[
https://issues.apache.org/jira/browse/AMQ-5854?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14904713#comment-14904713
]
Gary Tully commented on AMQ-5854:
---------------------------------
[~michael.s] the intent of the tracker to to maintain the transaction state,
and recover it, if it can. If the commit request gets a failure then it is
impossible to know if the message got there or if the reply was lost so
notifying the client is the only option. Otherwise failover can try and
recreate the transaction state and allow it to complete.
In hindsight, trying to maintain transaction semantics across failover
(especially in the context of multiple consumers) seems too complex.
I want to investigate the fix for this issue and see if it is better to
implement the fix or remove this feature altogether and just rollback any
indoubt transaction on failover.
> Duplicate messages when failover is done during prepare phase of two phase
> commit.
> ----------------------------------------------------------------------------------
>
> Key: AMQ-5854
> URL: https://issues.apache.org/jira/browse/AMQ-5854
> Project: ActiveMQ
> Issue Type: Bug
> Components: Broker, JMS client
> Affects Versions: 5.9.1, 5.10.2, 5.11.1
> Environment: Ubuntu or AIX
> ActiveMQ 5.9.1 (problem is reproduced in 5.10.2 and 5.11.1)
> Spring DMLC
> XA transactions with atomikos 3.7.0 (problem is also reproduced with 3.9.15)
> Persistent messages
> Multithreading (this problem occur when there is at least 2 consummers on a
> queue)
> Reporter: Michael
> Assignee: Gary Tully
> Attachments:
> ActiveMQMessageConsumer-5.10.2-ModifyWithThreadSleep.java,
> ActiveMQMessageConsumer-5.11.1-ModifyWithThreadSleep.java, amq5854.tar.gz
>
>
> Use case :
> With Spring DMLC, Read a jms message in a queue, produce a
> jms message in an output queue and write data in database.
> Problem description :
> Due to hight CPU usage, the inactity monitor closes
> connections between clients and broker while 16 messages were processed.
> {noformat}
> 2015-06-01 04:39:01,130 | WARN | Transport Connection to: tcp://*** failed:
> org.apache.activemq.transport.InactivityIOException: Channel was inactive for
> too (>30000) long: tcp://*** |
> org.apache.activemq.broker.TransportConnection.Transport | ActiveMQ
> InactivityMonitor Worker
> {noformat}
> 15 messages are rolled back and redilevered to another
> consummer.
> In the log we got 15 warnings :
> {noformat}
> ActiveMQMessageConsumer |WARN |jmsContainer-173|rolling back transaction
> (XID:***) post failover recovery. 1 previously delivered message(s) not
> replayed to consumer: ***
> {noformat}
> But one message is not rolled back (the transaction commit)
> and is also redileverd to another consummer. So it's processed twice by two
> different consummers (two inserts in database and two output JMS messages
> generated) and is not deduplicated.
> In the activeMq log we got the message :
> {noformat}
> WARN | Async error occurred: |
> org.apache.activemq.broker.TransportConnection.Service | ActiveMQ Transport:
> tcp:///***
> javax.jms.JMSException: Unmatched acknowledge:
> MessageAck {commandId = 6665, responseRequired = false, ackType = 2,
> consumerId = ID:***, firstMessageId =
> ID:***-50800-1433109620591-1:2:31356:1:1, lastMessageId =
> ID:***-50800-1433109620591-1:2:31356:1:1, destination = queue://***,
> transactionId =
> XID:[1096044365,globalId=47524f55505f3030303038736572766963657472616974656d656e7431363536373030343133,branchId=47524f55505f3030303038736572766963657472616974656d656e743137343737],
> messageCount = 1, poisonCause = null}; Could not find Message-ID
> ID:***-50800-1433109620591-1:2:31356:1:1 in dispatched-list (start of ack)
> {noformat}
> For this duplicated message, the failover occur during
> prepare phase of commit :
> {noformat}
> [{2015/06/01 04:39:50,322 |FailoverTransport |WARN
