Two-Stage Query Optimization in MPP Database Clusters

The query optimization process is segmented into two independent stages: global optimization at the coordinator level and local optimization at each data node level. The global optimizer generates an initial execution plan using generalized assumptions, while local optimizers at each data node independently refine the plan segments applicable to their local context using local statistics and configuration information. This segmentation allows each stage to operate with appropriate information scope without requiring complete information exchange.

The patent applies local quality by enabling each data node to optimize its local execution plan segments using locally available information such as local data distribution statistics and local resource configuration. Instead of requiring all nodes to use identical generalized assumptions, each node tailors the execution plan to its specific local characteristics, improving execution efficiency while maintaining consistency with the global query requirements.

The optimization system is segmented into global and local components. The global coordinator maintains a simplified architecture by generating execution plans based on generalized assumptions, while local optimizers at each data node handle the precision requirements by using local statistics and configuration information. This segmentation allows the central coordinator to remain simple while local nodes achieve high optimality.

Each data node performs self-service optimization by independently refining its local execution plan segments using its own local statistics and configuration information. The local optimizer at each node automatically adjusts the global execution plan to better suit local conditions without requiring the coordinator to have or manage detailed local information, thus maintaining architectural simplicity while improving plan quality.