Jumping the ORDER BY Barrier in Large-Scale Pattern Matching

Event-series pattern matching is a major component of large-scale data analytics pipelines enabling a wide range of system diagnostics tasks. A precursor to pattern matching is an expensive ``shuffle the world'' stage wherein data are ordered by time and shuffled across the network. Because many existing systems treat the pattern matching engine as a black box, they are unable to optimizing the entire data analytics pipeline, and in particular, this costly shuffle. This paper demonstrates how to optimize such queries. We first translate an expressive class of regular-expression like patterns to relational queries such that they can benefit from decades of progress in relational optimizers, and then we introduce the technique of abstract pattern matching, a linear time preprocessing step which, adapting ideas from symbolic execution and abstract interpretation, discards events from the input guaranteed not to appear in successful matches. Abstract pattern matching first computes a conservative representation of the output-relevant domain of every transition in a pattern based on the (unary) predicates of that transition. It then further refines these domains based on the structure of the pattern (i.e., paths through the pattern) as well as any of the pattern's join predicates across transitions. The outcome is an abstract filter that when applied to the original stream excludes events that are guaranteed not to participate in a match. We implemented and applied abstract pattern matching in COSMOS/Scope to an industrial benchmark where we obtained up to 3 orders of magnitude reduction in shuffled data and 1.23x average speedup in total processing time.


 Record created 2017-07-05, last modified 2018-09-13

Download fulltext

Rate this document:

Rate this document:
(Not yet reviewed)