No data left behind: real-time insights from a complex data ecosystem

The typical enterprise data architecture consists of several actively updated data sources (e.g., NoSQL systems, data warehouses), and a central data lake such as HDFS, in which all the data is periodically loaded through ETL processes. To simplify query processing, state-of-the-art data analysis approaches solely operate on top of the local, historical data in the data lake, and ignore the fresh tail end of data that resides in the original remote sources. However, as many business operations depend on real-time analytics, this approach is no longer viable. The alternative is hand-crafting the analysis task to explicitly consider the characteristics of the various data sources and identify optimization opportunities, rendering the overall analysis non-declarative and convoluted. Based on our experiences operating in data lake environments, we design System-PV, a real-time analytics system that masks the complexity of dealing with multiple data sources while offering minimal response times. System-PV extends Spark with a sophisticated data virtualization module that supports multiple applications - from SQL queries to machine learning. The module features a location-aware compiler that considers source complexity, and a two-phase optimizer that produces and refines the query plans, not only for SQL queries but for all other types of analysis as well. The experiments show that System-PV is often faster than Spark by more than an order of magnitude. In addition, the experiments show that the approach of accessing both the historical and the remote fresh data is viable, as it performs comparably to solely operating on top of the local, historical data.

ReCache: Reactive Caching for Fast Analytics over Heterogeneous Data

Anastasia Ailamaki, Tahir Azim, Manolis Karpathiotakis

As data continues to be generated at exponentially growing rates in heterogeneous formats, fast analytics to extract meaningful information is becoming increasingly important. Systems widely use in-memory caching as one of their primary techniques to speed up data analytics. However, caches in data analytics systems cannot rely on simple caching policies and a fixed data layout to achieve good performance. Different datasets and workloads require different layouts and policies to achieve optimal performance. This paper presents ReCache, a cache-based performance accelerator that is reactive to the cost and heterogeneity of diverse raw data formats. Using timing measurements of caching operations and selection operators in a query plan, ReCache accounts for the widely varying costs of reading, parsing, and caching data in nested and tabular formats. Combining these measurements with information about frequently accessed data fields in the workload, ReCache automatically decides whether a nested or relational column-oriented layout would lead to better query performance. Furthermore, ReCache keeps track of commonly utilized operators to make informed cache admission and eviction decisions. Experiments on synthetic and real-world datasets show that our caching techniques decrease caching overhead for individual queries by an average of 59%. Furthermore, over the entire workload, ReCache reduces execution time by 19-75% compared to existing techniques.

2017

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.

EPFL2015

Just-in-time Analytics Over Heterogeneous Data and Hardware

Manolis Karpathiotakis

Industry and academia are continuously becoming more data-driven and data-intensive, relying on the analysis of a wide variety of datasets to gain insights. At the same time, data variety increases continuously across multiple axes. First, data comes in multiple formats, such as the binary tabular data of a DBMS, raw textual files, and domain-specific formats. Second, different datasets follow different data models, such as the relational and the hierarchical one. Data location also varies: Some datasets reside in a central "data lake", whereas others lie in remote data sources. In addition, users execute widely different analysis tasks over all these data types. Finally, the process of gathering and integrating diverse datasets introduces several inconsistencies and redundancies in the data, such as duplicate entries for the same real-world concept. In summary, heterogeneity significantly affects the way data analysis is performed. In this thesis, we aim for data virtualization: Abstracting data out of its original form and manipulating it regardless of the way it is stored or structured, without a performance penalty. To achieve data virtualization, we design and implement systems that i) mask heterogeneity through the use of heterogeneity-aware, high-level building blocks and ii) offer fast responses through on-demand adaptation techniques. Regarding the high-level building blocks, we use a query language and algebra to handle multiple collection types, such as relations and hierarchies, express transformations between these collection types, as well as express complex data cleaning tasks over them. In addition, we design a location-aware compiler and optimizer that masks away the complexity of accessing multiple remote data sources. Regarding on-demand adaptation, we present a design to produce a new system per query. The design uses customization mechanisms that trigger runtime code generation to mimic the system most appropriate to answer a query fast: Query operators are thus created based on the query workload and the underlying data models; the data access layer is created based on the underlying data formats. In addition, we exploit emerging hardware by customizing the system implementation based on the available heterogeneous processors â CPUs and GPGPUs. We thus pair each workload with its ideal processor type. The end result is a just-in-time database system that is specific to the query, data, workload, and hardware instance. This thesis redesigns the data management stack to natively cater for data heterogeneity and exploit hardware heterogeneity. Instead of centralizing all relevant datasets, converting them to a single representation, and loading them in a monolithic, static, suboptimal system, our design embraces heterogeneity. Overall, our design decouples the type of performed analysis from the original data layout; users can perform their analysis across data stores, data models, and data formats, but at the same time experience the performance offered by a custom system that has been built on demand to serve their specific use case.