[cdi-dev] Proposal for global CDI enablement

[Pete Muir]

I've been working with the Bill and Linda, the Java EE spec leads, as well as with Jason, Stuart and Emmanuel at Red Hat, to come up with a proposal for global enablement for CDI in Java EE 7. Based on Linda and Bill's poll of the community, this appears to be much more popular than we had previously thought, so we have decided to propose it, despite it being quite late into the spec development. I think what we have come up with represents a good approach to achieve it, and also addresses a long requested feature request, that CDI should not consider every class a bean. Please let us know your thoughts. If you aren't on the Java EE EG, or the observers list, or the same for CDI, then please say, and I'll forward your thoughts.

--------------------------------

Background
-----------------

Globally enabling CI would allow other specifications in the Java EE platform to rely on CDI, and not have to provide an abstraction over DI services, such as those introduced by JSF managed beans, and considered by JAX-RS, Batch and others.


Challenges
----------------

* The startup time of Weld is O(n) with the number of beans. We assert that it is not possible for a CDI impl to attain an O(1) startup time, therefore globally enabling CDI will increase the startup time of Java EE servers. For Red Hat, low startup time is a top priority.
* Compatibility with other users of JSR-330. By considering all deployments as CDI deployments without any enablement marker, we will likely cause deployments which used another JSR-330 implementation to fail. This would have worked on Java EE 6. Other JSR-330 impls that fall into this camp are Spring and Guice.
* Backwards compatibility. Some users may have intentionally not placed a beans.xml into a deployment, but used CDI annotations, and enabled the beans some other way, eg. via a portable extension.
* A large proportion of Java EE users 800/1000 indicate they want this feature, thus meaning we should try to address it


Proposed solution
-------------------------

We introduce the concept of a "bean defining annotation" and define that any class in any deployment (including those with no beans.xml) with a bean defining annotation is discovered and may be a CDI bean, and can participate fully in the application. Any archive with a beans.xml continues to work in the same way, such that all classes in the archive are discovered and may be CDI beans.

This addresses the startup time problem. Whilst a scan of classes is still required, the impact on startup time is negligible:

* A Java EE server must scan all classes to discover other component defining annotations such as EJBs, Servlet's, JAX-RS resources etc.
* This scan can be done at the bytecode level, with no need to classload the class, which our research shows is the costly part of CDI startup

Any scope (normal scope or pseudo-scope) applied to a bean at source level is a bean defining annotation (so you must add @Dependent to your class in order to get it to be picked up as a dependent bean).

Only classes with a bean defining annotation, or with an annotation, or meta-annotation, present specified by @WithAnnotations are passed to ProcessAnnotatedType observers (the exact semantics are defined by CDI 1.1 PRD for @WithAnnotations). As mentioned above, if a ProcessAnnotatedType is observed for a type without a bean defining annotation, as a result of having an annotation present that is specified by @WithAnnotations, it may instruct the container to add discover the class as a bean.

Every archive in a deployment would be considered a bean archive, simply with differing contents depending on the presence of beans.xml

If a developer adds a beans.xml to their archive, behavior is as CDI 1.0. We will add an attribute to beans.xml "auto-discover=true", which the user may set to false in order to add a beans.xml and only have classes with a bean defining annotation discovered, which allows beans.xml to be used as a deployment descriptor but still limit the classes discovered.

OPEN ISSUE: Should auto-discover be false by default for beans.xml with version 1.1. This would mean that adding a beans.xml would have no impact on discovery for 1.1 apps, however it is a significant change from 1.0.

OPEN ISSUE: Should only scopes for which a CDI context exists be considered component defining? This could introduce some thorny edge cases, but would address the JSR-330 compatibility issue better.

OPEN ISSUE: Should we extend auto-discover in beans.xml to allow complete disablement of scanning e.g. auto-discover="all|bean-defining-annotations-only|none" ?

OPEN ISSUE: How should the ProcessAnnotatedType event instruct the container to discover a class as a bean? Perhaps something like event.discover(clazz)?

OPEN ISSUE: Should we integrate this with the package level scanning control we have proposed for CDI 1.1?

---------------------------------