Community
Participate
Working Groups
Created attachment 193388 [details] Remove a stopping lazy bundle from the lazyBundles collection I faced the following problem: - I have a persistence unit bundle with the "Bundle-ActivationPolicy: lazy" header in the manifest - Sometimes after the OSGi framework starts (equinox) I had the problem that there is no EntityManagerFactoryBuilder registered for the unit. Sometimes everything worked fine. I figured out that if the bundle is first activated lazily (p.u. is then registered) and restarted afterwards (e.g. by a refreshPackages call somewhere in the system), then I don't get the registration of the p.unit after the restart, because the bundle doesn't go through LAZY_ACTIVATION again. If I have properly debuged the coding, then the reason for the problem is that when STOPPING event is processed in PersistenceBundleExtender the lazy bundle is not removed from the field "lazyBundles". I will attach a patch which just remvoes the stopping bundle from lazyBundles. For me this solves the problem. But I was wondering why the LAZY_ACTIVATION event is at all tackled separately. After it there is anyway also a STARTING event fired for the same bundle. To me it seems that the LAZY_ACTIVATION handling can be fully removed... If you also think so, I could test it and send another patch.
Thanks for filing the bug, Krum. Looks like the STOPPING event handling did not get changed to properly handle bundle stopping. The problem with not handling the LAZY_ACTIVATION event is that the STARTING event does not get fired until the first class load occurs, and this would typically happen when a client looks up and uses the service. Problem is that the register would need to have already happened in order for clients to access the service and cause a class to be loaded (and thereby cause the STARTING event to be fired), so you have a chicken and egg problem. The solution to me seemed to just do the registration in the LAZY_ACTIVATION event. There might be other solutions to the problem, but that seemed like the best one to me.
Fix checked into RC2.
