Discuss cache named actor in core worker locally.

Bulk users use ray.get_actor(name=xxx) temporarily, and it’s necessary to enable caching named actor in core worker locally.

Before support it, we’d better to discuss the behavior of ray.get_actor().

There 're 2 facts of the current named actor now:

• one is ray.get_actor() supports the same name for different actors serially:

a = ActorA.options(name="named_actor").remote()
ray.exit_actor(a, no_restart=True)
b = ActorB.options(name="named_actor").remote()       # It does work!!!

• another one is that ray.get_actor() will raise a ValueError if the name doesn’t exist.

I’d like to propose these: （1） named actor doesn’t support duplicated name with different actors. Because this is not a strong requirement and if we enable this, users may use it incorrectly as well. (2) ray.get_actor() returning None indicates there doesn’t exist the name, and returning not None indicates we found a handle for the name but we don’t promise the actor is valid. If that’s an invalid handle, remote call will raise the corresponding exception.

---

BTW, the reason for proposing no duplicated name for different actors is local core worker has no way to be notified the deletion event.(GCS pubs the deletion event is not a good idea.)

After those changes, we can support the TTL for the named actor cache locally.

@edoakes @kfstorm @rkooo567 @simon-mo @raulchen cc

[jovany-wang]

Tao is working in progress: https://github.com/ray-project/ray/pull/17339

[jovany-wang]

@WangTaoTheTonic will follow up this item.