[jbosscache-issues] [JBoss JIRA] Created: (JBCACHE-1559) 0 ms MVCC timeout for invalidation commands

[Brian Stansberry (JIRA)]

0 ms MVCC timeout for invalidation commands
-------------------------------------------

                 Key: JBCACHE-1559
                 URL: https://jira.jboss.org/jira/browse/JBCACHE-1559
             Project: JBoss Cache
          Issue Type: Bug
      Security Level: Public (Everyone can see)
          Components: Locking
    Affects Versions: 3.2.1.GA
            Reporter: Brian Stansberry
            Assignee: Manik Surtani


User reports getting timeout exceptions w/ 0 ms timeouts despite configuring JBC for longer timeouts. See forum thread for specific details. Key bit is here:

> The 0 ms is coming from MVCCLockInterceptor:
> 
> Code:
> 
>    @Override
>    public Object handleInvalidateCommand(InvocationContext ctx, InvalidateCommand command) throws Throwable
>    {
>       // this should be handled the same as a recursive evict command.
>       ctx.getOptionOverrides().setLockAcquisitionTimeout(0);
>       if (!command.getFqn().isRoot()) helper.wrapNodesRecursivelyForRemoval(ctx, command.getFqn());
>       return invokeNextInterceptor(ctx, command);
>    }	
> 
> 
> 
> I personally don't understand the reasoning behind that; the semantics of an invalidation are not the same as an evict, since an invalidation is concerned with data correctness while an evict is concerned with the > lesser issue of managing memory.
> 
> Perhaps some of the folks who wrote this piece can comment?
> 
> This exception is thrown when a lock is attempted on a remote node. One thing I can think of is that with INVALIDATION_SYNC there is no point waiting more than 0 ms to acquire the lock since whatever is holding > the lock on this node is before releasing it also going to send an invalidation message. Which won't succeed because whatever sent this message will be holding a lock. So might as well fail promptly.
> 
> The above logic doesn't hold true with INVALIDATION_ASYNC though.

Galder and I just had a further discussion:

Brian: ok, so i think the logical flaw is just that 0 ms isn't correct for ASYNC
Galder: hmmm
Galder: why?
Galder: oh, because they could cross earch other?
Brian: well, first off, w/ ASYNC the TimeoutException won't propagate back to the sender, so only issue is what's the best state for the caches to end up in
Brian: so, yeah, they'll cross each other and both will invalidate
Brian: result is both caches are invalidated
Galder: but that's fine
Galder: they'll go back to db and get stuff from there
Brian: yep
Brian: w/ 0 ms timeout, whatever was holding the lock will succeed
Brian: because the invalidation msg from the other node will not succeed
Brian: so on one node you have potentially inconsistent data left in the cache
Galder: oh yeah, the remaining node has stale data
Galder: but u have the same problem with sync
Brian: w/ sync the TimeoutException propagates, allowing the caller to deal with it
Galder: yeah, the failure is at least known
Galder: the 2nd node could still succeed but you know the 1st did not
Galder: whereas with async, you don't know which one won
Brian: right


-- 
This message is automatically generated by JIRA.
-
If you think it was sent incorrectly contact one of the administrators: https://jira.jboss.org/jira/secure/Administrators.jspa
-
For more information on JIRA, see: http://www.atlassian.com/software/jira