On 08/20/2012 11:19 PM, Steve Ebersole wrote:
This ties together a few different discussions that have been going
on
simultaneously on the mailing list that I think are all related.
Right now to configure certain services (select one impl over another)
users generally give the FQN for that impl Class. For example to use
C3P0 connection pooling users would say:
hibernate.connection.provider_class =
org.hibernate.service.jdbc.connections.internal.C3P0ConnectionProvider
We have discussed why this is bad even before any of the OSGi
discussions and the solution we wanted to shoot for was that of naming
"selectors" such that the user would instead say:
hibernate.connection.provider_class = c3p0
And "c3p0" here would be interpreted to mean "instantiate and configure
the
org.hibernate.service.jdbc.connections.internal.C3P0ConnectionProvider
Class". But that still means a limited set of short name values *and*
still gives us a problem (iiuc) under OSGi due to visibility.
So what I propose instead is a way for service implementors to be
registered under a short name via discovery. The main piece to this is
the "registry" (which is also a service under the BootstrapServiceRegistry):
interface AvailableServiceRegistry extends Service {
public <T> Class<? extends T>
findAvailableServiceImplementor(Class<T> serviceRole, String selector);
}
class AvailableServiceRegistryImpl
implements AvailableServiceRegistry,
ServiceRegistryAwareService {
private Map<Class,Map<String,Class>> availableImplementors = ...;
@Override
public <T> Class<? extends T>
findAvailableServiceImplementor(Class<T> serviceRole, String selector) {
// purposely simplistic :)
return availableImplementors.get( serviceRole ).get( selector );
}
@Override
public void injectServices(ServiceRegistryImplementor
serviceRegistry) {
final LinkedHashSet<ServiceAvailabililtyNotifier> notifiers =
serviceRegistry.getService( ClassLoaderService.class ).loadJavaServices(
ServiceAvailabililtyNotifier.class );
for ( ServiceAvailabililtyNotifier notifier : notifiers ) {
for ( ServiceAvailabililty availability :
notifier.getAvailabilities() ) {
// again, purposely simplistic
Map<String,Class> serviceImplementors =
availableImplementors.get( availability.getRole() );
serviceImplementors.put(
availability.getSelector(),
availability.getImplementor()
);
}
}
}
}
Outstanding question... Especially in OSGi, where service bundles can be
added/removed, how do we best account for cleaning up no-longer valid
references (even more importantly perhaps, what the hell does it mean to
Hibernate when a ConnectionProvider implementation, for example, that is
in use gets taken away?!?). Perhaps this is just where an OSGi-specific
Hibernate ServiceRegistry implementation would come into play.
Adding Jesper as we were talking about how to handle "quiescence
shutdown" at the AS level, which sounds related. Once we take the
ConnectionProvider away, I would expect the Hibernate session(s)/session
factory to be broken. If/when the ConnectionProvider comes back,
Hibernate would need to re-establish it. I'm thinking that we need a
neutral (autonomic) API/SPI for attempting to re-establish the
ConnectionProvider.
For the most part, a "quiescence shutdown" of the AS, would mean keeping
the ConnetionProvider alive until the end (of the planned shutdown).
I'm thinking that being able to re-establish the ConnectionProvider
would still be useful (for AS "quiescence shutdown"), especially if
something goes wrong during the shutdown and manual intervention is needed.
To me, the process of re-establishing the ConnectionProvider, could be
labeled "self healing" (with the help of an autonomic API/SPI).