Data-Intensive Text Processing with MapReduce. Jimmy Lin

jobs). Section 2.4 fills in additional details by introducing partitioners and combiners, which provide greater control over data flow. MapReduce would not be practical without a tightly-integrated distributed file system that manages the data being processed; Section 2.5 covers this in detail. Tying everything together, a complete cluster architecture is described in Section 2.6 before the chapter ends with a summary.

2.1 FUNCTIONAL PROGRAMMING ROOTS

      MapReduce has its roots in functional programming, which is exemplified in languages such as Lisp and ML.⁴ A key feature of functional languages is the concept of higher-order functions, or functions that can accept other functions as arguments. Two common built-in higher order functions are map and fold, illustrated in Figure 2.1. Given a list, map takes as an argument a function f (that takes a single argument) and applies it to all elements in a list (the top part of the diagram). Given a list, fold takes as arguments a function g (that takes two arguments) and an initial value: g is first applied to the initial value and the first item in the list, the result of which is stored in an intermediate variable. This intermediate variable and the next item in the list serve as the arguments to a second application of g, the results of which are stored in the intermediate variable. This process repeats until all items in the list have been consumed; fold then returns the final value of the intermediate variable. Typically, map and fold are used in combination. For example, to compute the sum of squares of a list of integers, one could map a function that squares its argument (i.e., λx.x²) over the input list, and then fold the resulting list with the addition function (more precisely, λxλy.x + y) using an initial value of zero.

Figure 2.1: Illustration of map and fold, two higher-order functions commonly used together in functional programming: map takes a function f and applies it to every element in a list, while fold iteratively applies a function g to aggregate results.

      We can view map as a concise way to represent the transformation of a dataset (as defined by the function f). In the same vein, we can view fold as an aggregation operation, as defined by the function g. One immediate observation is that the application of f to each item in a list (or more generally, to elements in a large dataset) can be parallelized in a straightforward manner, since each functional application happens in isolation. In a cluster, these operations can be distributed across many different machines. The fold operation, on the other hand, has more restrictions on data locality—elements in the list must be “brought together” before the function g can be applied. However, many real-world applications do not require g to be applied to all elements of the list. To the extent that elements in the list can be divided into groups, the fold aggregations can also proceed in parallel. Furthermore, for operations that are commutative and associative, significant efficiencies can be gained in the fold operation through local aggregation and appropriate reordering.

      In a nutshell, we have described MapReduce. The map phase in MapReduce roughly corresponds to the map operation in functional programming, whereas the reduce phase in MapReduce roughly corresponds to the fold operation in functional programming. As we will discuss in detail shortly, the MapReduce execution framework coordinates the map and reduce phases of processing over large amounts of data on large clusters of commodity machines.

      Viewed from a slightly different angle, MapReduce codifies a generic “recipe” for processing large datasets that consists of two stages. In the first stage, a user-specified computation is applied over all input records in a dataset. These operations occur in parallel and yield intermediate output that is then aggregated by another user-specified computation. The programmer defines these two types of computations, and the execution framework coordinates the actual processing (very loosely, MapReduce provides a functional abstraction). Although such a two-stage processing structure may appear to be very restrictive, many interesting algorithms can be expressed quite concisely—especially if one decomposes complex algorithms into a sequence of MapReduce jobs. Subsequent chapters in this book focus on how a number of algorithms can be implemented in MapReduce.

      To be precise, MapReduce can refer to three distinct but related concepts. First, MapReduce is a programming model, which is the sense discussed above. Second, MapReduce can refer to the execution framework (i.e., the “runtime”) that coordinates the execution of programs written in this particular style. Finally, MapReduce can refer to the software implementation of the programming model and the execution framework: for example, Google’s proprietary implementation vs. the open-source Hadoop implementation in Java. And, in fact, there are many implementations of MapReduce, e.g., targeted specifically for multi-core processors [127], for GPGPUs [71], for the CELL architecture [126], etc. There are some differences between the MapReduce programming model implemented in Hadoop and Google’s proprietary implementation, which we will explicitly discuss throughout the book. However, we take a rather Hadoop-centric view of MapReduce, since Hadoop remains the most mature and accessible implementation to date, and therefore the one most developers are likely to use.

2.2 MAPPERS AND REDUCERS

      Key-value pairs form the basic data structure in MapReduce. Keys and values may be primitives such as integers, floating point values, strings, and raw bytes, or they may be arbitrarily complex structures (lists, tuples, associative arrays, etc.). Programmers typically need to define their own custom data types, although a number of libraries such as Protocol Buffers,⁵ Thrift,⁶ and Avro⁷ simplify the task.

      Part of the design of MapReduce algorithms involves imposing the key-value structure on arbitrary datasets. For a collection of web pages, keys may be URLs and values may be the actual HTML content. For a graph, keys may represent node ids and values may contain the adjacency lists of those nodes (see Chapter 5 for more details). In some algorithms, input keys are not particularly meaningful and are simply ignored during processing, while in other cases input keys are used to uniquely identify a datum (such as a record id). In Chapter 3, we discuss the role of complex keys and values in the design of various algorithms.

      In MapReduce, the programmer defines a mapper and a reducer with the following signatures:

      map: (k₁,v₁) → [(k₂,v₂)]

      reduce: (k₂, [v₂]) → [(k₃,v₃)]

      The convention […] is used throughout this book to denote a list. The input to a MapReduce job starts as data stored on the underlying distributed file system (see Section 2.5). The mapper is applied to every input key-value pair (split across an arbitrary number of files) to generate an arbitrary number of intermediate key-value pairs. The reducer is applied to all values associated with the same intermediate key to generate output key-value pairs.⁸ Implicit between the map and reduce phases is a distributed “group by” operation on intermediate keys. Intermediate data arrive at each reducer in order, sorted by the key. However, no ordering relationship is guaranteed for keys across different reducers. Output key-value pairs from each reducer are written persistently back onto the distributed file system (whereas intermediate key-value pairs are transient and not preserved). The output ends up in r files on the distributed file system, where r is the number of reducers. For the most part, there is no need to consolidate reducer output, since the r files often serve as input to yet another MapReduce job. Figure 2.2 illustrates this two-stage processing structure.

      A