Atomikos Forum

All feedback will be investigated, but priority support, confidential support, and product priority updates are reserved for customers. Interested in becoming a customer? Subscribe here...

Only registered users can post (registration is free). Problems registering? Send an email to support@atomikos.com to get help.


PostgreSQL XA: prepare() on different connection than start()

Hello everybody,

I'm trying to get XA transactions running with PostgreSQL and ActiveMQ. I am using Spring Framework 3.0.0 (declarative transactions), Hibernate 3.3.2, PostgreSQL 8.4 + JDBC4 Driver 8.4-701 (also tested with PostgreSQL 8.3 + JDBC3 Driver 8.3-603), ActiveMQ 5.3.0, Atomikos TransactionEssentials 3.6.0 (AtomikosDataSourceBean)

I am getting an error with the following workflow because of a call to prepare() on a different PGXAConnection instance than the calls to start() and end():

Query database and send JMS message with JmsTemplate in XA transaction.

1. Start transaction - UserTransactionImp.begin() (atomikos/spring)
2. Query database for the current transaction timestamp
  2.1. PGXAConnection.start() (postgres)
  2.2. Select timestamp from db
  2.3. PGXAConnection.end() (postgres)
3. Send JMS message with JmsTemplate (activemq/spring)
4. Commit transaction (2PC) - UserTransactionImp.commit() (atomikos/spring)
  4.1. PGXAConnection.prepare(): currentXid is null because prepare() is called on a different PGXAConnection instance than start() and end() -> NullPointerException, transaction is NOT prepared in database
  4.2. PGXAConnection.rollback():  because of the NPE in 4.1 rollback() is called on the same PGXAConnection instance as prepare()
  4.3. ROLLBACK PREPARED 'xid' is executed in the database but this xid DOES NOT EXIST because the transaction was started on a different connection and PREPARE TRANSACTION failed

Database log:

2010-05-12 16:59:41.761 CEST - appuser@10.10.1.3 testdb [4beac20c.396f|1/1027] LOG:  duration: 0.057 ms  bind S_1: BEGIN
2010-05-12 16:59:41.761 CEST - appuser@10.10.1.3 testdb [4beac20c.396f|1/1027] LOG:  duration: 0.017 ms  execute S_1: BEGIN
2010-05-12 16:59:41.761 CEST - appuser@10.10.1.3 testdb [4beac20c.396f|1/1027] LOG:  duration: 0.297 ms  parse <unnamed>: select cast(now() as timestamptz(3))
2010-05-12 16:59:41.761 CEST - appuser@10.10.1.3 testdb [4beac20c.396f|1/1027] LOG:  duration: 0.061 ms  bind <unnamed>: select cast(now() as timestamptz(3))
2010-05-12 16:59:41.762 CEST - appuser@10.10.1.3 testdb [4beac20c.396f|1/1027] LOG:  duration: 0.056 ms  execute <unnamed>: select cast(now() as timestamptz(3))
2010-05-12 16:59:48.257 CEST - appuser@10.10.1.3 testdb [4beac210.3971|2/90] LOG:  duration: 0.706 ms  parse <unnamed>: SELECT gid FROM pg_prepared_xacts
2010-05-12 16:59:48.257 CEST - appuser@10.10.1.3 testdb [4beac210.3971|2/90] LOG:  duration: 0.073 ms  bind <unnamed>: SELECT gid FROM pg_prepared_xacts
2010-05-12 16:59:48.257 CEST - appuser@10.10.1.3 testdb [4beac210.3971|2/90] LOG:  duration: 0.044 ms  execute <unnamed>: SELECT gid FROM pg_prepared_xacts
2010-05-12 17:00:03.007 CEST - appuser@10.10.1.3 testdb [4beac210.3971|2/91] LOG:  duration: 0.095 ms  parse <unnamed>: ROLLBACK PREPARED '1096044365_MTAuMTAuMS4zLnRtMDAwMDUwMDAwMQ==_MTAuMTAuMS4zLnRtNQ=='
2010-05-12 17:00:03.007 CEST - appuser@10.10.1.3 testdb [4beac210.3971|2/91] LOG:  duration: 0.010 ms  bind <unnamed>: ROLLBACK PREPARED '1096044365_MTAuMTAuMS4zLnRtMDAwMDUwMDAwMQ==_MTAuMTAuMS4zLnRtNQ=='
2010-05-12 17:00:03.007 CEST - appuser@10.10.1.3 testdb [4beac210.3971|2/91] ERROR:  prepared transaction with identifier "1096044365_MTAuMTAuMS4zLnRtMDAwMDUwMDAwMQ==_MTAuMTAuMS4zLnRtNQ==" does not exist
2010-05-12 17:00:03.007 CEST - appuser@10.10.1.3 testdb [4beac210.3971|2/91] STATEMENT:  ROLLBACK PREPARED '1096044365_MTAuMTAuMS4zLnRtMDAwMDUwMDAwMQ==_MTAuMTAuMS4zLnRtNQ=='

