Audited entity updates (lots of them) with no EntityManager clear

[drsaturno]

I was testing Envers performance.

I created an audited entity Mailmen, with a many-to-many relationship with another audited entyty, Address.

Mailmen has a simple field called register. (and an Address collection, but lets simplify the code)

I just created a new instance of Mailmen and edited it 100.000 times:

configurationOverrides = new HashMap<String, String>();
emf = Persistence.createEntityManagerFactory("jpa", configurationOverrides);
entityManager = emf.createEntityManager();

int MAX_EDITIONS = 1000000;
carteiroDefault = new Mailman();
carteiroDefault.setRegister(64737);
entityManager.persist(carteiroDefault);
  
for(int i = 1; i <= MAX_EDITIONS; i++) { 
          entityManager.getTransaction().begin();
          carteiroDefault.setRegister(i);
          entityManager.getTransaction().commit();
}

It tooks about 6 hours to run this code.

It gets better or worse depending on MAX_EDITIONS, in a non linear way. Bad.

But, if I change the for block to this:

     entityManager.getTransaction().begin();
     entityManager.find(Mailman.class, carteiroDefault.getId()).setRegister(i);
     entityManager.getTransaction().commit();
     entityManager.clear();

It tooks 30 minutes.

It gets better or worse depending on MAX_EDITIONS, in a linear way. Much better.

Is it recommended? Clearing entityManager after commits?

[adamw]

Did you also try running the test without Envers, to see if the huuuge performance hit because of the lack of clear() is because of Envers? If it turns out it is, please report a bug, as this should not have an impact.

Clear() essentialy detaches all entities and makes them available for GC; but usually you have one EntityManager for one transaction, max several transactions.

Adam

[drsaturno]

Yes, I "turned off" Envers for both scenarios. It took about 30 min for both. I will report a bug.

Thanks!

[drsaturno]

https://hibernate.onjira.com/browse/HHH-6614

[drsaturno]

Some corrections...

For 100,000 editions:

6h (without clear)

3min (with clear)

For 1,000,000 editions:

too slow to test (without clear)

30min (with clear)

[adamw]

Thanks for the data!

Now we only need to investigate what's wrong ...

Adam