def actor_def(self):
# Delayed import as replica depends on this file.
from ray.serve.replica import create_replica_wrapper
if self._cached_actor_def is None:
assert self.actor_name is not None
assert (
self.replica_config.import_path is not None
or self.serialized_deployment_def is not None
)
if self.replica_config.import_path is not None:
self._cached_actor_def = ray.remote(
create_replica_wrapper(
self.actor_name,
import_path=self.replica_config.import_path,
)
)
else:
self._cached_actor_def = ray.remote(
create_replica_wrapper(
self.actor_name,
serialized_deployment_def=self.serialized_deployment_def,
)
)
return self._cached_actor_def
def deploy(self,
name: str,
deployment_config_proto_bytes: bytes,
replica_config: ReplicaConfig,
version: Optional[str],
prev_version: Optional[str],
route_prefix: Optional[str],
deployer_job_id: "Optional[ray._raylet.JobID]" = None
) -> Tuple[Optional[GoalId], bool]:
if route_prefix is not None:
assert route_prefix.startswith("/")
deployment_config = DeploymentConfig.from_proto_bytes(
deployment_config_proto_bytes)
if prev_version is not None:
existing_deployment_info = (
self.deployment_state_manager.get_deployment(name))
if (existing_deployment_info is None
or not existing_deployment_info.version):
raise ValueError(
f"prev_version '{prev_version}' is specified but "
"there is no existing deployment.")
if existing_deployment_info.version != prev_version:
raise ValueError(
f"prev_version '{prev_version}' "
"does not match with the existing "
f"version '{existing_deployment_info.version}'.")
autoscaling_config = deployment_config.autoscaling_config
if autoscaling_config is not None:
# TODO: is this the desired behaviour? Should this be a setting?
deployment_config.num_replicas = autoscaling_config.min_replicas
autoscaling_policy = BasicAutoscalingPolicy(autoscaling_config)
else:
autoscaling_policy = None
deployment_info = DeploymentInfo(
actor_def=ray.remote(
create_replica_wrapper(
name, replica_config.serialized_deployment_def)),
version=version,
deployment_config=deployment_config,
replica_config=replica_config,
deployer_job_id=deployer_job_id,
start_time_ms=int(time.time() * 1000),
autoscaling_policy=autoscaling_policy)
# TODO(architkulkarni): When a deployment is redeployed, even if
# the only change was num_replicas, the start_time_ms is refreshed.
# Is this the desired behaviour?
goal_id, updating = self.deployment_state_manager.deploy(
name, deployment_info)
endpoint_info = EndpointInfo(route=route_prefix)
self.endpoint_state.update_endpoint(name, endpoint_info)
return goal_id, updating
def actor_def(self):
# Delayed import as replica depends on this file.
from ray.serve.replica import create_replica_wrapper
if self._cached_actor_def is None:
assert self.actor_name is not None
assert self.serialized_deployment_def is not None
self._cached_actor_def = ray.remote(
create_replica_wrapper(self.actor_name,
self.serialized_deployment_def))
return self._cached_actor_def
def actor_def(self):
# Delayed import as replica depends on this file.
from ray.serve.replica import create_replica_wrapper
if self._cached_actor_def is None:
assert self.actor_name is not None
self._cached_actor_def = ray.remote(
**REPLICA_DEFAULT_ACTOR_OPTIONS)(create_replica_wrapper(
self.actor_name))
return self._cached_actor_def
def deploy(
self,
name: str,
backend_config_proto_bytes: bytes,
replica_config: ReplicaConfig,
version: Optional[str],
prev_version: Optional[str],
route_prefix: Optional[str],
deployer_job_id: "Optional[ray._raylet.JobID]" = None
) -> Tuple[Optional[GoalId], bool]:
if route_prefix is not None:
assert route_prefix.startswith("/")
backend_config = BackendConfig.from_proto_bytes(
backend_config_proto_bytes)
if prev_version is not None:
existing_backend_info = self.backend_state_manager.get_backend(
name)
if (existing_backend_info is None
or not existing_backend_info.version):
raise ValueError(
f"prev_version '{prev_version}' is specified but "
"there is no existing deployment.")
if existing_backend_info.version != prev_version:
raise ValueError(f"prev_version '{prev_version}' "
"does not match with the existing "
f"version '{existing_backend_info.version}'.")
backend_info = BackendInfo(actor_def=ray.remote(
create_replica_wrapper(name,
replica_config.serialized_backend_def)),
version=version,
backend_config=backend_config,
replica_config=replica_config,
deployer_job_id=deployer_job_id,
start_time_ms=int(time.time() * 1000))
# TODO(architkulkarni): When a deployment is redeployed, even if
# the only change was num_replicas, the start_time_ms is refreshed.
# This is probably not the desired behavior for an autoscaling
# deployment, which redeploys very often to change num_replicas.
goal_id, updating = self.backend_state_manager.deploy_backend(
name, backend_info)
endpoint_info = EndpointInfo(route=route_prefix)
self.endpoint_state.update_endpoint(name, endpoint_info)
return goal_id, updating
