f99af8529b
As reported in https://spark-project.atlassian.net/browse/SPARK-1104 By @pwendell: "Sometimes due to large shuffles executors will take a long time shutting down. In particular this can happen if large numbers of shuffle files are around (this will be alleviated by SPARK-1103, but nonetheless...). The symptom is you have DEAD workers sitting around in the UI and the existing workers keep trying to re-register but can't because they've been assumed dead." In this patch, I add lines in the handler of InterruptedException in workerThread of executorRunner, so that the process.destroy() and process.waitFor() can only block the workerThread instead of blocking the worker Actor... --------- analysis: process.destroy() is a blocking method, i.e. it only returns when all shutdownHook threads return...so calling it in Worker thread will make Worker block for a long while.... about what will happen on the shutdown hooks when the JVM process is killed: http://www.tutorialspoint.com/java/lang/runtime_addshutdownhook.htm Author: CodingCat <zhunansjtu@gmail.com> Closes #35 from CodingCat/SPARK-1104 and squashes the following commits: 85767da [CodingCat] add null checking and remove unnecessary killProce 3107aeb [CodingCat] address Aaron's comments eb615ba [CodingCat] kill the process when the error happens 0accf2f [CodingCat] set process to null after killed it 1d511c8 [CodingCat] kill Process in workerThread |
||
|---|---|---|
| .. | ||
| main | ||
| test | ||