The database log shows that BEGIN (transaction) is executed with a different database connection/session (4beac20c.396f) than ROLLBACK PREPARED (4beac210.3971). It is OK to execute PREPARE TRANSACTION and COMMIT/ROLLBACK PREPARED with different connections, but PREPARE TRANSACTION was never executed because of the NPE in step 4.1 of the workflow.

When debugging the application with Eclipse I can see that poolSize == maxPoolSize connections (PGXAConnection) are allocated by AtomikosDataSourceBean. start() and end() are called on one of these PGXAConnection instances.

There is an additional connection allocated for recovery, prepare() is always called on that additional PGXAConnection instance and that causes the error.

The only way I can prevent the error from happening is setting max_connections in the database configuration file (postgresql.conf) to the same value as maxPoolSize of the AtomikosDataSourceBean configuration. This way the additional connection for recovery cannot be allocated and prepare() is called on the same connection as start() and end(). Btw. in postgresql.conf max_prepared_transactions is >= max_transactions in all scenarios.

I know that a lot of information is missing here to allow for a detailed analysis of the issue. I can provide further information like the complete spring configuration, atomikos log and application stack trace.

But for the moment I just want to ask if this bug rings a bell for anybody. Is this a known limitation of PostgreSQL XA support?

Any hints, comments and further questions appreciated!

Thanks a lot,
tscho
Tscho T Send private email
Thursday, May 13, 2010
 
 
Hi,

Does this also happen with the latest release (downloadable on our site)? We had a bug that looked like this one...

HTH
Guy Pardon Send private email
Tuesday, May 18, 2010
 
 
hello,

thanks for you reply and sorry for not reporting back earlier. an update to atomikos TransactionEssentials 3.6.5 solved this particular problem.

but now i have a somehow related problem with version 3.6.5. the application is running and i force a shutdown of the database system after BEGIN but before PREPARE (to simulate a failure). when i restart the database system a ROLLBACK PREPARED 'xid' is executed on the new connection.

the ROLLBACK PREPARED statement is executed over and over again. is there a way to limit the retry count? why is ROLLBACK PREPARED executed when PREPARE was never executed for the transaction?
Atomikos Log:

2010-05-28 14:25:14,946 [Atomikos:9] DEBUG atomikos  - XA resource 'dataSourceJTA': prepare for XID '31302E31302E312E332E746D30303034353030303031:31302E31302E312E332E746D3837' raised -3: the XA resource detected an internal error
org.postgresql.xa.PGXAException: Error preparing transaction
    at org.postgresql.xa.PGXAConnection.prepare(PGXAConnection.java:254)
    at com.atomikos.datasource.xa.XAResourceTransaction.prepare(XAResourceTransaction.java:591)
    at com.atomikos.icatch.imp.PrepareMessage.send(PrepareMessage.java:63)
    at com.atomikos.icatch.imp.PropagationMessage.submit(PropagationMessage.java:114)
    at com.atomikos.icatch.imp.Propagator$PropagatorThread.run(Propagator.java:90)
    at java.util.concurrent.ThreadPoolExecutor$Worker.runTask(ThreadPoolExecutor.java:886)
    at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:908)
    at java.lang.Thread.run(Thread.java:619)
Caused by: org.postgresql.util.PSQLException: FATAL: terminating connection due to administrator command
    at org.postgresql.core.v3.QueryExecutorImpl.receiveErrorResponse(QueryExecutorImpl.java:1592)
    at org.postgresql.core.v3.QueryExecutorImpl.processResults(QueryExecutorImpl.java:1327)
    at org.postgresql.core.v3.QueryExecutorImpl.execute(QueryExecutorImpl.java:192)
    at org.postgresql.jdbc2.AbstractJdbc2Statement.execute(AbstractJdbc2Statement.java:451)
    at org.postgresql.jdbc2.AbstractJdbc2Statement.executeWithFlags(AbstractJdbc2Statement.java:336)
    at org.postgresql.jdbc2.AbstractJdbc2Statement.executeUpdate(AbstractJdbc2Statement.java:282)
    at org.postgresql.xa.PGXAConnection.prepare(PGXAConnection.java:242)
    ... 7 more
2010-05-28 14:25:14,947 [DefaultMessageListenerContainer-1] DEBUG atomikos  - Coordinator 10.10.1.3.tm0004500001 entering state: ABORTING
2010-05-28 14:25:14,947 [DefaultMessageListenerContainer-1] DEBUG atomikos  - (1.5) executing task: com.atomikos.icatch.imp.Propagator$PropagatorThread@8e683a
2010-05-28 14:25:14,948 [Atomikos:9] INFO  atomikos  - XAResource.rollback ( 31302E31302E312E332E746D30303034353030303031:31302E31302E312E332E746D3837 ) on resource dataSourceJTA represented by XAResource instance org.postgresql.xa.PGXAConnection@da1c9f
2010-05-28 14:25:14,949 [Atomikos:9] WARN  atomikos  - XA resource 'dataSourceJTA': rollback for XID '31302E31302E312E332E746D30303034353030303031:31302E31302E312E332E746D3837' raised -3: the XA resource detected an internal error
org.postgresql.xa.PGXAException: Error rolling back prepared transaction
    at org.postgresql.xa.PGXAConnection.rollback(PGXAConnection.java:357)
    at com.atomikos.datasource.xa.XAResourceTransaction.rollback(XAResourceTransaction.java:659)
    at com.atomikos.icatch.imp.RollbackMessage.send(RollbackMessage.java:74)
    at com.atomikos.icatch.imp.PropagationMessage.submit(PropagationMessage.java:114)
    at com.atomikos.icatch.imp.Propagator$PropagatorThread.run(Propagator.java:90)
    at java.util.concurrent.ThreadPoolExecutor$Worker.runTask(ThreadPoolExecutor.java:886)
    at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:908)
    at java.lang.Thread.run(Thread.java:619)
Caused by: org.postgresql.util.PSQLException: An I/O error occured while sending to the backend.
    at org.postgresql.core.v3.QueryExecutorImpl.execute(QueryExecutorImpl.java:218)
    at org.postgresql.jdbc2.AbstractJdbc2Connection.executeTransactionCommand(AbstractJdbc2Connection.java:650)
    at org.postgresql.jdbc2.AbstractJdbc2Connection.commit(AbstractJdbc2Connection.java:670)
    at org.postgresql.jdbc2.AbstractJdbc2Connection.setAutoCommit(AbstractJdbc2Connection.java:633)
    at org.postgresql.xa.PGXAConnection.rollback(PGXAConnection.java:343)
    ... 7 more
Caused by: java.net.SocketException: Broken pipe
    at java.net.SocketOutputStream.socketWrite0(Native Method)
    at java.net.SocketOutputStream.socketWrite(SocketOutputStream.java:92)
    at java.net.SocketOutputStream.write(SocketOutputStream.java:136)
    at java.io.BufferedOutputStream.flushBuffer(BufferedOutputStream.java:65)
    at java.io.BufferedOutputStream.flush(BufferedOutputStream.java:123)
    at org.postgresql.core.PGStream.flush(PGStream.java:508)
    at org.postgresql.core.v3.QueryExecutorImpl.sendSync(QueryExecutorImpl.java:676)
    at org.postgresql.core.v3.QueryExecutorImpl.execute(QueryExecutorImpl.java:191)
    ... 11 more
    
