xoscar 0.3.1__cp311-cp311-macosx_11_0_arm64.whl

xoscar/backends/pool.py

@@ -0,0 +1,1593 @@
+# Copyright 2022-2023 XProbe Inc.
+# derived from copyright 1999-2021 Alibaba Group Holding Ltd.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import asyncio
+import concurrent.futures as futures
+import contextlib
+import itertools
+import logging
+import multiprocessing
+import os
+import threading
+import traceback
+from abc import ABC, ABCMeta, abstractmethod
+from typing import Any, Callable, Coroutine, Optional, Type, TypeVar
+
+import psutil
+
+from .._utils import TypeDispatcher, create_actor_ref, to_binary
+from ..api import Actor
+from ..core import ActorRef, BufferRef, FileObjectRef, register_local_pool
+from ..debug import debug_async_timeout, record_message_trace
+from ..errors import (
+    ActorAlreadyExist,
+    ActorNotExist,
+    CannotCancelTask,
+    SendMessageFailed,
+    ServerClosed,
+)
+from ..metrics import init_metrics
+from ..utils import implements, register_asyncio_task_timeout_detector
+from .allocate_strategy import AddressSpecified, allocated_type
+from .communication import (
+    Channel,
+    Server,
+    UCXChannel,
+    gen_local_address,
+    get_server_type,
+)
+from .communication.errors import ChannelClosed
+from .config import ActorPoolConfig
+from .core import ActorCaller, ResultMessageType
+from .message import (
+    DEFAULT_PROTOCOL,
+    ActorRefMessage,
+    CancelMessage,
+    ControlMessage,
+    ControlMessageType,
+    CreateActorMessage,
+    DestroyActorMessage,
+    ErrorMessage,
+    HasActorMessage,
+    MessageType,
+    ResultMessage,
+    SendMessage,
+    TellMessage,
+    _MessageBase,
+    new_message_id,
+)
+from .router import Router
+
+logger = logging.getLogger(__name__)
+
+
+@contextlib.contextmanager
+def _disable_log_temporally():
+    if os.getenv("CUDA_VISIBLE_DEVICES") == "-1":
+        # disable logging when CUDA_VISIBLE_DEVICES == -1
+        # many logging comes from ptxcompiler may distract users
+        try:
+            logging.disable(level=logging.ERROR)
+            yield
+        finally:
+            logging.disable(level=logging.NOTSET)
+    else:
+        yield
+
+
+class _ErrorProcessor:
+    def __init__(self, address: str, message_id: bytes, protocol):
+        self._address = address
+        self._message_id = message_id
+        self._protocol = protocol
+        self.result = None
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        if self.result is None:
+            self.result = ErrorMessage(
+                self._message_id,
+                self._address,
+                os.getpid(),
+                exc_type,
+                exc_val,
+                exc_tb,
+                protocol=self._protocol,
+            )
+            return True
+
+
+def _register_message_handler(pool_type: Type["AbstractActorPool"]):
+    pool_type._message_handler = dict()
+    for message_type, handler in [
+        (MessageType.create_actor, pool_type.create_actor),
+        (MessageType.destroy_actor, pool_type.destroy_actor),
+        (MessageType.has_actor, pool_type.has_actor),
+        (MessageType.actor_ref, pool_type.actor_ref),
+        (MessageType.send, pool_type.send),
+        (MessageType.tell, pool_type.tell),
+        (MessageType.cancel, pool_type.cancel),
+        (MessageType.control, pool_type.handle_control_command),
+        (MessageType.copy_to_buffers, pool_type.handle_copy_to_buffers_message),
+        (MessageType.copy_to_fileobjs, pool_type.handle_copy_to_fileobjs_message),
+    ]:
+        pool_type._message_handler[message_type] = handler  # type: ignore
+    return pool_type
+
+
+class AbstractActorPool(ABC):
+    __slots__ = (
+        "process_index",
+        "label",
+        "external_address",
+        "internal_address",
+        "env",
+        "_servers",
+        "_router",
+        "_config",
+        "_stopped",
+        "_actors",
+        "_caller",
+        "_process_messages",
+        "_asyncio_task_timeout_detector_task",
+    )
+
+    _message_handler: dict[MessageType, Callable]
+    _process_messages: dict[bytes, asyncio.Future | asyncio.Task | None]
+
+    def __init__(
+        self,
+        process_index: int,
+        label: str,
+        external_address: str,
+        internal_address: str,
+        env: dict,
+        router: Router,
+        config: ActorPoolConfig,
+        servers: list[Server],
+    ):
+        # register local pool for local actor lookup.
+        # The pool is weakrefed, so we don't need to unregister it.
+        register_local_pool(external_address, self)
+        self.process_index = process_index
+        self.label = label
+        self.external_address = external_address
+        self.internal_address = internal_address
+        self.env = env
+        self._router = router
+        self._config = config
+        self._servers = servers
+
+        self._stopped = asyncio.Event()
+
+        # states
+        # actor id -> actor
+        self._actors: dict[bytes, Actor] = dict()
+        # message id -> future
+        self._process_messages = dict()
+
+        # manage async actor callers
+        self._caller = ActorCaller()
+        self._asyncio_task_timeout_detector_task = (
+            register_asyncio_task_timeout_detector()
+        )
+        # init metrics
+        metric_configs = self._config.get_metric_configs()
+        metric_backend = metric_configs.get("backend")
+        init_metrics(metric_backend, config=metric_configs.get(metric_backend))
+
+    @property
+    def router(self):
+        return self._router
+
+    @abstractmethod
+    async def create_actor(self, message: CreateActorMessage) -> ResultMessageType:
+        """
+        Create an actor.
+
+        Parameters
+        ----------
+        message: CreateActorMessage
+            message to create an actor.
+
+        Returns
+        -------
+        result_message
+            result or error message.
+        """
+
+    @abstractmethod
+    async def has_actor(self, message: HasActorMessage) -> ResultMessage:
+        """
+        Check if an actor exists or not.
+
+        Parameters
+        ----------
+        message: HasActorMessage
+            message
+
+        Returns
+        -------
+        result_message
+            result message contains if an actor exists or not.
+        """
+
+    @abstractmethod
+    async def destroy_actor(self, message: DestroyActorMessage) -> ResultMessageType:
+        """
+        Destroy an actor.
+
+        Parameters
+        ----------
+        message: DestroyActorMessage
+            message to destroy an actor.
+
+        Returns
+        -------
+        result_message
+            result or error message.
