Efficient Online Processing for Advanced Analytics

With the advent of emerging technologies and the Internet of Things, the importance of online data analytics has become more pronounced. Businesses and companies are adopting approaches that provide responsive analytics to stay competitive in the global marketplace. Online analytics allow data analysts to promptly react to patterns or to gain preliminary insights from early results that aid in research, decision making, and effective strategy planning. The growth of data-velocity in a variety of domains including, high-frequency trading, social networks, infrastructure monitoring, and advertising require adopting online engines that can efficiently process continuous streams of data. This thesis presents foundations, techniques, and systems' design that extend the state-of-the-art in online query processing to efficiently support relational joins with arbitrary join-predicates (beyond traditional equi-joins); and to support other data models (beyond relational) that target machine learning and graph computations. The thesis is divided into two parts: We first present a brief overview of Squall, our open-source online query processing engine that supports SQL-like queries on top of streams. Then, we focus on extending Squall to support efficient theta-join processing. Scalable distributed join processing requires a partitioning policy that evenly distributes the processing load while minimizing the size of maintained state and duplicated messages. Efficient load-balance demands apriori-statistics which are not available in the online setting. We propose a novel operator that continuously adjusts itself to the data dynamics, through adaptive dataflow routing and state repartitioning. It is also resilient to data-skew, maintains high throughput rates, avoids blocking during state repartitioning, and behaves as a black-box dataflow operator with provable performance guarantees. Our evaluation demonstrates that the proposed operator outperforms the state-of-the-art static partitioning schemes in resource utilization, throughput, and execution time up to 7x. In the second part, we present a novel framework that supports the Incremental View Maintenance (IVM) of workloads expressed as linear algebra programs. Linear algebra represents a concrete substrate for advanced analytical tasks including, machine learning, scientific computation, and graph algorithms. Previous works on relational calculus IVM are not applicable to matrix algebra workloads. This is because a single entry change to an input-matrix results in changes all over the intermediate views, rendering IVM useless in comparison to re-evaluation. We present Lago, a unified modular compiler framework that supports the IVM of a broad class of linear algebra programs. Lago automatically derives and optimizes incremental trigger programs of analytical computations, while freeing the user from erroneous manual derivations, low-level implementation details, and performance tuning. We present a novel technique that captures $\Delta$ changes as low-rank matrices. Low-rank matrices are representable in a compressed factored form that enables cheaper computations. Lago automatically propagates the factored representation across program statements to derive an efficient trigger program. Moreover, Lago extends its support to other domains that use different semi-ring configurations, e.g., graph applications. Our evaluation results demonstrate orders of magnitude (10x-10

Efficient Incremental Data Analysis

Milos Nikolic

Many data-intensive applications require real-time analytics over streaming data. In a growing number of domains -- sensor network monitoring, social web applications, clickstream analysis, high-frequency algorithmic trading, and fraud detections to name a few -- applications continuously monitor stream events to promptly react to certain data conditions. These applications demand responsive analytics even when faced with high volume and velocity of incoming changes, large numbers of users, and complex processing requirements. Developing suitable online analytics engine that meets these requirements is challenging. In this thesis, we study techniques for efficient online processing of complex analytical queries, ranging from standard database queries to complex machine learning and digital signal processing workflows. First, we focus on the problem of efficient incremental computation for database queries. We have developed a system, called DBToaster, that compiles declarative queries into high-performance stream processing engines that keep query results (views) fresh at very high update rates. At the heart of our system is a recursive query compilation algorithm that materializes a set of supporting higher-order delta views to achieve a substantially lower view maintenance cost. We study the trade-offs between single-tuple and batch incremental processing in local execution, and we present a novel approach for compiling view maintenance code into data-parallel programs optimized for distributed execution. DBToaster supports millions of complete view refreshes per second for a broad range of queries and outperforms commercial database and stream engines by orders of magnitude. We also study the incremental computation for queries written as iterative linear algebra, which can capture many machine learning and scientific calculations. We have developed a framework, called LINVIEW, for capturing deltas of linear algebra programs and understanding their computational cost. Linear algebra operations tend to cause an avalanche effect where even very local changes to the input matrices spread out and infect all of the intermediate results and the final view, causing incremental view maintenance to lose its performance benefit over re-evaluation. We develop techniques based on matrix factorizations to contain such epidemics of change and make incremental view maintenance of linear algebra practical and usually substantially cheaper than re-evaluation. We show, both analytically and experimentally, the usefulness of these techniques when applied to standard analytics tasks. Our last research question concerns the integration of general-purpose query processors and domain-specific operations to enable deep data exploration in both online and offline analysis. We advocate a deep integration of signal processing operations and general-purpose query processors. We demonstrate that in-situ processing of tempo-relational and signal data through a unified query language empowers users to express end-to-end workflows more succinctly inside one system while at the same time offering orders of magnitude better performance than existing popular data management systems.

