I've been chasing classloader leaks that've been causing my app to fail to redeploy after a few iterations. A bit of digging using jmap and jhat has begun to strongly suggest that JBoss Logging is the problem.

The issue is exactly the same as the one used as an example in this article on tracing classloader leaks:
http://blogs.oracle.com/fkieviet/entry/classloader_leaks_the_dreaded_java

Specifically, org.jboss.logging.JDKLevel extends java.util.logging.Level . Unfortunately, java.util.logging.Level keeps a static array of all known logging levels. If JBoss Logging was deployed as part of an application war, that means there's a reference kept from an application server class (java.util.logging.Level) to a client application class (org.jboss.logging.JDKLevel). This keeps the whole reference chain intact and stops the application classloader from being destroyed.

This is a problem on Glassfish 3.1, because JBoss Logging isn't included in the application server so it must be deployed by applications if they want to use things like JBoss Seam.

It should be possible to add jboss-logging to glassfish/domains/domain1/lib to work around this, by making sure that org.jboss.logging.JDKLevel is loaded by the appserver classloader not the classloader of the deployed application.