Multi-Mode Stream Processing For Hopping Window Queries

Wei, Mingrui

Etd

Multi-Mode Stream Processing For Hopping Window Queries

Public

Window constraints are mechanisms to bound the tuples processed by continuous queries specified over unbounded data streams. While sliding window queries move the constraint window upon the arrival of each individual tuple, hopping window queries instead move the window by a fixed amount after some period, thus periodically refreshing their results. We observe that for large hops, techniques liked delta result updating may not be efficient -- as large portions of the tuples in the current window will be different from the previous window and thus must be maintained. On the other hand, the complete result updating technique, which has been found to be less suitable for sliding windows queries. Compute the next result based on the complete current window now can be shown to be superior in performance for some hopping windows queries. A trade-off emerges between the complete result method which has a lower per tuple processes cost but potentially processing redundant results versus the delta result method which has no redundant processing but pays a higher per tuple processing cost. On top of that, strict non-monotonic operators such as difference operator, cause premature expiration due to operator semantics. Negative tuples are needed for this kind of special expiration. Such negative tuples added extra burden to the stream engine. Thus, in streaming processing, the difference operator is typically suggested to be placed on top of the query plan despite its potential ability to reduce cardinality of the stream. With this thesis, we introduce a whole solution for hopping window query processing which includes an optimizer for generalized hopping window query optimization that exploits both processing techniques within one integrated query plan alone with query plan rewriting. First, we design the query operators to be multi-mode, that is, to be able to take either a delta or a complete result as input, and produce either a delta result or complete result as output. Then we design a cost model to be able to chose the optimal mode for each operator. Thirdly, our optimizer targets to configure each operator within a query plan to work in the suitable mode to achieve minimum overall processing costs. Last but not least, two query optimization techniques have been adopted. One explores all possibilities of pushing the difference down past joins using dynamic programming and assigning optimal mode at the same time, the other applies heuristic difference push down rule. The proposed techniques has been implemented within the WPI stream query engine, called CAPE. Finally, we show the benefit of our solution with a vast number of experimental results.

Creator