[
https://issues.apache.org/jira/browse/AMQ-4683?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13748380#comment-13748380
]
Andrey Zavodnik commented on AMQ-4683:
--------------------------------------
The current fix (revision 1516561) makes JobSchedulerImpl to wait on a
semaphore until the "main" thread executes either doStart() or addListener().
This fix has two big issues:
1. If the "main" thread manages to execute addListener() before the
JobSchedulerImpl thread arrives to semaphore wait() point, the JobSchedulerImpl
thread will be blocked indefinitely.
2. JobSchedulerImpl is designed to accept an arbitrary number of listeners.
This solution will only work when there is exactly one. If zero, worker will be
blocked. If more than one, there is a chance that listeners number 2 and over
will not receive a notification, just like the first one didn't in the original
bug.
> Scheduler discards overdue messages on startup
> ----------------------------------------------
>
> Key: AMQ-4683
> URL: https://issues.apache.org/jira/browse/AMQ-4683
> Project: ActiveMQ
> Issue Type: Bug
> Components: Broker
> Affects Versions: 5.8.0
> Environment: Windows 7 64bit, JDK 1.6.0_26
> Reporter: Andrey Zavodnik
> Assignee: Timothy Bish
> Priority: Critical
> Labels: scheduler
> Fix For: 5.9.0
>
> Attachments: LostScheduledMessagesTest.java
>
>
> When message is scheduled for a delivery (using AMQ_SCHEDULED_PERIOD
> property) it is not posted to the queue, but saved to the scheduler’s
> repository instead. If Broker instance is shut down at the time for which the
> message is scheduled for a delivery, the scheduler will try to process the
> message at the next time broker is started up. Due to a bug in the scheduler
> implementation such messages may be dismissed rather than posted to the queue
> depending on the racing conditions inside the JVM.
> Actual bug resides in
> org.apache.activemq.broker.scheduler.SchedulerBroker.getInternalScheduler()
> method.
> This method first calls JobSchedulerStoreImpl. getJobScheduler(String name)
> method which in turn creates an instance of runnable JobSchedulerImpl class
> and starts it.
> Then SchedulerBroker registers itselfas a listener of the newly created
> JobSchedulerImpl.
> Unfortunately this may happen after the JobSchedulerImpl have dismissed all
> the “missed” tasks as “fired” as there were no listeners configured.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
