Menu
▾
▴
springframework-developer
|
From: <jue...@we...> - 2004-03-25 19:51:46
|
Colin,
=20
Thanks for tracking this down. It's actually a problem with =
SessionFactoryUtils' inner class SessionSynchronization: It assumes that =
beforeCompletion is called in any case, even if beforeCommit has thrown =
an exception. However, this just applies if you specified a =
TransactionManagerLookup in the Hibernate configuration; else, =
afterCompletion will do the cleanup - which will be called in any case.
=20
When you remove the Hibernate TransactionManagerLookup, you shouldn't =
face the issue - the bug doesn't have any effects then. Note that you =
don't need that TransactionManagerLookup when using Spring's =
JtaTransactionManager, as Spring will properly apply cache callbacks =
anyway. A TransactionManagerLookup just adds value when used with EJB =
CMT or manual JTA, for ultra-correct cache callbacks.
=20
So essentially, everything should be fine if using =
HibernateTransactionManager, or JtaTransactionManager without a =
Hibernate TransactionManagerLookup. This is clearly something to fix, =
but I guess we don't need to do an immediate 1.0.1 followup release; I'd =
like to gather further bug reports first. For the time being, let's =
suggest to remove the TransactionManagerLookup from the Hibernate =
configuration.
=20
Juergen
=20
________________________________
Von: Colin Sampaleanu [mailto:col...@ex...]
Gesendet: Do 25.03.2004 20:32
An: spr...@li...; j=FCrgen h=F6ller =
[werk3AT]
Betreff: Re: [Springframework-developer] Hibernate resource management =
issue
Juergen,
I am almost 100% sure this block of code from
AbstractPlatformTransactionManager is wrong:
else {
try {
try {
triggerBeforeCommit(defStatus);
triggerBeforeCompletion(defStatus);
if (status.isNewTransaction()) {
logger.info("Initiating transaction commit");
doCommit(defStatus);
}
}
catch (UnexpectedRollbackException ex) {
triggerAfterCompletion(defStatus,
TransactionSynchronization.STATUS_ROLLED_BACK, ex);
throw ex;
}
catch (TransactionException ex) {
if (this.rollbackOnCommitFailure) {
doRollbackOnCommitException(defStatus, ex);
triggerAfterCompletion(defStatus,
TransactionSynchronization.STATUS_ROLLED_BACK, ex);
}
else {
triggerAfterCompletion(defStatus,
TransactionSynchronization.STATUS_UNKNOWN, ex);
}
throw ex;
}
catch (RuntimeException ex) {
doRollbackOnCommitException(defStatus, ex);
triggerAfterCompletion(defStatus,
TransactionSynchronization.STATUS_ROLLED_BACK, ex);
throw ex;
}
catch (Error err) {
doRollbackOnCommitException(defStatus, err);
triggerAfterCompletion(defStatus,
TransactionSynchronization.STATUS_UNKNOWN, err);
throw err;
}
triggerAfterCompletion(defStatus,
TransactionSynchronization.STATUS_COMMITTED, null);
}
finally {
cleanupAfterCompletion(defStatus);
}
}
triggerBeforeCommit() execute, which in the Hibernate case will force a
flush. However, if that flush throws an exception, then
triggerBeforeCompletion(defStatus);
never gets called. However, triggerBeforeCompletion is what is actually
supposed to release the Hibernate session holder from the current
thread! So in this case, the session (which is totally hosed of course),
gets left on the thread. I believe in some environments this wouldn't
matter that much, as the threads don't get resused. In the JBoss case,
new requests coming in will get the existing thread, and this time,
SessionFactoryUtils will see the session is there, and try to use it.
Bang, it all blows up...
So for this code to work properly, what needs to happen is that
triggerBeforeCompletion still needs to be called even if
triggerBeforeCommit fails. While I am ok with writing the code in this
method to handle this, I am not 100% sure this is safe in terms of all
the other interactions that will happen as a result; mot of this code is
your baby with me only having traced through it once in a while. So if
you would prefer to resolve this that would be great.
Unless I am mistaken about this bug, I think it is a pretty serious one,
and warrants an almost immediate release of a v1.0.1 of Spring...
Regards,
Colin
Colin Sampaleanu wrote:
> I am tracking down a possible Hibernate resource management issue.
>
> In a running app, some time yesterday, some code, running in a wrapped
> transaction with Hibernate handling ORM, encountered an Oracle
> constraint violation and threw an exception. Fine...
>
> But when I log into the app myself now via the web ui and then it gets
> a service object to read some data, the service object is wrapped with
> a transaction interceptor, and also a hibernate interceptor. The
> Hibernate interceptor is already seeing a Hibernate Session existing
> on the current thread, so it is not creating a new one. Then at the
> end of the transaction, when the Hibernate session is attempted to be
> flushed, Hibernate tries to write out the old bad data from yesterday.
>
> What this essentially means is that when the error from yesterday
> happened, the session did not get released from the thread, and has
> been sticking around all this time. When I came via struts, I was
> given the same thread as yesterday by the appserver, and the old
> invalid session was still on it. The problem is not that it's reusing
> that session, but why it was ever left that the day before.
>
> Will try to duplicate this...
|
|
From: <jue...@we...> - 2004-03-25 20:24:22
|
Odd - if there's no TransactionManagerLookup, the Session should =
definitely be closed in afterCompletion.
=20
Regarding beforeCompletion, the semantics indeed need to be clarified: I =
guess it's appropriate to always invoke it, even on beforeCommit =
failure.
=20
Juergen
=20
________________________________
Von: Colin Sampaleanu [mailto:col...@ex...]
Gesendet: Do 25.03.2004 21:09
An: j=FCrgen h=F6ller [werk3AT]
Cc: spr...@li...
Betreff: Re: [Springframework-developer] Hibernate resource management =
issue
We're on slightly different pages though. In my case, I am actually not
using TransactionManagerLookup... But in my case, because of the
exception in the flush in beforeCommit(), beforeCompletion() never gets
called (as it would normally). Now I do see that afterCompletion is also
supposed to call closeSessionIfNecessary() (I had actually missed this
before), but in my case, it's not getting there. I haven't traced it in
a debugger, which is what I will do now, as it seems to me it should get
to that code, certainly I do have a
"Triggering afterCompletion synchronization"
in the log which should happen right before afterCompletion() is called.
The other question is whether beforeCompletion still shouldn't be called
in any case even if beforeCommit fails. Ultimately, we are still before
completion of the transaction, unless you meant it to be only called in
the case of no failure.
Colin
j=FCrgen h=F6ller [werk3AT] wrote:
>Colin,
>
>Thanks for tracking this down. It's actually a problem with =
SessionFactoryUtils' inner class SessionSynchronization: It assumes that =
beforeCompletion is called in any case, even if beforeCommit has thrown =
an exception. However, this just applies if you specified a =
TransactionManagerLookup in the Hibernate configuration; else, =
afterCompletion will do the cleanup - which will be called in any case.
>
>When you remove the Hibernate TransactionManagerLookup, you shouldn't =
face the issue - the bug doesn't have any effects then. Note that you =
don't need that TransactionManagerLookup when using Spring's =
JtaTransactionManager, as Spring will properly apply cache callbacks =
anyway. A TransactionManagerLookup just adds value when used with EJB =
CMT or manual JTA, for ultra-correct cache callbacks.
>
>So essentially, everything should be fine if using =
HibernateTransactionManager, or JtaTransactionManager without a =
Hibernate TransactionManagerLookup. This is clearly something to fix, =
but I guess we don't need to do an immediate 1.0.1 followup release; I'd =
like to gather further bug reports first. For the time being, let's =
suggest to remove the TransactionManagerLookup from the Hibernate =
configuration.
>
>Juergen
>
>
>________________________________
>
>Von: Colin Sampaleanu [mailto:col...@ex...]
>Gesendet: Do 25.03.2004 20:32
>An: spr...@li...; j=FCrgen h=F6ller =
[werk3AT]
>Betreff: Re: [Springframework-developer] Hibernate resource management =
issue
>
>
>
>Juergen,
>
>I am almost 100% sure this block of code from
>AbstractPlatformTransactionManager is wrong:
> else {
> try {
> try {
> triggerBeforeCommit(defStatus);
> triggerBeforeCompletion(defStatus);
> if (status.isNewTransaction()) {
> logger.info("Initiating transaction commit");
> doCommit(defStatus);
> }
> }
> catch (UnexpectedRollbackException ex) {
> triggerAfterCompletion(defStatus,
>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
> throw ex;
> }
> catch (TransactionException ex) {
> if (this.rollbackOnCommitFailure) {
> doRollbackOnCommitException(defStatus, ex);
> triggerAfterCompletion(defStatus,
>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
> }
> else {
> triggerAfterCompletion(defStatus,
>TransactionSynchronization.STATUS_UNKNOWN, ex);
> }
> throw ex;
> }
> catch (RuntimeException ex) {
> doRollbackOnCommitException(defStatus, ex);
> triggerAfterCompletion(defStatus,
>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
> throw ex;
> }
> catch (Error err) {
> doRollbackOnCommitException(defStatus, err);
> triggerAfterCompletion(defStatus,
>TransactionSynchronization.STATUS_UNKNOWN, err);
> throw err;
> }
> triggerAfterCompletion(defStatus,
>TransactionSynchronization.STATUS_COMMITTED, null);
> }
> finally {
> cleanupAfterCompletion(defStatus);
> }
> }
>
>triggerBeforeCommit() execute, which in the Hibernate case will force a
>flush. However, if that flush throws an exception, then
> triggerBeforeCompletion(defStatus);
>never gets called. However, triggerBeforeCompletion is what is actually
>supposed to release the Hibernate session holder from the current
>thread! So in this case, the session (which is totally hosed of =
course),
>gets left on the thread. I believe in some environments this wouldn't
>matter that much, as the threads don't get resused. In the JBoss case,
>new requests coming in will get the existing thread, and this time,
>SessionFactoryUtils will see the session is there, and try to use it.
>Bang, it all blows up...
>
>So for this code to work properly, what needs to happen is that
>triggerBeforeCompletion still needs to be called even if
>triggerBeforeCommit fails. While I am ok with writing the code in this
>method to handle this, I am not 100% sure this is safe in terms of all
>the other interactions that will happen as a result; mot of this code =
is
>your baby with me only having traced through it once in a while. So if
>you would prefer to resolve this that would be great.
>
>Unless I am mistaken about this bug, I think it is a pretty serious =
one,
>and warrants an almost immediate release of a v1.0.1 of Spring...
>
>Regards,
>Colin
>
>
>Colin Sampaleanu wrote:
>
>=20
>
>>I am tracking down a possible Hibernate resource management issue.
>>
>>In a running app, some time yesterday, some code, running in a wrapped
>>transaction with Hibernate handling ORM, encountered an Oracle
>>constraint violation and threw an exception. Fine...
>>
>>But when I log into the app myself now via the web ui and then it gets
>>a service object to read some data, the service object is wrapped with
>>a transaction interceptor, and also a hibernate interceptor. The
>>Hibernate interceptor is already seeing a Hibernate Session existing
>>on the current thread, so it is not creating a new one. Then at the
>>end of the transaction, when the Hibernate session is attempted to be
>>flushed, Hibernate tries to write out the old bad data from yesterday.
>>
>>What this essentially means is that when the error from yesterday
>>happened, the session did not get released from the thread, and has
>>been sticking around all this time. When I came via struts, I was
>>given the same thread as yesterday by the appserver, and the old
>>invalid session was still on it. The problem is not that it's reusing
>>that session, but why it was ever left that the day before.
>>
>>Will try to duplicate this...
>> =20
>>
>
>
>
>
>=20
>
|
|
From: Colin S. <col...@ex...> - 2004-03-25 20:37:10
|
Things are simpler than they appear, and are actually as per my original=20
email. Look at this, from SessionFactoryUtils:
public void beforeCompletion() throws=20
CleanupFailureDataAccessException {
if (this.newSession) {
=20
TransactionSynchronizationManager.unbindResource(this.sessionFactory);
if (this.hibernateTransactionCompletion) {
=20
closeSessionIfNecessary(this.sessionHolder.getSession(),=20
this.sessionFactory);
}
}
}
public void afterCompletion(int status) {
if (!this.hibernateTransactionCompletion) {
Session session =3D this.sessionHolder.getSession();
if (session instanceof SessionImplementor) {
((SessionImplementor)=20
session).afterTransactionCompletion(status =3D=3D STATUS_COMMITTED);
}
if (this.newSession) {
closeSessionIfNecessary(session, this.sessionFactory)=
;
}
}
this.sessionHolder.setSynchronizedWithTransaction(false);
}
beforeCompletion is the only place that
TransactionSynchronizationManager.unbindResource
gets called for the SessionHolder, and in this case, beforeCompletion=20
never gets called.
So unfortunately this _is_ a serious bug. Any exception during the=20
beforeCommit call (which calls flush) means that the SessionHolder never=20
gets unbound from the thread. Nasty!
Colin
j=FCrgen h=F6ller [werk3AT] wrote:
>Odd - if there's no TransactionManagerLookup, the Session should definit=
ely be closed in afterCompletion.
>=20
>Regarding beforeCompletion, the semantics indeed need to be clarified: I=
guess it's appropriate to always invoke it, even on beforeCommit failure=
.
>=20
>Juergen
>=20
>
>________________________________
>
>Von: Colin Sampaleanu [mailto:col...@ex...]
>Gesendet: Do 25.03.2004 21:09
>An: j=FCrgen h=F6ller [werk3AT]
>Cc: spr...@li...
>Betreff: Re: [Springframework-developer] Hibernate resource management i=
ssue
>
>
>
>We're on slightly different pages though. In my case, I am actually not
>using TransactionManagerLookup... But in my case, because of the
>exception in the flush in beforeCommit(), beforeCompletion() never gets
>called (as it would normally). Now I do see that afterCompletion is also
>supposed to call closeSessionIfNecessary() (I had actually missed this
>before), but in my case, it's not getting there. I haven't traced it in
>a debugger, which is what I will do now, as it seems to me it should get
>to that code, certainly I do have a
> "Triggering afterCompletion synchronization"
>in the log which should happen right before afterCompletion() is called.
>
>The other question is whether beforeCompletion still shouldn't be called
>in any case even if beforeCommit fails. Ultimately, we are still before
>completion of the transaction, unless you meant it to be only called in
>the case of no failure.
>
>Colin
>
>
>j=FCrgen h=F6ller [werk3AT] wrote:
>
> =20
>
>>Colin,
>>
>>Thanks for tracking this down. It's actually a problem with SessionFact=
oryUtils' inner class SessionSynchronization: It assumes that beforeCompl=
etion is called in any case, even if beforeCommit has thrown an exception=
. However, this just applies if you specified a TransactionManagerLookup =
in the Hibernate configuration; else, afterCompletion will do the cleanup=
- which will be called in any case.
>>
>>When you remove the Hibernate TransactionManagerLookup, you shouldn't f=
ace the issue - the bug doesn't have any effects then. Note that you don'=
t need that TransactionManagerLookup when using Spring's JtaTransactionMa=
nager, as Spring will properly apply cache callbacks anyway. A Transactio=
nManagerLookup just adds value when used with EJB CMT or manual JTA, for =
ultra-correct cache callbacks.
