Dynamic Batch Interval Adjustment

[zhangqiang2]

The current Spark Streaming version cannot support the change of batch interval at runtime, given that the speed of input data streams may not highly dynamic from current Internet applications. If we have to do so, one must stop the program first, modify the corresponding code, and then restart the program. However, this will interrupt the execution of entire program, and may cause the data loss. Towards this end, our contribution is to implement a Dynamic Batch Interval Adjustment functionality that can help change the batch interval size at runtime.

This functionality contains two algorithms. One is dynamic adjustment, and the other is static adjustment. The former can predict the size of input data stream and as a result the processing time, by using the most recent processing time and used batch interval. In this way, one can decide whether the batch interval needs to be changed or not, to avoid the data backlog, and secure the system stability. On the other hand, the static adjustment needs the user to manually change the configuration file.

JIRA Issue: https://issues.apache.org/jira/browse/SPARK-19663
My report: https://github.com/floatingtony/System-Lever-Optimization-of-Spark-Streaming

[AmplabJenkins]

Can one of the admins verify this patch?

[srowen]

Please read http://spark.apache.org/contributing.html and start not with code but with a design doc. However, this has been discussed before on JIRA too IIRC and rejected.