Duplicate messages are an unfortunate possibility for any JMS solution, hence
the JMSRedelivered flag - which comes right out of the JMS specification.

At some point in a solution such as this, there must be a choice made
between allowing possible duplicates and possible lost messages.  JMS chose
possible duplicates.  Consider the following sequence of events which I hope
will make clear why this must be:

* Broker sends message to consumer
* Consumer receives message and processes
* Consumer sends ack
* Broker terminates before the ack is received or processed

As far as the consumer is concerned, it is done.  But, as you can see, the
broker cannot tell.

The typical solution is to implement idempotent processing.

Now, on to the pending message issue - something sounds very wrong here with
that scenario.  If the message stores between the two brokers are indeed
shared, there should be no way for the initially active broker to read
messages off disk that the second broker failed to read - they use the same
logic and code paths and are reading the same files.

Is it possible that somehow the two brokers don't actually share the same
files?  Or, could there be producers around that don't fail-over to the
second broker and therefore look like they have created pending messages on
the first broker that only succeed in being sent after it returns to
service?



--
View this message in context: 
http://activemq.2283324.n4.nabble.com/Master-slave-shared-FS-failover-resulting-in-pending-msgs-that-are-not-consumed-until-failed-broker-n-tp4691679p4691692.html
Sent from the ActiveMQ - User mailing list archive at Nabble.com.

Reply via email to