>>
>>So essentially, everything should be fine if using HibernateTransaction=
Manager, or JtaTransactionManager without a Hibernate TransactionManagerL=
ookup. This is clearly something to fix, but I guess we don't need to do =
an immediate 1.0.1 followup release; I'd like to gather further bug repor=
ts first. For the time being, let's suggest to remove the TransactionMana=
gerLookup from the Hibernate configuration.
>>
>>Juergen
>>
>>
>>________________________________
>>
>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>Gesendet: Do 25.03.2004 20:32
>>An: spr...@li...; j=FCrgen h=F6ller =
[werk3AT]
>>Betreff: Re: [Springframework-developer] Hibernate resource management =
issue
>>
>>
>>
>>Juergen,
>>
>>I am almost 100% sure this block of code from
>>AbstractPlatformTransactionManager is wrong:
>> else {
>> try {
>> try {
>> triggerBeforeCommit(defStatus);
>> triggerBeforeCompletion(defStatus);
>> if (status.isNewTransaction()) {
>> logger.info("Initiating transaction commit");
>> doCommit(defStatus);
>> }
>> }
>> catch (UnexpectedRollbackException ex) {
>> triggerAfterCompletion(defStatus,
>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>> throw ex;
>> }
>> catch (TransactionException ex) {
>> if (this.rollbackOnCommitFailure) {
>> doRollbackOnCommitException(defStatus, ex);
>> triggerAfterCompletion(defStatus,
>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>> }
>> else {
>> triggerAfterCompletion(defStatus,
>>TransactionSynchronization.STATUS_UNKNOWN, ex);
>> }
>> throw ex;
>> }
>> catch (RuntimeException ex) {
>> doRollbackOnCommitException(defStatus, ex);
>> triggerAfterCompletion(defStatus,
>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>> throw ex;
>> }
>> catch (Error err) {
>> doRollbackOnCommitException(defStatus, err);
>> triggerAfterCompletion(defStatus,
>>TransactionSynchronization.STATUS_UNKNOWN, err);
>> throw err;
>> }
>> triggerAfterCompletion(defStatus,
>>TransactionSynchronization.STATUS_COMMITTED, null);
>> }
>> finally {
>> cleanupAfterCompletion(defStatus);
>> }
>> }
>>
>>triggerBeforeCommit() execute, which in the Hibernate case will force a
>>flush. However, if that flush throws an exception, then
>> triggerBeforeCompletion(defStatus);
>>never gets called. However, triggerBeforeCompletion is what is actually
>>supposed to release the Hibernate session holder from the current
>>thread! So in this case, the session (which is totally hosed of course)=
,
>>gets left on the thread. I believe in some environments this wouldn't
>>matter that much, as the threads don't get resused. In the JBoss case,
>>new requests coming in will get the existing thread, and this time,
>>SessionFactoryUtils will see the session is there, and try to use it.
>>Bang, it all blows up...
>>
>>So for this code to work properly, what needs to happen is that
>>triggerBeforeCompletion still needs to be called even if
>>triggerBeforeCommit fails. While I am ok with writing the code in this
>>method to handle this, I am not 100% sure this is safe in terms of all
>>the other interactions that will happen as a result; mot of this code i=
s
>>your baby with me only having traced through it once in a while. So if
>>you would prefer to resolve this that would be great.
>>
>>Unless I am mistaken about this bug, I think it is a pretty serious one=
,
>>and warrants an almost immediate release of a v1.0.1 of Spring...
>>
>>Regards,
>>Colin
>>
>>
>>Colin Sampaleanu wrote:
>>
>>
>>
>> =20
>>
>>>I am tracking down a possible Hibernate resource management issue.
>>>
>>>In a running app, some time yesterday, some code, running in a wrapped
>>>transaction with Hibernate handling ORM, encountered an Oracle
>>>constraint violation and threw an exception. Fine...
>>>
>>>But when I log into the app myself now via the web ui and then it gets
>>>a service object to read some data, the service object is wrapped with
>>>a transaction interceptor, and also a hibernate interceptor. The
>>>Hibernate interceptor is already seeing a Hibernate Session existing
>>>on the current thread, so it is not creating a new one. Then at the
>>>end of the transaction, when the Hibernate session is attempted to be
>>>flushed, Hibernate tries to write out the old bad data from yesterday.
>>>
>>>What this essentially means is that when the error from yesterday
>>>happened, the session did not get released from the thread, and has
>>>been sticking around all this time. When I came via struts, I was
>>>given the same thread as yesterday by the appserver, and the old
>>>invalid session was still on it. The problem is not that it's reusing
>>>that session, but why it was ever left that the day before.
>>>
>>>Will try to duplicate this...
>>> =20
>>>
>>> =20
>>>
>>
>>
>>
>>
>> =20
>>
>
>
>
> =20
>
|
|
From: <jue...@we...> - 2004-03-26 07:23:28
|
I see - if flush fails, the Session is closed (if no =
TransactionManagerLookup) but not removed from the thread. I've already =
fixed the issue; will commit it promptly. Of course, none of this =
affects HibernateTransactionManager (which is assumably the primary =
choice for Spring apps), but it's still a nasty issue.
=20
Bad timing - one day earlier, and we would simply have delayed 1.0 final =
for this. I agree that we should do a quick 1.0.1 followup: I've also =
incorporated more sophisticated FieldError message resolution, as =
suggested recently, and will also look at the reported auto-proxy =
creator issue with multiple afterPropertiesSet calls.
=20
All things considered, I suggest 1.0.1 mid next week. Let's also =
incorporate all reported documentation inconsistencies and the like.
=20
Juergen
=20
________________________________
Von: Colin Sampaleanu [mailto:col...@ex...]
Gesendet: Do 25.03.2004 21:36
An: j=FCrgen h=F6ller [werk3AT]
Cc: spr...@li...
Betreff: Re: [Springframework-developer] Hibernate resource management =
issue
Things are simpler than they appear, and are actually as per my original
email. Look at this, from SessionFactoryUtils:
public void beforeCompletion() throws
CleanupFailureDataAccessException {
if (this.newSession) {
=20
TransactionSynchronizationManager.unbindResource(this.sessionFactory);
if (this.hibernateTransactionCompletion) {
=20
closeSessionIfNecessary(this.sessionHolder.getSession(),
this.sessionFactory);
}
}
}
public void afterCompletion(int status) {
if (!this.hibernateTransactionCompletion) {
Session session =3D this.sessionHolder.getSession();
if (session instanceof SessionImplementor) {
((SessionImplementor)
session).afterTransactionCompletion(status =3D=3D STATUS_COMMITTED);
}
if (this.newSession) {
closeSessionIfNecessary(session, =
this.sessionFactory);
}
}
this.sessionHolder.setSynchronizedWithTransaction(false);
}
beforeCompletion is the only place that
TransactionSynchronizationManager.unbindResource
gets called for the SessionHolder, and in this case, beforeCompletion
never gets called.
So unfortunately this _is_ a serious bug. Any exception during the
beforeCommit call (which calls flush) means that the SessionHolder never
gets unbound from the thread. Nasty!
Colin
j=FCrgen h=F6ller [werk3AT] wrote:
>Odd - if there's no TransactionManagerLookup, the Session should =
definitely be closed in afterCompletion.
>
>Regarding beforeCompletion, the semantics indeed need to be clarified: =
I guess it's appropriate to always invoke it, even on beforeCommit =
failure.
>
>Juergen
>
>
>________________________________
>
>Von: Colin Sampaleanu [mailto:col...@ex...]
>Gesendet: Do 25.03.2004 21:09
>An: j=FCrgen h=F6ller [werk3AT]
>Cc: spr...@li...
>Betreff: Re: [Springframework-developer] Hibernate resource management =
issue
>
>
>
>We're on slightly different pages though. In my case, I am actually not
>using TransactionManagerLookup... But in my case, because of the
>exception in the flush in beforeCommit(), beforeCompletion() never gets
>called (as it would normally). Now I do see that afterCompletion is =
also
>supposed to call closeSessionIfNecessary() (I had actually missed this
>before), but in my case, it's not getting there. I haven't traced it in
>a debugger, which is what I will do now, as it seems to me it should =
get
>to that code, certainly I do have a
> "Triggering afterCompletion synchronization"
>in the log which should happen right before afterCompletion() is =
called.
>
>The other question is whether beforeCompletion still shouldn't be =
called
>in any case even if beforeCommit fails. Ultimately, we are still before
>completion of the transaction, unless you meant it to be only called in
>the case of no failure.
>
>Colin
>
>
>j=FCrgen h=F6ller [werk3AT] wrote:
>
>=20
>
>>Colin,
>>
>>Thanks for tracking this down. It's actually a problem with =
SessionFactoryUtils' inner class SessionSynchronization: It assumes that =
beforeCompletion is called in any case, even if beforeCommit has thrown =
an exception. However, this just applies if you specified a =
TransactionManagerLookup in the Hibernate configuration; else, =
afterCompletion will do the cleanup - which will be called in any case.
>>
>>When you remove the Hibernate TransactionManagerLookup, you shouldn't =
face the issue - the bug doesn't have any effects then. Note that you =
don't need that TransactionManagerLookup when using Spring's =
JtaTransactionManager, as Spring will properly apply cache callbacks =
anyway. A TransactionManagerLookup just adds value when used with EJB =
CMT or manual JTA, for ultra-correct cache callbacks.
>>
>>So essentially, everything should be fine if using =
HibernateTransactionManager, or JtaTransactionManager without a =
Hibernate TransactionManagerLookup. This is clearly something to fix, =
but I guess we don't need to do an immediate 1.0.1 followup release; I'd =
like to gather further bug reports first. For the time being, let's =
suggest to remove the TransactionManagerLookup from the Hibernate =
configuration.
>>
>>Juergen
>>
>>
>>________________________________
>>
>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>Gesendet: Do 25.03.2004 20:32
>>An: spr...@li...; j=FCrgen h=F6ller =
[werk3AT]
>>Betreff: Re: [Springframework-developer] Hibernate resource management =
issue
>>
>>
>>
>>Juergen,
>>
>>I am almost 100% sure this block of code from
>>AbstractPlatformTransactionManager is wrong:
>> else {
>> try {
>> try {
>> triggerBeforeCommit(defStatus);
>> triggerBeforeCompletion(defStatus);
>> if (status.isNewTransaction()) {
>> logger.info("Initiating transaction commit");
>> doCommit(defStatus);
>> }
>> }
>> catch (UnexpectedRollbackException ex) {
>> triggerAfterCompletion(defStatus,
>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>> throw ex;
>> }
>> catch (TransactionException ex) {
>> if (this.rollbackOnCommitFailure) {
>> doRollbackOnCommitException(defStatus, ex);
>> triggerAfterCompletion(defStatus,
>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>> }
>> else {
>> triggerAfterCompletion(defStatus,
>>TransactionSynchronization.STATUS_UNKNOWN, ex);
>> }
>> throw ex;
>> }
>> catch (RuntimeException ex) {
>> doRollbackOnCommitException(defStatus, ex);
>> triggerAfterCompletion(defStatus,
>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>> throw ex;
>> }
>> catch (Error err) {
>> doRollbackOnCommitException(defStatus, err);
>> triggerAfterCompletion(defStatus,
>>TransactionSynchronization.STATUS_UNKNOWN, err);
>> throw err;
>> }
>> triggerAfterCompletion(defStatus,
>>TransactionSynchronization.STATUS_COMMITTED, null);
>> }
>> finally {
>> cleanupAfterCompletion(defStatus);
>> }
>> }
>>
>>triggerBeforeCommit() execute, which in the Hibernate case will force =
a
>>flush. However, if that flush throws an exception, then
>> triggerBeforeCompletion(defStatus);
>>never gets called. However, triggerBeforeCompletion is what is =
actually
>>supposed to release the Hibernate session holder from the current
>>thread! So in this case, the session (which is totally hosed of =
course),
>>gets left on the thread. I believe in some environments this wouldn't
>>matter that much, as the threads don't get resused. In the JBoss case,
>>new requests coming in will get the existing thread, and this time,
>>SessionFactoryUtils will see the session is there, and try to use it.
>>Bang, it all blows up...
>>
>>So for this code to work properly, what needs to happen is that
>>triggerBeforeCompletion still needs to be called even if
>>triggerBeforeCommit fails. While I am ok with writing the code in this
>>method to handle this, I am not 100% sure this is safe in terms of all
>>the other interactions that will happen as a result; mot of this code =
is
>>your baby with me only having traced through it once in a while. So if
>>you would prefer to resolve this that would be great.
>>
>>Unless I am mistaken about this bug, I think it is a pretty serious =
one,
>>and warrants an almost immediate release of a v1.0.1 of Spring...
>>
>>Regards,
>>Colin
>>
>>
>>Colin Sampaleanu wrote:
>>
>>
>>
>> =20
>>
>>>I am tracking down a possible Hibernate resource management issue.
>>>
>>>In a running app, some time yesterday, some code, running in a =
wrapped
>>>transaction with Hibernate handling ORM, encountered an Oracle
>>>constraint violation and threw an exception. Fine...
>>>
>>>But when I log into the app myself now via the web ui and then it =
gets
>>>a service object to read some data, the service object is wrapped =
with
>>>a transaction interceptor, and also a hibernate interceptor. The
>>>Hibernate interceptor is already seeing a Hibernate Session existing
>>>on the current thread, so it is not creating a new one. Then at the
>>>end of the transaction, when the Hibernate session is attempted to be
>>>flushed, Hibernate tries to write out the old bad data from =
yesterday.
>>>
>>>What this essentially means is that when the error from yesterday
>>>happened, the session did not get released from the thread, and has
>>>been sticking around all this time. When I came via struts, I was
>>>given the same thread as yesterday by the appserver, and the old
>>>invalid session was still on it. The problem is not that it's reusing
>>>that session, but why it was ever left that the day before.
>>>
>>>Will try to duplicate this...
>>>=20
>>>
>>> =20
>>>
>>
>>
>>
>>
>> =20
>>
>
>
>
>=20
>
|
|
From: Colin S. <col...@ex...> - 2004-03-26 12:50:52
|
I see you also changed the behaviour so that on a RuntimeException or
Error, triggerAfterCompletion does not get called any longer, whereas it
did with the previous code. I am not 100% sure why this went away,
considering it was there before, and it is still being done for
UnexpectedRollbackException and TransactionException. (However, I'm not
100% clear on the semantics of triggerAfterCompletion and whether it
always has to be called, I presume it does).
jürgen höller [werk3AT] wrote:
>I see - if flush fails, the Session is closed (if no TransactionManagerLookup) but not removed from the thread. I've already fixed the issue; will commit it promptly. Of course, none of this affects HibernateTransactionManager (which is assumably the primary choice for Spring apps), but it's still a nasty issue.
>
>Bad timing - one day earlier, and we would simply have delayed 1.0 final for this. I agree that we should do a quick 1.0.1 followup: I've also incorporated more sophisticated FieldError message resolution, as suggested recently, and will also look at the reported auto-proxy creator issue with multiple afterPropertiesSet calls.
