-problems and combines them in some way to form the output – the answer to the problem it was originally trying to solve.

MapReduce allows for distributed processing of the map and reduction operations. Provided that each mapping operation is independent of the others, all maps can be performed in parallel – though in practice this is limited by the number of independent data sources and/or the number of CPUs near each source. Similarly, a set of 'reducers' can perform the reduction phase, provided that all outputs of the map operation that share the same key are presented to the same reducer at the same time, or that the reduction function is associative. While this process can often appear inefficient compared to algorithms that are more sequential, MapReduce can be applied to significantly larger datasets than "commodity" servers can handle – a large server farm can use MapReduce to sort a petabyte of data in only a few hours.^{[citation needed]}The parallelism also offers some possibility of recovering from partial failure of servers or storage during the operation: if one mapper or reducer fails, the work can be rescheduled – assuming the input data is still available.

Another way to look at MapReduce is as a 5-step parallel and distributed computation:

1. Prepare the Map() input – the "MapReduce system" designates Map processors, assigns the K1 input key value each processor would work on, and provides that processor with all the input data associated with that key value.
2. Run the user-provided Map() code – Map() is run exactly once for each K1 key value, generating output organized by key values K2.
3. "Shuffle" the Map output to the Reduce processors – the MapReduce system designates Reduce processors, assigns the K2 key value each processor would work on, and provides that processor with all the Map-generated data associated with that key value.
4. Run the user-provided Reduce() code – Reduce() is run exactly once for each K2 key value produced by the Map step.
5. Produce the final output – the MapReduce system collects all the Reduce output, and sorts it by K2 to produce the final outcome.

Logically these 5 steps can be thought of as running in sequence – each step starts only after the previous step is completed – though in practice, of course, they can be intertwined, as long as the final result is not affected.

In many situations the input data might already be distributed ("sharded") among many different servers, in which case step 1 could sometimes be greatly simplified by assigning Map servers that would process the locally present input data. Similarly, step 3 could sometimes be sped up by assigning Reduce processors that are as much as possible local to the Map-generated data they need to process.

Logical view

The Map and Reduce functions of MapReduce are both defined with respect to data structured in (key, value) pairs. Map takes one pair of data with a type in one data domain, and returns a list of pairs in a different domain:

Map(k1,v1) → list(k2,v2)

The Map function is applied in parallel to every pair in the input dataset. This produces a list of pairs for each call. After that, the MapReduce framework collects all pairs with the same key from all lists and groups them together, creating one group for each key.

The Reduce function is then applied in parallel to each group, which in turn produces a collection of values in the same domain:

Reduce(k2, list (v2)) → list(v3)

Each Reduce call typically produces either one value v3 or an empty return, though one call is allowed to return more than one value. The returns of all calls are collected as the desired result list.

Thus the MapReduce framework transforms a list of (key, value) pairs into a list of values. This behavior is different from the typical functional programming map and reduce combination, which accepts a list of arbitrary values and returns one single value that combines all the values returned by map.

It is necessary but not sufficient to have implementations of the map and reduce abstractions in order to implement MapReduce. Distributed implementations of MapReduce require a means of connecting the processes performing the Map and Reduce phases. This may be a distributed file system. Other options are possible, such as direct streaming from mappers to reducers, or for the mapping processors to serve up their results to reducers that query them.

Examples

The prototypical MapReduce example counts the appearance of each word in a set of documents:^[4]

function map(String name, String document):
  // name: document name
  // document: document contents
  for each word w in document:
    emit (w, 1)

function reduce(String word, Iterator partialCounts):
  // word: a word
  // partialCounts: a list of aggregated partial counts
  sum = 0
  for each pc in partialCounts:
    sum += ParseInt(pc)
  emit (word, sum)

Here, each document is split into words, and each word is counted by the map function, using the word as the result key. The framework puts together all the pairs with the same key and feeds them to the same call to reduce, thus this function just needs to sum all of its input values to find the total appearances of that word.

As another example, imagine that for a database of 1.1 billion people, one would like to compute the average number of social contacts a person has according to age. InSQL such a query could be expressed as:

  SELECT age, AVG(contacts)
    FROM social.person
GROUP BY age
ORDER BY age

Using MapReduce, the K1 key values could be the integers 1 through 1,100, each representing a batch of 1 million records, the K2 key value could be a person’s age in years, and this computation could be achieved using the following functions:

function Map is
    input: integer K1 between 1 and 1100, representing a batch of 1 million social.person records
    for each social.person record in the K1 batch do
        let Y be the person's age
        let N be the number of contacts the person has
        produce one output record (Y,(N,1))
    repeat
end function

function Reduce is
    input: age (in years) Y
    for each input record (Y,(N,C)) do
        Accumulate in S the sum of N*C
        Accumulate in Cnew the sum of C
    repeat
    let A be S/Cnew
    produce one output record (Y,(A,Cnew))
end function

The MapReduce System would line up the 1,100 Map processors, and would provide each with its corresponding 1 million input records. The Map step would produce 1.1 billion(Y,(N,1)) records, with Y values ranging between, say, 8 and 103. The MapReduce System would then line up the 96 Reduce processors by performing shuffling operation of the key/value pairs due to the fact that we need average per age, and provide each with its millions of corresponding input records. The Reduce step would result in the much reduced set of only 96 output records (Y,A), which would be put in the final result file, sorted by Y.

The count info in the record is important if the processing is reduced more than one time. If we don't add the count of the records, the computed average would be wrong, for example:

-- map output #1: age, quantity of contacts
10, 9
10, 9
10, 9

-- map output #2: age, quantity of contacts
10, 9
10, 9

-- map output #3: age, quantity of contacts
10, 10

If we reduce files #1 and #2, we will have a new file with an average of 9 contacts for a 10 year old person ((9+9+9+9+9)/5):

-- reduce step #1: age, average of contacts
10, 9

If we reduce it with file #3, we lost the count of how many records we've already seen, so we would end up with an average of 9.5 contacts for a 10 year old person ((9+10)/2), wich is wrong. The correct answer is 9.17 ((9+9+9+9+9+10)/6).

Dataflow

The frozen part of the MapReduce framework is a large distributed sort. The hot spots, which the application defines, are:

• an input reader
• a Map function
• a partition function
• a compare function
• a Reduce function
• an output writer

Input reader

The input reader divides the input into appropriate size 'splits' (in practice typically 16 MB to 128 MB) and the framework assigns one split to each Map function. Theinput reader reads data from stable storage (typically a distributed file system) and generates key/value pairs.

A common example will read a directory full of text files and return each line as a record.

Map function

The Map function takes a series of key/value pairs, processes each, and generates zero or more output key/value pairs. The input and output types of the map can be (and often are) different from each other.

If the application is doing a word count, the map function would break the line into words and output a key/value pair for each word. Each output pair would contain the word as the key and the number of instances of that word in the line as the value.

Partition function

Each Map function output is allocated to a particular reducer by the application's partition function for sharding purposes. The partition function is given the key and the number of reducers and returns the index of the desired reducer.

A typical default is to hash the key and use the hash value modulo the number of reducers. It is important to pick a partition function that gives an approximately uniform distribution of data per shard for load-balancing purposes, otherwise the MapReduce operation can be held up waiting for slow reducers (reducers assigned more than their share of data) to finish.

Between the map and reduce stages, the data is shuffled (parallel-sorted / exchanged between nodes) in order to move the data from the map node that produced it to the shard in which it will be reduced. The shuffle can sometimes take longer than the computation time depending on network bandwidth, CPU speeds, data produced and time taken by map and reduce computations.

Comparison function

The input for each Reduce is pulled from the machine where the Map ran and sorted using the application's comparison function.

Reduce function

The framework calls the application's Reduce function once for each unique key in the sorted order. The Reduce can iterate through the values that are associated with that key and produce zero or more outputs.

In the word count example, the Reduce function takes the input values, sums them and generates a single output of the word and the final sum.

Output writer

The Output Writer writes the output of the Reduce to the stable storage, usually a distributed file system.

Distribution and reliability

MapReduce achieves reliability by parceling out a number of operations on the set of data to each node in the network. Each node is expected to report back periodically with completed work and status updates. If a node falls silent for longer than that interval, the master node (similar to the master server in the Google File System) records the node as dead and sends out the node's assigned work to other nodes. Individual operations use atomic operations for naming file outputs as a check to ensure that there are not parallel conflicting threads running. When files are renamed, it is possible to also copy them to another name in addition to the name of the task (allowing for side-effects).

The reduce operations operate much the same way. Because of their inferior properties with regard to parallel operations, the master node attempts to schedule reduce operations on the same node, or in the same rack as the node holding the data being operated on. This property is desirable as it conserves bandwidth across the backbone network of the datacenter.

Implementations are not necessarily highly reliable. For example, in older versions of Hadoop the NameNode was a single point of failure for the distributed filesystem. Later versions of Hadoop have high availability with an active/passive failover for the "NameNode."

Uses

MapReduce is useful in a wide range of applications, including distributed pattern-based searching, distributed sorting, web link-graph reversal, term-vector per host, web access log stats, inverted index construction, document clustering, machine learning,^[5] and statistical machine translation. Moreover, the MapReduce model has been adapted to several computing environments like multi-core and many-core systems,^[6][7][8] desktop grids,^[9] volunteer computing environments,^[10] dynamic cloud environments,^[11] and mobile environments.^[12]

At Google, MapReduce was used to completely regenerate Google's index of the World Wide Web. It replaced the old ad hoc programs that updated the index and ran the various analyses.^[13]

MapReduce's stable inputs and outputs are usually stored in a distributed file system. The transient data is usually stored on local disk and fetched remotely by the reducers.

Criticism

Lack of novelty

David DeWitt and Michael Stonebraker, computer scientists specializing in parallel databases and shared-nothing architectures, have been critical of the breadth of problems that MapReduce can be used for.^[14] They called its interface too low-level and questioned whether it really represents the paradigm shift its proponents have claimed it is.^[15] They challenged the MapReduce proponents' claims of novelty, citing Teradata as an example of prior art that has existed for over two decades. They also compared MapReduce programmers to Codasyl programmers, noting both are "writing in a low-level language performing low-level record manipulation."^[15] MapReduce's use of input files and lack of schema support prevents the performance improvements enabled by common database system features such as B-trees and hash partitioning, though projects such as Pig (or PigLatin), Sawzall, Apache Hive,^[16] YSmart,^[17] HBase^[18] and BigTable^[18][19] are addressing some of these problems.

Greg Jorgensen wrote an article rejecting these views.^[20] Jorgensen asserts that DeWitt and Stonebraker's entire analysis is groundless as MapReduce was never designed nor intended to be used as a database.

DeWitt and Stonebraker have subsequently published a detailed benchmark study in 2009 comparing performance of Hadoop's MapReduce and RDBMS approaches on several specific problems.^[21] They concluded that relational databases offer real advantages for many kinds of data use, especially on complex processing or where the data is used across an enterprise, but that MapReduce may be easier for users to adopt for simple or one-time processing tasks.

Google has been granted a patent on MapReduce.^[22] However, there have been claims that this patent should not have been granted because MapReduce is too similar to existing products. For example, map and reduce functionality can be very easily implemented in Oracle's PL/SQL database oriented language.^[23]

Restricted programming framework

MapReduce tasks must be written as acyclic dataflow programs, i.e. a stateless mapper followed by a stateless reducer, that are executed by a batch job scheduler. This paradigm makes repeated querying of datasets difficult and imposes limitations that are felt in fields such as machine learning, where iterative algorithms that revisit a single working set multiple times are the norm.^[24]