JPA is about mapping concrete classes to relational databases. I never
claimed otherwise. If that was the mis-assumption from suggesting
@Entity be allowed on interfaces, I apologize for not being more clear.
No, as Oliver points out, the proposal is more about aliasing. There is
a thread on the Hibernate developer list[1] that discuss this in detail.
Basically I think that trends in development prefer interface-driven
design. Whether you agree with that or not for a JPA model is not
really the discussion. The proposal is more about working with that
more seamlessly.
Personally I think this should be a straight-up re-aliasing;
one-for-one. That and another more esoteric reason is what led me to
propose @Entity on the interface itself, with a attribute pointing to
the class being physically mapped. But I also see the benefit in
Oliver's suggestion to having that "link" on the entity *class*. This
is a point we specifically discussed on that Hibernate developer
discussion. In retrospect, I think the approach Oliver mentions is
better approach. It allows the jar separation he mentions, plus other
benefits.
[1]
http://lists.jboss.org/pipermail/hibernate-dev/2012-January/007627.html
On 03/12/2012 07:08 AM, Oliver Gierke wrote:
> +1 on Mike's statement here. My initial intention (and I think Steve's as well) was what Mike explained in his other email: aliasing access to a JPA entity through an interface for the purpose of being able to separate code and e.g. have the interfaces in one JAR and the implementation class in another one.
>
>> The key part of what was being proposed (at least how I interpreted the
>> feature, Steve can correct me if I misinterpreted) was the ability to
>> "alias" an entity class. This just means that one might be able register
>> an entity class in the providers entity list keyed not only by the
>> concrete entity class, but also an (or more than one?) interface, so
>> that when the provider encounters that interface it gets loaded/saved by
>> the provider as the concrete entity class.
>
>
> No additional metadata, simply metadata aliasing. I think ideas like this originate from the trend for developers to build rich domain models, follow Domain Driven Design and don't treat domain classes as pure data containers. Although one might argue this is not what the JPA spec is about I think we should make sure we don't get in the way when following this approach so that devs don't get into a "I can't do this because JPA can't deal with it" situation from a class' design perspective.
>
> Looking at the extended proposal @Entity on an interface feels a bit weird as well, at least given the current semantics in the spec. I could imagine an @EntityAlias though with the following usage model:
>
> %<----------------------------------
> @EntityAlias
> public interface Person {
>
> }
>
> @Entity
> public class MyPerson implements Person {
>
> }
> %<----------------------------------
>
> Not sure I favor this over explicitly listing the alias interfaces on the @Entity (e.g. @Entity(typeAlias = { Person.class }) as with the original proposal you could implement the interface with a second class without creating the conflict of two implementation classes using the interface as alias.
>
> Mike Keith wrote:
>
>> Having said that, if we think that aliasing is one of those things that
>> could stop large companies from using JPA we might decide to close our
>> eyes, hold our noses and go ahead and add some kind of support.
>
>
> Could you elaborate on that? I don't quite get why this feature should stop companies from using JPA.
>
>> @Entity(alias=Employee.class)
>> public class EmployeeImpl implements Employee {
>> // mappings ...
>> }
>>
>> Where does one draw the line? Would we assume that a single alias is
>> enough and that an entity would not be allowed to specify a Set? For
>> example, that we would not need to support something like:
>>
>> @Entity(aliases={Employee.class, Worker.class}) ... or ...
>> @Entity(aliasInterfaces=true)
>> public class EmployeeImpl implements Employee, Worker {
>> // mappings ...
>> }
>
> The latter doesn't seem to complicate things, does it? At least given the assumption there has to be a single implementation class to be selected to back the type alias, right?
>
> Cheers,
> Ollie
>