> |jmsContainer-152|Transport (tcp://***) failed, reason: , attempting to
> automatically reconnect}]
> org.apache.activemq.transport.InactivityIOException: Cannot send, channel has
> already failed: ***
> at
> org.apache.activemq.transport.AbstractInactivityMonitor.doOnewaySend(AbstractInactivityMonitor.java:297)
> at
> org.apache.activemq.transport.AbstractInactivityMonitor.oneway(AbstractInactivityMonitor.java:286)
> at
> org.apache.activemq.transport.TransportFilter.oneway(TransportFilter.java:85)
> at
> org.apache.activemq.transport.WireFormatNegotiator.oneway(WireFormatNegotiator.java:104)
> at
> org.apache.activemq.transport.failover.FailoverTransport.oneway(FailoverTransport.java:658)
> at
> org.apache.activemq.transport.MutexTransport.oneway(MutexTransport.java:68)
> at
> org.apache.activemq.transport.ResponseCorrelator.oneway(ResponseCorrelator.java:60)
> at
> org.apache.activemq.ActiveMQConnection.doAsyncSendPacket(ActiveMQConnection.java:1321)
> at
> org.apache.activemq.ActiveMQConnection.asyncSendPacket(ActiveMQConnection.java:1315)
> at
> org.apache.activemq.ActiveMQSession.asyncSendPacket(ActiveMQSession.java:1933)
> at
> org.apache.activemq.ActiveMQSession.sendAck(ActiveMQSession.java:2099)
> at
> org.apache.activemq.ActiveMQSession.sendAck(ActiveMQSession.java:2094)
> at
> org.apache.activemq.ActiveMQMessageConsumer.acknowledge(ActiveMQMessageConsumer.java:1083)
> at
> org.apache.activemq.ActiveMQMessageConsumer$5.beforeEnd(ActiveMQMessageConsumer.java:1041)
> at
> org.apache.activemq.TransactionContext.beforeEnd(TransactionContext.java:202)
> at
> org.apache.activemq.TransactionContext.end(TransactionContext.java:409)
> at
> com.atomikos.datasource.xa.XAResourceTransaction.suspend(XAResourceTransaction.java:457)
> at
> com.atomikos.datasource.xa.XAResourceTransaction.prepare(XAResourceTransaction.java:608)
> at
> com.atomikos.icatch.imp.PrepareMessage.send(PrepareMessage.java:61)
> at
> com.atomikos.icatch.imp.PropagationMessage.submit(PropagationMessage.java:111)
> at
> com.atomikos.icatch.imp.Propagator$PropagatorThread.run(Propagator.java:87)
> at
> com.atomikos.icatch.imp.Propagator.submitPropagationMessage(Propagator.java:66)
> at
> com.atomikos.icatch.imp.ActiveStateHandler.prepare(ActiveStateHandler.java:173)
> at
> com.atomikos.icatch.imp.CoordinatorImp.prepare(CoordinatorImp.java:832)
> at
> com.atomikos.icatch.imp.CoordinatorImp.terminate(CoordinatorImp.java:1159)
> at
> com.atomikos.icatch.imp.CompositeTerminatorImp.commit(CompositeTerminatorImp.java:92)
> at
> com.atomikos.icatch.jta.TransactionImp.commit(TransactionImp.java:236)
> at
> com.atomikos.icatch.jta.TransactionManagerImp.commit(TransactionManagerImp.java:498)
> at
> com.atomikos.icatch.jta.UserTransactionImp.commit(UserTransactionImp.java:129)
> at
> org.springframework.transaction.jta.JtaTransactionManager.doCommit(JtaTransactionManager.java:1011)
> at
> org.springframework.transaction.support.AbstractPlatformTransactionManager.processCommit(AbstractPlatformTransactionManager.java:755)
> at
> org.springframework.transaction.support.AbstractPlatformTransactionManager.commit(AbstractPlatformTransactionManager.java:724)
> at
> org.springframework.jms.listener.AbstractPollingMessageListenerContainer.receiveAndExecute(AbstractPollingMessageListenerContainer.java:257)
> at
> org.springframework.jms.listener.DefaultMessageListenerContainer$AsyncMessageListenerInvoker.invokeListener(DefaultMessageListenerContainer.java:1101)
> at
> org.springframework.jms.listener.DefaultMessageListenerContainer$AsyncMessageListenerInvoker.run(DefaultMessageListenerContainer.java:995)
> at java.lang.Thread.run(Thread.java:761)
> {noformat}
> Our analysis :
> We think that the duplicate message is caused by the failover
> during the prepare phase of the commit so we modify the source code to
> reproduce the case.
> Our modifications in config to produce failovers:
> broker : transport.useKeepAlive=false
> client : wireFormat.maxInactivityDuration=5000
> We add Thread.sleep in the source code of
> org.apache.activemq.ActiveMQMessageConsumer to force failover to be done
> exactly where we think it causes problems :
> {code:title=org.apache.activemq.ActiveMQMessageConsumer#acknowledge()|borderStyle=solid}
>
> public void acknowledge() throws JMSException {
> clearDeliveredList();
> waitForRedeliveries();
> synchronized(deliveredMessages) {
> // BEGIN MODIFIED CODE
> LOG.warn("start sleeping 20 seconds to
> test failover");
> try{
> Thread.sleep(1000 * 20 );
> }catch (InterruptedException e){
> LOG.error("Exception :",e);
> }
> LOG.warn("end sleeping 20 seconds to test
> failover");
> // END MODIFIED CODE
> // Acknowledge all messages so far.
> MessageAck ack =
> makeAckForAllDeliveredMessages(MessageAck.STANDARD_ACK_TYPE);
> if (ack == null)
> return; // no msgs
> if (session.getTransacted()) {
> rollbackOnFailedRecoveryRedelivery();
> session.doStartTransaction();
>
> ack.setTransactionId(session.getTransactionContext().getTransactionId());
> }
> pendingAck = null;
> session.sendAck(ack);
> // Adjust the counters
> deliveredCounter = Math.max(0,
> deliveredCounter - deliveredMessages.size());
> additionalWindowSize = Math.max(0,
> additionalWindowSize - deliveredMessages.size());
> if (!session.getTransacted()) {
> deliveredMessages.clear();
> }
> }
> }
> {code}
>
> With these changes on the configuration and the code, the
> problem is easily reproduced.
> We also try with transactedIndividualAck=true, and we add a
> Thread.sleep in the code :
> {code:title=org.apache.activemq.ActiveMQMessageConsumer#registerSync()|borderStyle=solid}
>
> private void registerSync() throws JMSException {
> session.doStartTransaction();
> if (!synchronizationRegistered) {
> synchronizationRegistered = true;
>
> session.getTransactionContext().addSynchronization(new Synchronization() {
> @Override
> public void beforeEnd() throws
> Exception {
> if (transactedIndividualAck) {
> clearDeliveredList();
> waitForRedeliveries();
>
> synchronized(deliveredMessages) {
>
> // BEGIN MODIFIED CODE
> LOG.warn("start sleeping
> 20 seconds to test failover");
> try{
> Thread.sleep(1000 * 20
> );
> }catch
> (InterruptedException e){
> LOG.error("Exception
> :",e);
> }
> LOG.warn("end sleeping 20
> seconds to test failover");
> // END MODIFIED CODE
>
>
> rollbackOnFailedRecoveryRedelivery();
> }
> } else {
> acknowledge();
> }
> synchronizationRegistered = false;
> }
> @Override
> public void afterCommit() throws
> Exception {
> commit();
> synchronizationRegistered = false;
> }
> @Override
> public void afterRollback() throws
> Exception {
> rollback();
> synchronizationRegistered = false;
> }
> });
> }
> }
> {code}
> With these modifications, we still get duplicates messages.
> We think that the problem is that the statement
> synchronized(deliveredMessages) prevents the call of clearDeliveredList() by
> another ActiveMQConnection thread that clears messages in progress.
> By adding logs we observe that a thread is waiting
> deliveredMessages ‘s lock in clearDeliveredList() method.
>
> Question :
>
> We tried fixes described in
> https://issues.apache.org/jira/browse/AMQ-5068 and
> https://issues.apache.org/jira/browse/AMQ-3519 but it doesn’t help to solve
> our problem.
> Is there a workaround or a config parameter that can help to
> prevent this problem ?
>
> We are working on our side to find a correction. An option
> may be to force rolling back transaction if there is a failover during the
> prepare phase of commit in ConnectionStateTracker.restoreTransactions().
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
