Andreas, Your spec added two configuration elements your previous post didn't mention, and I'd like to eliminate each of them in turn to see if it's causing/contributing to the problem.
1. Your networkConnectors are apparently multicast. Please see what happens if you configure them as static:(tcp://host:port?tcpOptions)?staticOptions, to take the multicast (and the broker discovery that it's presumably doing) out of the equation. I recently experimented with what happens when the failover is allowed to perform a reconnect in a broker-to-broker networkConnector, and the result is duplicate and/or stale subscriptions between the brokers. That behavior could explain what you're seeing, if multicast is similarly performing reconnects without notifying the static wrapper so it can recreate the network bridge, so let's take it out of the equation to see if the behavior changes. (I've never used multicast, so this might not make sense; if someone knows that this can't be the issue, please say so.) 2. I don't know how gracefully conduitSubscriptions reacts to consumers moving around the network of brokers; I don't believe this should be the problem, but if #1 doesn't produce any change in behavior, can you set conduitSubscriptions=false and see if anything changes? I'm not clear on how Virtual Topics will solve the problem; can you explain? To me this feels like a problem with broker-to-broker management of subscriptions made on behalf of clients (most likely duplicate subscriptions for a client, one from each broker, after a failover), and I'm not sure how a Virtual Topic would make it any better if that's the case. But if you know of a way that it would, that might help me to understand what's going on. Tim On Sat, Nov 1, 2014 at 9:50 AM, Andreas Gies <andr...@wayofquality.de> wrote: > Hello Tim, > > thanks for your answer. It took me a bit to digest it - so my apologies for > the delay in my answer. > > I have come up with a test case that shows - and unfortunately confirms > my observations.The test case is located at [1]. > > Here is the excerpt of my problem descriptions & observations: > > /** > * This specification shall help to investigate the duplicate delivery of > messages for durable subscribers > * within a network of brokers. The problem has been posted on the > ActiveMQ mailing list on Oct. 18th 2014 > * and was described as follows: > * > * Suppose you have a network of brokers consisting of two members > discovering each other via multicast. > * The network bridge is set up using conduit subscriptions. Now assume > that we have a durable subscriber > * named "S" that connects to the network of brokers using a failover uri > pointing to both brokers. > * > * First, the subscriber connects to Broker A. It will consume all > messages published to either Broker A or B. > * Now the subscriber disconnects and stays offline for a bit, then it > reconnects to Broker B. Now it will pick > * up all messages that have been published while it was offline. > * > * Let's say then 10 messages are published. All is well as the subscriber > consumes those messages. > * If the subscriber then disconnects and reconnects to Broker A, these 10 > messages will be consumed > * again by the reconnected subscriber. > * > * According to Tim Bain on Oct., 20th 2014 this indicates a bug rather > than a missing feature in ActiveMQ > * and this Spec shall pinpoint the behavior. > * * > * The test is based on ActiveMQ 5.10 > * > * Observations: > * ------------- > * Depending on when the durable subscriber is known to the members of the > NWOB, messages can be either left pending > * or delivered repeatedly (see the last 2 test cases). Message gaps can > occur, if the DS has only connected > * to one broker so far. If the DS then disconnects and after a while > reconnects to the other broker it wasn't > * connected to so far, it will not see the messages that have been > produced while it was offline (it will see > * those messages after reconnecting to broker 1). > * > * Dupilcate delivery will happen if the DS was already connected to both > brokers. From the broker's perspective > * it seems that those DS are handled as two distinct subscribers, so > effectively all messages that are published > * will eventually be delivered to both subscribers. > */ > > I know that Virtual Topics could solve the problem - however we in the > middleware team are not in control of that > particular client application and therefor we cannot change the consumer > from a DS to a queue consumer. > > Can you confirm that we are indeed looking at a missing feature or a bug > in ActiveMQ 5.10 ? - Otherwise i would > need to get my thinking cap back on and see how I could solve the problem > without changing the client code. > > [1] > https://github.com/woq-blended/blended/blob/master/blended-testing/blended-testing-activemq/src/test/scala/de/woq/blended/testing/activemq/DurableSubscriberSpec.scala > < > https://github.com/woq-blended/blended/blob/master/blended-testing/blended-testing-activemq/src/test/scala/de/woq/blended/testing/activemq/DurableSubscriberSpec.scala > > > > Thanks and best regards > Andreas > > > > On 20 Oct 2014, at 17:40, Tim Bain <tb...@alumni.duke.edu> wrote: > > > > If you have a network of brokers, messages on topics will be forwarded to > > whichever broker the consumer connects to, without duplicate delivery of > > any messages so long as no messages were processed by the consumer > without > > being ack'ed. If you were using queues, there's the potential for > messages > > to get stranded on a broker if no consumers are left, but this isn't > > possible for topics. (I'm not clear on the reason that topics can't get > > messages stranded even when consumers bounce between brokers, and > > unfortunately http://activemq.apache.org/networks-of-brokers.html > doesn't > > describe why that is.) > > > > So I think that ActiveMQ's base capabilities will do exactly what you > want, > > and if you're seeing redelivery of messages that were successfully acked > > when the consumer bounces to another broker, I think that would indicate > a > > bug in ActiveMQ rather than a missing feature. > > > > On Sun, Oct 19, 2014 at 6:05 PM, Noel OConnor <noel.ocon...@gmail.com> > > wrote: > > > >> Take a look at idempotent consumers in camel. This may help you out as a > >> basis for your plugin if you decide to go with it. > >> On Oct 18, 2014 5:47 PM, "Andreas Gies" <andr...@wayofquality.de> > wrote: > >> > >>> Hi > >>> > >>> I am using ActiveMQ 5.10 in an application. So far the requirement for > >> the > >>> remote locations has been for pure store and forward capabilities, > >>> so that a single AMQ broker was sufficient. This has changed in a way > >> that > >>> now 2 nodes should be present in the remote location for > >>> resilience and load balancing. I had considered a master/configuration > as > >>> the requirement for resilience is stronger than that for load > balancing. > >>> > >>> However, the situation in those locations is that I don’t have a shared > >> db > >>> nor a shared filesystem. As far as I have understood the replicated > >> level db > >>> would require at least 3 nodes ? > >>> > >>> This is why I have chosen a network of brokers in the end, which works > >>> well for any Queue based communication. > >>> > >>> Now my problem is that there is one client application that is provided > >> by > >>> a 3rd party and uses durable subscriptions. It would be quite an effort > >>> to change that application towards using queues, so that I could > consider > >>> virtual destinations. > >>> > >>> The problem occurs two-fold: > >>> > >>> Assume a Subscriber connects to BrokerA, then disconnects and > reconnects > >>> to Broker B. It consumes messages for a while, than disconnects > >>> and reconnects to Broker A. All messages that have already been > consumed > >>> while it was connected to Broker B will be delivered again. > >>> > >>> My question is now whether this could be avoided by means of ActiveMQ > >>> alone ? - I was contemplating a broker plugin to track messages that > >>> have been consumed on other nodes so that I could avoid redelivering > them > >>> again. > >>> > >>> Sorry if thats a bit vague - I am fishing for ideas …. > >>> > >>> > >>> Thanks and best regards > >>> Andreas > >> > >
