that is a good question in my opinion. I did some research and found this :
Entity beans aren't permitted to use bean-managed transactions (presumably to encourage reuse of entity beans). In fact, the EJB specification states that any transaction in force when a BMT method is called should be suspended by the EJB container before the business logic is invoked. Thus, using BMT beans in various transactional contexts isn't possible. Bean reuse is really maximized by using container-managed transactions.
In Mastering EJB book i have found following but i am not able to understand it.
Perhaps one of the most misunderstood concepts in EJB is how transactions relate to entity beans. Let�s explore this concept. When you call an entity bean in a transaction, the first action that happens is the entity bean needs to load database data through the ejbLoad() method, which acquires locks in the database and ensures the entity bean cache is consistent. Then one or more business methods are called. When the transaction is committed, the entity bean�s ejbStore() method is called, which writes all updates to the database and releases the locks. A transaction should thus span the ejbLoad() business methods, and the final ejbStore(), so that if any one of those operations fail, they all fail. If we were to use bean-managed transactions, we would write code to perform begin() and commit() methods inside our bean (perhaps around the JDBC code). Perhaps we would start the transaction in ejbLoad(), and then commit the transaction in ejbStore(). The problem, though, is that you do not call your own ejbLoad() or ejbStore() methods�the container does. The bean cannot enforce that these methods happen in this order, if at all. Therefore if you started a transaction in ejbLoad(), the transaction may never complete. Because of this, bean-managed transactions are illegal for entity beans. Entity beans must use declarative transactions. Session beans or messagedriven beans can use bean-managed transactions because a session bean can load database data, perform operations on that data, and then store that data; all in a single method call, and thus is in direct control over the transaction. A corollary of this discussion is that entity beans do not load and store their data on every method call; rather, they load and store their data on every transaction. If your entity beans are not performing well, it could be because a transaction is happening on each method call, and thus a database read/write is happening on every get/set method. The solution is to make sure your transactions begin earlier and end later, perhaps encompassing many entity bean method calls. By properly controlling the duration of your transactions with transaction attributes (as we will see later in this chapter), you can control when database reads and writes happen with entity beans.
I am not sure if this is entirely accurate, but these are my thoughts.
Entity Beans exist because there is an entity in the DB that they represent. CMP fields representing the complete entity are declared in the DD, along with abstract getters and setters.
On the other hand, with BMT, the bean manages DB access code by itself upon notification from the container. That means, we could do DB access code or do nothing in the BMT container callback methods. But with CMT, container is in charge of putting the required code for DB access.
This makes me think that a BMT bean is not tied to any real entity. And if its not tied to any real entity, it cannot be an entity bean.
Another point to note is that if a session bean uses an entity bean and if the session bean is in a transaction and wants the entity bean (DB) updates to happen as a part of the same transaction, only CMT would support it. With BMT, the entity bean methods would run in a seperate transaction. So, a scenario like "Insert New Customer through a session bean which uses entity bean" will run into issues.
Not sure if I made sense. Correct me if I am wrong.