Consider a distributed experiment where data is being gathered at multiple sites and a query involves picking up data items from each location and combining the data to produce a composite result. Such an application may be structured as a master server front end and a multiple slave back ends. The front end gets the query and farms it out to to each of the participating sites. The slave sites process sub-queries locally, gathering results and returning them to the master site which then returns the combined result to the remote caller. If the combination operation is an involved one such as a database join, this could place an excessive burden on the master. Alternatively, this operation could be offloaded to the client or one of the slaves. When the query is received by the master, it creates an MStream at the client or one of the slaves to receive data from the data sources and and process the join.
If the data can be shipped incrementally from the data source, and results can be produced incrementally, the operation that receives results can be moved dynamically between the slaves and the client depending on available bandwidth and other machine resources. Such techniques are useful for optimizing dynamic query execution in client-server database systems. The ordering guarantee provided by MStreams ensures that the incremental join results are received in order by the join operator and the output operator regardless of the physical location of the join operator. In particular the join operator could be dynamically moving between sites as the join is being processed. Several positioning strategies may be considered in dynamically moving the join operator around. Some of these strategies are considered in [#!Ranga98Adapting!#] where we consider the more complex case of a join tree and adapt the operator placement to bandwidth variations. Figure 7 shows the overall organization of the system.