>
>All things considered, I suggest 1.0.1 mid next week. Let's also incorporate all reported documentation inconsistencies and the like.
>
>Juergen
>
>
>________________________________
>
>Von: Colin Sampaleanu [mailto:col...@ex...]
>Gesendet: Do 25.03.2004 21:36
>An: jürgen höller [werk3AT]
>Cc: spr...@li...
>Betreff: Re: [Springframework-developer] Hibernate resource management issue
>
>
>
>Things are simpler than they appear, and are actually as per my original
>email. Look at this, from SessionFactoryUtils:
> public void beforeCompletion() throws
>CleanupFailureDataAccessException {
> if (this.newSession) {
>
>TransactionSynchronizationManager.unbindResource(this.sessionFactory);
> if (this.hibernateTransactionCompletion) {
>
>closeSessionIfNecessary(this.sessionHolder.getSession(),
>this.sessionFactory);
> }
> }
> }
>
> public void afterCompletion(int status) {
> if (!this.hibernateTransactionCompletion) {
> Session session = this.sessionHolder.getSession();
> if (session instanceof SessionImplementor) {
> ((SessionImplementor)
>session).afterTransactionCompletion(status == STATUS_COMMITTED);
> }
> if (this.newSession) {
> closeSessionIfNecessary(session, this.sessionFactory);
> }
> }
> this.sessionHolder.setSynchronizedWithTransaction(false);
> }
>
>beforeCompletion is the only place that
> TransactionSynchronizationManager.unbindResource
>gets called for the SessionHolder, and in this case, beforeCompletion
>never gets called.
>
>So unfortunately this _is_ a serious bug. Any exception during the
>beforeCommit call (which calls flush) means that the SessionHolder never
>gets unbound from the thread. Nasty!
>
>Colin
>
>
>jürgen höller [werk3AT] wrote:
>
>
>
>>Odd - if there's no TransactionManagerLookup, the Session should definitely be closed in afterCompletion.
>>
>>Regarding beforeCompletion, the semantics indeed need to be clarified: I guess it's appropriate to always invoke it, even on beforeCommit failure.
>>
>>Juergen
>>
>>
>>________________________________
>>
>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>Gesendet: Do 25.03.2004 21:09
>>An: jürgen höller [werk3AT]
>>Cc: spr...@li...
>>Betreff: Re: [Springframework-developer] Hibernate resource management issue
>>
>>
>>
>>We're on slightly different pages though. In my case, I am actually not
>>using TransactionManagerLookup... But in my case, because of the
>>exception in the flush in beforeCommit(), beforeCompletion() never gets
>>called (as it would normally). Now I do see that afterCompletion is also
>>supposed to call closeSessionIfNecessary() (I had actually missed this
>>before), but in my case, it's not getting there. I haven't traced it in
>>a debugger, which is what I will do now, as it seems to me it should get
>>to that code, certainly I do have a
>> "Triggering afterCompletion synchronization"
>>in the log which should happen right before afterCompletion() is called.
>>
>>The other question is whether beforeCompletion still shouldn't be called
>>in any case even if beforeCommit fails. Ultimately, we are still before
>>completion of the transaction, unless you meant it to be only called in
>>the case of no failure.
>>
>>Colin
>>
>>
>>jürgen höller [werk3AT] wrote:
>>
>>
>>
>>
>>
>>>Colin,
>>>
>>>Thanks for tracking this down. It's actually a problem with SessionFactoryUtils' inner class SessionSynchronization: It assumes that beforeCompletion is called in any case, even if beforeCommit has thrown an exception. However, this just applies if you specified a TransactionManagerLookup in the Hibernate configuration; else, afterCompletion will do the cleanup - which will be called in any case.
>>>
>>>When you remove the Hibernate TransactionManagerLookup, you shouldn't face the issue - the bug doesn't have any effects then. Note that you don't need that TransactionManagerLookup when using Spring's JtaTransactionManager, as Spring will properly apply cache callbacks anyway. A TransactionManagerLookup just adds value when used with EJB CMT or manual JTA, for ultra-correct cache callbacks.
>>>
>>>So essentially, everything should be fine if using HibernateTransactionManager, or JtaTransactionManager without a Hibernate TransactionManagerLookup. This is clearly something to fix, but I guess we don't need to do an immediate 1.0.1 followup release; I'd like to gather further bug reports first. For the time being, let's suggest to remove the TransactionManagerLookup from the Hibernate configuration.
>>>
>>>Juergen
>>>
>>>
>>>________________________________
>>>
>>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>>Gesendet: Do 25.03.2004 20:32
>>>An: spr...@li...; jürgen höller [werk3AT]
>>>Betreff: Re: [Springframework-developer] Hibernate resource management issue
>>>
>>>
>>>
>>>Juergen,
>>>
>>>I am almost 100% sure this block of code from
>>>AbstractPlatformTransactionManager is wrong:
>>> else {
>>> try {
>>> try {
>>> triggerBeforeCommit(defStatus);
>>> triggerBeforeCompletion(defStatus);
>>> if (status.isNewTransaction()) {
>>> logger.info("Initiating transaction commit");
>>> doCommit(defStatus);
>>> }
>>> }
>>> catch (UnexpectedRollbackException ex) {
>>> triggerAfterCompletion(defStatus,
>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>> throw ex;
>>> }
>>> catch (TransactionException ex) {
>>> if (this.rollbackOnCommitFailure) {
>>> doRollbackOnCommitException(defStatus, ex);
>>> triggerAfterCompletion(defStatus,
>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>> }
>>> else {
>>> triggerAfterCompletion(defStatus,
>>>TransactionSynchronization.STATUS_UNKNOWN, ex);
>>> }
>>> throw ex;
>>> }
>>> catch (RuntimeException ex) {
>>> doRollbackOnCommitException(defStatus, ex);
>>> triggerAfterCompletion(defStatus,
>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>> throw ex;
>>> }
>>> catch (Error err) {
>>> doRollbackOnCommitException(defStatus, err);
>>> triggerAfterCompletion(defStatus,
>>>TransactionSynchronization.STATUS_UNKNOWN, err);
>>> throw err;
>>> }
>>> triggerAfterCompletion(defStatus,
>>>TransactionSynchronization.STATUS_COMMITTED, null);
>>> }
>>> finally {
>>> cleanupAfterCompletion(defStatus);
>>> }
>>> }
>>>
>>>triggerBeforeCommit() execute, which in the Hibernate case will force a
>>>flush. However, if that flush throws an exception, then
>>> triggerBeforeCompletion(defStatus);
>>>never gets called. However, triggerBeforeCompletion is what is actually
>>>supposed to release the Hibernate session holder from the current
>>>thread! So in this case, the session (which is totally hosed of course),
>>>gets left on the thread. I believe in some environments this wouldn't
>>>matter that much, as the threads don't get resused. In the JBoss case,
>>>new requests coming in will get the existing thread, and this time,
>>>SessionFactoryUtils will see the session is there, and try to use it.
>>>Bang, it all blows up...
>>>
>>>So for this code to work properly, what needs to happen is that
>>>triggerBeforeCompletion still needs to be called even if
>>>triggerBeforeCommit fails. While I am ok with writing the code in this
>>>method to handle this, I am not 100% sure this is safe in terms of all
>>>the other interactions that will happen as a result; mot of this code is
>>>your baby with me only having traced through it once in a while. So if
>>>you would prefer to resolve this that would be great.
>>>
>>>Unless I am mistaken about this bug, I think it is a pretty serious one,
>>>and warrants an almost immediate release of a v1.0.1 of Spring...
>>>
>>>Regards,
>>>Colin
>>>
>>>
>>>Colin Sampaleanu wrote:
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>>I am tracking down a possible Hibernate resource management issue.
>>>>
>>>>In a running app, some time yesterday, some code, running in a wrapped
>>>>transaction with Hibernate handling ORM, encountered an Oracle
>>>>constraint violation and threw an exception. Fine...
>>>>
>>>>But when I log into the app myself now via the web ui and then it gets
>>>>a service object to read some data, the service object is wrapped with
>>>>a transaction interceptor, and also a hibernate interceptor. The
>>>>Hibernate interceptor is already seeing a Hibernate Session existing
>>>>on the current thread, so it is not creating a new one. Then at the
>>>>end of the transaction, when the Hibernate session is attempted to be
>>>>flushed, Hibernate tries to write out the old bad data from yesterday.
>>>>
>>>>What this essentially means is that when the error from yesterday
>>>>happened, the session did not get released from the thread, and has
>>>>been sticking around all this time. When I came via struts, I was
>>>>given the same thread as yesterday by the appserver, and the old
>>>>invalid session was still on it. The problem is not that it's reusing
>>>>that session, but why it was ever left that the day before.
>>>>
>>>>Will try to duplicate this...
>>>>
>>>>
|
|
From: <jue...@we...> - 2004-03-26 14:39:14
|
triggerAfterCompletion is now called by doRollbackOnCommitException =
(that the catch blocks delegate to), with the correct completion status. =
I've just added catch blocks for RuntimeException and Error there, =
calling triggerAfterCompletion too - although a rollback should never =
throw an exception other than TransactionException, so that's just to be =
on the safe side.
Coincidentally, we're just debugging a ThreadLocal problem at werk3AT, =
but with HibernateTransactionManager: In case of some =
StackOverflowErrors thrown by data access operations, the catch/finally =
blocks of our transaction management do not get called (though they =
catch Error), thus the Hibernate Session stays attached to the thread. =
I'm not sure if I understand this: Even in case of an Error, catch and =
finally blocks should get called by the VM... We're currently trying to =
reproduce this, as it just occurs on some machines and just in special =
scenarios.
Juergen
-----Original Message-----
From: Colin Sampaleanu [mailto:col...@ex...]
Sent: Friday, March 26, 2004 1:54 PM
To: spr...@li...; j=FCrgen h=F6ller
[werk3AT]
Subject: Re: [Springframework-developer] Hibernate resource management
issue
I see you also changed the behaviour so that on a RuntimeException or=20
Error, triggerAfterCompletion does not get called any longer, whereas it =
did with the previous code. I am not 100% sure why this went away,=20
considering it was there before, and it is still being done for=20
UnexpectedRollbackException and TransactionException. (However, I'm not =
100% clear on the semantics of triggerAfterCompletion and whether it=20
always has to be called, I presume it does).
j=FCrgen h=F6ller [werk3AT] wrote:
>I see - if flush fails, the Session is closed (if no =
TransactionManagerLookup) but not removed from the thread. I've already =
fixed the issue; will commit it promptly. Of course, none of this =
affects HibernateTransactionManager (which is assumably the primary =
choice for Spring apps), but it's still a nasty issue.
>=20
>Bad timing - one day earlier, and we would simply have delayed 1.0 =
final for this. I agree that we should do a quick 1.0.1 followup: I've =
also incorporated more sophisticated FieldError message resolution, as =
suggested recently, and will also look at the reported auto-proxy =
creator issue with multiple afterPropertiesSet calls.
>=20
>All things considered, I suggest 1.0.1 mid next week. Let's also =
incorporate all reported documentation inconsistencies and the like.
>=20
>Juergen
>=20
>
>________________________________
>
>Von: Colin Sampaleanu [mailto:col...@ex...]
>Gesendet: Do 25.03.2004 21:36
>An: j=FCrgen h=F6ller [werk3AT]
>Cc: spr...@li...
>Betreff: Re: [Springframework-developer] Hibernate resource management =
issue
>
>
>
>Things are simpler than they appear, and are actually as per my =
original
>email. Look at this, from SessionFactoryUtils:
> public void beforeCompletion() throws
>CleanupFailureDataAccessException {
> if (this.newSession) {
> =20
>TransactionSynchronizationManager.unbindResource(this.sessionFactory);
> if (this.hibernateTransactionCompletion) {
> =20
>closeSessionIfNecessary(this.sessionHolder.getSession(),
>this.sessionFactory);
> }
> }
> }
>
> public void afterCompletion(int status) {
> if (!this.hibernateTransactionCompletion) {
> Session session =3D this.sessionHolder.getSession();
> if (session instanceof SessionImplementor) {
> ((SessionImplementor)
>session).afterTransactionCompletion(status =3D=3D STATUS_COMMITTED);
> }
> if (this.newSession) {
> closeSessionIfNecessary(session, =
this.sessionFactory);
> }
> }
> this.sessionHolder.setSynchronizedWithTransaction(false);
> }
>
>beforeCompletion is the only place that
> TransactionSynchronizationManager.unbindResource
>gets called for the SessionHolder, and in this case, beforeCompletion
>never gets called.
>
>So unfortunately this _is_ a serious bug. Any exception during the
>beforeCommit call (which calls flush) means that the SessionHolder =
never
>gets unbound from the thread. Nasty!
>
>Colin
>
>
>j=FCrgen h=F6ller [werk3AT] wrote:
>
> =20
>
>>Odd - if there's no TransactionManagerLookup, the Session should =
definitely be closed in afterCompletion.
>>
>>Regarding beforeCompletion, the semantics indeed need to be clarified: =
I guess it's appropriate to always invoke it, even on beforeCommit =
failure.
>>
>>Juergen
>>
>>
>>________________________________
>>
>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>Gesendet: Do 25.03.2004 21:09
>>An: j=FCrgen h=F6ller [werk3AT]
>>Cc: spr...@li...
>>Betreff: Re: [Springframework-developer] Hibernate resource management =
issue
>>
>>
>>
>>We're on slightly different pages though. In my case, I am actually =
not
>>using TransactionManagerLookup... But in my case, because of the
>>exception in the flush in beforeCommit(), beforeCompletion() never =
gets
>>called (as it would normally). Now I do see that afterCompletion is =
also
>>supposed to call closeSessionIfNecessary() (I had actually missed =
this
>>before), but in my case, it's not getting there. I haven't traced it =
in
>>a debugger, which is what I will do now, as it seems to me it should =
get
>>to that code, certainly I do have a
>> "Triggering afterCompletion synchronization"
>>in the log which should happen right before afterCompletion() is =
called.
>>
>>The other question is whether beforeCompletion still shouldn't be =
called
>>in any case even if beforeCommit fails. Ultimately, we are still =
before
>>completion of the transaction, unless you meant it to be only called =
in
>>the case of no failure.
>>
>>Colin
>>
>>
>>j=FCrgen h=F6ller [werk3AT] wrote:
>>
>>
>>
>> =20
>>
>>>Colin,
>>>
>>>Thanks for tracking this down. It's actually a problem with =
SessionFactoryUtils' inner class SessionSynchronization: It assumes that =
beforeCompletion is called in any case, even if beforeCommit has thrown =
an exception. However, this just applies if you specified a =
TransactionManagerLookup in the Hibernate configuration; else, =
afterCompletion will do the cleanup - which will be called in any case.
>>>
>>>When you remove the Hibernate TransactionManagerLookup, you shouldn't =
face the issue - the bug doesn't have any effects then. Note that you =
don't need that TransactionManagerLookup when using Spring's =
JtaTransactionManager, as Spring will properly apply cache callbacks =
anyway. A TransactionManagerLookup just adds value when used with EJB =
CMT or manual JTA, for ultra-correct cache callbacks.
