| Databases are distributed and replicated to increase the efficiency and resiliency of a database management system; unfortunately, mechanisms to improve resiliency can have a substantial negative impact on the overall efficiency of the system. Current approaches to resiliency extend transaction execution to include multiple copies of entire (or key portions of) transactions to increase the likelihood that transactions survive intervening failures, adding substantial overhead to the system in anticipation of failures that may or may not occur. My research introduces a new, optimistic approach to fault tolerance: reconstructing transactions interrupted by failures. Reconstruction duplicates only those components of active transactions that are actually lost when failures occur, offering superior efficiency while retaining a level of resiliency comparable to that achieved by replication.;Successful reconstruction of transactions requires that control information, in addition to data operations, be recoverable. In contrast to a centralized control module vulnerable to failure, my system model is based on migrating control mechanisms that distribute control information to conform to the natural flow of execution among the transaction's participants and allow distributed algorithms to respond to intervening failures. My design also incorporates dynamic selection among multiple transaction execution strategies to utilize efficient optimized strategies when the system is stable, switching to flexible, fault-tolerant strategies only when dictated by the current state of the system. |
