Thesis (Ph.D.) - Indiana University, Computer Sciences, 2015As Big Data processing problems evolve, many modern applications demonstrate special characteristics. Data exists in the form of both large historical datasets and high-speed real-time streams, and many analysis pipelines require integrated parallel batch processing and stream processing. Despite the large size of the whole dataset, most analyses focus on specific subsets according to certain criteria. Correspondingly, integrated support for efficient queries and post- query analysis is required. To address the system-level requirements brought by such characteristics, this dissertation proposes a scalable architecture for integrated queries, batch analysis, and streaming analysis of Big Data in the cloud. We verify its effectiveness using a representative application domain - social media data analysis - and tackle related research challenges emerging from each module of the architecture by integrating and extending multiple state-of-the-art Big Data storage and processing systems. In the storage layer, we reveal that existing text indexing techniques do not work well for the unique queries of social data, which put constraints on both textual content and social context. To address this issue, we propose a flexible indexing framework over NoSQL databases to support fully customizable index structures, which can embed necessary social context information for efficient queries. The batch analysis module demonstrates that analysis workflows consist of multiple algorithms with different computation and communication patterns, which are suitable for different processing frameworks. To achieve efficient workflows, we build an integrated analysis stack based on YARN, and make novel use of customized indices in developing sophisticated analysis algorithms. In the streaming analysis module, the high-dimensional data representation of social media streams poses special challenges to the problem of parallel stream clustering. Due to the sparsity of the high-dimensional data, traditional synchronization method becomes expensive and severely impacts the scalability of the algorithm. Therefore, we design a novel strategy that broadcasts the incremental changes rather than the whole centroids of the clusters to achieve scalable parallel stream clustering algorithms. Performance tests using real applications show that our solutions for parallel data loading/indexing, queries, analysis tasks, and stream clustering all significantly outperform implementations using current state-of-the-art technologies.