Distributed Adaptive Windowed Stream Join Processing

This paper presents an adaptive framework for processing a window-based multi-way join query over distributed data streams. The framework integrates distributed plan modification and distributed plan migration within the same scope by using a building block called the node operator set (NOS). An NOS is housed in each node that participates in the join execution, and specifies the set of atomic operations to be performed locally at the host node to execute its share of the global execution plan. The plan modification and migration techniques presented are for the case of updating the NOSs centralized at a single node and the case of updating them distributed at each node. The plan modification is triggered by the change of stream statistics and adjusts the join execution order and placement greedily to satisfy a cost invariant. The plan migration uses the distributed track strategy to accelerate the migration of window extents to new nodes. The migration of all window extents is synchronized. Experiments confirm the effectiveness of the developed adaptive framework on reducing the join execution cost and indicate a small additional adaptation-overhead for distributing the NOS update.