[infinispan-dev] SysAdmin operations for recovering transactions

[Manik Surtani]

On 18 Mar 2011, at 12:13, Mircea Markus wrote:

> Hi,
> 
> It's about the stage where TM's recovery  process finds a in-doubt transaction and notifies the sys admin about it: what hooks does ISPN provide to the sys admin in order to "fix" the tx.
> E.g. step >= 3.3 : http://community.jboss.org/servlet/JiveServlet/showImage/102-16552-14-11811/3_non_originator_failure.png
> 
> Here is what I have in mind:
> 
> Expose (JMX) two operations:
> 
>   //all the params together fully describe a xid.
>   replayTx(byte[] txBranch, byte[] txId, int formatId); 
>   forceRollbackTx(byte[] txBranch, byte[] txId, int formatId);

You expect a sysadmin to type a byte array into a JMX console?  :-)  You might get death threats from sysadmins... 

> Here is how these two ops would work:
> A. replayTx 
>    1. the node has locally the PrepareCommand associated with that XID
> 	- re-issues a prepare: TransactionXAResource.prepare
> 	- if successful re-issues a commit: TransactionXAResource.commit
>        -if failure happens at any step the user is informed and she/he can re-do the JMX call
> 	- if success the recovery information is removed from the cluster (async)
>    2. the node doesn't have the PrepareCommand associated with that XID
> 	- broadcast ReplayTxCommand (Xid)
>        - when a node receives ReplayTxCommand
> 		- if doesn't have a PreparedCommand associated with the Xid ignores it
> 		- if has a PreparedCommand...
> 			- is it the first in the view that has it [1]? 

How does a node know the answer to this question?  Is the list of nodes that holds the prepare replay info stored on the PrepareCommand?

> 				- yes. Execute A.1then returns result to node that broadcasted ReplayTxCommand. This is guaranteed to happen on at most[2] one node in the cluster
> 				- no. Ignores it.
> 	- if success the recovery information is removed from the cluster (async)
> B.rollbackTx
>   - node broadcasts RollbackCommand
>   - each node that has the PrepareCommand forces a rollback
>   - each node that doesn't have the PreparedCommand ignores it
>   - if success the recovery information is removed from the cluster (async)
> 
> Cheers,
> Mircea
> 
> [1] this is determined by building the set of nodes on which tx spreads, based on tx's state. Then determine the first in the view. 
> [2] it is possible not to happen on any node as the PrepareCommand might had been removed from all nodes in between (node failures, expiration from the recovery cache). 
> 
> 
> 
> 
> 
> _______________________________________________
> infinispan-dev mailing list
> infinispan-dev at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/infinispan-dev

--
Manik Surtani
manik at jboss.org
twitter.com/maniksurtani

Lead, Infinispan
http://www.infinispan.org



-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.jboss.org/pipermail/infinispan-dev/attachments/20110318/1ee27317/attachment.html