Source code for distributed.nanny

from __future__ import annotations

import asyncio
import contextlib
import errno
import functools
import logging
import multiprocessing
import os
import shutil
import threading
import uuid
import warnings
import weakref
from collections.abc import Callable, Collection
from inspect import isawaitable
from queue import Empty
from typing import ClassVar, Literal, cast

from toolz import merge
from tornado.ioloop import IOLoop

import dask
from dask.system import CPU_COUNT
from dask.utils import parse_timedelta

from distributed import preloading
from distributed._async_taskgroup import AsyncTaskGroupClosedError
from distributed.comm import get_address_host
from distributed.comm.addressing import address_from_user_args
from distributed.compatibility import asyncio_run
from distributed.config import get_loop_factory
from distributed.core import (
    CommClosedError,
    ErrorMessage,
    OKMessage,
    RPCClosed,
    Status,
    coerce_to_address,
    error_message,
)
from distributed.diagnostics.plugin import NannyPlugin, _get_plugin_name
from distributed.metrics import time
from distributed.node import ServerNode
from distributed.process import AsyncProcess
from distributed.proctitle import enable_proctitle_on_children
from distributed.protocol import pickle
from distributed.protocol.serialize import _is_dumpable
from distributed.security import Security
from distributed.utils import (
    get_ip,
    get_mp_context,
    json_load_robust,
    log_errors,
    parse_ports,
    silence_logging_cmgr,
    wait_for,
)
from distributed.worker import Worker, run
from distributed.worker_memory import (
    DeprecatedMemoryManagerAttribute,
    DeprecatedMemoryMonitor,
    NannyMemoryManager,
)

logger = logging.getLogger(__name__)


