I have been migrating from hibernate 3.3.x to hibernate 4.1.3-final using Spring 3.1.1-release.
Besides the hibernate specific refactorings (due to API changes) I thought the only changes for the Spring integration were changing packages names:
- Use the LocalSessionFactoryBean from the hibernate4 package
- Use the HibernateTransactionManger (which we only use for testing) from the hibernate4 package.
As it turned out the migration when smooth and everything was working on a local tomcat.
However, once we ran our app on glassfish with the JtaTransactionManager (I tested it on GF3.1.2) we got a "No Session found for current thread" when obtaining sessionFactory.currentSession().
After checking SpringSessionContext, we learned that if TransactionSynchronizationManager does not return a existing session or SessionHolder, a check
is performed for a jtaSessionContext, which was null.
The fact that no SessionHolder is registered made also sense as this done by the HibernateTransactionManager and we are using the JtaTransactionManager.
So then learned that in case of JTA you have to specify manually how the tx manager/user transaction can be found.
This was done automatically for you in the hibernate3 LocalSessionFactoryBean, but no longer in the hibernate4 LocalSessionFactoryBean.
So to solve this we configured:
hibernate.transaction.jta.platform and set it to SunOneJtaPlatform.
This resolved the "No Session found for current thread" as it initialized the jtaSessionContext with the txmanager provided by the configured JtaPlatform.
However, now it turns out that the hibernate session is not flushed before the transaction is commited and hence no modifications are written to database.
In the supplied sample we have a basic hibernate setup with H2. Next we have a JtaTransactionManager and a transactional facade.
Next we have a test entity having a single property. The facade has two methods, one to store the entity and one to retrieve the entity.
They both marked as @Transactional and if called will run in there own transaction.
The trigger is a plain JEE servlet which retrieves the facade from application context.
First the store method is called (tx1) then the retrieve method is called (tx2).
As you will see with the Spring hibernate4 integration there was nothing saved.
With the hibernate3 integration everything works as expected and the record is saved
(it can be retrieved by the subsequent retrieval transaction)
What is also bizarre is that in hibernate3 modus everything goes via the Spring TransactionSynchronizationManager (even in JTA mode).
Also the current session is bound via a thread local referenced via the synchronization.
This is bound using a SpringSessionSynchronization which will call flush before transaction completion.
All of this is gone with the hibernate4 integration from the moment a JTA environment is detected.
As of then everything goes via the JTA transaction manager, as there where no Spring Transacion management involved.
This could be normal to a certain extend, but it feels odd compared to the way is was done with hibernate3.
I supplied two samples:
- hibernate3.zip : this is the working one, deploy it on GF and goto "http://localhost:8080/hibernate3/Persist"
You will see that it stores a record and is able to retrieve it again
- hibernate4.zip : the exact same sample as above, but now with hibernate4 and using LocalSessionFactoryBean from the hibernate4 package and the hibernate.transaction.jta.platform set.
You will see that it stores a record and is NOT able to retrieve it.
Both samples have a POM so it should be trivial to build them.