[cdi-dev] [JBoss JIRA] (CDI-565) Unsatisfied dependency should be null?

[Martin Andersson (JIRA)]

    [ https://issues.jboss.org/browse/CDI-565?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13116971#comment-13116971 ] 

Martin Andersson edited comment on CDI-565 at 10/9/15 9:39 AM:
---------------------------------------------------------------

Emily, good to have you here. You wrote:

{quote}
.. optional injection forces the end users to check for null, which is boilerplate code in most scenarios and confusing.
{quote}

I must say that this statement is just wrong. The user doesn't have to do anything. With the proposed change, all CDI applications remain the same. We are talking about trading a "premature" CDI specific exception type, {{DeploymentException}} for the more specialized and well known {{NullPointerException}} during runtime. The fallacy in the CDI specification is that it treat an unsatisfied dependency as a problem. Section "5.2.2. Unsatisfied and ambiguous dependencies"

{quote}
If an unsatisfied .. dependency exists, the container automatically detects the problem and treats it as a deployment problem.
{quote}

But we know an unsatisfied dependency cannot always be treated as a problem. Contrary to what you stated, I had hoped all experienced Java developers abide to the principle of not writing "paranoid" null checks all over the code unless null is expected. Because +if+ and +when+ null is a problem, we will face a NPE exactly +where+ that problem surface. In a sense, nothing could ever be more clear or easily debugged and understood. If you don't agree with me, please share your thoughts. Until then, I say that checking for null, which is what all CDI implementations currently do, is the boilerplate code. It saddens me that they are enforced to act in this way with quite possibly the only result achieved being a reduced feature set offered to the application developer.

Today, when the application want to have an optional dependency, then he has to write, yeah that's right: *boilerplate code*.. to circumvent the normal lookup mechanism and take the route through {{Instance}} which is not the first API most people learn about. Even more so, {{Instance}} will provide you with all the available injection targets. If your application only expected one to be present, any more than so would be ambiguous, then the application has to add it's own {{@PostContruct}} method that crash bean initialization if more than just one target was found. Compare all this *boilerplate code* to the simple use of one annotation {{@Inject}} which automatically, and with all right, treat many targets of the same kind being ambiguous.

So if this proposition goes through, the "intuitive" approach would enrich CDI:s service offer and vastly reduce boilerplate code that all those applications in need of optional dependencies has to write. Like any other normal Java application, those that do require their dependency to be present will face the world's most known exception type: NPE. I say we reach out a hand to them by introducing the {{@Required}} annotation. Today, I believe that in this context, CDI work in an unintuitive way only to destroy its own usability and API simplicity. Again, the real question for us to answer, or vote on, is what the most intuitive approach is. If the most intuitive approach break backward compatibility, then the change has to wait to the next major release. Me personally, I would also like someone to give their view as to why CDI:s injection mechanism is semantically different from {{Map.get()}}?


was (Author: martin.andersson):
Emily, good to have you here. You wrote:

{quote}
.. optional injection forces the end users to check for null, which is boilerplate code in most scenarios and confusing.
{quote}

I must say that is just wrong. The user doesn't have to do anything. With the proposed change, all CDI applications remain the same. We are talking about trading a "premature" CDI specific exception type, {{DeploymentException}} for the more specialized and well known {{NullPointerException}} during runtime. The fallacy in the CDI specification is that it treat an unsatisfied dependency as a problem. Section "5.2.2. Unsatisfied and ambiguous dependencies"

{quote}
If an unsatisfied .. dependency exists, the container automatically detects the problem and treats it as a deployment problem.
{quote}

But we know an unsatisfied dependency cannot always be treated as a problem. Contrary to what you stated, I had hoped all experienced Java developers abide to the principle of not writing "paranoid" null checks all over the code unless null is expected. Because +if+ and +when+ null is a problem, we will face a NPE exactly +where+ that problem surface. In a sense, nothing could ever be more clear or easily debugged and understood. If you don't agree with me, please share your thoughts. Until then, I say that checking for null, which is what all CDI implementations currently do, is the boilerplate code. It saddens me that they are enforced to act in this way with quite possibly the only result achieved being a reduced feature set offered to the application developer.