+        """
+
+    @abstractmethod
+    async def actor_ref(self, message: ActorRefMessage) -> ResultMessageType:
+        """
+        Get an actor's ref.
+
+        Parameters
+        ----------
+        message: ActorRefMessage
+            message to get an actor's ref.
+
+        Returns
+        -------
+        result_message
+            result or error message.
+        """
+
+    @abstractmethod
+    async def send(self, message: SendMessage) -> ResultMessageType:
+        """
+        Send a message to some actor.
+
+        Parameters
+        ----------
+        message: SendMessage
+            Message to send.
+
+        Returns
+        -------
+        result_message
+            result or error message.
+        """
+
+    @abstractmethod
+    async def tell(self, message: TellMessage) -> ResultMessageType:
+        """
+        Tell message to some actor.
+
+        Parameters
+        ----------
+        message: TellMessage
+            Message to tell.
+
+        Returns
+        -------
+        result_message
+            result or error message.
+        """
+
+    @abstractmethod
+    async def cancel(self, message: CancelMessage) -> ResultMessageType:
+        """
+        Cancel message that sent
+
+        Parameters
+        ----------
+        message: CancelMessage
+            Cancel message.
+
+        Returns
+        -------
+        result_message
+            result or error message
+        """
+
+    def _sync_pool_config(self, actor_pool_config: ActorPoolConfig):
+        self._config = actor_pool_config
+        # remove router from global one
+        global_router = Router.get_instance()
+        global_router.remove_router(self._router)  # type: ignore
+        # update router
+        self._router.set_mapping(actor_pool_config.external_to_internal_address_map)
+        # update global router
+        global_router.add_router(self._router)  # type: ignore
+
+    async def handle_control_command(
+        self, message: ControlMessage
+    ) -> ResultMessageType:
+        """
+        Handle control command.
+
+        Parameters
+        ----------
+        message: ControlMessage
+            Control message.
+
+        Returns
+        -------
+        result_message
+            result or error message.
+        """
+        with _ErrorProcessor(
+            self.external_address, message.message_id, protocol=message.protocol
+        ) as processor:
+            content: bool | ActorPoolConfig = True
+            if message.control_message_type == ControlMessageType.stop:
+                await self.stop()
+            elif message.control_message_type == ControlMessageType.sync_config:
+                self._sync_pool_config(message.content)
+            elif message.control_message_type == ControlMessageType.get_config:
+                if message.content == "main_pool_address":
+                    main_process_index = self._config.get_process_indexes()[0]
+                    content = self._config.get_pool_config(main_process_index)[
+                        "external_address"
+                    ][0]
+                else:
+                    content = self._config
+            else:  # pragma: no cover
+                raise TypeError(
+                    f"Unable to handle control message "
+                    f"with type {message.control_message_type}"
+                )
+            processor.result = ResultMessage(
+                message.message_id, content, protocol=message.protocol
+            )
+
+        return processor.result
+
+    async def _run_coro(self, message_id: bytes, coro: Coroutine):
+        self._process_messages[message_id] = asyncio.tasks.current_task()
+        try:
+            return await coro
+        finally:
+            self._process_messages.pop(message_id, None)
+
+    async def _send_channel(
+        self, result: _MessageBase, channel: Channel, resend_failure: bool = True
+    ):
+        try:
+            await channel.send(result)
+        except (ChannelClosed, ConnectionResetError):
+            if not self._stopped.is_set():
+                raise
+        except Exception as ex:
+            logger.exception(
+                "Error when sending message %s from %s to %s",
+                result.message_id.hex(),
+                channel.local_address,
+                channel.dest_address,
+            )
+            if not resend_failure:  # pragma: no cover
+                raise
+
+            with _ErrorProcessor(
+                self.external_address, result.message_id, result.protocol
+            ) as processor:
+                error_msg = (
+                    f"Error when sending message {result.message_id.hex()}. "
+                    f"Caused by {ex!r}. "
+                )
+                if isinstance(result, ErrorMessage):
+                    format_tb = "\n".join(traceback.format_tb(result.traceback))
+                    error_msg += (
+                        f"\nOriginal error: {result.error!r}"
+                        f"Traceback: \n{format_tb}"
+                    )
+                else:
+                    error_msg += "See server logs for more details"
+                raise SendMessageFailed(error_msg) from None
+            await self._send_channel(processor.result, channel, resend_failure=False)
+
+    async def process_message(self, message: _MessageBase, channel: Channel):
+        handler = self._message_handler[message.message_type]
+        with _ErrorProcessor(
+            self.external_address, message.message_id, message.protocol
+        ) as processor:
+            # use `%.500` to avoid print too long messages
+            with debug_async_timeout(
+                "process_message_timeout",
+                "Process message %.500s of channel %s timeout.",
+                message,
+                channel,
+            ):
+                processor.result = await self._run_coro(
+                    message.message_id, handler(self, message)
+                )
+
+        await self._send_channel(processor.result, channel)
+
+    async def call(self, dest_address: str, message: _MessageBase) -> ResultMessageType:
+        return await self._caller.call(self._router, dest_address, message)  # type: ignore
+
+    @staticmethod
+    def _parse_config(config: dict, kw: dict) -> dict:
+        actor_pool_config: ActorPoolConfig = config.pop("actor_pool_config")
+        kw["config"] = actor_pool_config
+        kw["process_index"] = process_index = config.pop("process_index")
+        curr_pool_config = actor_pool_config.get_pool_config(process_index)
+        kw["label"] = curr_pool_config["label"]
+        external_addresses = curr_pool_config["external_address"]
+        kw["external_address"] = external_addresses[0]
+        kw["internal_address"] = curr_pool_config["internal_address"]
+        kw["router"] = Router(
+            external_addresses,
+            gen_local_address(process_index),
+            actor_pool_config.external_to_internal_address_map,
+            comm_config=actor_pool_config.get_comm_config(),
+        )
+        kw["env"] = curr_pool_config["env"]
+
+        if config:  # pragma: no cover
+            raise TypeError(
+                f"Creating pool got unexpected " f'arguments: {",".join(config)}'
+            )
+
+        return kw
+
+    @classmethod
+    @abstractmethod
+    async def create(cls, config: dict) -> "AbstractActorPool":
+        """
+        Create an actor pool.
+
+        Parameters
+        ----------
+        config: dict
+            configurations.
+
+        Returns
+        -------
+        actor_pool:
+            Actor pool.
+        """
+
+    async def start(self):
+        if self._stopped.is_set():
+            raise RuntimeError("pool has been stopped, cannot start again")
+        start_servers = [server.start() for server in self._servers]
+        await asyncio.gather(*start_servers)
+
+    async def join(self, timeout: float | None = None):
+        wait_stopped = asyncio.create_task(self._stopped.wait())
+
+        try:
+            await asyncio.wait_for(wait_stopped, timeout=timeout)
+        except (futures.TimeoutError, asyncio.TimeoutError):  # pragma: no cover
+            wait_stopped.cancel()
+
+    async def stop(self):
+        try:
+            # clean global router
+            router = Router.get_instance()
+            if router is not None:
+                router.remove_router(self._router)
+            stop_tasks = []
+            # stop all servers
+            stop_tasks.extend([server.stop() for server in self._servers])
+            # stop all clients
+            stop_tasks.append(self._caller.stop())
+            await asyncio.gather(*stop_tasks)
+
+            self._servers = []
+            if self._asyncio_task_timeout_detector_task:  # pragma: no cover
+                self._asyncio_task_timeout_detector_task.cancel()
+        finally:
+            self._stopped.set()
+
+    async def handle_copy_to_buffers_message(self, message) -> ResultMessage:
+        for addr, uid, start, _len, data in message.content:
+            buffer = BufferRef.get_buffer(BufferRef(addr, uid))
+            buffer[start : start + _len] = data
+        return ResultMessage(message_id=message.message_id, result=True)
+
+    async def handle_copy_to_fileobjs_message(self, message) -> ResultMessage:
+        for addr, uid, data in message.content:
+            file_obj = FileObjectRef.get_local_file_object(FileObjectRef(addr, uid))
+            await file_obj.write(data)
+        return ResultMessage(message_id=message.message_id, result=True)
+
+    @property
+    def stopped(self) -> bool:
+        return self._stopped.is_set()
+
+    async def _handle_ucx_meta_message(
+        self, message: _MessageBase, channel: Channel
+    ) -> bool:
+        if (
+            isinstance(message, ControlMessage)
+            and message.message_type == MessageType.control
+            and message.control_message_type == ControlMessageType.switch_to_copy_to
+            and isinstance(channel, UCXChannel)
+        ):
+            with _ErrorProcessor(
+                self.external_address, message.message_id, message.protocol
+            ) as processor:
+                # use `%.500` to avoid print too long messages
+                with debug_async_timeout(
+                    "process_message_timeout",
+                    "Process message %.500s of channel %s timeout.",
+                    message,
+                    channel,
+                ):
+                    buffers = [
+                        BufferRef.get_buffer(BufferRef(addr, uid))
+                        for addr, uid in message.content
+                    ]
+                    await channel.recv_buffers(buffers)
+                    processor.result = ResultMessage(
+                        message_id=message.message_id, result=True
+                    )
+            asyncio.create_task(self._send_channel(processor.result, channel))
+            return True
+        return False
+
+    async def on_new_channel(self, channel: Channel):
+        while not self._stopped.is_set():
+            try:
+                message = await channel.recv()
+            except EOFError:
+                # no data to read, check channel
+                try:
+                    await channel.close()
+                except (ConnectionError, EOFError):
+                    # close failed, ignore
+                    pass
+                return
+            if await self._handle_ucx_meta_message(message, channel):
+                continue
+            asyncio.create_task(self.process_message(message, channel))
+            # delete to release the reference of message
+            del message
+            await asyncio.sleep(0)
+
+    async def __aenter__(self):
+        await self.start()
+        return self
+
+    async def __aexit__(self, exc_type, exc_val, exc_tb):
+        await self.stop()
+
+
+class ActorPoolBase(AbstractActorPool, metaclass=ABCMeta):
+    __slots__ = ()
+
+    @implements(AbstractActorPool.create_actor)
+    async def create_actor(self, message: CreateActorMessage) -> ResultMessageType:
+        with _ErrorProcessor(
+            self.external_address, message.message_id, message.protocol
+        ) as processor:
+            actor_id = message.actor_id
+            if actor_id in self._actors:
+                raise ActorAlreadyExist(
+                    f"Actor {actor_id!r} already exist, cannot create"
+                )
+
+            actor = message.actor_cls(*message.args, **message.kwargs)
+            actor.uid = actor_id
+            actor.address = address = self.external_address
+            self._actors[actor_id] = actor
+            await self._run_coro(message.message_id, actor.__post_create__())
+
+            result = ActorRef(address, actor_id)
+            # ensemble result message
+            processor.result = ResultMessage(
+                message.message_id, result, protocol=message.protocol
+            )
+        return processor.result
+
+    @implements(AbstractActorPool.has_actor)
+    async def has_actor(self, message: HasActorMessage) -> ResultMessage:
+        result = ResultMessage(
+            message.message_id,
+            message.actor_ref.uid in self._actors,
+            protocol=message.protocol,
+        )
+        return result
+
+    @implements(AbstractActorPool.destroy_actor)
+    async def destroy_actor(self, message: DestroyActorMessage) -> ResultMessageType:
+        with _ErrorProcessor(
+            self.external_address, message.message_id, message.protocol
+        ) as processor:
+            actor_id = message.actor_ref.uid
+            try:
+                actor = self._actors[actor_id]
+            except KeyError:
+                raise ActorNotExist(f"Actor {actor_id} does not exist")
+            await self._run_coro(message.message_id, actor.__pre_destroy__())
+            del self._actors[actor_id]
+
+            processor.result = ResultMessage(
+                message.message_id, actor_id, protocol=message.protocol
+            )
+        return processor.result
+
+    @implements(AbstractActorPool.actor_ref)
+    async def actor_ref(self, message: ActorRefMessage) -> ResultMessageType:
+        with _ErrorProcessor(
+            self.external_address, message.message_id, message.protocol
+        ) as processor:
+            actor_id = message.actor_ref.uid
+            if actor_id not in self._actors:
+                raise ActorNotExist(f"Actor {actor_id} does not exist")
+            result = ResultMessage(
+                message.message_id,
+                ActorRef(self.external_address, actor_id),
+                protocol=message.protocol,
+            )
+            processor.result = result
+        return processor.result
+
+    @implements(AbstractActorPool.send)
+    async def send(self, message: SendMessage) -> ResultMessageType:
+        with _ErrorProcessor(
+            self.external_address, message.message_id, message.protocol
+        ) as processor, record_message_trace(message):
+            actor_id = message.actor_ref.uid
+            if actor_id not in self._actors:
+                raise ActorNotExist(f"Actor {actor_id} does not exist")
+            coro = self._actors[actor_id].__on_receive__(message.content)
+            result = await self._run_coro(message.message_id, coro)
+            processor.result = ResultMessage(
+                message.message_id,
+                result,
+                protocol=message.protocol,
+                profiling_context=message.profiling_context,
+            )
+        return processor.result
+
+    @implements(AbstractActorPool.tell)
+    async def tell(self, message: TellMessage) -> ResultMessageType:
+        with _ErrorProcessor(
+            self.external_address, message.message_id, message.protocol
+        ) as processor:
+            actor_id = message.actor_ref.uid
+            if actor_id not in self._actors:  # pragma: no cover
+                raise ActorNotExist(f"Actor {actor_id} does not exist")
+            call = self._actors[actor_id].__on_receive__(message.content)
+            # asynchronously run, tell does not care about result
+            asyncio.create_task(call)
+            await asyncio.sleep(0)
+            processor.result = ResultMessage(
+                message.message_id,
+                None,
+                protocol=message.protocol,
+                profiling_context=message.profiling_context,
+            )
+        return processor.result
+
+    @implements(AbstractActorPool.cancel)
+    async def cancel(self, message: CancelMessage) -> ResultMessageType:
+        with _ErrorProcessor(
+            self.external_address, message.message_id, message.protocol
+        ) as processor:
+            future = self._process_messages.get(message.cancel_message_id)
+            if future is None or future.done():  # pragma: no cover
+                raise CannotCancelTask(
+                    "Task not exists, maybe it is done or cancelled already"
+                )
+            future.cancel()
+            processor.result = ResultMessage(
+                message.message_id, True, protocol=message.protocol
+            )
+        return processor.result
+
+    @staticmethod
+    def _set_global_router(router: Router):
+        # be cautious about setting global router
+        # for instance, multiple main pool may be created in the same process
+
+        # get default router or create an empty one
+        default_router = Router.get_instance_or_empty()
+        Router.set_instance(default_router)
+        # append this router to global
+        default_router.add_router(router)
+
+    @staticmethod
+    def _update_stored_addresses(
+        servers: list[Server],
+        raw_addresses: list[str],
+        actor_pool_config: ActorPoolConfig,
+        kw: dict,
+    ):
+        process_index = kw["process_index"]
+        curr_pool_config = actor_pool_config.get_pool_config(process_index)
+        external_addresses = curr_pool_config["external_address"]
+        external_address_set = set(external_addresses)
+
+        kw["servers"] = servers
+
+        new_external_addresses = [
+            server.address
+            for server, raw_address in zip(servers, raw_addresses)
+            if raw_address in external_address_set
+        ]
+
+        if external_address_set != set(new_external_addresses):
+            external_addresses = new_external_addresses
+            actor_pool_config.reset_pool_external_address(
+                process_index, external_addresses
+            )
+            external_addresses = curr_pool_config["external_address"]
+
+            logger.debug(
+                "External address of process index %s updated to %s",
+                process_index,
+                external_addresses[0],
+            )
+            if kw["internal_address"] == kw["external_address"]:
+                # internal address may be the same as external address in Windows
+                kw["internal_address"] = external_addresses[0]
+            kw["external_address"] = external_addresses[0]
+
+            kw["router"] = Router(
+                external_addresses,
+                gen_local_address(process_index),
+                actor_pool_config.external_to_internal_address_map,
+                comm_config=actor_pool_config.get_comm_config(),
+            )
+
+    @classmethod
+    async def _create_servers(
+        cls, addresses: list[str], channel_handler: Callable, config: dict
+    ):
+        assert len(set(addresses)) == len(addresses)
+        # create servers
+        create_server_tasks = []
+        for addr in addresses:
+            server_type = get_server_type(addr)
+            extra_config = server_type.parse_config(config)
+            server_config = dict(address=addr, handle_channel=channel_handler)
+            server_config.update(extra_config)
+            task = asyncio.create_task(server_type.create(server_config))
+            create_server_tasks.append(task)
+
+        await asyncio.gather(*create_server_tasks)
+        return [f.result() for f in create_server_tasks]
+
+    @classmethod
+    @implements(AbstractActorPool.create)
+    async def create(cls, config: dict) -> "ActorPoolType":
+        config = config.copy()
+        kw: dict[str, Any] = dict()
+        cls._parse_config(config, kw)
+        process_index: int = kw["process_index"]
+        actor_pool_config = kw["config"]  # type: ActorPoolConfig
+        cur_pool_config = actor_pool_config.get_pool_config(process_index)
+        external_addresses = cur_pool_config["external_address"]
+        internal_address = kw["internal_address"]
+
+        # import predefined modules
+        modules = cur_pool_config["modules"] or []
+        for mod in modules:
+            __import__(mod, globals(), locals(), [])
+        # make sure all lazy imports loaded
+        with _disable_log_temporally():
+            TypeDispatcher.reload_all_lazy_handlers()
+
+        def handle_channel(channel):
+            return pool.on_new_channel(channel)
+
+        # create servers
+        server_addresses = list(external_addresses)
+        if internal_address:
+            server_addresses.append(internal_address)
+        server_addresses.append(gen_local_address(process_index))
+        server_addresses = sorted(set(server_addresses))
+        servers = await cls._create_servers(
+            server_addresses, handle_channel, actor_pool_config.get_comm_config()
+        )
+        cls._update_stored_addresses(servers, server_addresses, actor_pool_config, kw)
+
+        # set default router
+        # actor context would be able to use exact client
+        cls._set_global_router(kw["router"])
+
+        # create pool
+        pool = cls(**kw)
+        return pool  # type: ignore
+
+
+ActorPoolType = TypeVar("ActorPoolType", bound=AbstractActorPool)
+MainActorPoolType = TypeVar("MainActorPoolType", bound="MainActorPoolBase")
+SubProcessHandle = multiprocessing.Process
+
+
+class SubActorPoolBase(ActorPoolBase):
+    __slots__ = ("_main_address", "_watch_main_pool_task")
+    _watch_main_pool_task: Optional[asyncio.Task]
+
+    def __init__(
+        self,
+        process_index: int,
+        label: str,
+        external_address: str,
+        internal_address: str,
+        env: dict,
+        router: Router,
+        config: ActorPoolConfig,
+        servers: list[Server],
+        main_address: str,
+        main_pool_pid: Optional[int],
+    ):
+        super().__init__(
+            process_index,
+            label,
+            external_address,
+            internal_address,
+            env,
+            router,
+            config,
+            servers,
+        )
+        self._main_address = main_address
+        if main_pool_pid:
+            self._watch_main_pool_task = asyncio.create_task(
+                self._watch_main_pool(main_pool_pid)
+            )
+        else:
+            self._watch_main_pool_task = None
+
+    async def _watch_main_pool(self, main_pool_pid: int):
+        main_process = psutil.Process(main_pool_pid)
+        while not self.stopped:
+            try:
+                await asyncio.to_thread(main_process.status)
+                await asyncio.sleep(0.1)
+                continue
+            except (psutil.NoSuchProcess, ProcessLookupError, asyncio.CancelledError):
+                # main pool died
+                break
+
+        if not self.stopped:
+            await self.stop()
+
+    async def notify_main_pool_to_destroy(
+        self, message: DestroyActorMessage
+    ):  # pragma: no cover
+        await self.call(self._main_address, message)
+
+    async def notify_main_pool_to_create(self, message: CreateActorMessage):
+        reg_message = ControlMessage(
+            new_message_id(),
+            self.external_address,
+            ControlMessageType.add_sub_pool_actor,
+            (self.external_address, message.allocate_strategy, message),
+        )
+        await self.call(self._main_address, reg_message)
+
+    @implements(AbstractActorPool.create_actor)
+    async def create_actor(self, message: CreateActorMessage) -> ResultMessageType:
+        result = await super().create_actor(message)
+        if not message.from_main:
+            await self.notify_main_pool_to_create(message)
+        return result
+
+    @implements(AbstractActorPool.actor_ref)
+    async def actor_ref(self, message: ActorRefMessage) -> ResultMessageType:
+        result = await super().actor_ref(message)
+        if isinstance(result, ErrorMessage):
+            # need a new message id to call main actor
+            main_message = ActorRefMessage(
+                new_message_id(),
+                create_actor_ref(self._main_address, message.actor_ref.uid),
+            )
+            result = await self.call(self._main_address, main_message)
+            # rewrite to message_id of the original request
+            result.message_id = message.message_id
+        return result
+
+    @implements(AbstractActorPool.destroy_actor)
+    async def destroy_actor(self, message: DestroyActorMessage) -> ResultMessageType:
+        result = await super().destroy_actor(message)
+        if isinstance(result, ResultMessage) and not message.from_main:
+            # sync back to main actor pool
+            await self.notify_main_pool_to_destroy(message)
+        return result
+
+    @implements(AbstractActorPool.handle_control_command)
+    async def handle_control_command(
+        self, message: ControlMessage
+    ) -> ResultMessageType:
+        if message.control_message_type == ControlMessageType.sync_config:
+            self._main_address = message.address
+        return await super().handle_control_command(message)
+
+    @staticmethod
+    def _parse_config(config: dict, kw: dict) -> dict:
+        main_pool_pid = config.pop("main_pool_pid", None)
+        kw = AbstractActorPool._parse_config(config, kw)
+        pool_config: ActorPoolConfig = kw["config"]
+        main_process_index = pool_config.get_process_indexes()[0]
+        kw["main_address"] = pool_config.get_pool_config(main_process_index)[
+            "external_address"
+        ][0]
+        kw["main_pool_pid"] = main_pool_pid
+        return kw
+
+    async def stop(self):
+        await super().stop()
+        if self._watch_main_pool_task:
+            self._watch_main_pool_task.cancel()
+            await self._watch_main_pool_task
+
+
+class MainActorPoolBase(ActorPoolBase):
+    __slots__ = (
+        "_subprocess_start_method",
+        "_allocated_actors",
+        "sub_actor_pool_manager",
+        "_auto_recover",
+        "_monitor_task",
+        "_on_process_down",
+        "_on_process_recover",
+        "_recover_events",
+        "_allocation_lock",
+        "sub_processes",
+    )
+
+    def __init__(
+        self,
+        process_index: int,
+        label: str,
+        external_address: str,
+        internal_address: str,
+        env: dict,
+        router: Router,
+        config: ActorPoolConfig,
+        servers: list[Server],
+        subprocess_start_method: str | None = None,
+        auto_recover: str | bool = "actor",
+        on_process_down: Callable[[MainActorPoolType, str], None] | None = None,
+        on_process_recover: Callable[[MainActorPoolType, str], None] | None = None,
+    ):
+        super().__init__(
+            process_index,
+            label,
+            external_address,
+            internal_address,
+            env,
+            router,
+            config,
+            servers,
+        )
+        self._subprocess_start_method = subprocess_start_method
+
+        # auto recovering
+        self._auto_recover = auto_recover
+        self._monitor_task: Optional[asyncio.Task] = None
+        self._on_process_down = on_process_down
+        self._on_process_recover = on_process_recover
+        self._recover_events: dict[str, asyncio.Event] = dict()
+
+        # states
+        self._allocated_actors: allocated_type = {
+            addr: dict() for addr in self._config.get_external_addresses()
+        }
+        self._allocation_lock = threading.Lock()
+
+        self.sub_processes: dict[str, SubProcessHandle] = dict()
+
+    _process_index_gen = itertools.count()
+
+    @classmethod
+    def process_index_gen(cls, address):
+        # make sure different processes does not share process indexes
+        pid = os.getpid()
+        for idx in cls._process_index_gen:
+            yield pid << 16 + idx
+
+    @property
+    def _sub_processes(self):
+        return self.sub_processes
+
+    @implements(AbstractActorPool.create_actor)
+    async def create_actor(self, message: CreateActorMessage) -> ResultMessageType:
+        with _ErrorProcessor(
+            address=self.external_address,
+            message_id=message.message_id,
+            protocol=message.protocol,
+        ) as processor:
+            allocate_strategy = message.allocate_strategy
+            with self._allocation_lock:
+                # get allocated address according to corresponding strategy
+                address = allocate_strategy.get_allocated_address(
+                    self._config, self._allocated_actors
+                )
+                # set placeholder to make sure this label is occupied
+                self._allocated_actors[address][None] = (allocate_strategy, message)
+            if address == self.external_address:
+                # creating actor on main actor pool
+                result = await super().create_actor(message)
+                if isinstance(result, ResultMessage):
+                    self._allocated_actors[self.external_address][result.result] = (
+                        allocate_strategy,
+                        message,
+                    )
+                processor.result = result
+            else:
+                # creating actor on sub actor pool
+                # rewrite allocate strategy to AddressSpecified
+                new_allocate_strategy = AddressSpecified(address)
+                new_create_actor_message = CreateActorMessage(
+                    message.message_id,
+                    message.actor_cls,
+                    message.actor_id,
+                    message.args,
+                    message.kwargs,
+                    allocate_strategy=new_allocate_strategy,
+                    from_main=True,
+                    protocol=message.protocol,
+                    message_trace=message.message_trace,
+                )
+                result = await self.call(address, new_create_actor_message)
+                if isinstance(result, ResultMessage):
+                    self._allocated_actors[address][result.result] = (
+                        allocate_strategy,
