[core] _ActorClassMetadata.last_export_session_and_job should not be pickled

[rynewang]

What happened + What you expected to happen

In actor.py, _ActorClassMetadata is a simple data class and is automatically pickled/unpickled. This means the field last_export_session_and_job is passed across workers. However it should not - the session index is a worker-local concept and does not mean anything when passed around.

This has a real issue because, it's used to determine if we need to upload actor class to GCS. If in a first session we upload it and thus setting session = 0; then in second session, we pass the Actor to a task (fresh worker, session = 0) and invoke the Actor, the worker thinks it's already uploaded and skip the upload, causing actor creation failure.

Versions / Dependencies

master

Reproduction script

import ray

@ray.remote
class A:
    def __init__(self):
        print("A.__init__")
    def hello(self):
        print("A.hello")

ray.init()
# Here: A session = None, worker session = 0; uploaded.
a = A.remote()
# Here: A session = 0
ray.get(a.hello.remote())
ray.shutdown()

@ray.remote
def f():
    # A is unpickled with session = 0 (should be = None)
    # Here: A session = 0, worker.session = 0, does not upload A.
    # actor creation failed.
    a = A.remote()
    ray.get(a.hello.remote())

ray.get(f.remote())

Issue Severity

Medium: It is a significant difficulty but I can work around it.

[rynewang]

Same things happen for Ray Tasks - class RemoteFunction is having a similar field _last_export_session_and_job and if we replicate the repro script for task, it hangs.