Orchestrating safe streaming computations with precise control
Peng Li, Kunal Agrawal, Jeremy D. Buhler, Roger D. Chamberlain · 2014
Streaming computing is a paradigm of distributed computing that features networked nodes connected by first-in-first-out data channels. Communication between nodes may include not only high-volume data tokens but also infrequent and unpredictable control messages carrying control information, such as data set boundaries, exceptions, or reconfiguration requests. In many applications, it is necessary to order delivery of control messages precisely relative to data tokens, which can be especially challenging when nodes can filter data tokens. Existing approaches, mainly data serialization protocols, do not exploit the low-volume nature of control messages and may not guarantee that synchronization of these messages with data will be free of deadlock. In this paper, we propose an efficient messaging system for adding precisely ordered control messages to streaming applications. We use a credit-based protocol to avoid the need to tag data tokens and control messages. For potential deadlocks caused by filtering behavior and global synchronization, we propose deadlock avoidance solutions and prove their correctness.