Efficient Pattern Detection over a Distributed Framework

Abstract. In recent past, work has been done to parallelize pattern detection queries over event stream, by partitioning the event stream on certain keys or attributes. In such partitioning schemes the degree of parallelization totally relies on the available partition keys. A limited number of partitioning keys, or unavailability of such partitioning attributes noticeably affect the distribution of data among multiple nodes, and is a reason of potential data skew and improper resource utilization. Moreover, majority of the past implementations of complex event detection are based on a single machine, hence, they are immune to potential data skew that could be seen in a real distributed environment. In this study, we propose an event stream partitioning scheme that without considering any key attributes partitions the stream over time-windows. This scheme efficiently distributes the event stream partitions across network, and detects pattern sequences in distributed fashion. Our scheme also provides an effective means to minimize potential data skew and handles a substantial number of pattern queries across network.