+                        new_create_actor_message,
+                    )
+                processor.result = result
+
+            # revert placeholder
+            self._allocated_actors[address].pop(None, None)
+
+        return processor.result
+
+    @implements(AbstractActorPool.has_actor)
+    async def has_actor(self, message: HasActorMessage) -> ResultMessage:
+        actor_ref = message.actor_ref
+        # lookup allocated
+        for address, item in self._allocated_actors.items():
+            ref = create_actor_ref(address, to_binary(actor_ref.uid))
+            if ref in item:
+                return ResultMessage(
+                    message.message_id, True, protocol=message.protocol
+                )
+
+        return ResultMessage(message.message_id, False, protocol=message.protocol)
+
+    @implements(AbstractActorPool.destroy_actor)
+    async def destroy_actor(self, message: DestroyActorMessage) -> ResultMessageType:
+        actor_ref_message = ActorRefMessage(
+            message.message_id, message.actor_ref, protocol=message.protocol
+        )
+        result = await self.actor_ref(actor_ref_message)
+        if not isinstance(result, ResultMessage):
+            return result
+        real_actor_ref = result.result
+        if real_actor_ref.address == self.external_address:
+            result = await super().destroy_actor(message)
+            if result.message_type == MessageType.error:
+                return result
+            del self._allocated_actors[self.external_address][real_actor_ref]
+            return ResultMessage(
+                message.message_id, real_actor_ref.uid, protocol=message.protocol
+            )
+        # remove allocated actor ref
+        self._allocated_actors[real_actor_ref.address].pop(real_actor_ref, None)
+        new_destroy_message = DestroyActorMessage(
+            message.message_id,
+            real_actor_ref,
+            from_main=True,
+            protocol=message.protocol,
+        )
+        return await self.call(real_actor_ref.address, new_destroy_message)
+
+    @implements(AbstractActorPool.send)
+    async def send(self, message: SendMessage) -> ResultMessageType:
+        if message.actor_ref.uid in self._actors:
+            return await super().send(message)
+        actor_ref_message = ActorRefMessage(
+            message.message_id, message.actor_ref, protocol=message.protocol
+        )
+        result = await self.actor_ref(actor_ref_message)
+        if not isinstance(result, ResultMessage):
+            return result
+        actor_ref = result.result
+        new_send_message = SendMessage(
+            message.message_id,
+            actor_ref,
+            message.content,
+            protocol=message.protocol,
+            message_trace=message.message_trace,
+        )
+        return await self.call(actor_ref.address, new_send_message)
+
+    @implements(AbstractActorPool.tell)
+    async def tell(self, message: TellMessage) -> ResultMessageType:
+        if message.actor_ref.uid in self._actors:
+            return await super().tell(message)
+        actor_ref_message = ActorRefMessage(
+            message.message_id, message.actor_ref, protocol=message.protocol
+        )
+        result = await self.actor_ref(actor_ref_message)
+        if not isinstance(result, ResultMessage):
+            return result
+        actor_ref = result.result
+        new_tell_message = TellMessage(
+            message.message_id,
+            actor_ref,
+            message.content,
+            protocol=message.protocol,
+            message_trace=message.message_trace,
+        )
+        return await self.call(actor_ref.address, new_tell_message)
+
+    @implements(AbstractActorPool.actor_ref)
+    async def actor_ref(self, message: ActorRefMessage) -> ResultMessageType:
+        actor_ref = message.actor_ref
+        actor_ref.uid = to_binary(actor_ref.uid)
+        if actor_ref.address == self.external_address and actor_ref.uid in self._actors:
+            return ResultMessage(
+                message.message_id, actor_ref, protocol=message.protocol
+            )
+
+        # lookup allocated
+        for address, item in self._allocated_actors.items():
+            ref = create_actor_ref(address, actor_ref.uid)
+            if ref in item:
+                return ResultMessage(message.message_id, ref, protocol=message.protocol)
+
+        with _ErrorProcessor(
+            self.external_address, message.message_id, protocol=message.protocol
+        ) as processor:
+            raise ActorNotExist(
+                f"Actor {actor_ref.uid} does not exist in {actor_ref.address}"
+            )
+
+        return processor.result
+
+    @implements(AbstractActorPool.cancel)
+    async def cancel(self, message: CancelMessage) -> ResultMessageType:
+        if message.address == self.external_address:
+            # local message
+            return await super().cancel(message)
+        # redirect to sub pool
+        return await self.call(message.address, message)
+
+    @implements(AbstractActorPool.handle_control_command)
+    async def handle_control_command(
+        self, message: ControlMessage
+    ) -> ResultMessageType:
+        with _ErrorProcessor(
+            self.external_address, message.message_id, message.protocol
+        ) as processor:
+            if message.address == self.external_address:
+                if message.control_message_type == ControlMessageType.sync_config:
+                    # sync config, need to notify all sub pools
+                    tasks = []
+                    for addr in self.sub_processes:
+                        control_message = ControlMessage(
+                            new_message_id(),
+                            message.address,
+                            message.control_message_type,
+                            message.content,
+                            protocol=message.protocol,
+                            message_trace=message.message_trace,
+                        )
+                        tasks.append(
+                            asyncio.create_task(self.call(addr, control_message))
+                        )
+                    # call super
+                    task = asyncio.create_task(super().handle_control_command(message))
+                    tasks.append(task)
+                    await asyncio.gather(*tasks)
+                    processor.result = await task
+                else:
+                    processor.result = await super().handle_control_command(message)
+            elif message.control_message_type == ControlMessageType.stop:
+                timeout, force = (
+                    message.content if message.content is not None else (None, False)
+                )
+                await self.stop_sub_pool(
+                    message.address,
+                    self.sub_processes[message.address],
+                    timeout=timeout,
+                    force=force,
+                )
+                processor.result = ResultMessage(
+                    message.message_id, True, protocol=message.protocol
+                )
+            elif message.control_message_type == ControlMessageType.wait_pool_recovered:
+                if self._auto_recover and message.address not in self._recover_events:
+                    self._recover_events[message.address] = asyncio.Event()
+
+                event = self._recover_events.get(message.address, None)
+                if event is not None:
+                    await event.wait()
+                processor.result = ResultMessage(
+                    message.message_id, True, protocol=message.protocol
+                )
+            elif message.control_message_type == ControlMessageType.add_sub_pool_actor:
+                address, allocate_strategy, create_message = message.content
+                create_message.from_main = True
+                ref = create_actor_ref(address, to_binary(create_message.actor_id))
+                self._allocated_actors[address][ref] = (
+                    allocate_strategy,
+                    create_message,
+                )
+                processor.result = ResultMessage(
+                    message.message_id, True, protocol=message.protocol
+                )
+            else:
+                processor.result = await self.call(message.address, message)
+        return processor.result
+
+    @staticmethod
+    def _parse_config(config: dict, kw: dict) -> dict:
+        kw["subprocess_start_method"] = config.pop("start_method", None)
+        kw["auto_recover"] = config.pop("auto_recover", "actor")
+        kw["on_process_down"] = config.pop("on_process_down", None)
+        kw["on_process_recover"] = config.pop("on_process_recover", None)
+        kw = AbstractActorPool._parse_config(config, kw)
+        return kw
+
+    @classmethod
+    @implements(AbstractActorPool.create)
+    async def create(cls, config: dict) -> MainActorPoolType:
+        config = config.copy()
+        actor_pool_config: ActorPoolConfig = config.get("actor_pool_config")  # type: ignore
+        start_method = config.get("start_method", None)
+        if "process_index" not in config:
+            config["process_index"] = actor_pool_config.get_process_indexes()[0]
+        curr_process_index = config.get("process_index")
+        old_config_addresses = set(actor_pool_config.get_external_addresses())
+
+        tasks = []
+        subpool_process_idxes = []
+        # create sub actor pools
+        n_sub_pool = actor_pool_config.n_pool - 1
+        if n_sub_pool > 0:
+            process_indexes = actor_pool_config.get_process_indexes()
+            for process_index in process_indexes:
+                if process_index == curr_process_index:
+                    continue
+                create_pool_task = asyncio.create_task(
+                    cls.start_sub_pool(actor_pool_config, process_index, start_method)
+                )
+                await asyncio.sleep(0)
+                # await create_pool_task
+                tasks.append(create_pool_task)
+                subpool_process_idxes.append(process_index)
+
+        processes, ext_addresses = await cls.wait_sub_pools_ready(tasks)
+        if ext_addresses:
+            for process_index, ext_address in zip(subpool_process_idxes, ext_addresses):
+                actor_pool_config.reset_pool_external_address(
+                    process_index, ext_address
+                )
+
+        # create main actor pool
+        pool: MainActorPoolType = await super().create(config)
+        addresses = actor_pool_config.get_external_addresses()[1:]
+
+        assert len(addresses) == len(
+            processes
+        ), f"addresses {addresses}, processes {processes}"
+        for addr, proc in zip(addresses, processes):
+            pool.attach_sub_process(addr, proc)
+
+        new_config_addresses = set(actor_pool_config.get_external_addresses())
+        if old_config_addresses != new_config_addresses:
+            control_message = ControlMessage(
+                message_id=new_message_id(),
+                address=pool.external_address,
+                control_message_type=ControlMessageType.sync_config,
+                content=actor_pool_config,
+            )
+            await pool.handle_control_command(control_message)
+
+        return pool
+
+    async def start_monitor(self):
+        if self._monitor_task is None:
+            self._monitor_task = asyncio.create_task(self.monitor_sub_pools())
+        return self._monitor_task
+
+    @implements(AbstractActorPool.stop)
+    async def stop(self):
+        global_router = Router.get_instance()
+        if global_router is not None:
+            global_router.remove_router(self._router)
+
+        # turn off auto recover to avoid errors
+        self._auto_recover = False
+        self._stopped.set()
+        if self._monitor_task and not self._monitor_task.done():
+            await self._monitor_task
+            self._monitor_task = None
+        await self.stop_sub_pools()
+        await super().stop()
+
+    @classmethod
+    @abstractmethod
+    async def start_sub_pool(
+        cls,
+        actor_pool_config: ActorPoolConfig,
+        process_index: int,
+        start_method: str | None = None,
+    ):
+        """Start a sub actor pool"""
+
+    @classmethod
+    @abstractmethod
+    async def wait_sub_pools_ready(cls, create_pool_tasks: list[asyncio.Task]):
+        """Wait all sub pools ready"""
+
+    def attach_sub_process(self, external_address: str, process: SubProcessHandle):
+        self.sub_processes[external_address] = process
+
+    async def stop_sub_pools(self):
+        to_stop_processes: dict[str, SubProcessHandle] = dict()  # type: ignore
+        for address, process in self.sub_processes.items():
+            if not await self.is_sub_pool_alive(process):
+                continue
+            to_stop_processes[address] = process
+
+        tasks = []
+        for address, process in to_stop_processes.items():
+            tasks.append(self.stop_sub_pool(address, process))
+        await asyncio.gather(*tasks)
+
+    async def stop_sub_pool(
+        self,
+        address: str,
+        process: SubProcessHandle,
+        timeout: float | None = None,
+        force: bool = False,
+    ):
+        if force:
+            await self.kill_sub_pool(process, force=True)
+            return
+
+        stop_message = ControlMessage(
+            new_message_id(),
+            address,
+            ControlMessageType.stop,
+            None,
+            protocol=DEFAULT_PROTOCOL,
+        )
+        try:
+            if timeout is None:
+                message = await self.call(address, stop_message)
+                if isinstance(message, ErrorMessage):  # pragma: no cover
+                    raise message.as_instanceof_cause()
+            else:
+                call = asyncio.create_task(self.call(address, stop_message))
+                try:
+                    await asyncio.wait_for(call, timeout)
+                except (futures.TimeoutError, asyncio.TimeoutError):  # pragma: no cover
+                    # timeout, just let kill to finish it
+                    force = True
+        except (ConnectionError, ServerClosed):  # pragma: no cover
+            # process dead maybe, ignore it
+            pass
+        # kill process
+        await self.kill_sub_pool(process, force=force)
+
+    @abstractmethod
+    async def kill_sub_pool(self, process: SubProcessHandle, force: bool = False):
+        """Kill a sub actor pool"""
+
+    @abstractmethod
+    async def is_sub_pool_alive(self, process: SubProcessHandle):
+        """
+        Check whether sub pool process is alive
+        Parameters
+        ----------
+        process : SubProcessHandle
+            sub pool process handle
+        Returns
+        -------
+        bool
+        """
+
+    @abstractmethod
+    def recover_sub_pool(self, address):
+        """Recover a sub actor pool"""
+
+    def process_sub_pool_lost(self, address: str):
+        if self._auto_recover in (False, "process"):
+            # process down, when not auto_recover
+            # or only recover process, remove all created actors
+            self._allocated_actors[address] = dict()
+
+    async def monitor_sub_pools(self):
+        try:
+            while not self._stopped.is_set():
+                # Copy sub_processes to avoid changes during recover.