>>>
>>>So essentially, everything should be fine if using =
HibernateTransactionManager, or JtaTransactionManager without a =
Hibernate TransactionManagerLookup. This is clearly something to fix, =
but I guess we don't need to do an immediate 1.0.1 followup release; I'd =
like to gather further bug reports first. For the time being, let's =
suggest to remove the TransactionManagerLookup from the Hibernate =
configuration.
>>>
>>>Juergen
>>>
>>>
>>>________________________________
>>>
>>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>>Gesendet: Do 25.03.2004 20:32
>>>An: spr...@li...; j=FCrgen =
h=F6ller [werk3AT]
>>>Betreff: Re: [Springframework-developer] Hibernate resource =
management issue
>>>
>>>
>>>
>>>Juergen,
>>>
>>>I am almost 100% sure this block of code from
>>>AbstractPlatformTransactionManager is wrong:
>>> else {
>>> try {
>>> try {
>>> triggerBeforeCommit(defStatus);
>>> triggerBeforeCompletion(defStatus);
>>> if (status.isNewTransaction()) {
>>> logger.info("Initiating transaction commit");
>>> doCommit(defStatus);
>>> }
>>> }
>>> catch (UnexpectedRollbackException ex) {
>>> triggerAfterCompletion(defStatus,
>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>> throw ex;
>>> }
>>> catch (TransactionException ex) {
>>> if (this.rollbackOnCommitFailure) {
>>> doRollbackOnCommitException(defStatus, ex);
>>> triggerAfterCompletion(defStatus,
>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>> }
>>> else {
>>> triggerAfterCompletion(defStatus,
>>>TransactionSynchronization.STATUS_UNKNOWN, ex);
>>> }
>>> throw ex;
>>> }
>>> catch (RuntimeException ex) {
>>> doRollbackOnCommitException(defStatus, ex);
>>> triggerAfterCompletion(defStatus,
>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>> throw ex;
>>> }
>>> catch (Error err) {
>>> doRollbackOnCommitException(defStatus, err);
>>> triggerAfterCompletion(defStatus,
>>>TransactionSynchronization.STATUS_UNKNOWN, err);
>>> throw err;
>>> }
>>> triggerAfterCompletion(defStatus,
>>>TransactionSynchronization.STATUS_COMMITTED, null);
>>> }
>>> finally {
>>> cleanupAfterCompletion(defStatus);
>>> }
>>> }
>>>
>>>triggerBeforeCommit() execute, which in the Hibernate case will force =
a
>>>flush. However, if that flush throws an exception, then
>>> triggerBeforeCompletion(defStatus);
>>>never gets called. However, triggerBeforeCompletion is what is =
actually
>>>supposed to release the Hibernate session holder from the current
>>>thread! So in this case, the session (which is totally hosed of =
course),
>>>gets left on the thread. I believe in some environments this wouldn't
>>>matter that much, as the threads don't get resused. In the JBoss =
case,
>>>new requests coming in will get the existing thread, and this time,
>>>SessionFactoryUtils will see the session is there, and try to use it.
>>>Bang, it all blows up...
>>>
>>>So for this code to work properly, what needs to happen is that
>>>triggerBeforeCompletion still needs to be called even if
>>>triggerBeforeCommit fails. While I am ok with writing the code in =
this
>>>method to handle this, I am not 100% sure this is safe in terms of =
all
>>>the other interactions that will happen as a result; mot of this code =
is
>>>your baby with me only having traced through it once in a while. So =
if
>>>you would prefer to resolve this that would be great.
>>>
>>>Unless I am mistaken about this bug, I think it is a pretty serious =
one,
>>>and warrants an almost immediate release of a v1.0.1 of Spring...
>>>
>>>Regards,
>>>Colin
>>>
>>>
>>>Colin Sampaleanu wrote:
>>>
>>>
>>>
>>> =20
>>>
>>> =20
>>>
>>>>I am tracking down a possible Hibernate resource management issue.
>>>>
>>>>In a running app, some time yesterday, some code, running in a =
wrapped
>>>>transaction with Hibernate handling ORM, encountered an Oracle
>>>>constraint violation and threw an exception. Fine...
>>>>
>>>>But when I log into the app myself now via the web ui and then it =
gets
>>>>a service object to read some data, the service object is wrapped =
with
>>>>a transaction interceptor, and also a hibernate interceptor. The
>>>>Hibernate interceptor is already seeing a Hibernate Session existing
>>>>on the current thread, so it is not creating a new one. Then at the
>>>>end of the transaction, when the Hibernate session is attempted to =
be
>>>>flushed, Hibernate tries to write out the old bad data from =
yesterday.
>>>>
>>>>What this essentially means is that when the error from yesterday
>>>>happened, the session did not get released from the thread, and has
>>>>been sticking around all this time. When I came via struts, I was
>>>>given the same thread as yesterday by the appserver, and the old
>>>>invalid session was still on it. The problem is not that it's =
reusing
>>>>that session, but why it was ever left that the day before.
>>>>
>>>>Will try to duplicate this...
>>>> =20
>>>>
|
|
From: Colin S. <col...@ex...> - 2004-03-26 15:00:38
|
Well, I've never seen a case of the VM not calling catch and finally=20
blocks, so for the 2nd problem you have, so my feeling is that it's=20
almost certainly got to be a code issue, i.e. the code as written does=20
not to do exactly what you think it does... Could there be another=20
catch Throwable somewhere catching even the Error before it gets up to=20
that layer?
j=FCrgen h=F6ller [werk3AT] wrote:
>triggerAfterCompletion is now called by doRollbackOnCommitException (tha=
t the catch blocks delegate to), with the correct completion status. I've=
just added catch blocks for RuntimeException and Error there, calling tr=
iggerAfterCompletion too - although a rollback should never throw an exce=
ption other than TransactionException, so that's just to be on the safe s=
ide.
>
>Coincidentally, we're just debugging a ThreadLocal problem at werk3AT, b=
ut with HibernateTransactionManager: In case of some StackOverflowErrors =
thrown by data access operations, the catch/finally blocks of our transac=
tion management do not get called (though they catch Error), thus the Hib=
ernate Session stays attached to the thread. I'm not sure if I understand=
this: Even in case of an Error, catch and finally blocks should get call=
ed by the VM... We're currently trying to reproduce this, as it just occu=
rs on some machines and just in special scenarios.
>
>Juergen
>
>
>-----Original Message-----
>From: Colin Sampaleanu [mailto:col...@ex...]
>Sent: Friday, March 26, 2004 1:54 PM
>To: spr...@li...; j=FCrgen h=F6ller
>[werk3AT]
>Subject: Re: [Springframework-developer] Hibernate resource management
>issue
>
>
>I see you also changed the behaviour so that on a RuntimeException or=20
>Error, triggerAfterCompletion does not get called any longer, whereas it=
=20
>did with the previous code. I am not 100% sure why this went away,=20
>considering it was there before, and it is still being done for=20
>UnexpectedRollbackException and TransactionException. (However, I'm not=
=20
>100% clear on the semantics of triggerAfterCompletion and whether it=20
>always has to be called, I presume it does).
>
>
>j=FCrgen h=F6ller [werk3AT] wrote:
>
> =20
>
>>I see - if flush fails, the Session is closed (if no TransactionManager=
Lookup) but not removed from the thread. I've already fixed the issue; wi=
ll commit it promptly. Of course, none of this affects HibernateTransacti=
onManager (which is assumably the primary choice for Spring apps), but it=
's still a nasty issue.
>>
>>Bad timing - one day earlier, and we would simply have delayed 1.0 fina=
l for this. I agree that we should do a quick 1.0.1 followup: I've also i=
ncorporated more sophisticated FieldError message resolution, as suggeste=
d recently, and will also look at the reported auto-proxy creator issue w=
ith multiple afterPropertiesSet calls.
>>
>>All things considered, I suggest 1.0.1 mid next week. Let's also incorp=
orate all reported documentation inconsistencies and the like.
>>
>>Juergen
>>
>>
>>________________________________
>>
>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>Gesendet: Do 25.03.2004 21:36
>>An: j=FCrgen h=F6ller [werk3AT]
>>Cc: spr...@li...
>>Betreff: Re: [Springframework-developer] Hibernate resource management =
issue
>>
>>
>>
>>Things are simpler than they appear, and are actually as per my origina=
l
>>email. Look at this, from SessionFactoryUtils:
>> public void beforeCompletion() throws
>>CleanupFailureDataAccessException {
>> if (this.newSession) {
>> =20
>>TransactionSynchronizationManager.unbindResource(this.sessionFactory);
>> if (this.hibernateTransactionCompletion) {
>> =20
>>closeSessionIfNecessary(this.sessionHolder.getSession(),
>>this.sessionFactory);
>> }
>> }
>> }
>>
>> public void afterCompletion(int status) {
>> if (!this.hibernateTransactionCompletion) {
>> Session session =3D this.sessionHolder.getSession();
>> if (session instanceof SessionImplementor) {
>> ((SessionImplementor)
>>session).afterTransactionCompletion(status =3D=3D STATUS_COMMITTED);
>> }
>> if (this.newSession) {
>> closeSessionIfNecessary(session, this.sessionFactory=
);
>> }
>> }
>> this.sessionHolder.setSynchronizedWithTransaction(false);
>> }
>>
>>beforeCompletion is the only place that
>> TransactionSynchronizationManager.unbindResource
>>gets called for the SessionHolder, and in this case, beforeCompletion
>>never gets called.
>>
>>So unfortunately this _is_ a serious bug. Any exception during the
>>beforeCommit call (which calls flush) means that the SessionHolder neve=
r
>>gets unbound from the thread. Nasty!
>>
>>Colin
>>
>>
>>j=FCrgen h=F6ller [werk3AT] wrote:
>>
>>=20
>>
>> =20
>>
>>>Odd - if there's no TransactionManagerLookup, the Session should defin=
itely be closed in afterCompletion.
>>>
>>>Regarding beforeCompletion, the semantics indeed need to be clarified:=
I guess it's appropriate to always invoke it, even on beforeCommit failu=
re.
>>>
>>>Juergen
>>>
>>>
>>>________________________________
>>>
>>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>>Gesendet: Do 25.03.2004 21:09
>>>An: j=FCrgen h=F6ller [werk3AT]
>>>Cc: spr...@li...
>>>Betreff: Re: [Springframework-developer] Hibernate resource management=
issue
>>>
>>>
>>>
>>>We're on slightly different pages though. In my case, I am actually no=
t
>>>using TransactionManagerLookup... But in my case, because of the
>>>exception in the flush in beforeCommit(), beforeCompletion() never get=
s
>>>called (as it would normally). Now I do see that afterCompletion is al=
so
>>>supposed to call closeSessionIfNecessary() (I had actually missed thi=
s
>>>before), but in my case, it's not getting there. I haven't traced it i=
n
>>>a debugger, which is what I will do now, as it seems to me it should g=
et
>>>to that code, certainly I do have a
>>> "Triggering afterCompletion synchronization"
>>>in the log which should happen right before afterCompletion() is calle=
d.
>>>
>>>The other question is whether beforeCompletion still shouldn't be call=
ed
>>>in any case even if beforeCommit fails. Ultimately, we are still befor=
e
>>>completion of the transaction, unless you meant it to be only called i=
n
>>>the case of no failure.
>>>
>>>Colin
>>>
>>>
>>>j=FCrgen h=F6ller [werk3AT] wrote:
>>>
>>>
>>>
>>> =20
>>>
>>> =20
>>>
>>>>Colin,
>>>>
>>>>Thanks for tracking this down. It's actually a problem with SessionFa=
ctoryUtils' inner class SessionSynchronization: It assumes that beforeCom=
pletion is called in any case, even if beforeCommit has thrown an excepti=
on. However, this just applies if you specified a TransactionManagerLooku=
p in the Hibernate configuration; else, afterCompletion will do the clean=
up - which will be called in any case.
>>>>
>>>>When you remove the Hibernate TransactionManagerLookup, you shouldn't=
face the issue - the bug doesn't have any effects then. Note that you do=
n't need that TransactionManagerLookup when using Spring's JtaTransaction=
Manager, as Spring will properly apply cache callbacks anyway. A Transact=
ionManagerLookup just adds value when used with EJB CMT or manual JTA, fo=
r ultra-correct cache callbacks.
>>>>
>>>>So essentially, everything should be fine if using HibernateTransacti=
onManager, or JtaTransactionManager without a Hibernate TransactionManage=
rLookup. This is clearly something to fix, but I guess we don't need to d=
o an immediate 1.0.1 followup release; I'd like to gather further bug rep=
orts first. For the time being, let's suggest to remove the TransactionMa=
nagerLookup from the Hibernate configuration.
>>>>
>>>>Juergen
>>>>
>>>>
>>>>________________________________
>>>>
>>>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>>>Gesendet: Do 25.03.2004 20:32
>>>>An: spr...@li...; j=FCrgen h=F6lle=
r [werk3AT]
>>>>Betreff: Re: [Springframework-developer] Hibernate resource managemen=
t issue
>>>>
>>>>
>>>>
>>>>Juergen,
>>>>
>>>>I am almost 100% sure this block of code from
>>>>AbstractPlatformTransactionManager is wrong:
>>>> else {
>>>> try {
>>>> try {
>>>> triggerBeforeCommit(defStatus);
>>>> triggerBeforeCompletion(defStatus);
>>>> if (status.isNewTransaction()) {
>>>> logger.info("Initiating transaction commit");
>>>> doCommit(defStatus);
>>>> }
>>>> }
>>>> catch (UnexpectedRollbackException ex) {
>>>> triggerAfterCompletion(defStatus,
>>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>>> throw ex;
>>>> }
>>>> catch (TransactionException ex) {
>>>> if (this.rollbackOnCommitFailure) {
>>>> doRollbackOnCommitException(defStatus, ex);
>>>> triggerAfterCompletion(defStatus,
>>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>>> }
>>>> else {
>>>> triggerAfterCompletion(defStatus,
>>>>TransactionSynchronization.STATUS_UNKNOWN, ex);
>>>> }
>>>> throw ex;
>>>> }
>>>> catch (RuntimeException ex) {
>>>> doRollbackOnCommitException(defStatus, ex);
>>>> triggerAfterCompletion(defStatus,
>>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>>> throw ex;
>>>> }
>>>> catch (Error err) {
>>>> doRollbackOnCommitException(defStatus, err);
>>>> triggerAfterCompletion(defStatus,
>>>>TransactionSynchronization.STATUS_UNKNOWN, err);
>>>> throw err;
>>>> }
>>>> triggerAfterCompletion(defStatus,
>>>>TransactionSynchronization.STATUS_COMMITTED, null);
>>>> }
>>>> finally {
>>>> cleanupAfterCompletion(defStatus);
>>>> }
>>>> }
>>>>
>>>>triggerBeforeCommit() execute, which in the Hibernate case will force=
a
>>>>flush. However, if that flush throws an exception, then
>>>>triggerBeforeCompletion(defStatus);
>>>>never gets called. However, triggerBeforeCompletion is what is actual=
ly
>>>>supposed to release the Hibernate session holder from the current
>>>>thread! So in this case, the session (which is totally hosed of cours=
e),
>>>>gets left on the thread. I believe in some environments this wouldn't
>>>>matter that much, as the threads don't get resused. In the JBoss case=
,
>>>>new requests coming in will get the existing thread, and this time,
>>>>SessionFactoryUtils will see the session is there, and try to use it.
