Distributed systems are often used to collect, access, and manipulate large data sets. For example, the database systems described earlier in the chapter can operate over datasets that are stored across multiple machines. No single machine may contain the data necessary to respond to a query, and so communication is required to service requests.

This section investigates a typical big data processing scenario in which a data set too large to be processed by a single machine is instead distributed among many machines, each of which process a portion of the dataset. The result of processing must often be aggregated across machines, so that results from one machine's computation can be combined with others. To coordinate this distributed data processing, we will discuss a programming framework called MapReduce.

Creating a distributed data processing application with MapReduce combines many of the ideas presented throughout this text. An application is expressed in terms of pure functions that are used to map over a large dataset and then to reduce the mapped sequences of values into a final result.

Familiar concepts from functional programming are used to maximal advantage in a MapReduce program. MapReduce requires that the functions used to map and reduce the data be pure functions. In general, a program expressed only in terms of pure functions has considerable flexibility in how it is executed. Sub-expressions can be computed in arbitrary order and in parallel without affecting the final result. A MapReduce application evaluates many pure functions in parallel, reordering computations to be executed efficiently in a distributed system.

The principal advantage of MapReduce is that it enforces a separation of concerns between two parts of a distributed data processing application:

The map and reduce functions that process data and combine results.
The communication and coordination between machines.

The coordination mechanism handles many issues that arise in distributed computing, such as machine failures, network failures, and progress monitoring. While managing these issues introduces some complexity in a MapReduce application, none of that complexity is exposed to the application developer. Instead, building a MapReduce application only requires specifying the map and reduce functions in (1) above; the challenges of distributed computation are hidden via abstraction.

Original Chapter Link

Introduction

Implicit Sequences

Declarative Programming

Logic Programming

Unification

Distributed Computing

Distributed Data Processing

Parallel Computing