CS262A Reading Summary 24

Eddies: Continuously Adaptive Query Processing

R. Avnur and J. M. Hellerstein
Summary by Feng Zhou
10/31/2002

1. A noval architecture is proposed to do dynamic reoptmization in decentralized dataflow query processing. Instead of determine an static optimal global query plan before execution based on knowledge of costs of each operation, it relies on simple dynamic local decisions to get a good global strategy. This is useful when cost knowledge is hard to obtain or data characteristics are continuously changing.
2. The concepts of "Syncrhonization Barrier" and "Moments of Symmetry" regarding join algorithms provide insights into the possibility of reoptimization of different algorithms. From this point of view, the authors convincingly argue that ripple joins are a group of good algorithms that provide near-optimal static performance and allow flexible dynamic reoptimization.
3. "Lottery Scheduling" is utilized to do local scheduling based on selectivity. By adding a "feedback" to the control path, it effectively favors the low-selectivity operators and thus corresponds to the selectivity heuristic in static query optimization.

1 flaw:

No formal demonstration or analysis is given to prove that the "eddy" scheme will give optimal or near-optimal query performance, or on the other hand, under what circumstance it will perform badly.