>>>>Bang, it all blows up...
>>>>
>>>>So for this code to work properly, what needs to happen is that
>>>>triggerBeforeCompletion still needs to be called even if
>>>>triggerBeforeCommit fails. While I am ok with writing the code in thi=
s
>>>>method to handle this, I am not 100% sure this is safe in terms of al=
l
>>>>the other interactions that will happen as a result; mot of this code=
is
>>>>your baby with me only having traced through it once in a while. So i=
f
>>>>you would prefer to resolve this that would be great.
>>>>
>>>>Unless I am mistaken about this bug, I think it is a pretty serious o=
ne,
>>>>and warrants an almost immediate release of a v1.0.1 of Spring...
>>>>
>>>>Regards,
>>>>Colin
>>>>
>>>>
>>>>Colin Sampaleanu wrote:
>>>>
>>>>
>>>>
>>>>=20
>>>>
>>>> =20
>>>>
>>>> =20
>>>>
>>>>>I am tracking down a possible Hibernate resource management issue.
>>>>>
>>>>>In a running app, some time yesterday, some code, running in a wrapp=
ed
>>>>>transaction with Hibernate handling ORM, encountered an Oracle
>>>>>constraint violation and threw an exception. Fine...
>>>>>
>>>>>But when I log into the app myself now via the web ui and then it ge=
ts
>>>>>a service object to read some data, the service object is wrapped wi=
th
>>>>>a transaction interceptor, and also a hibernate interceptor. The
>>>>>Hibernate interceptor is already seeing a Hibernate Session existing
>>>>>on the current thread, so it is not creating a new one. Then at the
>>>>>end of the transaction, when the Hibernate session is attempted to b=
e
>>>>>flushed, Hibernate tries to write out the old bad data from yesterda=
y.
>>>>>
>>>>>What this essentially means is that when the error from yesterday
>>>>>happened, the session did not get released from the thread, and has
>>>>>been sticking around all this time. When I came via struts, I was
>>>>>given the same thread as yesterday by the appserver, and the old
>>>>>invalid session was still on it. The problem is not that it's reusin=
g
>>>>>that session, but why it was ever left that the day before.
>>>>>
>>>>>Will try to duplicate this...
>>>>> =20
>>>>>
>>>>> =20
>>>>>
>
>
> =20
>
|
|
From: <jue...@we...> - 2004-03-26 21:40:00
|
The bug was pretty obscure: It was caused by an exception that came =
across the wire via Hessian. (We had to track down that scenario first.) =
The exception was a custom subclass of Spring's NestedRuntimeException. =
For whatever reason, Hessian put a reference to the exception itself =
into the "cause" instance variable. (This situation cannot occur through =
normal usage of NestedRuntimeException, as it just allows to specify the =
cause in a constructor.) This caused NestedRuntimeException's getMessage =
and printStackTrace to recurse indefinitely, resulting in a =
StackOverflowError. As this was happening in logging code within a catch =
block of TransactionInterceptor, rollback was never invoked, thus no =
cleanup of the associated transactional resources.
=20
A candidate for the most obscure bug in quite a while...
=20
Juergen
=20
________________________________
Von: spr...@li... im Auftrag =
von Colin Sampaleanu
Gesendet: Fr 26.03.2004 16:00
An: spr...@li...
Betreff: Re: [Springframework-developer] Hibernate resource management =
issue
Well, I've never seen a case of the VM not calling catch and finally
blocks, so for the 2nd problem you have, so my feeling is that it's
almost certainly got to be a code issue, i.e. the code as written does
not to do exactly what you think it does... Could there be another
catch Throwable somewhere catching even the Error before it gets up to
that layer?
j=FCrgen h=F6ller [werk3AT] wrote:
>triggerAfterCompletion is now called by doRollbackOnCommitException =
(that the catch blocks delegate to), with the correct completion status. =
I've just added catch blocks for RuntimeException and Error there, =
calling triggerAfterCompletion too - although a rollback should never =
throw an exception other than TransactionException, so that's just to be =
on the safe side.
>
>Coincidentally, we're just debugging a ThreadLocal problem at werk3AT, =
but with HibernateTransactionManager: In case of some =
StackOverflowErrors thrown by data access operations, the catch/finally =
blocks of our transaction management do not get called (though they =
catch Error), thus the Hibernate Session stays attached to the thread. =
I'm not sure if I understand this: Even in case of an Error, catch and =
finally blocks should get called by the VM... We're currently trying to =
reproduce this, as it just occurs on some machines and just in special =
scenarios.
>
>Juergen
>
>
>-----Original Message-----
>From: Colin Sampaleanu [mailto:col...@ex...]
>Sent: Friday, March 26, 2004 1:54 PM
>To: spr...@li...; j=FCrgen h=F6ller
>[werk3AT]
>Subject: Re: [Springframework-developer] Hibernate resource management
>issue
>
>
>I see you also changed the behaviour so that on a RuntimeException or
>Error, triggerAfterCompletion does not get called any longer, whereas =
it
>did with the previous code. I am not 100% sure why this went away,
>considering it was there before, and it is still being done for
>UnexpectedRollbackException and TransactionException. (However, I'm =
not
>100% clear on the semantics of triggerAfterCompletion and whether it
>always has to be called, I presume it does).
>
>
>j=FCrgen h=F6ller [werk3AT] wrote:
>
>=20
>
>>I see - if flush fails, the Session is closed (if no =
TransactionManagerLookup) but not removed from the thread. I've already =
fixed the issue; will commit it promptly. Of course, none of this =
affects HibernateTransactionManager (which is assumably the primary =
choice for Spring apps), but it's still a nasty issue.
>>
>>Bad timing - one day earlier, and we would simply have delayed 1.0 =
final for this. I agree that we should do a quick 1.0.1 followup: I've =
also incorporated more sophisticated FieldError message resolution, as =
suggested recently, and will also look at the reported auto-proxy =
creator issue with multiple afterPropertiesSet calls.
>>
>>All things considered, I suggest 1.0.1 mid next week. Let's also =
incorporate all reported documentation inconsistencies and the like.
>>
>>Juergen
>>
>>
>>________________________________
>>
>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>Gesendet: Do 25.03.2004 21:36
>>An: j=FCrgen h=F6ller [werk3AT]
>>Cc: spr...@li...
>>Betreff: Re: [Springframework-developer] Hibernate resource management =
issue
>>
>>
>>
>>Things are simpler than they appear, and are actually as per my =
original
>>email. Look at this, from SessionFactoryUtils:
>> public void beforeCompletion() throws
>>CleanupFailureDataAccessException {
>> if (this.newSession) {
>> =20
>>TransactionSynchronizationManager.unbindResource(this.sessionFactory);
>> if (this.hibernateTransactionCompletion) {
>> =20
>>closeSessionIfNecessary(this.sessionHolder.getSession(),
>>this.sessionFactory);
>> }
>> }
>> }
>>
>> public void afterCompletion(int status) {
>> if (!this.hibernateTransactionCompletion) {
>> Session session =3D this.sessionHolder.getSession();
>> if (session instanceof SessionImplementor) {
>> ((SessionImplementor)
>>session).afterTransactionCompletion(status =3D=3D STATUS_COMMITTED);
>> }
>> if (this.newSession) {
>> closeSessionIfNecessary(session, =
this.sessionFactory);
>> }
>> }
>> this.sessionHolder.setSynchronizedWithTransaction(false);
>> }
>>
>>beforeCompletion is the only place that
>> TransactionSynchronizationManager.unbindResource
>>gets called for the SessionHolder, and in this case, beforeCompletion
>>never gets called.
>>
>>So unfortunately this _is_ a serious bug. Any exception during the
>>beforeCommit call (which calls flush) means that the SessionHolder =
never
>>gets unbound from the thread. Nasty!
>>
>>Colin
>>
>>
>>j=FCrgen h=F6ller [werk3AT] wrote:
>>
>>
>>
>> =20
>>
>>>Odd - if there's no TransactionManagerLookup, the Session should =
definitely be closed in afterCompletion.
>>>
>>>Regarding beforeCompletion, the semantics indeed need to be =
clarified: I guess it's appropriate to always invoke it, even on =
beforeCommit failure.
>>>
>>>Juergen
>>>
>>>
>>>________________________________
>>>
>>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>>Gesendet: Do 25.03.2004 21:09
>>>An: j=FCrgen h=F6ller [werk3AT]
>>>Cc: spr...@li...
>>>Betreff: Re: [Springframework-developer] Hibernate resource =
management issue
>>>
>>>
>>>
>>>We're on slightly different pages though. In my case, I am actually =
not
>>>using TransactionManagerLookup... But in my case, because of the
>>>exception in the flush in beforeCommit(), beforeCompletion() never =
gets
>>>called (as it would normally). Now I do see that afterCompletion is =
also
>>>supposed to call closeSessionIfNecessary() (I had actually missed =
this
>>>before), but in my case, it's not getting there. I haven't traced it =
in
>>>a debugger, which is what I will do now, as it seems to me it should =
get
>>>to that code, certainly I do have a
>>> "Triggering afterCompletion synchronization"
>>>in the log which should happen right before afterCompletion() is =
called.
>>>
>>>The other question is whether beforeCompletion still shouldn't be =
called
>>>in any case even if beforeCommit fails. Ultimately, we are still =
before
>>>completion of the transaction, unless you meant it to be only called =
in
>>>the case of no failure.
>>>
>>>Colin
>>>
>>>
>>>j=FCrgen h=F6ller [werk3AT] wrote:
>>>
>>>
>>>
>>> =20
>>>
>>> =20
>>>
>>>>Colin,
>>>>
>>>>Thanks for tracking this down. It's actually a problem with =
SessionFactoryUtils' inner class SessionSynchronization: It assumes that =
beforeCompletion is called in any case, even if beforeCommit has thrown =
an exception. However, this just applies if you specified a =
TransactionManagerLookup in the Hibernate configuration; else, =
afterCompletion will do the cleanup - which will be called in any case.
>>>>
>>>>When you remove the Hibernate TransactionManagerLookup, you =
shouldn't face the issue - the bug doesn't have any effects then. Note =
that you don't need that TransactionManagerLookup when using Spring's =
JtaTransactionManager, as Spring will properly apply cache callbacks =
anyway. A TransactionManagerLookup just adds value when used with EJB =
CMT or manual JTA, for ultra-correct cache callbacks.
>>>>
>>>>So essentially, everything should be fine if using =
HibernateTransactionManager, or JtaTransactionManager without a =
Hibernate TransactionManagerLookup. This is clearly something to fix, =
but I guess we don't need to do an immediate 1.0.1 followup release; I'd =
like to gather further bug reports first. For the time being, let's =
suggest to remove the TransactionManagerLookup from the Hibernate =
configuration.
>>>>
>>>>Juergen
>>>>
>>>>
>>>>________________________________
>>>>
>>>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>>>Gesendet: Do 25.03.2004 20:32
>>>>An: spr...@li...; j=FCrgen =
h=F6ller [werk3AT]
>>>>Betreff: Re: [Springframework-developer] Hibernate resource =
management issue
>>>>
>>>>
>>>>
>>>>Juergen,
>>>>
>>>>I am almost 100% sure this block of code from
>>>>AbstractPlatformTransactionManager is wrong:
>>>> else {
>>>> try {
>>>> try {
>>>> triggerBeforeCommit(defStatus);
>>>> triggerBeforeCompletion(defStatus);
>>>> if (status.isNewTransaction()) {
>>>> logger.info("Initiating transaction commit");
>>>> doCommit(defStatus);
>>>> }
>>>> }
>>>> catch (UnexpectedRollbackException ex) {
>>>> triggerAfterCompletion(defStatus,
>>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>>> throw ex;
>>>> }
>>>> catch (TransactionException ex) {
>>>> if (this.rollbackOnCommitFailure) {
>>>> doRollbackOnCommitException(defStatus, ex);
>>>> triggerAfterCompletion(defStatus,
>>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>>> }
>>>> else {
>>>> triggerAfterCompletion(defStatus,
>>>>TransactionSynchronization.STATUS_UNKNOWN, ex);
>>>> }
>>>> throw ex;
>>>> }
>>>> catch (RuntimeException ex) {
>>>> doRollbackOnCommitException(defStatus, ex);
>>>> triggerAfterCompletion(defStatus,
>>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>>> throw ex;
>>>> }
>>>> catch (Error err) {
>>>> doRollbackOnCommitException(defStatus, err);
>>>> triggerAfterCompletion(defStatus,
>>>>TransactionSynchronization.STATUS_UNKNOWN, err);
>>>> throw err;
>>>> }
>>>> triggerAfterCompletion(defStatus,
>>>>TransactionSynchronization.STATUS_COMMITTED, null);
>>>> }
>>>> finally {
>>>> cleanupAfterCompletion(defStatus);
>>>> }
>>>> }
>>>>
>>>>triggerBeforeCommit() execute, which in the Hibernate case will =
force a
>>>>flush. However, if that flush throws an exception, then
>>>>triggerBeforeCompletion(defStatus);
>>>>never gets called. However, triggerBeforeCompletion is what is =
actually
>>>>supposed to release the Hibernate session holder from the current
>>>>thread! So in this case, the session (which is totally hosed of =
course),
>>>>gets left on the thread. I believe in some environments this =
wouldn't
>>>>matter that much, as the threads don't get resused. In the JBoss =
case,
>>>>new requests coming in will get the existing thread, and this time,
>>>>SessionFactoryUtils will see the session is there, and try to use =
it.
>>>>Bang, it all blows up...
>>>>
>>>>So for this code to work properly, what needs to happen is that
>>>>triggerBeforeCompletion still needs to be called even if
>>>>triggerBeforeCommit fails. While I am ok with writing the code in =
this
>>>>method to handle this, I am not 100% sure this is safe in terms of =
all
>>>>the other interactions that will happen as a result; mot of this =
code is
>>>>your baby with me only having traced through it once in a while. So =
if
>>>>you would prefer to resolve this that would be great.
>>>>
>>>>Unless I am mistaken about this bug, I think it is a pretty serious =
one,
>>>>and warrants an almost immediate release of a v1.0.1 of Spring...
>>>>
>>>>Regards,
>>>>Colin
>>>>
>>>>
>>>>Colin Sampaleanu wrote:
>>>>
>>>>
>>>>
>>>>
>>>>
>>>> =20
>>>>
>>>> =20
>>>>
>>>>>I am tracking down a possible Hibernate resource management issue.
>>>>>
>>>>>In a running app, some time yesterday, some code, running in a =
wrapped
>>>>>transaction with Hibernate handling ORM, encountered an Oracle
>>>>>constraint violation and threw an exception. Fine...
