Failure Recovery Alternatives in Grid-Based Distributed Query Processing: A Case Study

Fault-tolerance has long been a feature of database systems, with transactions supporting the structuring of applications so as to ensure continuation of updating applications in spite of machine failures. For read-only queries the perceived wisdom has been that support for faulttolerance is too expensive to be worthwhile. Distributed query processing (DQP) is coming to be seen as a promising way of implementing applications that combine structured data and analysis operations in dynamic distributed settings such as computational grids. Accordingly, a number of protocols have been described that support tolerance to failure of intermediate machines, so as to permit continuation from surviving intermediate state. However, a distributed query can have a non-trivial mapping onto hardware resources. Because of this it is often possible to choose between a number of possible recovery strategies in the event of a failure. The work described here makes an initial investigation in this area in the context of an example query expressed over distributed resources in a Grid and shows that it can be worthwhile to make this choice between recovery alternatives dynamically, at the point a query is detected rather than statically beforehand.