[jira] [Comment Edited] (HIVE-15879) Fix HiveMetaStoreChecker.checkPartitionDirs method

["Rajesh Balamohan (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HIVE-15879?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15881827#comment-15881827 ] 

Rajesh Balamohan edited comment on HIVE-15879 at 2/24/17 2:53 AM:
------------------------------------------------------------------

Thanks for sharing the details [~vihangk1]. I have a different point of view here.

I agree that ThreadPoolExecutor.getActiveCount() is approximate. It is approximate because,  by the time {{getActiveCount()}} iterates over the running threads in the worker list, it is possible that some of the threads which were executing are complete. http://hg.openjdk.java.net/jdk8/jdk8/jdk/file/687fd7c7986d/src/share/classes/java/util/concurrent/ThreadPoolExecutor.java#l1818. So the reported numbers could be slightly higher than the actually running threads. But it would never be less, as new Worker in ThreadPoolExecutor is added with {{mainLock}}. 

In the context of MSCK logic, this approximation should not be a problem. 

This is due to the check of "(pool.getActiveCount() < pool.getMaximumPoolSize())". In case threadpool executor reports approximate value (i.e higher than the actual number of threads), thread pool would not be used as per current logic. So in corner cases there can be instances where in threadpool executor could have been used, but failed due to the approximate (higher values)  reported by ThreadPoolExecutor. 


was (Author: rajesh.balamohan):
Thanks for sharing the details [~vihangk1]. I have a different point of view here.

I agree that ThreadPoolExecutor.getActiveCount() is approximate. It is approximate because,  by the time {{getActiveCount()}} iterates over the running threads in the worker list, it is possible that some of the threads which were executing are complete. http://hg.openjdk.java.net/jdk8/jdk8/jdk/file/687fd7c7986d/src/share/classes/java/util/concurrent/ThreadPoolExecutor.java#l1818. So the reported numbers could be slightly higher than the actually running threads. But it would never be less, as new Worker in ThreadPoolExecutor is added with {{mainLock}}. 

In the context of MSCK logic, this approximation should not be a problem. This is due to the check of "(pool.getActiveCount() < pool.getMaximumPoolSize())". In case threadpool executor reports approximate value (i.e higher than the actual number of threads), thread pool would not be used as per current logic. So in corner cases there can be instances where in threadpool executor could have been used, but failed due to the approximate (higher values)  reported by ThreadPoolExecutor. 

> Fix HiveMetaStoreChecker.checkPartitionDirs method
> --------------------------------------------------
>
>                 Key: HIVE-15879
>                 URL: https://issues.apache.org/jira/browse/HIVE-15879
>             Project: Hive
>          Issue Type: Bug
>            Reporter: Vihang Karajgaonkar
>            Assignee: Vihang Karajgaonkar
>         Attachments: HIVE-15879.01.patch
>
>
> HIVE-15803 fixes the msck hang issue in HiveMetaStoreChecker.checkPartitionDirs method by adding a check to see if the Threadpool has any spare threads. If not it uses single threaded listing of the files.
> {noformat}
>     if (pool != null) {
>       synchronized (pool) {
>         // In case of recursive calls, it is possible to deadlock with TP. Check TP usage here.
>         if (pool.getActiveCount() < pool.getMaximumPoolSize()) {
>           useThreadPool = true;
>         }
>         if (!useThreadPool) {
>           if (LOG.isDebugEnabled()) {
>             LOG.debug("Not using threadPool as active count:" + pool.getActiveCount()
>                 + ", max:" + pool.getMaximumPoolSize());
>           }
>         }
>       }
>     }
> {noformat}
> Based on the java doc of getActiveCount() below 
> bq. Returns the approximate number of threads that are actively executing tasks.
> it returns only approximate number of threads and it cannot be guaranteed that it always returns the exact number of active threads. This still exposes the method implementation to the msck hang bug in rare corner cases.
> We could either:
> 1. Use a atomic counter to track exactly how many threads are actively running
> 2. Relook at the method itself to make it much simpler. Like eg, look into the possibility of changing the recursive implementation to an iterative implementation where worker threads pick tasks from a queue until the queue is empty.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)