Abstract
Data stream processing systems have become ubiquitous in academic and commercial sectors, with application areas that include financial services, network traffic analysis, battlefield monitoring and traffic control. The append-only model of streams implies that input data is immutable and therefore always correct. But in practice, streaming data sources often contend with noise (e.g., embedded sensors) or data entry errors (e.g., financial data feeds) resulting in erroneous inputs and by implication, erroneous query results. Many data stream sources (e.g., Reuters ticker feeds) issue "revision tuples" ( revisions ) that amend previously issued tuples (e.g. erroneous share prices). A stream processing engine might reasonably respond to revision inputs by generating revision outputs that correct previously emitted query results. We know of no stream processing system that presently has this capability.
In this paper, we describe how a stream processing engine can be extended to support revision processing via replay. Replay-based revision processing techniques assume that a stream engine maintains an archive of recent data seen on each of its input streams. These archives are then queried in response to a revision, with the resulting tuples replayed through the system so as to generate corrected query outputs. We first present the design and implementation of the revision processing engine for the Borealis stream processing engine [1]. We then compare techniques for archiving streams to support replay, and then compare the performance and overhead of two revision processing techniques that replay input tuples to recompute and thereby revise previously output query results. These experiments reveal scalability issues due to the overhead required to maintain stream archives, and has motivated our current research on using sampling and data summarization (e.g., histograms) to reduce the data that must be stored in a stream archive.