=== shutdown/restart ===

2010-05-28 14:25:24,949 [Atomikos:9] DEBUG atomikos  - Propagator: retrying message: RollbackMessage to XAResourceTransaction: 31302E31302E312E332E746D30303034353030303031:31302E31302E312E332E746D3837
2010-05-28 14:25:24,949 [Atomikos:9] DEBUG atomikos  - dataSourceJTA: forcing refresh of XAConnection...
2010-05-28 14:25:24,972 [Atomikos:9] INFO  atomikos  - XAResource.rollback ( 31302E31302E312E332E746D30303034353030303031:31302E31302E312E332E746D3837 ) on resource dataSourceJTA represented by XAResource instance org.postgresql.xa.PGXAConnection@f1036f
2010-05-28 14:25:24,973 [Atomikos:9] WARN  atomikos  - XA resource 'dataSourceJTA': rollback for XID '31302E31302E312E332E746D30303034353030303031:31302E31302E312E332E746D3837' raised -3: the XA resource detected an internal error
org.postgresql.xa.PGXAException: Error rolling back prepared transaction
    at org.postgresql.xa.PGXAConnection.rollback(PGXAConnection.java:357)
    at com.atomikos.datasource.xa.XAResourceTransaction.rollback(XAResourceTransaction.java:659)
    at com.atomikos.icatch.imp.RollbackMessage.send(RollbackMessage.java:74)
    at com.atomikos.icatch.imp.PropagationMessage.submit(PropagationMessage.java:114)
    at com.atomikos.icatch.imp.Propagator$PropagatorThread.run(Propagator.java:90)
    at java.util.concurrent.ThreadPoolExecutor$Worker.runTask(ThreadPoolExecutor.java:886)
    at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:908)
    at java.lang.Thread.run(Thread.java:619)
Caused by: org.postgresql.util.PSQLException: ERROR: prepared transaction with identifier "1096044365_MTAuMTAuMS4zLnRtMDAwNDUwMDAwMQ==_MTAuMTAuMS4zLnRtODc=" does not exist
    at org.postgresql.core.v3.QueryExecutorImpl.receiveErrorResponse(QueryExecutorImpl.java:1592)
    at org.postgresql.core.v3.QueryExecutorImpl.processResults(QueryExecutorImpl.java:1327)
    at org.postgresql.core.v3.QueryExecutorImpl.execute(QueryExecutorImpl.java:192)
    at org.postgresql.jdbc2.AbstractJdbc2Statement.execute(AbstractJdbc2Statement.java:451)
    at org.postgresql.jdbc2.AbstractJdbc2Statement.executeWithFlags(AbstractJdbc2Statement.java:336)
    at org.postgresql.jdbc2.AbstractJdbc2Statement.executeUpdate(AbstractJdbc2Statement.java:282)
    at org.postgresql.xa.PGXAConnection.rollback(PGXAConnection.java:347)
    ... 7 more

    
Database Log:
    
2010-05-28 14:25:05.216 CEST - postgres@10.10.1.3 testdb [4bffb5e5.214b|1/1409] LOG:  duration: 0.016 ms  bind S_1: BEGIN
2010-05-28 14:25:05.216 CEST - postgres@10.10.1.3 testdb [4bffb5e5.214b|1/1409] LOG:  duration: 0.004 ms  execute S_1: BEGIN
2010-05-28 14:25:05.216 CEST - postgres@10.10.1.3 testdb [4bffb5e5.214b|1/1409] LOG:  duration: 0.074 ms  parse <unnamed>: select cast(now() as timestamptz(3))
2010-05-28 14:25:05.216 CEST - postgres@10.10.1.3 testdb [4bffb5e5.214b|1/1409] LOG:  duration: 0.018 ms  bind <unnamed>: select cast(now() as timestamptz(3))
2010-05-28 14:25:05.216 CEST - postgres@10.10.1.3 testdb [4bffb5e5.214b|1/1409] LOG:  duration: 0.011 ms  execute <unnamed>: select cast(now() as timestamptz(3))
2010-05-28 14:25:06.213 CEST - @  [4bffae25.e61|] LOG:  received fast shutdown request
2010-05-28 14:25:06.213 CEST - @  [4bffae25.e61|] LOG:  aborting any active transactions
2010-05-28 14:25:06.213 CEST - postgres@10.10.1.3 testdb [4bffb5f3.216b|12/0] FATAL:  terminating connection due to administrator command
2010-05-28 14:25:06.213 CEST - postgres@10.10.1.3 testdb [4bffb5e5.2152|8/0] FATAL:  terminating connection due to administrator command
2010-05-28 14:25:06.213 CEST - postgres@10.10.1.3 testdb [4bffb5f2.2162|11/0] FATAL:  terminating connection due to administrator command
2010-05-28 14:25:06.213 CEST - postgres@10.10.1.3 testdb [4bffb5e5.2154|10/0] FATAL:  terminating connection due to administrator command

