Remembrance of streams past: overload-sensitive management of archived streams

This paper studies Data Stream Management Systems that combine real-time data streams with historical data, and hence access incoming streams and archived data simultaneously. A significant problem for these systems is the I/O cost of fetching historical data which inhibits processing of the live data streams. Our solution is to reduce the I/O cost for accessing the archive by retrieving only a reduced (summarized or sampled) version of the historical data. This paper does not propose new summarization or sampling techniques, but rather a framework in which multiple resolutions of summarization/sampling can be generated efficiently. The query engine can select the appropriate level of summarization to use depending on the resources currently available. The central research problem studied is whether to generate the multiple representations of archived data eagerly upon data-arrival, lazily at query-time, or in a hybrid fashion. Concrete techniques for each approach are presented, which are tied to a specific data reduction technique (random sampling). The tradeoffs among the three approaches are studied both analytically and experimentally.