[wildfly-dev] duel between JPA use of ClassFileTransformer and CDI implicit/explicit Bean Manager (why can't we just get along)...

Jason Greene jason.greene at redhat.com
Wed Jun 12 11:15:33 EDT 2013 

Let me clarify after the recent posts. I am talking about how to solve the problem generally for all JPA providers. I still believe that we need a custom integration point (2-phase etc) for Hibernate because the spec SPIs have always been insufficient to meeting our needs.

So basically what I am saying is that we can make it all work generically by doing the following:

1. The deployers pre-construct the module classloader which will support the new bytecode loaded by the transformers as we already do today. (The CL returned by the JPA SPI getClassloader).

2. JPA providers can still use getTempClassloader to preload classes for analysis purposes *BEFORE* the CDI bean manager loads

3. To satisfy 2 and the new SPIs that allow obtaining a BM prematurely, we instead provide a proxied, lazy-loading BM 

4. The JPA provider adds all necessary transformers via the JPA SPIs *before* using the BM

5. After the transformers are added, the BM is called to initialize entity listeners, and the BM lazily initializes itself using the module class loader in 1.


On Jun 10, 2013, at 6:16 PM, Jason Greene <jason.greene at redhat.com> wrote:

> Well it just needs to load it after the transformer is added, and from the non-temp class-loader. So basically I was saying the spec forces the JPA provider to follow that kind of order.
> 
> 
> On Jun 10, 2013, at 5:35 PM, Stuart Douglas <stuart.w.douglas at gmail.com> wrote:
> 
>> The thing is that the CDI specification *requires* us to load the entity classes, there is simply no way around it. Because we have to fire a ProcessAnnotatedType<T> for the entity class (with T being the entity class) there is no way to fire this event without loading the entity class. Any CDI implementation will run into this, it is not specific to Wildfly.
>> 
>> The only way to get around this is basically making sure that entities are not made into CDI beans, so vetoing them as Jozef said, or packing them into their own JAR, neither of which is a good solution.
>> 
>> Basically the JPA spec is broken, it can't deal with both CDI and class transformation, and we need to hack around it somehow.
>> 
>> Stuart
>> 
>> 
>> 
>> Steve Ebersole wrote:
>>> So you'd like JPA providers to read the JPA spec in a special "WildFly
>>> light" wrt these quoted sections?  Put aside your understanding that
>>> WildFly suffers from this particular problem and point out the JPA spec
>>> passages that say I need to delay using that BeanManager.  In fact, I'd
>>> argue the wording implies that the BeanManager should be usable
>>> immediately upon being passed to the JPA provider.
>>> 
>>> On the flip side, is it possible that other containers might *only*
>>> make the BeanManager available during bootstrap?
>>> 
>>> As for WildFly+Hibernate, I have already provided hooks to make sure
>>> this can be done properly (albeit in non-spec way) as outlined in
>>> earlier response.
>>> 
>>> 
>>> On Mon 10 Jun 2013 04:54:22 PM CDT, Jason Greene wrote:
>>>> Right but my point is if the entities aren't instantiated, there is nothing for the listener to listen for, and it can be delayed up until that point in the deployment process.
>>>> 
>>>> It seems that the only hard ordering, is that the bean manager be created on top of PersistenceUnitInfo.getClassLoader().
>>>> 
>>>> On Jun 10, 2013, at 4:41 PM, Steve Ebersole<steve at hibernate.org>  wrote:
>>>> 
>>>>> We are talking about entity listeners, not entities.  Entities can be annotated with callback methods as well, but those are not supported (by spec) for CDI injection.
>>>>> 
>>>>> BTW, for those maybe not familiar... The relevant sections in the JPA spec are:
>>>>> 
>>>>> 9.1 Java EE Deployment.  In this section, it discusses the expectations of the container.  In the part that discusses the call to createContainerEntityManagerFactory (aka, bootstrapping an EntityManagerFactory) it says:
>>>>> 
>>>>> <quote>
>>>>> If CDI is enabled, a BeanManager instance must be made available by the container. The
>>>>> container is responsible for passing this BeanManager instance via the map that is passed as
>>>>> an argument to the createContainerEntityManagerFactory call. The map key
>>>>> used must be the standard property name javax.persistence.bean.manager.
>>>>> </quote>
>>>>> 
>>>>> And then, 3.5.1 Entity Listeners, as partially quote by Scott earlier.
>>>>> 
>>>>> 
>>>>> On Mon 10 Jun 2013 03:46:01 PM CDT, Jason Greene wrote:
>>>>>> I guess the point I am still not clear on is exactly when the entities are instantiated. My understanding is that the earliest a user can cause a JPA impl load them is inside a PostConstruct method. That is unless JPA providers load instances for some kind of caching purposes?
>>>>>> 
>>>>>> On Jun 10, 2013, at 12:25 PM, Steve Ebersole<steve at hibernate.org>  wrote:
>>>>>> 
>>>>>>> As Scott stated, Hibernate is doing the CDI calls as part of the
>>>>>>> EntityManagerFactory bootstrap (however that translates to WildFly
>>>>>>> deployment phases...).  There is a difficulty with delaying these CDI
>>>>>>> calls.  Getting around that would require either:
>>>>>>> 1) a 2-phase bootstrap process for the EntityManagerFactory.  Even
>>>>>>> though that proposal was not accepted into the JPA spec, I went ahead
>>>>>>> and broke out Hibernate's bootstrapping into a 2-phase process, so
>>>>>>> WildFly (via jipijapa) could leverage that.
>>>>>>> 2) Creating proxies of the listeners that delegate to the real listeners
>>>>>>> and perform the init on first call.  Ugh.  Proxies calling proxies
>>>>>>> calling proxies... not a recipe for great performance.  Not to mention
>>>>>>> don't call me to debug through that mess
>>>>>>> 
>>>>>>> There is a third option.  How entity listeners are built is hidden
>>>>>>> behind an interface (called ListenerFactory oddly enough).  I can expose
>>>>>>> that as an integration point for WildFly to be able to plug in whatever
>>>>>>> "delayed init" scheme it deems appropriate.  The only reason I'd suggest
>>>>>>> this is if y'all feel strongly about option 2 above (I'm not comfortable
>>>>>>> implementing that in the Hibernate code base).  But my vote would be for
>>>>>>> the first option.
>>>>>>> 
>>>>>>> I am open to other suggestions.
>>>>>>> 
>>>>>>> 
>>>>>>> On 06/10/2013 02:10 AM, Jozef Hartinger wrote:
>>>>>>>> Weld can be started before a JPA impl without a risk of suppressing
>>>>>>>> ClassFileTransformers under condition that all entities are annotated
>>>>>>>> with @Vetoed. We could document that as a requirement.
>>>>>>>> 
>>>>>>>> On 06/07/2013 06:20 PM, Scott Marlow wrote:
>>>>>>>>> For application deployments that use ClassFileTransformer to
>>>>>>>>> enhance/rewrite entity classes, we start the persistence unit service
>>>>>>>>> (PersistenceProvider.createContainerEntityManagerFactory()) during the
>>>>>>>>> Phase.FIRST_MODULE_USE (before any application classes have been loaded).
>>>>>>>>> 
>>>>>>>>> For application deployments that have an explicit CDI Bean Manager,
>>>>>>>>> there is a beans.xml that means the ClassFileTransformer will not work,
>>>>>>>>> since the CDI Bean Manager will scan all of the application classes
>>>>>>>>> (loading them), before the persistence unit service is started (so that
>>>>>>>>> the persistence provider can use CDI in entity listeners).
>>>>>>>>> 
>>>>>>>>> The same is also true for implicit CDI Bean manager support [1], expect
>>>>>>>>> all application deployments that contain an ejb3 module, will be wired
>>>>>>>>> for CDI (meaning JPA ClassFileTransformer support will work even less).
>>>>>>>>> 
>>>>>>>>> I raised this on the JPA 2.1 EG [2] in response to an earlier
>>>>>>>>> discussion, about switching to a two phase approach to address problems
>>>>>>>>> like this (didn't discuss CDI implicit support then but am raising that
>>>>>>>>> now).
>>>>>>>>> 
>>>>>>>>> [3] talks about why we don't create the CDI bean managers before the
>>>>>>>>> Install phase (would cause all application classes to be read which
>>>>>>>>> breaks JPA ClassFileTransformer use).
>>>>>>>>> 
>>>>>>>>> [4] is for adding implicit CDI support but is blocked currently by [5].
>>>>>>>>> 
>>>>>>>>> We can add persistence unit flags (jboss.as.jpa.classtransformer=false)
>>>>>>>>> for disabling JPA ClassFileTransformer support as a workaround but that
>>>>>>>>> doesn't help enough since many deployments will have implicit CDI
>>>>>>>>> support enabled (since they contain EJB modules).  We could add a way to
>>>>>>>>> disable implicit CDI support as another workaround for deployments that
>>>>>>>>> want to use ClassFileTransformer.
>>>>>>>>> 
>>>>>>>>> I'm not yet seeing a proper fix for this.  Anyone else?
>>>>>>>>> 
>>>>>>>>> Scott
>>>>>>>>> 
>>>>>>>>> [1] http://docs.jboss.org/cdi/spec/1.1/cdi-spec.html#bean_archive
>>>>>>>>> [2]
>>>>>>>>> https://java.net/projects/jpa-spec/lists/jsr338-experts/archive/2013-06/message/0
>>>>>>>>> [3] https://issues.jboss.org/browse/WFLY-1322
>>>>>>>>> [4] https://issues.jboss.org/browse/WFLY-476
>>>>>>>>> [5] https://issues.jboss.org/browse/WFLY-1463
>>>>>>>>> _______________________________________________
>>>>>>>>> wildfly-dev mailing list
>>>>>>>>> wildfly-dev at lists.jboss.org
>>>>>>>>> https://lists.jboss.org/mailman/listinfo/wildfly-dev
>>>>>>>> _______________________________________________
>>>>>>>> wildfly-dev mailing list
>>>>>>>> wildfly-dev at lists.jboss.org
>>>>>>>> https://lists.jboss.org/mailman/listinfo/wildfly-dev
>>>>>>> _______________________________________________
>>>>>>> wildfly-dev mailing list
>>>>>>> wildfly-dev at lists.jboss.org
>>>>>>> https://lists.jboss.org/mailman/listinfo/wildfly-dev
>>>>>> --
>>>>>> Jason T. Greene
>>>>>> WildFly Lead / JBoss EAP Platform Architect
>>>>>> JBoss, a division of Red Hat
>>>>>> 
>>>> --
>>>> Jason T. Greene
>>>> WildFly Lead / JBoss EAP Platform Architect
>>>> JBoss, a division of Red Hat
>>>> 
>>> _______________________________________________
>>> wildfly-dev mailing list
>>> wildfly-dev at lists.jboss.org
>>> https://lists.jboss.org/mailman/listinfo/wildfly-dev
> 
> --
> Jason T. Greene
> WildFly Lead / JBoss EAP Platform Architect
> JBoss, a division of Red Hat
> 
> 
> _______________________________________________
> wildfly-dev mailing list
> wildfly-dev at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/wildfly-dev

--
Jason T. Greene
WildFly Lead / JBoss EAP Platform Architect
JBoss, a division of Red Hat

More information about the wildfly-dev mailing list