[
https://issues.apache.org/activemq/browse/AMQ-2736?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=62560#action_62560
]
Eduardo Costa commented on AMQ-2736:
------------------------------------
After some evil debugging tricks, I found that there are some pending inflight
transactions (even after restarting the server) in my database files. In this
scenario, the gcCandidadeSet is empty, since
"metadata.firstInProgressTransactionLocation" points to the first log file
(MessageDatabase.java, lines 1140-1154):
{code}// Don't GC files after the first in progress tx
Location firstTxLocation = metadata.lastUpdate;
if( metadata.firstInProgressTransactionLocation!=null ) {
firstTxLocation = metadata.firstInProgressTransactionLocation;
}
if( firstTxLocation!=null ) {
while( !gcCandidateSet.isEmpty() ) {
Integer last = gcCandidateSet.last();
if( last >= firstTxLocation.getDataFileId() ) {
gcCandidateSet.remove(last);
} else {
break;
}
}
}
{code}
This leads to the question: why it must discard every log file starting from
the one containing the first pending transaction? Any malfunctioning operation
(even an one-time little mistake) will trash all log files from that point!
I believe this should happens also on "kill -9" or unexpected power down.
> KahaDB doesn't clean up old files
> ---------------------------------
>
> Key: AMQ-2736
> URL: https://issues.apache.org/activemq/browse/AMQ-2736
> Project: ActiveMQ
> Issue Type: Bug
> Affects Versions: 5.3.2
> Reporter: Adrian Trenaman
> Priority: Critical
> Attachments: AMQ-2736.zip, amq-2987-testcase.patch, amq-2987.patch,
> AMQ2736Test_should_with_this_diff.txt
>
>
> Over time, we're seeing that kahadb doesn't clean up old journal files. As a
> result, we eventually run out of disk space, or rather, we hit our usage
> limits and our producers are slowed down by the producer flow control
> mechanism. Others have experienced this problem too (for example, see
> http://mail-archives.apache.org/mod_mbox/activemq-users/201002.mbox/%3c27502591.p...@talk.nabble.com%3e)
> For now, we're moving back to the old amqPersistenceStore.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