[docs]class Nanny(ServerNode): """A process to manage worker processes The nanny spins up Worker processes, watches them, and kills or restarts them as necessary. It is necessary if you want to use the ``Client.restart`` method, or to restart the worker automatically if it gets to the terminate fraction of its memory limit. The parameters for the Nanny are mostly the same as those for the Worker with exceptions listed below. Parameters ---------- env: dict, optional Environment variables set at time of Nanny initialization will be ensured to be set in the Worker process as well. This argument allows to overwrite or otherwise set environment variables for the Worker. It is also possible to set environment variables using the option ``distributed.nanny.environ``. Precedence as follows 1. Nanny arguments 2. Existing environment variables 3. Dask configuration .. note:: Some environment variables, like ``OMP_NUM_THREADS``, must be set before importing numpy to have effect. Others, like ``MALLOC_TRIM_THRESHOLD_`` (see :ref:`memtrim`), must be set before starting the Linux process. Such variables would be ineffective if set here or in ``distributed.nanny.environ``; they must be set in ``distributed.nanny.pre-spawn-environ`` so that they are set before spawning the subprocess, even if this means poisoning the process running the Nanny. For the same reason, be warned that changing ``distributed.worker.multiprocessing-method`` from ``spawn`` to ``fork`` or ``forkserver`` may inhibit some environment variables; if you do, you should set the variables yourself in the shell before you start ``dask-worker``. See Also -------- Worker """ _instances: ClassVar[weakref.WeakSet[Nanny]] = weakref.WeakSet() process: WorkerProcess | None memory_manager: NannyMemoryManager env: dict[str, str] pre_spawn_env: dict[str, str] # Inputs to parse_ports() _given_worker_port: int | str | Collection[int] | None _start_port: int | str | Collection[int] | None def __init__( # type: ignore[no-untyped-def] self, scheduler_ip=None, scheduler_port=None, scheduler_file=None, worker_port: int | str | Collection[int] | None = 0, nthreads=None, loop=None, local_directory=None, services=None, name=None, memory_limit="auto", reconnect=True, validate=False, quiet=False, resources=None, silence_logs=None, death_timeout=None, preload=None, preload_argv=None, preload_nanny=None, preload_nanny_argv=None, security=None, contact_address=None, listen_address=None, worker_class=None, env=None, interface=None, host=None, port: int | str | Collection[int] | None = None, protocol=None, config=None, **worker_kwargs, ): if loop is not None: warnings.warn( "the `loop` kwarg to `Nanny` is ignored, and will be removed in a future release. " "The Nanny always binds to the current loop.", DeprecationWarning, stacklevel=2, ) self.__exit_stack = stack = contextlib.ExitStack() self.process = None self._setup_logging(logger) self.loop = self.io_loop = IOLoop.current() if isinstance(security, dict): security = Security(**security) self.security = security or Security() assert isinstance(self.security, Security) self.connection_args = self.security.get_connection_args("worker") self.preload = preload if self.preload is None: self.preload = dask.config.get("distributed.worker.preload") self.preload_argv = preload_argv if self.preload_argv is None: self.preload_argv = dask.config.get("distributed.worker.preload-argv") if preload_nanny is None: preload_nanny = dask.config.get("distributed.nanny.preload") if preload_nanny_argv is None: preload_nanny_argv = dask.config.get("distributed.nanny.preload-argv") handlers = { "instantiate": self.instantiate, "kill": self.kill, "restart": self.restart, "get_logs": self.get_logs, # cannot call it 'close' on the rpc side for naming conflict "terminate": self.close, "close_gracefully": self.close_gracefully, "run": self.run, "plugin_add": self.plugin_add, "plugin_remove": self.plugin_remove, } super().__init__( handlers=handlers, connection_args=self.connection_args, local_directory=local_directory, needs_workdir=False, ) self.preloads = preloading.process_preloads( self, preload_nanny, preload_nanny_argv, file_dir=self.local_directory ) self.death_timeout = parse_timedelta(death_timeout) if scheduler_file: cfg = json_load_robust(scheduler_file, timeout=self.death_timeout) self.scheduler_addr = cfg["address"] elif scheduler_ip is None and dask.config.get("scheduler-address"): self.scheduler_addr = dask.config.get("scheduler-address") elif scheduler_port is None: self.scheduler_addr = coerce_to_address(scheduler_ip) else: self.scheduler_addr = coerce_to_address((scheduler_ip, scheduler_port)) if protocol is None: protocol_address = self.scheduler_addr.split("://") if len(protocol_address) == 2: protocol = protocol_address[0] self._given_worker_port = worker_port self.nthreads: int = nthreads or CPU_COUNT self.reconnect = reconnect self.validate = validate self.resources = resources self.Worker = Worker if worker_class is None else worker_class self.pre_spawn_env = _get_env_variables("distributed.nanny.pre-spawn-environ") # To get consistent hashing on subprocesses, we need to set a consistent seed for # the Python hash algorithm; xref https://github.com/dask/distributed/pull/8400 if self.pre_spawn_env.get("PYTHONHASHSEED") in (None, "0"): # This number is arbitrary; it was chosen to commemorate # https://github.com/dask/dask/issues/6640. self.pre_spawn_env.update({"PYTHONHASHSEED": "6640"}) self.env = merge( self.pre_spawn_env, _get_env_variables("distributed.nanny.environ"), {k: str(v) for k, v in env.items()} if env else {}, ) self.config = merge(dask.config.config, config or {}) worker_kwargs.update( { "port": worker_port, "interface": interface, "protocol": protocol, "host": host, } ) self.worker_kwargs = worker_kwargs self.contact_address = contact_address self.services = services self.name = name self.quiet = quiet if silence_logs: stack.enter_context(silence_logging_cmgr(level=silence_logs)) self.silence_logs = silence_logs self.plugins: dict[str, NannyPlugin] = {} self.scheduler = self.rpc(self.scheduler_addr) self.memory_manager = NannyMemoryManager(self, memory_limit=memory_limit) if ( not host and not interface and not self.scheduler_addr.startswith("inproc://") ): host = get_ip(get_address_host(self.scheduler.address)) self._start_port = port self._start_host = host self._interface = interface self._protocol = protocol self._listen_address = listen_address Nanny._instances.add(self) # Deprecated attributes; use Nanny.memory_manager.