Adaptive selection of query execution strategies by learning automata (Q908718)

The traditional approach to evaluate query execution strategies using approximate cost models may be inadequate for particular environments. For instance, if the environment does not satisfy the assumptions made by the cost model, the cost estimates can be so distorted that expensive strategies will be chosen. We propose a new approach for choosing execution strategies based on the actual history of query execution under various strategies, rather than on assumption-loaded estimates of these costs. Adaptive selection automatically changes the strategies selected, tracking cost variations caused by changes in the database state and query load. Furthermore, it does not require any assumptions about internal database structures, data characteristics, or distribution of queries. Queries are divided into query classes, where all queries in a class share the same execution strategies. A learning automaton is then used for each class to infer over time which are the current best strategies, based on actual query execution costs. We show the results of running the adaptive selector using real query loads for an existing database.