Adaptive Query Processing in the Looking Glass

Abstract

A great deal of work on adaptive query processing has been done over the last few years: Adaptive query processing has been used to detect and correct optimizer errors due to incorrect statistics or simplified cost metrics; it has been applied to long-running continuous queries over data streams whose characteristics change over time; and routing-based adaptive query processing does away with the optimizer altogether. Despite this large body of interrelated work, no unifying comparison of adaptive query processing techniques or systems has been attempted; we tackle this problem. We identify three families of systems (plan-based, CQ-based, and routing-based), and compare them in detail with respect to the most important aspects of adaptive query processing: plan quality, statistics monitoring and re-optimization, plan migration, and scalability. We also suggest two new approaches to adaptive query processing that address some of the shortcomings revealed by our in-depth analysis: (1) "Proactive re-optimization," in which the optimizer chooses initial query plans with the expectation of re-optimization; and (2) "Plan logging," in which optimizer decisions under different conditions are logged over time, enabling plan re-use as well as analysis of relevant statistics and benefits of adaptivity.