activemq-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Gary Tully (JIRA)" <j...@apache.org>
Subject [jira] Updated: (AMQ-2149) Shared Filesystem Master Slave: missing messages
Date Thu, 26 Mar 2009 15:10:43 GMT

     [ https://issues.apache.org/activemq/browse/AMQ-2149?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Gary Tully updated AMQ-2149:
----------------------------

    Attachment: amq2149.patch

here is a patch for this issue. There were a bunch of related problems. Duplicate messages
from that failover transport are replayed when the connection does down during a message send,
the message is replayed. This can then be a duplicate.

One problem with the persisted hashindex that backs the reference store could result in a
spurious message reference and "could not be recovered, already dispatched message". 
In addition, if the duplicate message is received before the existing message is dispatched
and acked, the duplicate reference remains and will eventually get dispatched on restart or
when the duplicate checker exceeds its range.
Eliminating the duplicate reference at source in the reference store resolves both these issues.
When memory limits are reached, it was possible for a stores messages to be exhausted (cursor
gets to the end) causing dispatch to halt until a restart. Dealing with recovery failure due
to no space resolves this.
A consumer ack of a duplicate message could be replaced with a delivery ack and get lost,
thus leaving the duplicate to be redsipatched on restart and delivered to the consumer when
the client side message audit is exceeded. This explains redelivery of old messages. Ensuring
each duplicate is acked in turn resolves this.
Finally, an unmatched ack can occur if recovery dispatch has not yet happened after a restart
and an ack is received from a consumer as it was outstanding. In this case, the subscription
can wait for recovery and a dispatch to kick in such that it will have a record of the target
message. 


> Shared Filesystem Master Slave: missing messages
> ------------------------------------------------
>
>                 Key: AMQ-2149
>                 URL: https://issues.apache.org/activemq/browse/AMQ-2149
>             Project: ActiveMQ
>          Issue Type: Bug
>    Affects Versions: 5.2.0
>         Environment: Ubuntu Linux 8.10 AMD64, Sun JDK 1.6.0.10
>            Reporter: Aaron Riekenberg
>            Assignee: Gary Tully
>         Attachments: activemq.log, activemq.log.2009_03_12_1, activemq.log.2009_03_12_2,
activemq.xml, AMQ-2149.zip, amq2149.patch, MasterSlaveTest.java, MasterSlaveTestWithTransactions.java,
run_master_slave_brokers.sh
>
>
> I'm finding occasionally messages are not delivered in order in a shared filesystem master
slave setup when the master fails and the slave takes over.  I'm running a simple test on
one physical machine where the shared filesystem is on a single disk (no SAN currently involved).
> I'm attaching a shell script (run_master_slave_brokers.sh) that starts a master and slave
broker in the same directory, sleeps 20 seconds, kills the master, sleeps 20 seconds, starts
a new slave, sleeps 20 seconds, kills the master, etc.
> Also attached is a small java test program (MasterSlaveTest.java)  The program starts
10 JMS senders that send 75kb text messages every 25 ms to unique queues.  These messages
contain a sequence number header (a long).  The program also starts 10 receivers (1 for each
queue) that keep track of the next expected sequence number and validate each incoming sequence
number.  If a receiver gets an unexpected sequence number, the test program exits (System.exit(1)).
 Both the senders and receivers use the failover transport to connect to the broker.  Messages
being sent are persistent, so in theory there should be no message loss when the master fails
and slave takes over.
> I run the script to start the brokers, then run my test program.  Most times when the
script kills the master and the slave is promoted, things work fine - the test program reconnects,
and messages continue to be delivered in order.  If I run this long enough though, eventually
my test program fails just after a slave broker is promoted to master with output similar
to this:
> Mar 6, 2009 11:58:12 AM org.apache.activemq.transport.failover.FailoverTransport doReconnect
> INFO: Successfully reconnected to tcp://localhost:61616
> Mar 6, 2009 11:58:12 AM org.aaron.MasterSlaveTest$Receiver onMessage
> WARNING: test.queue.3 received 630 expected 629
> This indicates the receiver for test.queue.3 received message 630 after the slave broker
took over and missed message 629.
> This seems to happen more often when more senders and receivers are running and more
queues are in use.  If I run a single sender/receiver pair on 1 queue, it is very difficult
to make this happen.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


Mime
View raw message