>>>>>
>>>>>But when I log into the app myself now via the web ui and then it =
gets
>>>>>a service object to read some data, the service object is wrapped =
with
>>>>>a transaction interceptor, and also a hibernate interceptor. The
>>>>>Hibernate interceptor is already seeing a Hibernate Session =
existing
>>>>>on the current thread, so it is not creating a new one. Then at the
>>>>>end of the transaction, when the Hibernate session is attempted to =
be
>>>>>flushed, Hibernate tries to write out the old bad data from =
yesterday.
>>>>>
>>>>>What this essentially means is that when the error from yesterday
>>>>>happened, the session did not get released from the thread, and has
>>>>>been sticking around all this time. When I came via struts, I was
>>>>>given the same thread as yesterday by the appserver, and the old
>>>>>invalid session was still on it. The problem is not that it's =
reusing
>>>>>that session, but why it was ever left that the day before.
>>>>>
>>>>>Will try to duplicate this...
>>>>> =20
>>>>>
>>>>> =20
>>>>>
>
>
>=20
>
-------------------------------------------------------
This SF.Net email is sponsored by: IBM Linux Tutorials
Free Linux tutorial presented by Daniel Robbins, President and CEO of
GenToo technologies. Learn everything from fundamentals to system
administration.http://ads.osdn.com/?ad_id=1470&alloc_id638&op=3Dick
_______________________________________________
Springframework-developer mailing list
Spr...@li...
https://lists.sourceforge.net/lists/listinfo/springframework-developer
|
|
From: Ronald H. <ro...@co...> - 2004-03-26 22:11:12
|
Wow, while you are busy doing search marvelous research Juergen, could you try and find out who really killed JFK? :) Obscure bug award +1 On a side note, I saw some mentions to a new book that you are writing? Is this correct and if so, where is it pre-ordable? I can handle a bit of spring now, but you keep amazing me with your knowledge of every little obscure detail of the inner workings of spring, and I wouldnt mind to read about this. Cheers Ronald jürgen höller [werk3AT] wrote: >The bug was pretty obscure: It was caused by an exception that came across the wire via Hessian. (We had to track down that scenario first.) The exception was a custom subclass of Spring's NestedRuntimeException. For whatever reason, Hessian put a reference to the exception itself into the "cause" instance variable. (This situation cannot occur through normal usage of NestedRuntimeException, as it just allows to specify the cause in a constructor.) This caused NestedRuntimeException's getMessage and printStackTrace to recurse indefinitely, resulting in a StackOverflowError. As this was happening in logging code within a catch block of TransactionInterceptor, rollback was never invoked, thus no cleanup of the associated transactional resources. > >A candidate for the most obscure bug in quite a while... > >Juergen > > |
|
From: Rod J. <rod...@in...> - 2004-03-26 22:20:54
|
http://www.amazon.com/exec/obidos/ASIN/0764558315/ ----- Original Message ----- From: "Ronald Haring" <ro...@co...> To: <spr...@li...> Sent: Friday, March 26, 2004 10:11 PM Subject: Re: [Springframework-developer] OT: Hibernate resource managemen= t issue > Wow, > > while you are busy doing search marvelous research Juergen, could you > try and find out who really killed JFK? :) > > Obscure bug award +1 > > On a side note, I saw some mentions to a new book that you are writing? > Is this correct and if so, where is it pre-ordable? I can handle a bit > of spring now, but you keep amazing me with your knowledge of every > little obscure detail of the inner workings of spring, and I wouldnt > mind to read about this. > > Cheers > Ronald > > > > j=FCrgen h=F6ller [werk3AT] wrote: > > >The bug was pretty obscure: It was caused by an exception that came across the wire via Hessian. (We had to track down that scenario first.) = The exception was a custom subclass of Spring's NestedRuntimeException. For whatever reason, Hessian put a reference to the exception itself into the "cause" instance variable. (This situation cannot occur through normal us= age of NestedRuntimeException, as it just allows to specify the cause in a constructor.) This caused NestedRuntimeException's getMessage and printStackTrace to recurse indefinitely, resulting in a StackOverflowErro= r. As this was happening in logging code within a catch block of TransactionInterceptor, rollback was never invoked, thus no cleanup of th= e associated transactional resources. > > > >A candidate for the most obscure bug in quite a while... > > > >Juergen > > > > > > > > ------------------------------------------------------- > This SF.Net email is sponsored by: IBM Linux Tutorials > Free Linux tutorial presented by Daniel Robbins, President and CEO of > GenToo technologies. Learn everything from fundamentals to system > administration.http://ads.osdn.com/?ad_id=3D1470&alloc_id=3D3638&op=3Dc= lick > _______________________________________________ > Springframework-developer mailing list > Spr...@li... > https://lists.sourceforge.net/lists/listinfo/springframework-developer |
|
From: James C. <jim...@do...> - 2004-03-27 15:04:14
|
Rod, Juergen, Despite the title's hinting, does this book fully explore the Spring Framework? > -----Original Message----- > From: spr...@li... > [mailto:spr...@li...] On Behalf > Of Rod Johnson > Sent: Friday, March 26, 2004 5:21 PM > To: spr...@li... > Subject: Re: [Springframework-developer] OT: Hibernate resource management > issue > > http://www.amazon.com/exec/obidos/ASIN/0764558315/ |
|
From: Rod J. <rod...@in...> - 2004-03-27 17:06:26
|
It's not a book about Spring, but a book about J2EE architecture in which all examples use Spring and which covers most of the important Spring functionality. Essential reading for Spring developers, or anyone interested in simpler, better J2EE architecture. ----- Original Message ----- From: "James Cook" <jim...@do...> To: <spr...@li...> Sent: Saturday, March 27, 2004 3:03 PM Subject: RE: [Springframework-developer] OT: Hibernate resource management issue > Rod, Juergen, > > Despite the title's hinting, does this book fully explore the Spring > Framework? > > > -----Original Message----- > > From: spr...@li... > > [mailto:spr...@li...] On Behalf > > Of Rod Johnson > > Sent: Friday, March 26, 2004 5:21 PM > > To: spr...@li... > > Subject: Re: [Springframework-developer] OT: Hibernate resource management > > issue > > > > http://www.amazon.com/exec/obidos/ASIN/0764558315/ > > > > > > ------------------------------------------------------- > This SF.Net email is sponsored by: IBM Linux Tutorials > Free Linux tutorial presented by Daniel Robbins, President and CEO of > GenToo technologies. Learn everything from fundamentals to system > administration.http://ads.osdn.com/?ad_id=1470&alloc_id=3638&op=click > _______________________________________________ > Springframework-developer mailing list > Spr...@li... > https://lists.sourceforge.net/lists/listinfo/springframework-developer |
|
From: Matthew E. P. <ma...@me...> - 2004-03-27 17:21:34
|
FYI: For those attending TSS Symposium, this book is one of the free book choices. Cheers, mattthew On Mar 27, 2004, at 11:06 AM, Rod Johnson wrote: > It's not a book about Spring, but a book about J2EE architecture in > which > all examples use Spring and which covers most of the important Spring > functionality. > > Essential reading for Spring developers, or anyone interested in > simpler, > better J2EE architecture. > > ----- Original Message ----- > From: "James Cook" <jim...@do...> > To: <spr...@li...> > Sent: Saturday, March 27, 2004 3:03 PM > Subject: RE: [Springframework-developer] OT: Hibernate resource > management > issue > > >> Rod, Juergen, >> >> Despite the title's hinting, does this book fully explore the Spring >> Framework? >> >>> -----Original Message----- >>> From: spr...@li... >>> [mailto:spr...@li...] On >>> Behalf >>> Of Rod Johnson >>> Sent: Friday, March 26, 2004 5:21 PM >>> To: spr...@li... >>> Subject: Re: [Springframework-developer] OT: Hibernate resource > management >>> issue >>> >>> http://www.amazon.com/exec/obidos/ASIN/0764558315/ >> >> >> >> >> >> ------------------------------------------------------- >> This SF.Net email is sponsored by: IBM Linux Tutorials >> Free Linux tutorial presented by Daniel Robbins, President and CEO of >> GenToo technologies. Learn everything from fundamentals to system >> administration.http://ads.osdn.com/?ad_id=1470&alloc_id=3638&op=click >> _______________________________________________ >> Springframework-developer mailing list >> Spr...@li... >> https://lists.sourceforge.net/lists/listinfo/springframework-developer > > > > > ------------------------------------------------------- > This SF.Net email is sponsored by: IBM Linux Tutorials > Free Linux tutorial presented by Daniel Robbins, President and CEO of > GenToo technologies. Learn everything from fundamentals to system > administration.http://ads.osdn.com/?ad_id=1470&alloc_id=3638&op=click > _______________________________________________ > Springframework-developer mailing list > Spr...@li... > https://lists.sourceforge.net/lists/listinfo/springframework-developer |
|
From: Colin S. <col...@ex...> - 2004-03-26 22:20:08
|
We should probably wrap all logging calls in critical sections (like=20
catch blocks) in the tx and thread resource mgmt code with empty=20
try/catch blocks, to avoid this sort of thing in the future.
j=FCrgen h=F6ller [werk3AT] wrote:
>The bug was pretty obscure: It was caused by an exception that came acro=
ss the wire via Hessian. (We had to track down that scenario first.) The =
exception was a custom subclass of Spring's NestedRuntimeException. For w=
hatever reason, Hessian put a reference to the exception itself into the =
"cause" instance variable. (This situation cannot occur through normal us=
age of NestedRuntimeException, as it just allows to specify the cause in =
a constructor.) This caused NestedRuntimeException's getMessage and print=
StackTrace to recurse indefinitely, resulting in a StackOverflowError. As=
this was happening in logging code within a catch block of TransactionIn=
terceptor, rollback was never invoked, thus no cleanup of the associated =
transactional resources.
>=20
>A candidate for the most obscure bug in quite a while...
>=20
>Juergen
>=20
>
>________________________________
>
>Von: spr...@li... im Auftrag vo=
n Colin Sampaleanu
>Gesendet: Fr 26.03.2004 16:00
>An: spr...@li...
>Betreff: Re: [Springframework-developer] Hibernate resource management i=
ssue
>
>
>
>Well, I've never seen a case of the VM not calling catch and finally
>blocks, so for the 2nd problem you have, so my feeling is that it's
>almost certainly got to be a code issue, i.e. the code as written does
>not to do exactly what you think it does... Could there be another
>catch Throwable somewhere catching even the Error before it gets up to
>that layer?
>
>j=FCrgen h=F6ller [werk3AT] wrote:
>
> =20
>
>>triggerAfterCompletion is now called by doRollbackOnCommitException (th=
at the catch blocks delegate to), with the correct completion status. I'v=
e just added catch blocks for RuntimeException and Error there, calling t=
riggerAfterCompletion too - although a rollback should never throw an exc=
eption other than TransactionException, so that's just to be on the safe =
side.
>>
>>Coincidentally, we're just debugging a ThreadLocal problem at werk3AT, =
but with HibernateTransactionManager: In case of some StackOverflowErrors=
thrown by data access operations, the catch/finally blocks of our transa=
ction management do not get called (though they catch Error), thus the Hi=
bernate Session stays attached to the thread. I'm not sure if I understan=
d this: Even in case of an Error, catch and finally blocks should get cal=
led by the VM... We're currently trying to reproduce this, as it just occ=
urs on some machines and just in special scenarios.
>>
>>Juergen
>>
>>
>>-----Original Message-----
>>From: Colin Sampaleanu [mailto:col...@ex...]
>>Sent: Friday, March 26, 2004 1:54 PM
>>To: spr...@li...; j=FCrgen h=F6ller
>>[werk3AT]
>>Subject: Re: [Springframework-developer] Hibernate resource management
>>issue
>>
>>
>>I see you also changed the behaviour so that on a RuntimeException or
>>Error, triggerAfterCompletion does not get called any longer, whereas i=
t
>>did with the previous code. I am not 100% sure why this went away,
>>considering it was there before, and it is still being done for
>>UnexpectedRollbackException and TransactionException. (However, I'm no=
t
>>100% clear on the semantics of triggerAfterCompletion and whether it
>>always has to be called, I presume it does).
>>
>>
>>j=FCrgen h=F6ller [werk3AT] wrote:
>>
>>
>>
>> =20
>>
>>>I see - if flush fails, the Session is closed (if no TransactionManage=
rLookup) but not removed from the thread. I've already fixed the issue; w=
ill commit it promptly. Of course, none of this affects HibernateTransact=
ionManager (which is assumably the primary choice for Spring apps), but i=
t's still a nasty issue.
>>>
>>>Bad timing - one day earlier, and we would simply have delayed 1.0 fin=
al for this. I agree that we should do a quick 1.0.1 followup: I've also =
incorporated more sophisticated FieldError message resolution, as suggest=
ed recently, and will also look at the reported auto-proxy creator issue =
with multiple afterPropertiesSet calls.
>>>
>>>All things considered, I suggest 1.0.1 mid next week. Let's also incor=
porate all reported documentation inconsistencies and the like.
>>>
>>>Juergen
>>>
>>>
>>>________________________________
>>>
>>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>>Gesendet: Do 25.03.2004 21:36
>>>An: j=FCrgen h=F6ller [werk3AT]
>>>Cc: spr...@li...
>>>Betreff: Re: [Springframework-developer] Hibernate resource management=
issue
>>>
>>>
>>>
>>>Things are simpler than they appear, and are actually as per my origin=
al
>>>email. Look at this, from SessionFactoryUtils:
>>> public void beforeCompletion() throws
>>>CleanupFailureDataAccessException {
>>> if (this.newSession) {
>>> =20
>>>TransactionSynchronizationManager.unbindResource(this.sessionFactory);
>>> if (this.hibernateTransactionCompletion) {
>>> =20
>>>closeSessionIfNecessary(this.sessionHolder.getSession(),
>>>this.sessionFactory);
>>> }
>>> }
>>> }
>>>
>>> public void afterCompletion(int status) {
>>> if (!this.hibernateTransactionCompletion) {
>>> Session session =3D this.sessionHolder.getSession();
>>> if (session instanceof SessionImplementor) {
>>> ((SessionImplementor)
>>>session).afterTransactionCompletion(status =3D=3D STATUS_COMMITTED);
>>> }
>>> if (this.newSession) {
>>> closeSessionIfNecessary(session, this.sessionFactory=
);
>>> }
>>> }
>>> this.sessionHolder.setSynchronizedWithTransaction(false);
>>> }
>>>
>>>beforeCompletion is the only place that
>>>TransactionSynchronizationManager.unbindResource
>>>gets called for the SessionHolder, and in this case, beforeCompletion
>>>never gets called.
>>>
>>>So unfortunately this _is_ a serious bug. Any exception during the
>>>beforeCommit call (which calls flush) means that the SessionHolder nev=
er
>>>gets unbound from the thread. Nasty!
>>>
>>>Colin
>>>
>>>
>>>j=FCrgen h=F6ller [werk3AT] wrote:
>>>
>>>
>>>
>>> =20
>>>
>>> =20
>>>
>>>>Odd - if there's no TransactionManagerLookup, the Session should defi=
nitely be closed in afterCompletion.
