Abstract: A variable checkpoint mechanism in a streams manager checkpoints a streaming application based on periodic time periods for checkpoints. The variable checkpoint mechanism can take a checkpoint early before a periodic time period ends or late after the periodic time period ends based on predicted size of one or more tuple windows in the streaming application. The time for taking the checkpoint can be selected based on multiple checkpoint timing criteria, which include storage requirement for the checkpoint and predicted backpressure in the flow graph. In this manner the checkpoint timing of the variable checkpoint mechanism can be adjusted real-time to minimize the negative impact of checkpointing on the performance of the streaming application.

Abstract: A variable checkpoint mechanism in a streams manager checkpoints a streaming application based on periodic time periods for checkpoints. The variable checkpoint mechanism can take a checkpoint before a periodic time period ends when a spike is coming, or can take a checkpoint after the periodic time period ends when there is backpressure in a consistent region of the streaming application. When there is no anticipated spike coming and when there is no backpressure in a consistent region of the streaming application, the checkpoint is performed at the normal end of the periodic time period for checkpoints. In this manner the checkpoint timing of the variable checkpoint mechanism can be adjusted real-time to minimize the negative impact of checkpointing on the performance of the streaming application.

Abstract: A latency reconfiguration mechanism in a streams manager detects latency mismatches in a streaming application that affect the performance of the streaming application, and reconfigures the streaming application to minimize the latency mismatches. The reconfiguration can include fusing adjacent processing elements into a single processing element, co-locating adjacent processing elements on the same machine, moving processing elements and/or network links, adding parallel processing elements, and rearranging processing elements and/or network links to create more consistent latency paths. By minimizing latency mismatches and providing more consistent latency paths, the performance of a streaming application is increased by not having to perform significant reordering of data tuples.