On Atomic Batch Executions in Stream Processing

Abstract

Stream processing is about processing continuous streams of data by programs in a workflow. Continuous execution is discretized by grouping input stream tuples into batches and using one batch at a time for the execution of programs. As source input batches arrive continuously, several batches may be processed in the workflow simultaneously. A general requirement is that each batch be processed completely in the workflow. That is, all the programs triggered by the batch, directly and transitively, in the workflow must be executed successfully. Executing only a prefix of the workflow amounts to dropping (discarding) the batches that were derived by the executed part and were supposed to be input to the rest of the workflow. In some cases, such partial executions may not be acceptable and may have to be rolled back, amounting to dropping the source input batches that were processed by the partial execution. We refer to this property of processing the batches either completely or not at all as atomic execution of the batches. We also attribute the property to the batches themselves, calling them atomic batches, meaning that the property applies to the set of transactions that are executed due to that batch. If batches are processed in isolation in the workflow, preserving atomicity is fairly straightforward. When batches are split or merged along the workflow computation, the problem becomes complicated. In this paper, we study issues relating to the atomicity of batches. We illustrate that, in general, preserving atomicity of some batches may affect the atomicity of some other batches, and suggest trade-offs.