Dokument

Summary:
Due to a growing number of data producers and their ever-increasing data volume, the ability to ingest, analyze, and store potentially never-ending streams of data is a mission-critical task in today's data processing landscape. A widespread form of data streams are event streams, which consist of continuously arriving notifications about some real-world phenomena. For example, a temperature sensor naturally generates an event stream by periodically measuring the temperature and reporting it with measurement time in case of a substantial change to the previous measurement. In this thesis, we consider two kinds of event stream processing: online and offline. Online refers to processing events solely in main memory as soon as they arrive, while offline means processing event data previously persisted to non-volatile storage. Both modes are supported by widely used scale-out general-purpose stream processing engines (SPEs) like Apache Flink or Spark Streaming. However, such engines suffer from two significant deficiencies that severely limit their processing performance. First, for offline processing, they load the entire stream from non-volatile secondary storage and replay all data items into the associated online engine in order of their original arrival. While this naturally ensures unified query semantics for on- and offline processing, the costs for reading the entire stream from non-volatile storage quickly dominate the overall processing costs. Second, modern SPEs focus on scaling out computations across the nodes of a cluster, but use only a fraction of the available resources of individual nodes. This thesis tackles those problems with three different approaches. First, we present novel techniques for the offline processing of two important query types (windowed aggregation and sequential pattern matching). Our methods utilize well-understood indexing techniques to reduce the total amount of data to read from non-volatile storage. We show that this improves the overall query runtime significantly. In particular, this thesis develops the first index-based algorithms for pattern queries expressed with the Match_Recognize clause, a new and powerful language feature of SQL that has received little attention so far. Second, we show how to maximize resource utilization of single nodes by exploiting the capabilities of modern hardware. Therefore, we develop a prototypical shared-memory CPU-GPU-enabled event processing system. The system provides implementations of all major event processing operators (filtering, windowed aggregation, windowed join, and sequential pattern matching). Our experiments reveal that regarding resource utilization and processing throughput, such a hardware-enabled system is superior to hardware-agnostic general-purpose engines. Finally, we present TPStream, a new operator for pattern matching over temporal intervals. TPStream achieves low processing latency and, in contrast to sequential pattern matching, is easily parallelizable even for unpartitioned input streams. This results in maximized resource utilization, especially for modern CPUs with multiple cores.

Das Dokument ist im Internet frei zugänglich - Hinweise zu den Nutzungsrechten