[jira] [Issue Comment Edited] (BOOKKEEPER-215) Deadlock occurs under high load

["Aniruddha (Issue Comment Edited) (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/BOOKKEEPER-215?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13251924#comment-13251924 ] 

Aniruddha edited comment on BOOKKEEPER-215 at 4/11/12 9:23 PM:
---------------------------------------------------------------

Hi Ivan and Flavio, 

I've attached the thread dump of the hedwig-server when it is deadlocked. As you can see, both pool-1-thread-1 and pool-52-thread-1 are blocked on the semaphore. pool-52-thread-1 is one of the threads in the mainWorkerPool. It seems that PendingReadOp is executed in pool-1-thread-1 always, except for just before I get the error, when it's executed in pool-52-thread-1. Now, when the channel is disconnected, the errorOut* functions enqueue all their SafeRunnables in this same thread (using OrderedSafeExecutor.submitOrdered with ledgerId as the key) - pool-52-thread-1. Even when PerChannelBookieClient closes the channel after errorOutOutstandingEntries, the callbacks that would release the semaphore are never called because all of them are being queued in the same thread. So, now you have a PendingReadOp waiting to acquire a lock at the front of the queue while operations that will release the lock are queued behind it. 

In short, there is a PendingReadOp being queued in pool-52-thread-1 which shouldn't be there. The only operations in that thread should be PendingAddOps. 
                
      was (Author: i0exception):
    Hi Ivan and Flavio, 

I've attached the thread dump of the hedwig-server when it is deadlocked. As you can see, both pool-1-thread-1 and pool-52-thread-1 are blocked on the semaphore. pool-52-thread-1 is one of the threads in the mainWorkerPool. It seems that PendingReadOp is executed in pool-1-thread-1 always, except for just before I get the error, when it's executed in pool-52-thread-1. Now, when the channel is disconnected, the errorOut* functions enqueue all their SafeRunnables in this same thread (using OrderedSafeExecutor.submitOrdered with ledgerId as the key) - pool-52-thread-1. Even when PerChannelBookieClient closes the channel after errorOutOutstandingEntries, the callbacks that would release the semaphore are never called because all of them are being queued in the same thread. So, now you have a PendingReadOp waiting to acquire a lock at the front of the queue while operations that will release the lock are queued behind it. 
                  
> Deadlock occurs under high load
> -------------------------------
>
>                 Key: BOOKKEEPER-215
>                 URL: https://issues.apache.org/jira/browse/BOOKKEEPER-215
>             Project: Bookkeeper
>          Issue Type: Bug
>          Components: hedwig-server
>    Affects Versions: 4.1.0
>            Reporter: Aniruddha
>            Priority: Critical
>         Attachments: hedwig_ts.log
>
>
> LedgerHandle uses a Semaphore(opCounterSem) with a default value of 5000 permits to implement throttling for outstanding requests. This is causing a deadlock under high load. What I've observed is the following - There are a fixed number of threads created by OrderedSafeExecutor(mainWorkerPool in BookKeeper) and this is used to execute operations by PerChannelBookieClient. Under high load, the bookies are not able to satisfy requests at the rate at which they are being generated. This exhausts all permits in the Semaphore and any further operations block on lh.opCounterSem.acquire(). In this scenario, if the connection to the bookies is shut down, channelDisconnected in PerChannelBookieClient tries to error out all outstanding entries. The errorOutReadKey and errorOutAddKey functions enqueue these operations in the same mainWorkerPool, all threads in which are blocked on acquire. So, handleBookieFailure is never executed and the server stops responding. 
> Blocking operations in a fixed size thread pool doesn't sound quite right. Temporarily, I fixed this by having another ExecutorService for every PerChannelBookieClient and queuing the operations from the errorOut* functions in it, but this is just a quick fix. I feel that the server shouldn't rely on LedgerHandle to throttle connections, but do this itself. Any other ideas on how to fix this? I'd be happy to contribute a patch. 

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira