When sending, for example, to a predefined anycast address and queue
from a multicast (JMS topic) producer, the routed count on the address
is incremented, but the message count on the matching queue is not. No
indication is given at the client end that the messages failed to get
routed - the messages are just silently dropped.
Fixing this problem requires a slight semantic change. The broker is now
more strict in what it allows specifically with regards to
auto-creation. If, for example, a JMS application attempts to send a
message to a topic and the corresponding multicast address doesn't exist
already or the broker cannot automatically create it or update it then
sending the message will fail.
Also, part of this commit moves a chunk of auto-create logic into
ServerSession and adds an enum for auto-create results. Aside from
helping fix this specific issue this can serve as a foundation for
de-duplicating the auto-create logic spread across many of the protocol
implementations.
- activemq.notifications are being transferred to the target node, unless an ignore is setup
- topics are being duplicated after redistribution
- topics sends are being duplicated when a 2 node cluster mirrors to another 2 node cluster, and both nodes are mirrored.
- interrupted message breaking reference counting
After the server writing to the client is interrupted in AMQP, the reference counting was broken what would require the server restarted
in order to cleanup the files of any interrupted sends.
- Removed consumer during large message delivery damaging large messages
If the consumer failed to deliver messages for any reason, the message on the queue would be duplicated. what would wipe out the body of the message
and other journal errors would happen because of this.
extra debug capabilities added into RefCountMessage as part of ARTEMIS-4206 in order to identify these issues
This commit fixes the following things:
- Moves connection audit logging to the resource audit logger instead
of using a dedicated logger as that would adversely impact upgrading
users, and arguably didn't make sense in the first place.
- Mitigates an potential NPE w.r.t. connection ID.
- Updates the "dummy" management connection to return a valid
connection ID.
This fix will delay the message.copy to the redistributor itself.
Meaning no copy would be performed if the redistribution itself failed.
No need to remove a copy any longer
The federated queue consumer has to generate a new id for the messages
received from the upstream broker because they have an id generated by
the store manager of the upstream broker.
Co-authored-by: Clebert Suconic <clebertsuconic@apache.org>
- redistribute received the handle call, it then copies the message
- the routing table changes
- the message is left behind
With the new version of the server these messages will be removed. But we should remove these right away
Basically I started the testsuite and attached check leak with "java -jar check-leak.jar --pid <pid> --report testsuite-report --sleep 1000" and saw the allocations of this were pretty high.
I have seen a werid intermittent failure in the testsuite
caused by some race where the EMPTY_SET ends up Not Empty! Causing weird failures that were really difficult to be investigated
This fix is scanning journal and paging for existing large messages. We will remove any large messages that do not have a corresponding record in journals or paging.
This is just an annoying thing. as the functionality would work properly if there was a previous value.
No other harm other than the cannot find TX record on startup for this, hence I'm fixing it.
There are certain use-cases where addresses will be auto-created and
never have a direct binding created on them. Because of this they will
never be auto-deleted. If a large number of these addresses build up
they will consume a problematic amount of heap space.
One specific example of this use-case is an MQTT subscriber with a
wild-card subscription and a large number of MQTT producers sending one
or two messages a large number of different MQTT topics covered by the
wild-card. Since no bindings are ever created on any of these individual
addresses (e.g. from a subscription queue) they will never be
auto-deleted, but they will eventually consume a large amount of heap.
The only way to deal with these addresses is to manually delete them.
There are also situations where queues may be created and never have
any messages sent to them or never have a consumer connect. These
queues will never be auto-deleted so they must be deleted manually.
This commit adds the ability to configure the broker to skip the usage
check so that these kinds of addresses and queues can be deleted
automatically.
there are two leaks here:
* QueueImpl::delivery might create a new iterator if a delivery happens right after a consumer was removed, and that iterator might belog to a consumer that was already closed
as a result of that, the iterator may leak messages and hold references until a reboot is done. I have seen scenarios where messages would not be dleivered because of this.
* ProtonTransaction holding references: the last transaction might hold messages in the memory longer than expected. In tests I have performed the messages were accumulating in memory. and I cleared it here.
The issue identified with AMQP was under Transaction usage, and while opening and closing sessions.
It seems the leak would be released once the connection is closed.
We added a new testsuite under ./tests/leak-tests To fix and validate these issues
Configurations employing shared-storage with NFS are susceptible to
split-brain in certain scenarios. For example:
1) Primary loses network connection to NFS.
2) Backup activates.
3) Primary reconnects to NFS.
4) Split-brain.
In reality this situation is pretty unlikely due to the timing involved,
but the possibility still exists. Currently the file lock held by the
primary broker on the NFS share is essentially worthless in this
situation. This commit adds logic by which the timestamp of the lock
file is updated during activation and then routinely checked during
runtime to ensure consistency. This effectively mitigates split-brain in
this situation (and likely others). Here's how it works now.
1) Primary loses network connection to NFS.
2) Backup activates.
3) Primary reconnects to NFS.
4) Primary detects that the lock file's timestamp has been updated and
shuts itself down.
When the primary shuts down in step #4 the Topology on the backup can be
damaged. Protections were added for this via ARTEMIS-2868 but only for
the replicated use-case. This commit applies the protection for
removeMember() so that the Topology remains intact.
There are no tests for these changes as I cannot determine how to
properly simulate this use-case. However, there have never been robust,
automated tests for these kinds of NFS use-cases so this is not a
departure from the norm.
Justin Bertram
Domenico Francesco Bruscino
Clebert Suconic
Alexey Markevich
Gary Tully
Rakhi Kumari