<name> instead memory_limit = DeprecatedMemoryManagerAttribute() memory_terminate_fraction = DeprecatedMemoryManagerAttribute() memory_monitor = DeprecatedMemoryMonitor() def __repr__(self): return "<Nanny: %s, threads: %d>" % (self.worker_address, self.nthreads) async def _unregister(self, timeout=10): if self.process is None: return worker_address = self.process.worker_address if worker_address is None: return try: await wait_for( self.scheduler.unregister( address=self.worker_address, stimulus_id=f"nanny-close-{time()}" ), timeout, ) except (asyncio.TimeoutError, CommClosedError, OSError, RPCClosed): pass @property def worker_address(self): return None if self.process is None else self.process.worker_address @property def worker_dir(self): return None if self.process is None else self.process.worker_dir
[docs] async def start_unsafe(self): """Start nanny, start local process, start watching""" await super().start_unsafe() ports = parse_ports(self._start_port) for port in ports: start_address = address_from_user_args( host=self._start_host, port=port, interface=self._interface, protocol=self._protocol, security=self.security, ) try: await self.listen( start_address, **self.security.get_listen_args("worker") ) except OSError as e: if len(ports) > 1 and e.errno == errno.EADDRINUSE: continue else: raise else: self._start_address = start_address break else: raise ValueError( f"Could not start Nanny on host {self._start_host} " f"with port {self._start_port}" ) self.ip = get_address_host(self.address) await self.preloads.start() saddr = self.scheduler.addr comm = await self.rpc.connect(saddr) comm.name = "Nanny->Scheduler (registration)" try: await comm.write({"op": "register_nanny", "address": self.address}) msg = await comm.read() try: for name, plugin in msg["nanny-plugins"].items(): await self.plugin_add(plugin=plugin, name=name) logger.info(" Start Nanny at: %r", self.address) response = await self.instantiate() if response != Status.running: raise RuntimeError("Nanny failed to start worker process") except Exception: try: await comm.write({"status": "error"}) # If self.instantiate() failed, the comm will already be closed. except CommClosedError: pass await self.close(reason="nanny-start-failed") raise else: await comm.write({"status": "ok"}) finally: await comm.close() assert self.worker_address self.start_periodic_callbacks() return self
[docs] async def kill(self, timeout: float = 5, reason: str = "nanny-kill") -> None: """Kill the local worker process Blocks until both the process is down and the scheduler is properly informed """ if self.process is not None: await self.process.kill(reason=reason, timeout=timeout)
[docs] async def instantiate(self) -> Status: """Start a local worker process Blocks until the process is up and the scheduler is properly informed """ if self.process is None: worker_kwargs = dict( scheduler_ip=self.scheduler_addr, nthreads=self.nthreads, local_directory=self._original_local_dir, services=self.services, nanny=self.address, name=self.name, memory_limit=self.memory_manager.memory_limit, resources=self.resources, validate=self.validate, silence_logs=self.silence_logs, death_timeout=self.death_timeout, preload=self.preload, preload_argv=self.preload_argv, security=self.security, contact_address=self.contact_address, ) worker_kwargs.update(self.worker_kwargs) self.process = WorkerProcess( worker_kwargs=worker_kwargs, silence_logs=self.silence_logs, on_exit=self._on_worker_exit_sync, worker=self.Worker, env=self.env, pre_spawn_env=self.pre_spawn_env, config=self.config, ) if self.death_timeout: try: result = await wait_for(self.process.start(), self.death_timeout) except asyncio.TimeoutError: logger.error( "Timed out connecting Nanny '%s' to scheduler '%s'", self, self.scheduler_addr, ) await self.close( timeout=self.death_timeout, reason="nanny-instantiate-timeout" ) raise else: try: result = await self.process.start() except Exception: logger.error("Failed to start process", exc_info=True) await self.close(reason="nanny-instantiate-failed") raise return result
@log_errors async def plugin_add( self, plugin: NannyPlugin | bytes, name: str | None = None ) -> ErrorMessage | OKMessage: if isinstance(plugin, bytes): plugin = pickle.loads(plugin) if not isinstance(plugin, NannyPlugin): warnings.warn( "Registering duck-typed plugins has been deprecated. " "Please make sure your plugin inherits from `NannyPlugin`.", DeprecationWarning, stacklevel=2, ) plugin = cast(NannyPlugin, plugin) if name is None: name = _get_plugin_name(plugin) assert name self.plugins[name] = plugin logger.info("Starting Nanny plugin %s", name) if hasattr(plugin, "setup"): try: result = plugin.setup(nanny=self) if isawaitable(result): result = await result except Exception as e: logger.exception("Nanny plugin %s failed to setup", name) return error_message(e) if getattr(plugin, "restart", False): await self.restart(reason=f"nanny-plugin-{name}-restart") return {"status": "OK"} @log_errors async def plugin_remove(self, name: str) -> ErrorMessage | OKMessage: logger.info(f"Removing Nanny plugin {name}") try: plugin = self.plugins.pop(name) if hasattr(plugin, "teardown"): result = plugin.teardown(nanny=self) if isawaitable(result): result = await result except Exception as e: logger.exception("Nanny plugin %s failed to teardown", name) msg = error_message(e) return