If I change the repro script to:

import ray


@ray.remote
class Actor:
    pass


ray.init()

actors = []
for _ in range(0, 1000):
    actors.append(Actor.options(
        name="actor",
        get_if_exists=True,
        max_restarts=1,
    ).remote())

It succeeds.

I mis-titled this initially, it's when max_restarts != 0

What is your expected behavior for the script below? Should only the first Actor.options call create the actor, while all the following 999 calls to Actor.options retrieve the existing actor?

The expected behavior is somewhat difficult to define because there is no reference to the actor. As a result, the reference count will be 0, leading to the actor being destroyed.

import ray

@ray.remote
class Actor:
    pass

ray.init()

for _ in range(0, 1000):
    Actor.options(
        name="actor",
        get_if_exists=True,
        max_restarts=-1,
    ).remote()

@kevin85421 like I said, this is not exactly our use case - in our case we do have a reference to the Actor, but the reference goes away and is then re-created (we're seeing this in tests, where a test is fine when run on its own, but fails when run right after a previous test. The previous test will have created the Actor, used it, then all references disappear, and then we get this exception when the subsequent test tries to create the Actor).

So I don't really have a good answer about what this test case should do - I guess I'd expect it to either create or retrieve the Actor in each iteration, depending on how async actor destruction is, and not to throw an exception.