In the attempt to design & implement & test a distributed capabilities system, Remote Promises[1][2][3], bit identical between Squeak & Java, there are shortcomings. I am seeking work-arounds.
With Remote Promises, proxies can change state, which changes the class implementing the proxy. In Squeak this is done with #becomeForward:, while in Java, it requires a secondary proxy, one that can change it's implemention. This does work.
Exceptions should be non-blocking to allow the event loop to continue, yet also display the problem stack for debugging, out of a quarantine. This is good in Squeak but an open issue with Java. I suppose the answer is do all your logging and then close the exception, allowing the event loop to proceed: it is server-style log debugging.
Using a meta repository, it should be possible to demand load consumers of a particular event type. Dynamically load the latest released code into the consumer servers and spread out the load to speed up the throughput. Update the system at runtime for continuous, seemless operations. I suppose the solution here is to build a dynamic jar classLoader system. Are there any examples of this? An Apache project perhaps?
Use cloud technologies made for that kind of usecases
I would say that in today world, to get the latest version of a code, you don't use a class loader or any advanced capability of your programming langage. You would user likely some kind of cloud service.
That's may be serverless cloud implementation or a container/kubernetes (https://kubernetes.io/) implementation. You can then perfectly when the new release is loaded, control if you want to do Canary, Blue/Green or progressive rollout or even implement your own strategy.
Because it would work with containers, that would be fine whatever the langage be it C++, java, python, shell, Squeak or anything.
That layer would also provide auto scaling of your various services, redundancy and load balancing and distribute the workload on your cluster.
You can go to the next step with gitops. A PR merge in git automatically trigger the load of the new version in production (https://www.weave.works/technologies/gitops/)
Dynamically loading of jars in Java
Still for sure java thanks to its class loaded API allows to load classes dynamically. This is what web servers are doing and several implementations of that do exist like OSGI or check the response of dimo414.
Conclusion
It would seems that the java route make more sense for a generic plugin system like the one of Eclipse (OSGI), and that the containers solution make more sense for a globally distributed system, auto scaling & resiliance in clusters.
Kubernetes scale to thousand of nodes and provides a whole echosystem to deal with distributed system and it can scale and operate any linux or windows process. This is the de-facto standard pushed by Google and used by thousand of companies over the world.