Explainable and Resource-Efficient Stream Processing Through Provenance and Scheduling

Doktorsavhandling, 2022

In our era of big data, information is captured at unprecedented volumes and velocities, with technologies such as Cyber-Physical Systems making quick decisions based on the processing of streaming, unbounded datasets. In such scenarios, it can be beneficial to process the data in an online manner, using the stream processing paradigm implemented by Stream Processing Engines (SPEs). While SPEs enable high-throughput, low-latency analysis, they are faced with challenges connected to evolving deployment scenarios, like the increasing use of heterogeneous, resource-constrained edge devices together with cloud resources and the increasing user expectations for usability, control, and resource-efficiency, on par with features provided by traditional databases.

This thesis tackles open challenges regarding making stream processing more user-friendly, customizable, and resource-efficient. The first part outlines our work, providing high-level background information, descriptions of the research problems, and our contributions. The second part presents our three state-of-the-art frameworks for explainable data streaming using data provenance, which can help users of streaming queries to identify important data points, explain unexpected behaviors, and aid query understanding and debugging. (A) GeneaLog provides backward provenance allowing users to identify the inputs that contributed to the generation of each output of a streaming query. (B) Ananke is the first framework to provide a duplicate-free graph of live forward provenance, enabling easy bidirectional tracing of input-output relationships in streaming queries and identifying data points that have finished contributing to results. (C) Erebus is the first framework that allows users to define expectations about the results of a streaming query, validating whether these expectations are met or providing explanations in the form of why-not provenance otherwise. The third part presents techniques for execution efficiency through custom scheduling, introducing our state-of-the-art scheduling frameworks that control resource allocation and achieve user-defined performance goals. (D) Haren is an SPE-agnostic user-level scheduler that can efficiently enforce user-defined scheduling policies. (E) Lachesis is a standalone scheduling middleware that requires no changes to SPEs but, instead, directly guides the scheduling decisions of the underlying Operating System. Our extensive evaluations using real-world SPEs and workloads show that our work significantly improves over the state-of-the-art while introducing only small performance overheads.