To cope with bursty data arrivals, a stream query processor may perform load shedding to cut the system load by discarding some portion of tuples kept in memory. Load shedding can be conducted in a stateful operator such as join or aggregation in a query plan tree or in a dedicated operator, called load shedder, which is typically placed at the entry of the query processor. In this paper, we show that the load shedding can also be performed in window operators. In general, window operators are placed in the initial phase of a query plan and designed to ignore tuples whose arrivals are not in a predefined order. With the functionality to control the number of tuple discards, they can play the role of load shedding. In the proposed method, the conventional load shedder is not necessary and the number of query processing steps can be reduced, which leads to performance improvement of the continuous query processing.
展开▼
