def _connect(
_override_controller_namespace: Optional[str] = None,
) -> ServeControllerClient:
"""Connect to an existing Serve instance on this Ray cluster.
If calling from the driver program, the Serve instance on this Ray cluster
must first have been initialized using `serve.start(detached=True)`.
If called from within a replica, this will connect to the same Serve
instance that the replica is running in.
Args:
_override_controller_namespace (Optional[str]): The namespace to use
when looking for the controller. If None, Serve recalculates the
controller's namespace using get_controller_namespace().
Raises:
RayServeException: if there is no Serve controller actor in the
expected namespace.
"""
# Initialize ray if needed.
ray.worker.global_worker.filter_logs_by_job = False
if not ray.is_initialized():
ray.init(namespace="serve")
# When running inside of a replica, _INTERNAL_REPLICA_CONTEXT is set to
# ensure that the correct instance is connected to.
if _INTERNAL_REPLICA_CONTEXT is None:
controller_name = SERVE_CONTROLLER_NAME
controller_namespace = get_controller_namespace(
detached=True,
_override_controller_namespace=_override_controller_namespace)
else:
controller_name = _INTERNAL_REPLICA_CONTEXT._internal_controller_name
controller_namespace = _INTERNAL_REPLICA_CONTEXT._internal_controller_namespace
# Try to get serve controller if it exists
try:
controller = ray.get_actor(controller_name,
namespace=controller_namespace)
except ValueError:
raise RayServeException("There is no "
"instance running on this Ray cluster. Please "
"call `serve.start(detached=True) to start "
"one.")
client = ServeControllerClient(
controller,
controller_name,
detached=True,
_override_controller_namespace=_override_controller_namespace,
)
set_global_client(client)
return client
def start(
detached: bool = False,
http_options: Optional[Union[dict, HTTPOptions]] = None,
dedicated_cpu: bool = False,
_checkpoint_path: str = DEFAULT_CHECKPOINT_PATH,
_override_controller_namespace: Optional[str] = None,
**kwargs,
) -> ServeControllerClient:
"""Initialize a serve instance.
By default, the instance will be scoped to the lifetime of the returned
Client object (or when the script exits). If detached is set to True, the
instance will instead persist until serve.shutdown() is called. This is
only relevant if connecting to a long-running Ray cluster (e.g., with
ray.init(address="auto") or ray.init("ray://<remote_addr>")).
Args:
detached (bool): Whether not the instance should be detached from this
script. If set, the instance will live on the Ray cluster until it is
explicitly stopped with serve.shutdown().
http_options (Optional[Dict, serve.HTTPOptions]): Configuration options
for HTTP proxy. You can pass in a dictionary or HTTPOptions object
with fields:
- host(str, None): Host for HTTP servers to listen on. Defaults to
"127.0.0.1". To expose Serve publicly, you probably want to set
this to "0.0.0.0".
- port(int): Port for HTTP server. Defaults to 8000.
- root_path(str): Root path to mount the serve application
(for example, "/serve"). All deployment routes will be prefixed
with this path. Defaults to "".
- middlewares(list): A list of Starlette middlewares that will be
applied to the HTTP servers in the cluster. Defaults to [].
- location(str, serve.config.DeploymentMode): The deployment
location of HTTP servers:
- "HeadOnly": start one HTTP server on the head node. Serve
assumes the head node is the node you executed serve.start
on. This is the default.
- "EveryNode": start one HTTP server per node.
- "NoServer" or None: disable HTTP server.
- num_cpus (int): The number of CPU cores to reserve for each
internal Serve HTTP proxy actor. Defaults to 0.
dedicated_cpu (bool): Whether to reserve a CPU core for the internal
Serve controller actor. Defaults to False.
"""
usage_lib.record_library_usage("serve")
http_deprecated_args = ["http_host", "http_port", "http_middlewares"]
for key in http_deprecated_args:
if key in kwargs:
raise ValueError(
f"{key} is deprecated, please use serve.start(http_options="
f'{{"{key}": {kwargs[key]}}}) instead.'
)
# Initialize ray if needed.
ray.worker.global_worker.filter_logs_by_job = False
if not ray.is_initialized():
ray.init(namespace="serve")
controller_namespace = get_controller_namespace(
detached, _override_controller_namespace=_override_controller_namespace
)
try:
client = get_global_client(
_override_controller_namespace=_override_controller_namespace,
_health_check_controller=True,
)
logger.info(
"Connecting to existing Serve instance in namespace "
f"'{controller_namespace}'."
)
_check_http_and_checkpoint_options(client, http_options, _checkpoint_path)
return client
except RayServeException:
pass
if detached:
controller_name = SERVE_CONTROLLER_NAME
else:
controller_name = format_actor_name(get_random_letters(), SERVE_CONTROLLER_NAME)
if isinstance(http_options, dict):
http_options = HTTPOptions.parse_obj(http_options)
if http_options is None:
http_options = HTTPOptions()
controller = ServeController.options(
num_cpus=1 if dedicated_cpu else 0,
name=controller_name,
lifetime="detached" if detached else None,
max_restarts=-1,
max_task_retries=-1,
# Pin Serve controller on the head node.
resources={get_current_node_resource_key(): 0.01},
namespace=controller_namespace,
max_concurrency=CONTROLLER_MAX_CONCURRENCY,
).remote(
controller_name,
http_options,
_checkpoint_path,
detached=detached,
_override_controller_namespace=_override_controller_namespace,
)
proxy_handles = ray.get(controller.get_http_proxies.remote())
if len(proxy_handles) > 0:
try:
ray.get(
[handle.ready.remote() for handle in proxy_handles.values()],
timeout=HTTP_PROXY_TIMEOUT,
)
except ray.exceptions.GetTimeoutError:
raise TimeoutError(
f"HTTP proxies not available after {HTTP_PROXY_TIMEOUT}s."
)
client = ServeControllerClient(
controller,
controller_name,
detached=detached,
_override_controller_namespace=_override_controller_namespace,
)
set_global_client(client)
logger.info(
f"Started{' detached ' if detached else ' '}Serve instance in "
f"namespace '{controller_namespace}'."
)
return client