EPFL2016

LINVIEW: Incremental View Maintenance for Complex Analytical Queries

Mohamed Elsayed Mohamed Ahmed El Seidy, Christoph Koch, Milos Nikolic

Many analytics tasks and machine learning problems can be naturally expressed by iterative linear algebra programs. In this paper, we study the incremental view maintenance problem for such complex analytical queries. We develop a framework, called LINVIEW, for capturing deltas of linear algebra programs and understanding their computational cost. Linear algebra operations tend to cause an avalanche effect where even very local changes to the input matrices spread out and infect all of the intermediate results and the final view, causing incremental view maintenance to lose its performance benefit over re-evaluation. We develop techniques based on matrix factorizations to contain such epidemics of change. As a consequence, our techniques make incremental view maintenance of linear algebra practical and usually substantially cheaper than re-evaluation. We show, both analytically and experimentally, the usefulness of these techniques when applied to standard analytics tasks. Our evaluation demonstrates the efficiency of LINVIEW in generating parallel incremental programs that outperform re-evaluation techniques by more than an order of magnitude.

2014

Runtime Prediction for Scale-Out Data Analytics

Adrian Daniel Popescu

Many analytics applications generate mixed workloads, i.e., workloads comprised of analytical tasks with different processing characteristics including data pre-processing, SQL, and iterative machine learning algorithms. Examples of such mixed workloads can be found in web data analysis, social media analysis, and graph analytics, where they are executed repetitively on large input datasets (e.g., "Find the average user time spent on the top 10 most popular web pages on the UK domain web graph."). Scale-out processing engines satisfy the needs of these applications by distributing the data and the processing task efficiently among multiple workers that are first reserved and then used to execute the task in parallel on a cluster of machines. Finding the resource allocation that can complete the workload execution within a given time constraint, and optimizing cluster resource allocations among multiple analytical workloads motivates the need for estimating the runtime of the workload before its actual execution. Predicting runtime of analytical workloads is a challenging problem as runtime depends on a large number of factors that are hard to model a priori execution. These factors can be summarized as workload characteristics (i.e., data statistics and processing costs), the execution configuration (i.e., deployment, resource allocation, and software settings), and the cost model that captures the interplay among all of the above parameters. While conventional cost models proposed in the context of query optimization can assess the relative order among alternative SQL query plans, they are not aimed to estimate absolute runtime. Additionally, conventional models are ill-equipped to estimate the runtime of iterative analytics that are executed repetitively until convergence and that of user defined data pre-processing operators which are not "owned" by the underlying data management system. This thesis demonstrates that runtime for data analytics can be predicted accurately by breaking the analytical tasks into multiple processing phases, collecting key input features during a reference execution on a sample of the dataset, and then using the features to build per-phase cost models. We develop prediction models for three categories of data analytics produced by social media applications: iterative machine learning, data pre-processing, and reporting SQL. The prediction framework for iterative analytics, PREDIcT, addresses the challenging problem of estimating the number of iterations, and per-iteration runtime for a class of iterative machine learning algorithms that are run repetitively until convergence. The hybrid prediction models we develop for data pre-processing tasks and for reporting SQL combine the benefits of analytical modeling with that of machine learning-based models. Through a training methodology and a pruning algorithm we reduce the cost of running training queries to a minimum while maintaining a good level of accuracy for the models.