Hi Kevin,
I tried adding the javaagent stuff, and it still ends up leaking. I
extracted the code into a tiny
war with just one servlet which persist the authToken. The build creates
a Tomcat distro backed by a
derby database. You can download it from:
http://people.apache.org/~kstam/leak.tgz
You can build it with: mvn install, you can find the final tomcat in
tomcat/target/tomcat-1.0.SNAPSHOT.zip
or if you don't want to build it:
http://people.apache.org/~kstam/tomcat-1.0.SNAPSHOT.zip
After you bring up tomcat it persistent an authToken when you hit
http://localhost:8080/leak/test
In the profiler I see an org.apache.juddi.AuthToken left behind for each
request.
Hopefully this little example will pinpoint what's going on.
Thx Kev for sticking with me :)
--Kurt
Kevin Sutter wrote:
Hi Kurt,
Since you have the environment to reproduce this, could you run with the
OpenJPA trace enabled? I'm not even sure what I'm looking for at this
point. :-) But, maybe there's some tracing that would point us in the
right direction.
<property name="openjpa.Log" value="DefaultLevel=TRACE"/>
We have been always been on the watch for memory leaks, so this one is
puzzling. You did mention blocked threads as well. At this point, you are
probably hitting blocked threads due to the GC and Memory problems. But, we
did recently resolve some locking situations [1] in our MetaDataRepository
that could affect scaling (under extreme load). This was resolved in the
1.2.x SNAPSHOT drivers (Mike is currently pushing for a 1.2.2 release, so
that should be coming soon).
Does your WebService version of the problem use EntityManager transactions,
or is the transaction service coming from Tomcat? And, does your WebService
version also do the em.clear() and em.close()? Basically, are you using
application-managed or container-managed persistence while running as a
WebService?
I'm very interested in helping to resolve this issue. Like I mentioned, we
do a lot of memory leakage and scalability testing of the OpenJPA solution
and we're not aware of anything. Although there is always the possibility
that something was resolved in the 1.2.x branch after the 1.2.1 release that
I'm not remembering...
Thanks for your help,
Kevin
[1] https://issues.apache.org/jira/browse/OPENJPA-250
On Wed, Jan 13, 2010 at 9:27 PM, Kurt T Stam <kurt.s...@gmail.com> wrote:
Thanks Kevin,
We're enhancing at build time:
http://svn.apache.org/repos/asf/webservices/juddi/trunk/juddi-core/pom.xml
Yeah we've been running load tests and things are nice and stable with
Hibernate but with Openjpa we see increasing memory use, blocking threads
and then an OOM. http://issues.apache.org/jira/browse/JUDDI-267. Our
preference would be to ship with openjpa by default; but our build supports
both hibernate and openjpa.
And yes we use openjpa 1.2.1 (latest stable version).
--Kurt
Kevin Sutter wrote:
Interesting detective work, Kurt. Thanks.
Why the WebService version of the app would behave differently as far as
GC
is concerned is a mystery. And, you said that plugging in Hibernate into
this scenario, everything works okay? Very confusing.
How are you performing the Entity enhancement processing? Are you
pre-enhancing via your build process? Or, are you using the -javaagent
mechanism? Or, are you falling back to the subclassing support within
OpenJPA? (See [1] for more information on these questions in case they
don't make sense.)
This would be one area that is different between Hibernate and OpenJPA --
enhancement processing.
In the Tomcat environment, you may be falling back to the subclassing
support (which we do not recommend) and hitting a memory leak with that.
You said OpenJPA 1.2.x, right?
Just a couple of thoughts on the subject...
Kevin
[1]
http://webspherepersistence.blogspot.com/2009/02/openjpa-enhancement.html
On Wed, Jan 13, 2010 at 4:25 PM, Kurt T Stam <kurt.s...@gmail.com> wrote:
The same code executed straight from a java client (inVM) shows no memory
leak.
So is the fact that it is WebService significant then? What else can be
different? I think one thread remains up, and somehow this causes openjpa
not being able to clean up after itself. What can I do to debug this
more? I
can actually see in the profiler that the objects are allocated by the
WebService, but why aren't they cleaned up?
Thx,
--Kurt
Kurt T Stam wrote:
Thanks Kevin, thanks for your response.
I just replaced the static call by:
apiAuthToken = new org.uddi.api_v3.AuthToken();
apiAuthToken.setAuthInfo(modelAuthToken.getAuthToken());
//MappingModelToApi.mapAuthToken(modelAuthToken,
apiAuthToken);
which did not make a difference.
I'm wondering if the fact that my class is a webservice makes a
difference. I'll try extracting it into
a regular class with a main method and profile that. At least I know
that
I didn't forget something
completely obvious..
--Kurt
Kevin Sutter wrote:
Kurt,
I agree that this is very common usage of the JPA programming model.
And,
we are not aware of any memory leaks. About the only thing that jumps
out
at me is the following two lines:
apiAuthToken = new org.uddi.api_v3.AuthToken();
MappingModelToApi.mapAuthToken(modelAuthToken,
apiAuthToken);
What do these do? Can you comment these out and see if the memory leak
still exists? Since you are passing the modelAuthToken into this
method,
I
don't know what it's doing with the reference and could it be holding
onto
something to prevent the GC from cleaning up?
The rest of your example seems very straight forward with creating and
persisting objects.
Kevin
On Wed, Jan 13, 2010 at 2:09 PM, Rick Curtis <curti...@gmail.com>
wrote:
If you change the 1000 to something like 1000000... does your
application
go
OOM? Are you running in a JSE environment? What is PersistenceManager?
On Wed, Jan 13, 2010 at 2:05 PM, Kurt T Stam <kurt.s...@gmail.com>
wrote:
BTW I'm running with the cache off
<property name="openjpa.DataCache" value="false"/>
(that turns it off right?)
--Kurt
Kurt T Stam wrote:
Hi guys,
[DESCRIPTION] The code below inserts a 1000 records in the database.
for (int i=1; i<1000; i++) {
EntityManager em = PersistenceManager.getEntityManager();
EntityTransaction tx = em.getTransaction();
try {
tx.begin();
// Generate auth token and store it!
String authInfo = AUTH_TOKEN_PREFIX + UUID.randomUUID();
org.apache.juddi.model.AuthToken modelAuthToken = new
org.apache.juddi.model.AuthToken();
if (authInfo != null) {
modelAuthToken.setAuthToken(authInfo);
modelAuthToken.setCreated(new Date());
modelAuthToken.setLastUsed(new Date());
modelAuthToken.setAuthorizedName(publisherId);
modelAuthToken.setNumberOfUses(0);
modelAuthToken.setTokenState(AUTHTOKEN_ACTIVE);
em.persist(modelAuthToken);
}
apiAuthToken = new org.uddi.api_v3.AuthToken();
MappingModelToApi.mapAuthToken(modelAuthToken,
apiAuthToken);
tx.commit();
} finally {
if (tx.isActive()) {
tx.rollback();
}
em.clear();
em.close();
}
}
[ISSUE]
After it leaving this code I end up with a 1000
org.apache.juddi.model.AuthToken objects in memory. I've been using
the
profiler, and these objects cannot be garbage collected.
This seems to be pretty the most common use case of using an
OR-mapping
tool, so I find it hard to believe openjpa has a memory leak here.
Does
anyone see what I'm doing wrong? Or can someone point me to an
example
that
does not exhibit this behavior? BTW same code using hibernate does
not
accumulate these objects.
We're using openjpa 1.2.1.
Thx,
Kurt
Apache jUDDI.
--
Thanks,
Rick