>>>>
>>>>Regarding beforeCompletion, the semantics indeed need to be clarified=
: I guess it's appropriate to always invoke it, even on beforeCommit fail=
ure.
>>>>
>>>>Juergen
>>>>
>>>>
>>>>________________________________
>>>>
>>>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>>>Gesendet: Do 25.03.2004 21:09
>>>>An: j=FCrgen h=F6ller [werk3AT]
>>>>Cc: spr...@li...
>>>>Betreff: Re: [Springframework-developer] Hibernate resource managemen=
t issue
>>>>
>>>>
>>>>
>>>>We're on slightly different pages though. In my case, I am actually n=
ot
>>>>using TransactionManagerLookup... But in my case, because of the
>>>>exception in the flush in beforeCommit(), beforeCompletion() never ge=
ts
>>>>called (as it would normally). Now I do see that afterCompletion is a=
lso
>>>>supposed to call closeSessionIfNecessary() (I had actually missed th=
is
>>>>before), but in my case, it's not getting there. I haven't traced it =
in
>>>>a debugger, which is what I will do now, as it seems to me it should =
get
>>>>to that code, certainly I do have a
>>>>"Triggering afterCompletion synchronization"
>>>>in the log which should happen right before afterCompletion() is call=
ed.
>>>>
>>>>The other question is whether beforeCompletion still shouldn't be cal=
led
>>>>in any case even if beforeCommit fails. Ultimately, we are still befo=
re
>>>>completion of the transaction, unless you meant it to be only called =
in
>>>>the case of no failure.
>>>>
>>>>Colin
>>>>
>>>>
>>>>j=FCrgen h=F6ller [werk3AT] wrote:
>>>>
>>>>
>>>>
>>>>=20
>>>>
>>>> =20
>>>>
>>>> =20
>>>>
>>>>>Colin,
>>>>>
>>>>>Thanks for tracking this down. It's actually a problem with SessionF=
actoryUtils' inner class SessionSynchronization: It assumes that beforeCo=
mpletion is called in any case, even if beforeCommit has thrown an except=
ion. However, this just applies if you specified a TransactionManagerLook=
up in the Hibernate configuration; else, afterCompletion will do the clea=
nup - which will be called in any case.
>>>>>
>>>>>When you remove the Hibernate TransactionManagerLookup, you shouldn'=
t face the issue - the bug doesn't have any effects then. Note that you d=
on't need that TransactionManagerLookup when using Spring's JtaTransactio=
nManager, as Spring will properly apply cache callbacks anyway. A Transac=
tionManagerLookup just adds value when used with EJB CMT or manual JTA, f=
or ultra-correct cache callbacks.
>>>>>
>>>>>So essentially, everything should be fine if using HibernateTransact=
ionManager, or JtaTransactionManager without a Hibernate TransactionManag=
erLookup. This is clearly something to fix, but I guess we don't need to =
do an immediate 1.0.1 followup release; I'd like to gather further bug re=
ports first. For the time being, let's suggest to remove the TransactionM=
anagerLookup from the Hibernate configuration.
>>>>>
>>>>>Juergen
>>>>>
>>>>>
>>>>>________________________________
>>>>>
>>>>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>>>>Gesendet: Do 25.03.2004 20:32
>>>>>An: spr...@li...; j=FCrgen h=F6ll=
er [werk3AT]
>>>>>Betreff: Re: [Springframework-developer] Hibernate resource manageme=
nt issue
>>>>>
>>>>>
>>>>>
>>>>>Juergen,
>>>>>
>>>>>I am almost 100% sure this block of code from
>>>>>AbstractPlatformTransactionManager is wrong:
>>>>> else {
>>>>> try {
>>>>> try {
>>>>> triggerBeforeCommit(defStatus);
>>>>> triggerBeforeCompletion(defStatus);
>>>>> if (status.isNewTransaction()) {
>>>>> logger.info("Initiating transaction commit");
>>>>> doCommit(defStatus);
>>>>> }
>>>>> }
>>>>> catch (UnexpectedRollbackException ex) {
>>>>> triggerAfterCompletion(defStatus,
>>>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>>>> throw ex;
>>>>> }
>>>>> catch (TransactionException ex) {
>>>>> if (this.rollbackOnCommitFailure) {
>>>>> doRollbackOnCommitException(defStatus, ex);
>>>>> triggerAfterCompletion(defStatus,
>>>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>>>> }
>>>>> else {
>>>>> triggerAfterCompletion(defStatus,
>>>>>TransactionSynchronization.STATUS_UNKNOWN, ex);
>>>>> }
>>>>> throw ex;
>>>>> }
>>>>> catch (RuntimeException ex) {
>>>>> doRollbackOnCommitException(defStatus, ex);
>>>>> triggerAfterCompletion(defStatus,
>>>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>>>> throw ex;
>>>>> }
>>>>> catch (Error err) {
>>>>> doRollbackOnCommitException(defStatus, err);
>>>>> triggerAfterCompletion(defStatus,
>>>>>TransactionSynchronization.STATUS_UNKNOWN, err);
>>>>> throw err;
>>>>> }
>>>>> triggerAfterCompletion(defStatus,
>>>>>TransactionSynchronization.STATUS_COMMITTED, null);
>>>>> }
>>>>> finally {
>>>>> cleanupAfterCompletion(defStatus);
>>>>> }
>>>>> }
>>>>>
>>>>>triggerBeforeCommit() execute, which in the Hibernate case will forc=
e a
>>>>>flush. However, if that flush throws an exception, then
>>>>>triggerBeforeCompletion(defStatus);
>>>>>never gets called. However, triggerBeforeCompletion is what is actua=
lly
>>>>>supposed to release the Hibernate session holder from the current
>>>>>thread! So in this case, the session (which is totally hosed of cour=
se),
>>>>>gets left on the thread. I believe in some environments this wouldn'=
t
>>>>>matter that much, as the threads don't get resused. In the JBoss cas=
e,
>>>>>new requests coming in will get the existing thread, and this time,
>>>>>SessionFactoryUtils will see the session is there, and try to use it=
.
>>>>>Bang, it all blows up...
>>>>>
>>>>>So for this code to work properly, what needs to happen is that
>>>>>triggerBeforeCompletion still needs to be called even if
>>>>>triggerBeforeCommit fails. While I am ok with writing the code in th=
is
>>>>>method to handle this, I am not 100% sure this is safe in terms of a=
ll
>>>>>the other interactions that will happen as a result; mot of this cod=
e is
>>>>>your baby with me only having traced through it once in a while. So =
if
>>>>>you would prefer to resolve this that would be great.
>>>>>
>>>>>Unless I am mistaken about this bug, I think it is a pretty serious =
one,
>>>>>and warrants an almost immediate release of a v1.0.1 of Spring...
>>>>>
>>>>>Regards,
>>>>>Colin
>>>>>
>>>>>
>>>>>Colin Sampaleanu wrote:
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> =20
>>>>>
>>>>> =20
>>>>>
>>>>> =20
>>>>>
>>>>>>I am tracking down a possible Hibernate resource management issue.
>>>>>>
>>>>>>In a running app, some time yesterday, some code, running in a wrap=
ped
>>>>>>transaction with Hibernate handling ORM, encountered an Oracle
>>>>>>constraint violation and threw an exception. Fine...
>>>>>>
>>>>>>But when I log into the app myself now via the web ui and then it g=
ets
>>>>>>a service object to read some data, the service object is wrapped w=
ith
>>>>>>a transaction interceptor, and also a hibernate interceptor. The
>>>>>>Hibernate interceptor is already seeing a Hibernate Session existin=
g
>>>>>>on the current thread, so it is not creating a new one. Then at the
>>>>>>end of the transaction, when the Hibernate session is attempted to =
be
>>>>>>flushed, Hibernate tries to write out the old bad data from yesterd=
ay.
>>>>>>
>>>>>>What this essentially means is that when the error from yesterday
>>>>>>happened, the session did not get released from the thread, and has
>>>>>>been sticking around all this time. When I came via struts, I was
>>>>>>given the same thread as yesterday by the appserver, and the old
>>>>>>invalid session was still on it. The problem is not that it's reusi=
ng
>>>>>>that session, but why it was ever left that the day before.
>>>>>>
>>>>>>Will try to duplicate this...
>>>>>> =20
>>>>>>
|
|
From: <jue...@we...> - 2004-03-27 10:48:57
|
I considered such wrapping of logging calls too, but I'm not sure =
whether to add it: The downside is that it clutters up the exception =
handling code. (AbstractPlatformTransactionManager has a quite elaborate =
try-try-catch-catch-catch-catch-finally block.)
=20
This sort of problem can just occur when the exception class messes up =
in getMessage or printStackTrace: Normally, this should not occur by any =
means. It certainly won't occur with J2SE 1.4 exception nesting, and it =
won't occur with the adapted version of our nested exception base =
classes.
=20
In our particular case, the deserialization mechanism of the remoting =
protocol was the cause: Nested exception base classes need to be =
programmed defensively here. Still adding try-catch blocks around =
logging statements might be overkill, although I don't have a strong =
opinion on this.
=20
Juergen
=20
________________________________
Von: spr...@li... im Auftrag =
von Colin Sampaleanu
Gesendet: Fr 26.03.2004 23:19
An: spr...@li...
Betreff: Re: [Springframework-developer] Hibernate resource management =
issue
We should probably wrap all logging calls in critical sections (like
catch blocks) in the tx and thread resource mgmt code with empty
try/catch blocks, to avoid this sort of thing in the future.
j=FCrgen h=F6ller [werk3AT] wrote:
>The bug was pretty obscure: It was caused by an exception that came =
across the wire via Hessian. (We had to track down that scenario first.) =
The exception was a custom subclass of Spring's NestedRuntimeException. =
For whatever reason, Hessian put a reference to the exception itself =
into the "cause" instance variable. (This situation cannot occur through =
normal usage of NestedRuntimeException, as it just allows to specify the =
cause in a constructor.) This caused NestedRuntimeException's getMessage =
and printStackTrace to recurse indefinitely, resulting in a =
StackOverflowError. As this was happening in logging code within a catch =
block of TransactionInterceptor, rollback was never invoked, thus no =
cleanup of the associated transactional resources.
>
>A candidate for the most obscure bug in quite a while...
>
>Juergen
>
>
>________________________________
>
>Von: spr...@li... im Auftrag =
von Colin Sampaleanu
>Gesendet: Fr 26.03.2004 16:00
>An: spr...@li...
>Betreff: Re: [Springframework-developer] Hibernate resource management =
issue
>
>
>
>Well, I've never seen a case of the VM not calling catch and finally
>blocks, so for the 2nd problem you have, so my feeling is that it's
>almost certainly got to be a code issue, i.e. the code as written does
>not to do exactly what you think it does... Could there be another
>catch Throwable somewhere catching even the Error before it gets up to
>that layer?
>
>j=FCrgen h=F6ller [werk3AT] wrote:
>
>=20
>
>>triggerAfterCompletion is now called by doRollbackOnCommitException =
(that the catch blocks delegate to), with the correct completion status. =
I've just added catch blocks for RuntimeException and Error there, =
calling triggerAfterCompletion too - although a rollback should never =
throw an exception other than TransactionException, so that's just to be =
on the safe side.
>>
>>Coincidentally, we're just debugging a ThreadLocal problem at werk3AT, =
but with HibernateTransactionManager: In case of some =
StackOverflowErrors thrown by data access operations, the catch/finally =
blocks of our transaction management do not get called (though they =
catch Error), thus the Hibernate Session stays attached to the thread. =
I'm not sure if I understand this: Even in case of an Error, catch and =
finally blocks should get called by the VM... We're currently trying to =
reproduce this, as it just occurs on some machines and just in special =
scenarios.
>>
>>Juergen
>>
>>
>>-----Original Message-----
>>From: Colin Sampaleanu [mailto:col...@ex...]
>>Sent: Friday, March 26, 2004 1:54 PM
>>To: spr...@li...; j=FCrgen h=F6ller
>>[werk3AT]
>>Subject: Re: [Springframework-developer] Hibernate resource management
>>issue
>>
>>
>>I see you also changed the behaviour so that on a RuntimeException or
>>Error, triggerAfterCompletion does not get called any longer, whereas =
it
>>did with the previous code. I am not 100% sure why this went away,
>>considering it was there before, and it is still being done for
>>UnexpectedRollbackException and TransactionException. (However, I'm =
not
>>100% clear on the semantics of triggerAfterCompletion and whether it
>>always has to be called, I presume it does).
>>
>>
>>j=FCrgen h=F6ller [werk3AT] wrote:
>>
>>
>>
>> =20
>>
>>>I see - if flush fails, the Session is closed (if no =
TransactionManagerLookup) but not removed from the thread. I've already =
fixed the issue; will commit it promptly. Of course, none of this =
affects HibernateTransactionManager (which is assumably the primary =
choice for Spring apps), but it's still a nasty issue.
>>>
>>>Bad timing - one day earlier, and we would simply have delayed 1.0 =
final for this. I agree that we should do a quick 1.0.1 followup: I've =
also incorporated more sophisticated FieldError message resolution, as =
suggested recently, and will also look at the reported auto-proxy =
creator issue with multiple afterPropertiesSet calls.
>>>
>>>All things considered, I suggest 1.0.1 mid next week. Let's also =
incorporate all reported documentation inconsistencies and the like.
>>>
>>>Juergen
>>>
>>>
>>>________________________________
>>>
>>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>>Gesendet: Do 25.03.2004 21:36
>>>An: j=FCrgen h=F6ller [werk3AT]
>>>Cc: spr...@li...
>>>Betreff: Re: [Springframework-developer] Hibernate resource =
management issue
>>>
>>>
>>>
>>>Things are simpler than they appear, and are actually as per my =
original
>>>email. Look at this, from SessionFactoryUtils:
>>> public void beforeCompletion() throws
>>>CleanupFailureDataAccessException {
>>> if (this.newSession) {
>>> =20
>>>TransactionSynchronizationManager.unbindResource(this.sessionFactory);=
>>> if (this.hibernateTransactionCompletion) {
>>> =20
>>>closeSessionIfNecessary(this.sessionHolder.getSession(),
>>>this.sessionFactory);
>>> }
>>> }
>>> }
>>>
>>> public void afterCompletion(int status) {
>>> if (!this.hibernateTransactionCompletion) {
>>> Session session =3D this.sessionHolder.getSession();
>>> if (session instanceof SessionImplementor) {
>>> ((SessionImplementor)
>>>session).afterTransactionCompletion(status =3D=3D STATUS_COMMITTED);
>>> }
>>> if (this.newSession) {
>>> closeSessionIfNecessary(session, =
this.sessionFactory);
>>> }
>>> }
>>> this.sessionHolder.setSynchronizedWithTransaction(false);
>>> }
>>>
>>>beforeCompletion is the only place that
>>>TransactionSynchronizationManager.unbindResource
>>>gets called for the SessionHolder, and in this case, beforeCompletion
>>>never gets called.
>>>
>>>So unfortunately this _is_ a serious bug. Any exception during the
>>>beforeCommit call (which calls flush) means that the SessionHolder =
never
>>>gets unbound from the thread. Nasty!
