def export_actor(actor_id, class_id, actor_method_names, num_cpus, num_gpus,
worker):
"""Export an actor to redis.
Args:
actor_id: The ID of the actor.
actor_method_names (list): A list of the names of this actor's methods.
num_cpus (int): The number of CPUs that this actor requires.
num_gpus (int): The number of GPUs that this actor requires.
"""
ray.worker.check_main_thread()
if worker.mode is None:
raise Exception("Actors cannot be created before Ray has been "
"started. You can start Ray with 'ray.init()'.")
key = b"Actor:" + actor_id.id()
# For now, all actor methods have 1 return value.
driver_id = worker.task_driver_id.id()
for actor_method_name in actor_method_names:
# TODO(rkn): When we create a second actor, we are probably overwriting
# the values from the first actor here. This may or may not be a
# problem.
function_id = get_actor_method_function_id(actor_method_name).id()
worker.function_properties[driver_id][function_id] = (
FunctionProperties(num_return_vals=1,
num_cpus=1,
num_gpus=0,
num_custom_resource=0,
max_calls=0))
# Select a local scheduler for the actor.
local_scheduler_id = select_local_scheduler(
worker.task_driver_id.id(), ray.global_state.local_schedulers(),
num_gpus, worker.redis_client)
assert local_scheduler_id is not None
# We must put the actor information in Redis before publishing the actor
# notification so that when the newly created actor attempts to fetch the
# information from Redis, it is already there.
worker.redis_client.hmset(
key, {
"class_id": class_id,
"driver_id": driver_id,
"local_scheduler_id": local_scheduler_id,
"num_gpus": num_gpus,
"removed": False
})
# TODO(rkn): There is actually no guarantee that the local scheduler that
# we are publishing to has already subscribed to the actor_notifications
# channel. Therefore, this message may be missed and the workload will
# hang. This is a bug.
ray.utils.publish_actor_creation(actor_id.id(), driver_id,
local_scheduler_id, False,
worker.redis_client)
def export_actor(actor_id, class_id, class_name, actor_method_names,
actor_method_num_return_vals, resources, worker):
"""Export an actor to redis.
Args:
actor_id (common.ObjectID): The ID of the actor.
class_id (str): A random ID for the actor class.
class_name (str): The actor class name.
actor_method_names (list): A list of the names of this actor's methods.
actor_method_num_return_vals: A list of the number of return values for
each of the actor's methods.
resources: A dictionary mapping resource name to the quantity of that
resource required by the actor.
"""
ray.worker.check_main_thread()
if worker.mode is None:
raise Exception("Actors cannot be created before Ray has been "
"started. You can start Ray with 'ray.init()'.")
driver_id = worker.task_driver_id.id()
register_actor_signatures(worker, driver_id, class_name,
actor_method_names, actor_method_num_return_vals)
# Select a local scheduler for the actor.
key = b"Actor:" + actor_id.id()
local_scheduler_id = select_local_scheduler(
worker.task_driver_id.id(), ray.global_state.local_schedulers(),
resources.get("GPU", 0), worker.redis_client)
assert local_scheduler_id is not None
# We must put the actor information in Redis before publishing the actor
# notification so that when the newly created actor attempts to fetch the
# information from Redis, it is already there.
driver_id = worker.task_driver_id.id()
worker.redis_client.hmset(
key, {
"class_id": class_id,
"driver_id": driver_id,
"local_scheduler_id": local_scheduler_id,
"num_gpus": resources.get("GPU", 0),
"removed": False
})
# TODO(rkn): There is actually no guarantee that the local scheduler that
# we are publishing to has already subscribed to the actor_notifications
# channel. Therefore, this message may be missed and the workload will
# hang. This is a bug.
ray.utils.publish_actor_creation(actor_id.id(), driver_id,
local_scheduler_id, False,
worker.redis_client)