A store directory is created by MessageDatabase#getPageFile which
is called in two cases:
1. KahaDBStore.start() when creating a queue
2. KahaDBStore.size() which is performed when sending any persistent message
If both methods are called concurrently it's possible to get an IOException
thrown from the IOHelper.mkdirs method.
(cherry picked from commit 7de7ba2aa9)
This is best practice and will prevent unlock from being attempted
inside of a finally block when the thread doesn't actually own the
lock which can happen when the lock attempt throws an exception
such as calling lockInterruptibly()
(cherry picked from commit ed924cddac)
This should improve test reliability for the unit tests so brokers don't
hang around after the end of a test on error. Also increase the surefire
re-run count to 3 times before failing.
(cherry picked from commit a083ff4d23)
(cherry picked from commit 393a696955)
Scenario on client:
1. Employing RedeliveryPolicy with exponential backoff (keeping maximum
redeliveries at default 6)
2. Enabled non-blocking redelivery
3. Receiving e.g. 100 consecutive poison messages (which eventually
should DLQ after max redeliveries)
This will result in massive redelivery delays due to a logic bug.
The reason is that redeliveryDelay is a field variable kept on the
ActiveMQMessageConsumer, instead of being a property on the message - or
that the redelivery delay was calculated per message based on the
redelivery count.
When consecutive messages rollbacks multiple times, the redeliveryDelay
field is continuously multiplied by the backoff multiplier, resulting in
enormous delays.
Fix: Ditch the field variable, instead calculating the redeliveryDelay
per delivery from the redelivery count. (This happens to be identical to
how it is done in afterRollback() in ActiveMQSession:1004.)
Test is added - which fails with the previous code, and passes with
this. Added a debug log line for the calculated delay.
Topic subscriptions expire a message
This fixes topic subs to send the right advisory type, if enabled, when
the server discards a message on dispatch to a topic sub. Also add some
more expiration tests for other subscription types
(cherry picked from commit 757a712890)
current dispatched count
The previous way of computing the count of using total dispatched minus
total dequeued didn't work in the case of destination removal and
messages were not acked. The counter is needed as the dispatched list is
optional unlike prefetch subs.
(cherry picked from commit 8554a1464c)
managed region broker
This new approach just looks matching Subscriptions from the region for the
destination which prevents having to store another map and falls back to
the old approach if something went wrong.
(cherry picked from commit d46b74d674)
Due to changes with Queues to check if consumers are full before adding
more messages to the subscription, the Queue dispatch logic needed to be
updated to mark subscriptions as slow and send advisories if configured
instead of relying on the subscription itself to do it.
(cherry picked from commit 596ee31687)
JB Onofré
Matt Pavlovich
Jean-Baptiste Onofré
Michał Janczykowski
Christopher L. Shannon (cshannon)
Dmytro Chebotarskyi
Endre Stølsvik
Jonathan Gallimore
Lucas Tétreault