Automated Testing and Debugging for Big Data Analytics
- Author(s): Gulzar, Muhammad Ali
- Advisor(s): Kim, Miryung
- et al.
The prevalence of big data analytics in almost every large-scale software system has generated a substantial push to build data-intensive scalable computing (DISC) frameworks such as Google MapReduce and Apache Spark that can fully harness the power of existing data centers. However, frameworks once used by domain experts are now being leveraged by data scientists, business analysts, and researchers. This shift in user demographics calls for immediate advancements in the development, debugging, and testing practices of big data applications, which are falling behind compared to the DISC framework design and implementation. In practice, big data applications often fail as users are unable to test all behaviors emerging from interleaving dataflow operators, user-defined functions, and framework's code. "Testing based on a random sample" rarely guarantees the reliability and "trial and error" and "print" debugging methods are expensive and time-consuming. Thus, the current practice of developing a big data application must be improved and the tools built to enhance the developer's productivity must adapt to the distinct characteristics of data-intensive scalable computing.
By synthesizing ideas from software engineering and database systems, our hypothesis is that we can design effective and scalable testing and debugging algorithms for big data analytics without compromising the performance and efficiency of the underlying DISC framework. To design such techniques, we investigate how we can build interactive and responsive debugging primitives that significantly reduce the debugging time, yet do not pose much performance overhead on big data applications. Furthermore, we investigate how we can leverage data provenance techniques from databases and fault-isolation algorithms from software engineering to pinpoint the minimal subset of failure-inducing inputs efficiently. To improve the reliability of big data analytics, we investigate how we can abstract the semantics of dataflow operators and use them in tandem with the semantics of user-defined functions to generate a minimum set of synthetic test inputs capable of revealing more defects than the entire input dataset.
To examine the first hypothesis, we introduce interactive, real-time debugging primitives for big data analytics through innovative and scalable debugging features such as simulated breakpoint, dynamic watchpoint, and crash culprit identification. Second, we design a new automated fault localization approach that combines insights from both the software engineering and database literature to bring delta debugging closer to a reality in the big data applications by leveraging data provenance and by constructing systems optimizations for debugging provenance queries. Lastly, we devise a new symbolic-execution based white-box testing algorithm for big data applications that abstracts the implementation of dataflow operators using logical specifications instead of modeling their implementations and combines them with the semantics of any arbitrary user-defined function.
We instantiate the idea of an interactive debugging algorithm as BigDebug, the idea of an automated debugging algorithm as BigSift, and the idea of symbolic execution-based testing as BigTest. Our investigation shows that the interactive debugging primitives can scale to terabytes---our record-level tracing incurs less than 25% overhead on average and provides up to 100% time saving compared to the baseline replay debugger. Second, we observe that by combining data provenance with delta debugging, we can identify the minimum faulty input in just under 30% of the original job execution time. Lastly, we verify that by abstracting dataflow operators using logical specifications, we can efficiently generate the most concise test data suitable for local testing while revealing twice as many faults as prior approaches. Our investigations collectively demonstrate that developer productivity can be significantly improved through effective and scalable testing and debugging techniques for big data analytics, without impacting the DISC framework's performance. This dissertation affirms the feasibility of automated debugging and testing techniques for big data analytics---techniques that were previously considered infeasible for large-scale data processing.