HornetQ JMS Topic & GroupId problem on failover

[mactex]

Hi all,

I have the following scenario.

We have a topic where the server places notifications for clients. We absolutely need to be sure of message ordering and we don't want to miss any messages in case the client get disconnected for any reason. So we use durable subscriptions and GroupIds. This works well, messages are received always in sequential order and after a reconnection messages are always received.

Now we want to enable clustering (symmetric cluster with colocated backup) and there come the problems. This works in clustering as well, until a failover happens.

When the node dies where our subscription is and a failover happens during the time the client is disconnected, then after reconnection the client will ALWAYS get connected to another live instance, since JMS connection to a backup instance is not supported. But messages are still sent to the backup instance, since the GroupId pins this durable subscription there. I.e. messages are not received until the dead node comes back again.

Here is our test case:

HA-JNDI client1 with url1,url2

1. client1 connected with id1
2. publish 10 msgs with order index – client1 should receive 10msgs in order
3. check which node has the subscriber
4. disconnect client1
5. publish 10 msgs with order index – msgs should be visible as durable messages
6. kill jboss instance on the node from step3
7. publish 10 msgs with order index - msgs should be visible as durable messages
8. reconnect client1 with id1 – client1 should receive 20msgs in order

Messages are not received in the last step.

We do think this is a problem with HornetQ and this should work.

Do you have any idea how this use case could be made to work?

Thanks for any input!

[ataylor]

Are you saying that your consumer is connecting to a different live server? this shouldnt happen, a consumer will either reconnect to the original live server or connect to its backup.

If this is the case can you provide a simple test so we can easily recreate the issue?

[mactex]

Yes, exactly this is happening. When a failover happened and the consumer is disconnected than the consumer is not reconnected to the backup of the original server.

I attached a test case.

[ataylor]

I will need a bit more than that, server configurations, topology, how you are starting the servers and instructions as to how to run the test etc etc.

[mactex]

Ok, I attach server configurations.

The server is JBoss  5 with HornetQ embedded. It's a symmetrical cluster with colocated backup. I have 2 nodes.

You need to create a topic named TestTopic1 and there are some parameters at the beginning of the testcase which you need to configure, especially: JNP_URL, CONNECTION_FACT_NAME.

Then just run the test cases, it outputs when you need to kill one jboss instance.

It does exactly what I wrote above in the testcase.

Thanks!

[ataylor]

ok, its a public holiday in the uk today but i ill take a look tomorrow. what HornetQ version are you using by the way?

[mactex]

Ok thanks. I'm using 2.2.24.

[ataylor]

Any chance you could provide me with a script or something to set up the env, Ive spent a couple of hours on this and can'tr really get it running.