[
https://issues.apache.org/jira/browse/AMQ-7067?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16643245#comment-16643245
]
Gary Tully commented on AMQ-7067:
---------------------------------
[~cshannon] We just need the lock for protection, no tx, b/c the index is not
updated.
On the compaction, I think that is the simplest approach.
Ideally we would have some sort of reference count on the referenced ids in the
ackMessageFileMap - allowing each ack and outcome to be recorded and having
some way to know there are just non ack references remaining.
We don't need to retain everything. i.e.: the prepare record can go if there is
an outcome, and if the outcome is rollback then both the prepare and rollback
can go b/c recovery will default to rollback. For commit, both xa and non xa,
it needs to be retained as long as the updates (add/ack commands).
It will get complex fast however.
Getting it correct first will suffice, which means retaining the transaction
related commands. I will see if I can wrangle a test to reproduce to get the
full context.
> KahaDB Recovery can experience a dangling transaction when prepare and commit
> occur on different data files.
> ------------------------------------------------------------------------------------------------------------
>
> Key: AMQ-7067
> URL: https://issues.apache.org/jira/browse/AMQ-7067
> Project: ActiveMQ
> Issue Type: Bug
> Components: KahaDB, XA
> Affects Versions: 5.15.6
> Reporter: Jamie goodyear
> Assignee: Gary Tully
> Priority: Critical
> Fix For: 5.16.0, 5.15.7
>
> Attachments: amq7067test.patch
>
>
> KahaDB Recovery can experience a dangling transaction when prepare and commit
> occur on different data files.
> Scenario:
> A XA Transaction is started, message is prepared and sent into Broker.
> We then send into broker enough messages to file page file (100 message with
> 512 * 1024 characters in message payload). This forces a new data file to be
> created.
> Commit the XA transaction. Commit will land on the new data file.
> Restart the Broker.
> Upon restart a KahaDB recovery is executed.
> The prepare in PageFile 1 is not matched to Commit on PageFile 2, as such, it
> will appear in recovered message state.
> Looking deeper into this scenario, it appears that the commit message is
> GC'd, hence the prepare & commit can not be matched.
> The MessageDatabase only checks the following for GC:
> {color:#808080}// Don't GC files referenced by in-progress
> tx{color}{color:#cc7832}if {color}(inProgressTxRange[{color:#6897bb}0{color}]
> != {color:#cc7832}null{color}) {
> {color:#cc7832}for {color}({color:#cc7832}int
> {color}pendingTx=inProgressTxRange[{color:#6897bb}0{color}].getDataFileId(){color:#cc7832};
> {color}pendingTx <=
> inProgressTxRange[{color:#6897bb}1{color}].getDataFileId(){color:#cc7832};
> {color}pendingTx++) {
> gcCandidateSet.remove(pendingTx){color:#cc7832};{color} }
> }
> We need to become aware of where the prepare & commits occur in pagefiles
> with respect to GCing files.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