Today, when the application want to have an optional dependency, then he has to write, yeah that's right: *boilerplate code*.. to circumvent the normal lookup mechanism and take the route through {{Instance}} which is not the first API most people learn about. Even more so, {{Instance}} will provide you with all the available injection targets. If your application only expected one to be present, any more than so would be ambiguous, then the application has to add it's own {{@PostContruct}} method that crash bean initialization if more than just one target was found. Compare all this *boilerplate code* to the simple use of one annotation {{@Inject}} which automatically, and with all right, treat many targets of the same kind being ambiguous.

So if this proposition goes through, the "intuitive" approach would enrich CDI:s service offer and vastly reduce boilerplate code that all those applications in need of optional dependencies has to write. Like any other normal Java application, those that do require their dependency to be present will face the world's most known exception type: NPE. I say we reach out a hand to them by introducing the {{@Required}} annotation. Today, I believe that in this context, CDI work in an unintuitive way only to destroy its own usability and API simplicity. Again, the real question for us to answer, or vote on, is what the most intuitive approach is. If the most intuitive approach break backward compatibility, then the change has to wait to the next major release. Me personally, I would also like someone to give their view as to why CDI:s injection mechanism is semantically different from {{Map.get()}}?

> Unsatisfied dependency should be null?
> --------------------------------------
>
>                 Key: CDI-565
>                 URL: https://issues.jboss.org/browse/CDI-565
>             Project: CDI Specification Issues
>          Issue Type: Feature Request
>          Components: Beans
>            Reporter: Martin Andersson
>
> What if I am writing a component that call a collaborator only if that class exist during runtime? It is amazingly intuitive to write code like this:
> {code:java}
> class MyComponent
> {
>     @Inject
>     SomeCollaborator collaborator;
>     public void someMethod() {
>         // .. do something
>         if (collaborator != null) {
>             collaborator.callback();
>         }
>     }
> }
> {code}
> But the specification currently says in section "5.2.2. Unsatisfied and ambiguous dependencies" that this code should cause a "deployment problem". IIRC, GlassFish 4.1 and WildFly 9 doesn't actually crash during deployment. They will inject {{null}}. WebLogic 12, I just learned, do crash during deployment.
> I believe it is unfortunate to have this null value logic for no apparent reason. It is inevitably so that the application code would crash anyways as soon as it tries to dereference a null reference. And you probably agree with me that there's a general guideline established in the developer community which say you shouldn't do null pointer checks all over the code because null pointers will crash exactly when and where the absence of a value really is a problem.
> The work around is to inject an {{Instance}} of my type and iterate through all of them, or do any other form of programmatic lookup. However, me personally, I've had this requirement far too many times now. It is often the case that a component I write has a "subframework" in place such that when I want to affect how the application performs, I can just add in new classes of a particular type and it is scoped up. Please don't even make the notion of a design smell out of your own lack of creativity, if you want to see a concrete example then of course I am more than happy to provide you with that. Just saying =)
> Something so intuitive and present in our every day coding life as a "null return value" should be present in the CDI specification too. I mean that is what the specification in essence is; one huge {{lookupInstanceOf(class)}}-method. Reading this method name, would you really expect it to crash, or return {{null}}?
> How about adding in a new annotation such that the injection point accept null values for unsatisfied dependencies, but if the injection point has an annotation {{@Required}}, then an unsatisfied dependency do crash for this injection point?
> You will probably say "let's do it the other way around so that we don't brake backward compatibility" by creating an {{@Optional}} annotation. Hey, if I am abusive to my wife and she divorce me. Should I continue to be abusive to my next wife so that I don't break backward compatibility? Our number one goal should be to define the most awesome specification and API possible - nothing else. If we keep leaving small piles of poo everywhere, then we will inevitably end up deep in shit.
> Thank you all for your hard work and time devoted to making all of our lives so much greater.



--
This message was sent by Atlassian JIRA
(v6.4.11#64026)