>>>
>>>Colin
>>>
>>>
>>>j=FCrgen h=F6ller [werk3AT] wrote:
>>>
>>>
>>>
>>>=20
>>>
>>> =20
>>>
>>>>Odd - if there's no TransactionManagerLookup, the Session should =
definitely be closed in afterCompletion.
>>>>
>>>>Regarding beforeCompletion, the semantics indeed need to be =
clarified: I guess it's appropriate to always invoke it, even on =
beforeCommit failure.
>>>>
>>>>Juergen
>>>>
>>>>
>>>>________________________________
>>>>
>>>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>>>Gesendet: Do 25.03.2004 21:09
>>>>An: j=FCrgen h=F6ller [werk3AT]
>>>>Cc: spr...@li...
>>>>Betreff: Re: [Springframework-developer] Hibernate resource =
management issue
>>>>
>>>>
>>>>
>>>>We're on slightly different pages though. In my case, I am actually =
not
>>>>using TransactionManagerLookup... But in my case, because of the
>>>>exception in the flush in beforeCommit(), beforeCompletion() never =
gets
>>>>called (as it would normally). Now I do see that afterCompletion is =
also
>>>>supposed to call closeSessionIfNecessary() (I had actually missed =
this
>>>>before), but in my case, it's not getting there. I haven't traced it =
in
>>>>a debugger, which is what I will do now, as it seems to me it should =
get
>>>>to that code, certainly I do have a
>>>>"Triggering afterCompletion synchronization"
>>>>in the log which should happen right before afterCompletion() is =
called.
>>>>
>>>>The other question is whether beforeCompletion still shouldn't be =
called
>>>>in any case even if beforeCommit fails. Ultimately, we are still =
before
>>>>completion of the transaction, unless you meant it to be only called =
in
>>>>the case of no failure.
>>>>
>>>>Colin
>>>>
>>>>
>>>>j=FCrgen h=F6ller [werk3AT] wrote:
>>>>
>>>>
>>>>
>>>>
>>>>
>>>> =20
>>>>
>>>> =20
>>>>
>>>>>Colin,
>>>>>
>>>>>Thanks for tracking this down. It's actually a problem with =
SessionFactoryUtils' inner class SessionSynchronization: It assumes that =
beforeCompletion is called in any case, even if beforeCommit has thrown =
an exception. However, this just applies if you specified a =
TransactionManagerLookup in the Hibernate configuration; else, =
afterCompletion will do the cleanup - which will be called in any case.
>>>>>
>>>>>When you remove the Hibernate TransactionManagerLookup, you =
shouldn't face the issue - the bug doesn't have any effects then. Note =
that you don't need that TransactionManagerLookup when using Spring's =
JtaTransactionManager, as Spring will properly apply cache callbacks =
anyway. A TransactionManagerLookup just adds value when used with EJB =
CMT or manual JTA, for ultra-correct cache callbacks.
>>>>>
>>>>>So essentially, everything should be fine if using =
HibernateTransactionManager, or JtaTransactionManager without a =
Hibernate TransactionManagerLookup. This is clearly something to fix, =
but I guess we don't need to do an immediate 1.0.1 followup release; I'd =
like to gather further bug reports first. For the time being, let's =
suggest to remove the TransactionManagerLookup from the Hibernate =
configuration.
>>>>>
>>>>>Juergen
>>>>>
>>>>>
>>>>>________________________________
>>>>>
>>>>>Von: Colin Sampaleanu [mailto:col...@ex...]
>>>>>Gesendet: Do 25.03.2004 20:32
>>>>>An: spr...@li...; j=FCrgen =
h=F6ller [werk3AT]
>>>>>Betreff: Re: [Springframework-developer] Hibernate resource =
management issue
>>>>>
>>>>>
>>>>>
>>>>>Juergen,
>>>>>
>>>>>I am almost 100% sure this block of code from
>>>>>AbstractPlatformTransactionManager is wrong:
>>>>> else {
>>>>> try {
>>>>> try {
>>>>> triggerBeforeCommit(defStatus);
>>>>> triggerBeforeCompletion(defStatus);
>>>>> if (status.isNewTransaction()) {
>>>>> logger.info("Initiating transaction commit");
>>>>> doCommit(defStatus);
>>>>> }
>>>>> }
>>>>> catch (UnexpectedRollbackException ex) {
>>>>> triggerAfterCompletion(defStatus,
>>>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>>>> throw ex;
>>>>> }
>>>>> catch (TransactionException ex) {
>>>>> if (this.rollbackOnCommitFailure) {
>>>>> doRollbackOnCommitException(defStatus, ex);
>>>>> triggerAfterCompletion(defStatus,
>>>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>>>> }
>>>>> else {
>>>>> triggerAfterCompletion(defStatus,
>>>>>TransactionSynchronization.STATUS_UNKNOWN, ex);
>>>>> }
>>>>> throw ex;
>>>>> }
>>>>> catch (RuntimeException ex) {
>>>>> doRollbackOnCommitException(defStatus, ex);
>>>>> triggerAfterCompletion(defStatus,
>>>>>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
>>>>> throw ex;
>>>>> }
>>>>> catch (Error err) {
>>>>> doRollbackOnCommitException(defStatus, err);
>>>>> triggerAfterCompletion(defStatus,
>>>>>TransactionSynchronization.STATUS_UNKNOWN, err);
>>>>> throw err;
>>>>> }
>>>>> triggerAfterCompletion(defStatus,
>>>>>TransactionSynchronization.STATUS_COMMITTED, null);
>>>>> }
>>>>> finally {
>>>>> cleanupAfterCompletion(defStatus);
>>>>> }
>>>>> }
>>>>>
>>>>>triggerBeforeCommit() execute, which in the Hibernate case will =
force a
>>>>>flush. However, if that flush throws an exception, then
>>>>>triggerBeforeCompletion(defStatus);
>>>>>never gets called. However, triggerBeforeCompletion is what is =
actually
>>>>>supposed to release the Hibernate session holder from the current
>>>>>thread! So in this case, the session (which is totally hosed of =
course),
>>>>>gets left on the thread. I believe in some environments this =
wouldn't
>>>>>matter that much, as the threads don't get resused. In the JBoss =
case,
>>>>>new requests coming in will get the existing thread, and this time,
>>>>>SessionFactoryUtils will see the session is there, and try to use =
it.
>>>>>Bang, it all blows up...
>>>>>
>>>>>So for this code to work properly, what needs to happen is that
>>>>>triggerBeforeCompletion still needs to be called even if
>>>>>triggerBeforeCommit fails. While I am ok with writing the code in =
this
>>>>>method to handle this, I am not 100% sure this is safe in terms of =
all
>>>>>the other interactions that will happen as a result; mot of this =
code is
>>>>>your baby with me only having traced through it once in a while. So =
if
>>>>>you would prefer to resolve this that would be great.
>>>>>
>>>>>Unless I am mistaken about this bug, I think it is a pretty serious =
one,
>>>>>and warrants an almost immediate release of a v1.0.1 of Spring...
>>>>>
>>>>>Regards,
>>>>>Colin
>>>>>
>>>>>
>>>>>Colin Sampaleanu wrote:
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> =20
>>>>>
>>>>> =20
>>>>>
>>>>> =20
>>>>>
>>>>>>I am tracking down a possible Hibernate resource management issue.
>>>>>>
>>>>>>In a running app, some time yesterday, some code, running in a =
wrapped
>>>>>>transaction with Hibernate handling ORM, encountered an Oracle
>>>>>>constraint violation and threw an exception. Fine...
>>>>>>
>>>>>>But when I log into the app myself now via the web ui and then it =
gets
>>>>>>a service object to read some data, the service object is wrapped =
with
>>>>>>a transaction interceptor, and also a hibernate interceptor. The
>>>>>>Hibernate interceptor is already seeing a Hibernate Session =
existing
>>>>>>on the current thread, so it is not creating a new one. Then at =
the
>>>>>>end of the transaction, when the Hibernate session is attempted to =
be
>>>>>>flushed, Hibernate tries to write out the old bad data from =
yesterday.
>>>>>>
>>>>>>What this essentially means is that when the error from yesterday
>>>>>>happened, the session did not get released from the thread, and =
has
>>>>>>been sticking around all this time. When I came via struts, I was
>>>>>>given the same thread as yesterday by the appserver, and the old
>>>>>>invalid session was still on it. The problem is not that it's =
reusing
>>>>>>that session, but why it was ever left that the day before.
>>>>>>
>>>>>>Will try to duplicate this...
>>>>>> =20
>>>>>>
-------------------------------------------------------
This SF.Net email is sponsored by: IBM Linux Tutorials
Free Linux tutorial presented by Daniel Robbins, President and CEO of
GenToo technologies. Learn everything from fundamentals to system
administration.http://ads.osdn.com/?ad_id=1470&alloc_id638&op=3Dick
_______________________________________________
Springframework-developer mailing list
Spr...@li...
https://lists.sourceforge.net/lists/listinfo/springframework-developer
|
|
From: Colin S. <col...@ex...> - 2004-03-25 20:09:09
|
We're on slightly different pages though. In my case, I am actually not=20
using TransactionManagerLookup... But in my case, because of the=20
exception in the flush in beforeCommit(), beforeCompletion() never gets=20
called (as it would normally). Now I do see that afterCompletion is also=20
supposed to call closeSessionIfNecessary() (I had actually missed this=20
before), but in my case, it's not getting there. I haven't traced it in=20
a debugger, which is what I will do now, as it seems to me it should get=20
to that code, certainly I do have a
"Triggering afterCompletion synchronization"
in the log which should happen right before afterCompletion() is called.
The other question is whether beforeCompletion still shouldn't be called=20
in any case even if beforeCommit fails. Ultimately, we are still before=20
completion of the transaction, unless you meant it to be only called in=20
the case of no failure.
Colin
j=FCrgen h=F6ller [werk3AT] wrote:
>Colin,
>=20
>Thanks for tracking this down. It's actually a problem with SessionFacto=
ryUtils' inner class SessionSynchronization: It assumes that beforeComple=
tion is called in any case, even if beforeCommit has thrown an exception.=
However, this just applies if you specified a TransactionManagerLookup i=
n the Hibernate configuration; else, afterCompletion will do the cleanup =
- which will be called in any case.
>=20
>When you remove the Hibernate TransactionManagerLookup, you shouldn't fa=
ce the issue - the bug doesn't have any effects then. Note that you don't=
need that TransactionManagerLookup when using Spring's JtaTransactionMan=
ager, as Spring will properly apply cache callbacks anyway. A Transaction=
ManagerLookup just adds value when used with EJB CMT or manual JTA, for u=
ltra-correct cache callbacks.
>=20
>So essentially, everything should be fine if using HibernateTransactionM=
anager, or JtaTransactionManager without a Hibernate TransactionManagerLo=
okup. This is clearly something to fix, but I guess we don't need to do a=
n immediate 1.0.1 followup release; I'd like to gather further bug report=
s first. For the time being, let's suggest to remove the TransactionManag=
erLookup from the Hibernate configuration.
>=20
>Juergen
>=20
>
>________________________________
>
>Von: Colin Sampaleanu [mailto:col...@ex...]
>Gesendet: Do 25.03.2004 20:32
>An: spr...@li...; j=FCrgen h=F6ller [=
werk3AT]
>Betreff: Re: [Springframework-developer] Hibernate resource management i=
ssue
>
>
>
>Juergen,
>
>I am almost 100% sure this block of code from
>AbstractPlatformTransactionManager is wrong:
> else {
> try {
> try {
> triggerBeforeCommit(defStatus);
> triggerBeforeCompletion(defStatus);
> if (status.isNewTransaction()) {
> logger.info("Initiating transaction commit");
> doCommit(defStatus);
> }
> }
> catch (UnexpectedRollbackException ex) {
> triggerAfterCompletion(defStatus,
>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
> throw ex;
> }
> catch (TransactionException ex) {
> if (this.rollbackOnCommitFailure) {
> doRollbackOnCommitException(defStatus, ex);
> triggerAfterCompletion(defStatus,
>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
> }
> else {
> triggerAfterCompletion(defStatus,
>TransactionSynchronization.STATUS_UNKNOWN, ex);
> }
> throw ex;
> }
> catch (RuntimeException ex) {
> doRollbackOnCommitException(defStatus, ex);
> triggerAfterCompletion(defStatus,
>TransactionSynchronization.STATUS_ROLLED_BACK, ex);
> throw ex;
> }
> catch (Error err) {
> doRollbackOnCommitException(defStatus, err);
> triggerAfterCompletion(defStatus,
>TransactionSynchronization.STATUS_UNKNOWN, err);
> throw err;
> }
> triggerAfterCompletion(defStatus,
>TransactionSynchronization.STATUS_COMMITTED, null);
> }
> finally {
> cleanupAfterCompletion(defStatus);
> }
> }
>
>triggerBeforeCommit() execute, which in the Hibernate case will force a
>flush. However, if that flush throws an exception, then
> triggerBeforeCompletion(defStatus);
>never gets called. However, triggerBeforeCompletion is what is actually
>supposed to release the Hibernate session holder from the current
>thread! So in this case, the session (which is totally hosed of course),
>gets left on the thread. I believe in some environments this wouldn't
>matter that much, as the threads don't get resused. In the JBoss case,
>new requests coming in will get the existing thread, and this time,
>SessionFactoryUtils will see the session is there, and try to use it.
>Bang, it all blows up...
>
>So for this code to work properly, what needs to happen is that
>triggerBeforeCompletion still needs to be called even if
>triggerBeforeCommit fails. While I am ok with writing the code in this
>method to handle this, I am not 100% sure this is safe in terms of all
>the other interactions that will happen as a result; mot of this code is
>your baby with me only having traced through it once in a while. So if
>you would prefer to resolve this that would be great.
>
>Unless I am mistaken about this bug, I think it is a pretty serious one,
>and warrants an almost immediate release of a v1.0.1 of Spring...
>
>Regards,
>Colin
>
>
>Colin Sampaleanu wrote:
>
> =20
>
>>I am tracking down a possible Hibernate resource management issue.
>>
>>In a running app, some time yesterday, some code, running in a wrapped
>>transaction with Hibernate handling ORM, encountered an Oracle
>>constraint violation and threw an exception. Fine...
>>
>>But when I log into the app myself now via the web ui and then it gets
>>a service object to read some data, the service object is wrapped with
>>a transaction interceptor, and also a hibernate interceptor. The
>>Hibernate interceptor is already seeing a Hibernate Session existing
>>on the current thread, so it is not creating a new one. Then at the
>>end of the transaction, when the Hibernate session is attempted to be
>>flushed, Hibernate tries to write out the old bad data from yesterday.
>>
>>What this essentially means is that when the error from yesterday
>>happened, the session did not get released from the thread, and has
>>been sticking around all this time. When I came via struts, I was
>>given the same thread as yesterday by the appserver, and the old
>>invalid session was still on it. The problem is not that it's reusing
>>that session, but why it was ever left that the day before.
>>
>>Will try to duplicate this...
>> =20
>>
>
>
>
>
> =20
>
|