+                for address, process in list(self.sub_processes.items()):
+                    try:
+                        recover_events_discovered = address in self._recover_events
+                        if not await self.is_sub_pool_alive(
+                            process
+                        ):  # pragma: no cover
+                            if self._on_process_down is not None:
+                                self._on_process_down(self, address)
+                            self.process_sub_pool_lost(address)
+                            if self._auto_recover:
+                                await self.recover_sub_pool(address)
+                                if self._on_process_recover is not None:
+                                    self._on_process_recover(self, address)
+                        if recover_events_discovered:
+                            event = self._recover_events.pop(address)
+                            event.set()
+                    except asyncio.CancelledError:
+                        raise
+                    except RuntimeError as ex:  # pragma: no cover
+                        if "cannot schedule new futures" not in str(ex):
+                            # to silence log when process exit, otherwise it
+                            # will raise "RuntimeError: cannot schedule new futures
+                            # after interpreter shutdown".
+                            logger.exception("Monitor sub pool %s failed", address)
+                    except Exception:
+                        # log the exception instead of stop monitoring the
+                        # sub pool silently.
+                        logger.exception("Monitor sub pool %s failed", address)
+
+                # check every half second
+                await asyncio.sleep(0.5)
+        except asyncio.CancelledError:  # pragma: no cover
+            # cancelled
+            return
+
+    @classmethod
+    @abstractmethod
+    def get_external_addresses(
+        cls,
+        address: str,
+        n_process: int | None = None,
+        ports: list[int] | None = None,
+        schemes: list[Optional[str]] | None = None,
+    ):
+        """Returns external addresses for n pool processes"""
+
+    @classmethod
+    @abstractmethod
+    def gen_internal_address(
+        cls, process_index: int, external_address: str | None = None
+    ) -> str | None:
+        """Returns internal address for pool of specified process index"""
+
+
+async def create_actor_pool(
+    address: str,
+    *,
+    pool_cls: Type[MainActorPoolType] | None = None,
+    n_process: int | None = None,
+    labels: list[str] | None = None,
+    ports: list[int] | None = None,
+    envs: list[dict] | None = None,
+    external_address_schemes: list[Optional[str]] | None = None,
+    enable_internal_addresses: list[bool] | None = None,
+    subprocess_start_method: str | None = None,
+    auto_recover: str | bool = "actor",
+    modules: list[str] | None = None,
+    suspend_sigint: bool | None = None,
+    use_uvloop: str | bool = "auto",
+    logging_conf: dict | None = None,
+    on_process_down: Callable[[MainActorPoolType, str], None] | None = None,
+    on_process_recover: Callable[[MainActorPoolType, str], None] | None = None,
+    extra_conf: dict | None = None,
+    **kwargs,
+) -> MainActorPoolType:
+    if n_process is None:
+        n_process = multiprocessing.cpu_count()
+    if labels and len(labels) != n_process + 1:
+        raise ValueError(
+            f"`labels` should be of size {n_process + 1}, got {len(labels)}"
+        )
+    if envs and len(envs) != n_process:
+        raise ValueError(f"`envs` should be of size {n_process}, got {len(envs)}")
+    if external_address_schemes and len(external_address_schemes) != n_process + 1:
+        raise ValueError(
+            f"`external_address_schemes` should be of size {n_process + 1}, "
+            f"got {len(external_address_schemes)}"
+        )
+    if enable_internal_addresses and len(enable_internal_addresses) != n_process + 1:
+        raise ValueError(
+            f"`enable_internal_addresses` should be of size {n_process + 1}, "
+            f"got {len(enable_internal_addresses)}"
+        )
+    elif not enable_internal_addresses:
+        enable_internal_addresses = [True] * (n_process + 1)
+    if auto_recover is True:
+        auto_recover = "actor"
+    if auto_recover not in ("actor", "process", False):
+        raise ValueError(
+            f'`auto_recover` should be one of "actor", "process", '
+            f"True or False, got {auto_recover}"
+        )
+    if use_uvloop == "auto":
+        try:
+            import uvloop  # noqa: F401 # pylint: disable=unused-variable
+
+            use_uvloop = True
+        except ImportError:
+            use_uvloop = False
+
+    assert pool_cls is not None
+    external_addresses = pool_cls.get_external_addresses(
+        address, n_process=n_process, ports=ports, schemes=external_address_schemes
+    )
+    actor_pool_config = ActorPoolConfig()
+    actor_pool_config.add_metric_configs(kwargs.get("metrics", {}))
+    # add main config
+    process_index_gen = pool_cls.process_index_gen(address)
+    main_process_index = next(process_index_gen)
+    main_internal_address = (
+        pool_cls.gen_internal_address(main_process_index, external_addresses[0])
+        if enable_internal_addresses[0]
+        else None
+    )
+    actor_pool_config.add_pool_conf(
+        main_process_index,
+        labels[0] if labels else None,
+        main_internal_address,
+        external_addresses[0],
+        modules=modules,
+        suspend_sigint=suspend_sigint,
+        use_uvloop=use_uvloop,  # type: ignore
+        logging_conf=logging_conf,
+        kwargs=kwargs,
+    )
+    # add sub configs
+    for i in range(n_process):
+        sub_process_index = next(process_index_gen)
+        internal_address = (
+            pool_cls.gen_internal_address(sub_process_index, external_addresses[i + 1])
+            if enable_internal_addresses[i + 1]
+            else None
+        )
+        actor_pool_config.add_pool_conf(
+            sub_process_index,
+            labels[i + 1] if labels else None,
+            internal_address,
+            external_addresses[i + 1],
+            env=envs[i] if envs else None,
+            modules=modules,
+            suspend_sigint=suspend_sigint,
+            use_uvloop=use_uvloop,  # type: ignore
+            logging_conf=logging_conf,
+            kwargs=kwargs,
+        )
+    actor_pool_config.add_comm_config(extra_conf)
+
+    pool: MainActorPoolType = await pool_cls.create(
+        {
+            "actor_pool_config": actor_pool_config,
+            "process_index": main_process_index,
+            "start_method": subprocess_start_method,
+            "auto_recover": auto_recover,
+            "on_process_down": on_process_down,
+            "on_process_recover": on_process_recover,
+        }
+    )
+    await pool.start()
+    return pool