Big data processing frameworks and architectures: a survey

In recent times, there has been rapid growth in data generated from autonomous sources. The existing data processing techniques are not suitable to deal with these large volumes of complex data that can be structured, semi-structured or unstructured. This large data is referred to as Big data because of its main characteristics: volume, variety velocity, value and veracity. Extensive research on Big data is ongoing, and the primary focus of this research is on processing massive amounts of data effectively and efficiently. However, researchers are paying little attention on how to store and analyze the large volumes of data to get useful insights from it. In this chapter, the authors examine existing Big data processing frameworks like MapReduce, Apache Spark, Storm and Flink. In this chapter, the architectures of MapReduce, iterative MapReduce frameworks and components of Apache Spark are discussed in detail. Most of the widely used classical machine learning techniques are implemented using these Big data frameworks in the form of Apache Mahout and Spark MLlib libraries and these need to be enhanced to support all existing machine learning techniques like formal concept analysis (FCA) and neural embedding. In this chapter, authors have taken FCA as an application and provided scalable FCA algorithms using the Big data processing frameworks like MapReduce and Spark. Streaming data processing frameworks like Apache Flink and Apache Storm is also examined. Authors also discuss about the storage architectures like Hadoop Distributed File System (HDFS), Dynamo and Amazon S3 in detail while processing large Big data applications. The survey concludes with a proposal for best practices related to the studied architectures and frameworks.