I think what this shows is that this TM at least does handle this in an appropriate manner. However, there is still a question about the CMT container and any responsibilities imposed by the EJB spec for handling "CMT commit failures". I asked Scott Marlow to comment as he has a lot of experience in this area in WildFly. I basically gave up attempting to use Arquillian to set up a WildFly or Glassfish container for testing this; that approach is completely unwieldy unfortunately I am open to suggestions for other ways to CI test CMT EJBs.