An optimistic protocol for partitioned distributed database systems

A protocol for transaction processing during partition failures is presented which guarantees mutual consistency between copies of data-items after repair is completed. The protocol is "optimistic" in that transactions are processed without restrictions during the failure; conflicts are detected at repair time using a precedence graph and are resolved by backing out transactions according to some backout strategy. The protocol is then evaluated using simulation and probabilistic modeling. In the simulation, several parameters are varied such as the number of transactions processed in a group, the type of transactions processed, the number of data-items present in the database, and the distribution of references to data-items. The simulation also uses different backout strategies. From these results we note conditions under which the protocol performs well, i.e. conditions under which the protocol backs out a small percentage of the transaction run. A probabilistic model is developed to estimate the expected number of transactions backed out using most of the above database and transaction parameters, and is shown to agree with simulation results. Suggestions are then made on how to improve the performance of the protocol. Insights gained from the simulation and probabilistic modeling are used to develop a backout strategy which takes into account individual transaction costs and attempts to minimize total backout cost. Although the problem of choosing transactions to minimize total backout cost is, in general, NP-complete, the backout strategy is efficient and produces very good results. Details of how to efficiently implement the optimistic protocol are also discussed. In particular, efficient algorithms are given for constructing the precedence graph in the context of several different concurrency control algorithms. Finally, the question of using the semantics of transactions to decrease the number of transactions backed out is considered, and guidelines for future research are given.