Adaptive On-the-Fly Changes in Distributed Processing Pipelines.

Abstract

Distributed data processing systems have become the standard means for big data analytics. These systems are based on processing pipelines where operations on data are performed in a chain of consecutive steps. Normally, the operations performed by these pipelines are set at design time, and any changes to their functionality require the applications to be restarted. This is not always acceptable, for example, when we cannot afford downtime or when a long-running calculation would lose significant progress. The introduction of variation points to distributed processing pipelines allows for on-the-fly updating of individual analysis steps. In this paper, we extend such basic variation point functionality to provide fully automated reconfiguration of the processing steps within a running pipeline through an automated planner. We have enabled pipeline modeling through constraints. Based on these constraints, we not only ensure that configurations are compatible with type but also verify that expected pipeline functionality is achieved. Furthermore, automating the reconfiguration process simplifies its use, in turn allowing users with less development experience to make changes. The system can automatically generate and validate pipeline configurations that achieve a specified goal, selecting from operation definitions available at planning time. It then automatically integrates these configurations into the running pipeline. We verify the system through the testing of a proof-of-concept implementation. The proof of concept also shows promising results when reconfiguration is performed frequently.