=== shutdown/restart ===

2010-05-28 14:25:16.554 CEST - [unknown]@ [unknown] [4bffb62c.222b|] LOG:  connection received: host=10.10.1.3 port=49073
2010-05-28 14:25:16.554 CEST - postgres@10.10.1.3 testdb [4bffb62c.222b|] LOG:  connection authorized: user=postgres database=testdb
2010-05-28 14:25:16.559 CEST - postgres@10.10.1.3 testdb [4bffb62c.222b|1/8] LOG:  duration: 0.185 ms  parse <unnamed>: ROLLBACK PREPARED '1096044365_MTAuMTAuMS4zLnRtMDAwNDUwMDAwMQ==_MTAuMTAuMS4zLnRtODc='
2010-05-28 14:25:16.559 CEST - postgres@10.10.1.3 testdb [4bffb62c.222b|1/8] LOG:  duration: 0.008 ms  bind <unnamed>: ROLLBACK PREPARED '1096044365_MTAuMTAuMS4zLnRtMDAwNDUwMDAwMQ==_MTAuMTAuMS4zLnRtODc='
2010-05-28 14:25:16.559 CEST - postgres@10.10.1.3 testdb [4bffb62c.222b|1/8] ERROR:  prepared transaction with identifier "1096044365_MTAuMTAuMS4zLnRtMDAwNDUwMDAwMQ==_MTAuMTAuMS4zLnRtODc=" does not exist
2010-05-28 14:25:16.559 CEST - postgres@10.10.1.3 testdb [4bffb62c.222b|1/8] STATEMENT:  ROLLBACK PREPARED '1096044365_MTAuMTAuMS4zLnRtMDAwNDUwMDAwMQ==_MTAuMTAuMS4zLnRtODc='
Tscho T Send private email
Friday, May 28, 2010
 
 
Hi,

If this happens across restart then this means that the transaction was prepared indeed. Looks like a problem with PostgreSQL: it seems to ignore prepared XA branches...

HTH
Guy Pardon Send private email
Friday, May 28, 2010
 
 
hello guy,

thanks for your reply! the transaction was definitely not prepared because the database system was shutdown before PREPARE.

do you mean PostgreSQL should support a PREPARE on a different connection than BEGIN? afaik PostgreSQL does support a COMMIT/ROLLBACK PREPARED on a different connection than PREPARE but not PREPARE or COMMIT/ROLLBACK on a different connection than BEGIN.

is there a way to limit the retry count of ROLLBACK PREPARED? the transaction will be rolled back due to a timeout anyway when no PREPARE is executed. Or am i missing something here!?

any further comments appreciated!

regards,
tscho
Tscho T Send private email
Monday, May 31, 2010
 
 
hello guy,

regarding your reply:

"If this happens across restart then this means that the transaction was prepared indeed."

The error occurs across (actually is caused by) a restart of the database system, not across a restart of the application or the atomikos transaction manager.

regards,
tscho
Tscho T Send private email
Monday, May 31, 2010
 
 
There are two things here:

-the fact we call rollback is due to the fact that prepare fails with an error, so we assume the worst case, being a prepared resource -> hence the rollback call after prepare

-the fact that prepare is called on a different connection than start: this is allowed by the JTA/XA specs, but you can override this by setting com.atomikos.icatch.threaded_2pc to false

Best
Guy Pardon Send private email
Tuesday, June 01, 2010
 
 
hello,

again, thanks for your quick reply!

There are two things here:

-the fact we call rollback is due to the fact that prepare fails with an error, so we assume the worst case, being a prepared resource -> hence the rollback call after prepare

> ok, fair enough, that makes sense! is there a way to limit the retry count of ROLLBACK PREPARED?

-the fact that prepare is called on a different connection than start: this is allowed by the JTA/XA specs, but you can override this by setting com.atomikos.icatch.threaded_2pc to false

> i've run this particular scenario with "true" and "false", makes no difference since PREPARE fails due to a broken connection. but this information is very important for me, because it seems that PostgreSQL does not support the JTA/XA specs in this case.

if would really appreciate any further comments on this issue, because after several unsuccessful calls of ROLLBACK PREPARED i get an hibernate exception, and other unrelated transactions fail too.

regards,
tscho
Tscho T Send private email
Tuesday, June 01, 2010
 
 

This topic is archived. No further replies will be accepted.

Other recent topics Other recent topics
 
Powered by FogBugz