On 26 Apr 2011, at 22:58, Rick Hightower wrote:
Inline...
On Tue, Apr 26, 2011 at 12:38 AM, Peter Muir
<pmuir@redhat.com> wrote:
On 26 Apr 2011, at 00:07, Rick Hightower wrote:
We need a general purpose way of finding a BeanManager.
+1 for me.
I agree, this is something we can make easier for CDI 1.0. In part we need to see what the javax.inject EG comes up with here, as we build on their foundations.
FTR this is a (possibly the?) reason CDI moved away from having a "Container" API -- this is supposed to be addressed by JSR-330 part 2, and so was intentionally left out of CDI 1.0 so that we didn't end up with an inconsistent approach. I need to check with the EE spec lead on the status of the 330-part2 work.
Ok.
This way should include JNDI lookup to a fallback BeanManagerLocator that uses a ServiceLoader
+1 for me.
Regarding design for this feature, we should keep it as low footprint in terms of specification and API class as possible, as (always) the impl can make a better decision about how to locate the BM than we can. Rick, can you elaborate on why you want all these additonal JNDI lookup, system property lookup, priority lookup stuff? Comments on each in line.
I think the behavior should look like this:
- Lookup BeanManager in JNDI
We should avoid requiring this. JNDI lookups are expensive (they require synchronization) compared to straight singleton lookups (which at least two CDI impls, Weld and OWB support today).
Ok. But don't singletons wreak havoc with classloaders. What if CDI was being used as library in a container that did not support CDI (jetty, tomcat). Wouldn't we want to avoid singletons?
Singletons can be implemented in many ways, they don't have to be static based (which I think you are referring to in your comment) - suggest you take a look at the way this is implemented in GlassFish or Weld for example (or I think OWB has a similar concept)
I am ok with dropping JNDI lookup. The CDI container should be in a well known locations stored in the ServletContext, i.e., application scope. There should be no classical singleton. There should be nothing that would prevent a web application restart. We should not assume that CDI will only be used by Java EE 6. It should be OSGI, web application, re-loadable friendly.
I mentioned JNDI because currently the BeanManager is stored in JNDI.
In general I think this is swerving too much into specifying implementation. We certainly musn't specify that the Bean Manager must be *stored* in the servlet context or in JNDI, this is an implementation detail.
What we can do is say that the BM is available for lookup from the ServletContext or JNDI or via a static method on an API class.
- If not found, try to load BeanManagerLocator from system property.
What is the use case for this?
Flexibility. Let CDI be used in places we have not thought of.
That's not a usecase, that's a get out for not having a usecase ;-) We need to stick to addressing concrete usecases that have real world applications. So, unless we can come up with a real use case for this, it shouldn't be put in the spec.
- If not found, try to load BeanManagerLocator from ServiceLoader.
- If more than one BeanManagerLocator found, sort by priority or die if there is more than one. (not sure about the sort, I think it should die, but can be convinced otherwise).
Why can it not just be that a service loader is used to find the BeanManagerLocator, and that there must not be more than one property name in use (if there is, it's an error).
I doubt more than one CDI impl (due to scanning) can be running at the same time, and I'm also not sure this is useful.
I agree. There should be only one. If there is more than one, it is an error. I was trying to include the logic that was in Solder. The stuff I have done, throws as exception if more than one is found.
Because Solder is an extension which is working with multiple impl's of CDI it needs to jump through hoops that a spec based impl doesn't need.
- Once the BeanManagerLocator is found, use it to return a BeanManager
In addition,
I think we need a simplified interface for BeanManager.
This could just be a utility class, but it should be included with CDI 1.1. I think the BeanManager interface is too low level.
--
Rick Hightower
(415) 968-9037
Profile