msg return {"status": "OK"} async def restart( self, timeout: float = 30, reason: str = "nanny-restart" ) -> Literal["OK", "timed out"] | ErrorMessage: async def _(): if self.process is not None: await self.kill(reason=reason) await self.instantiate() try: await wait_for(asyncio.shield(_()), timeout) except asyncio.TimeoutError: logger.error( f"Restart timed out after {timeout}s; returning before finished" ) return "timed out" except Exception as e: return error_message(e) else: return "OK" def is_alive(self): return self.process is not None and self.process.is_alive() def run(self, comm, *args, **kwargs): return run(self, comm, *args, **kwargs) def _on_worker_exit_sync(self, exitcode): try: self._ongoing_background_tasks.call_soon(self._on_worker_exit, exitcode) except ( AsyncTaskGroupClosedError ): # Async task group has already been closed, so the nanny is already clos(ed|ing). pass @log_errors async def _on_worker_exit(self, exitcode): if self.status not in ( Status.init, Status.closing, Status.closed, Status.closing_gracefully, Status.failed, ): try: await self._unregister() except OSError: logger.exception("Failed to unregister") if not self.reconnect: await self.close(reason="nanny-unregister-failed") return try: if self.status not in ( Status.closing, Status.closed, Status.closing_gracefully, Status.failed, ): logger.warning("Restarting worker") await self.instantiate() elif self.status == Status.closing_gracefully: await self.close(reason="nanny-close-gracefully") except Exception: logger.error( "Failed to restart worker after its process exited", exc_info=True ) @property def pid(self): return self.process and self.process.pid def _close(self, *args, **kwargs): warnings.warn("Worker._close has moved to Worker.close", stacklevel=2) return self.close(*args, **kwargs)
[docs] def close_gracefully(self, reason: str = "nanny-close-gracefully") -> None: """ A signal that we shouldn't try to restart workers if they go away This is used as part of the cluster shutdown process. """ self.status = Status.closing_gracefully logger.info( "Closing Nanny gracefully at %r. Reason: %s", self.address_safe, reason )
[docs] async def close( # type:ignore[override] self, timeout: float = 5, reason: str = "nanny-close" ) -> Literal["OK"]: """ Close the worker process, stop all comms. """ if self.status == Status.closing: await self.finished() assert self.status == Status.closed if self.status == Status.closed: return "OK" self.status = Status.closing logger.info("Closing Nanny at %r. Reason: %s", self.address_safe, reason) await self.preloads.teardown() await asyncio.gather(*(self.plugin_remove(name) for name in self.plugins)) self.stop() if self.process is not None: await self.kill(timeout=timeout, reason=reason) self.process = None await self.rpc.close() self.status = Status.closed await super().close() self.__exit_stack.__exit__(None, None, None) return "OK"
async def _log_event(self, topic, msg): await self.scheduler.log_event( topic=topic, msg=msg, )
[docs] def log_event(self, topic, msg): """Log an event under a given topic Parameters ---------- topic : str, list[str] Name of the topic under which to log an event. To log the same event under multiple topics, pass a list of topic names. msg Event message to log. Note this must be msgpack serializable. See also -------- Client.log_event """ if not _is_dumpable(msg): raise TypeError( f"Message must be msgpack serializable. Got {type(msg)=} instead." ) self._ongoing_background_tasks.call_soon(self._log_event, topic, msg)
class WorkerProcess: running: asyncio.Event stopped: asyncio.Event process: AsyncProcess | None env: dict[str, str] pre_spawn_env: dict[str, str] # The interval how often to check the msg queue for init _init_msg_interval = 0.05 def __init__( self, worker_kwargs, silence_logs, on_exit, worker, env, pre_spawn_env, config, ): self.status = Status.init self.silence_logs = silence_logs self.worker_kwargs = worker_kwargs self.on_exit = on_exit self.process = None self.Worker = worker self.env = env self.pre_spawn_env = pre_spawn_env self.config = config.copy() # Ensure default clients don't propagate to subprocesses try: from distributed.client import default_client default_client() self.config.pop("scheduler", None) self.config.pop("shuffle", None) except ValueError: pass # Initialized when worker is ready self.worker_dir = None self.worker_address = None async def start(self) -> Status: """ Ensure the worker process is started. """ enable_proctitle_on_children() if self.status == Status.running: return self.status if self.status == Status.starting: await self.running.wait() return self.status mp_ctx = get_mp_context() self.init_result_q = mp_ctx.Queue() self.child_stop_q = mp_ctx.Queue() # put an empty message to start the background self._child_stop_q._thread # otherwise calling .put({"op": "stop"}) later in an atexit handler # results in a RuntimeError starting the thread # get the message before passing child_stop_q to AsyncProcess to avoid # dealing with it in watch_stop_q self.child_stop_q.put(None) self.child_stop_q.get() uid = uuid.uuid4().hex self.process = AsyncProcess( target=functools.partial( self._run, silence_logs=self.silence_logs, init_result_q=self.init_result_q, child_stop_q=self.child_stop_q, uid=uid, worker_factory=functools.partial(self.Worker, **self.worker_kwargs), env=self.env, config=self.config, ), name="Dask Worker process (from Nanny)", kwargs=dict(), ) self.process.daemon = dask.config.get("distributed.worker.daemon", default=True) self.process.set_exit_callback(self._on_exit) self.running = asyncio.Event() self.stopped = asyncio.Event() self.status = Status.starting # Set selected environment variables before spawning the subprocess. # See note in Nanny docstring. os.environ.update(self.pre_spawn_env) try: try: await self.process.start() except OSError: # This can only happen if the actual process creation failed, e.g. # multiprocessing.Process.start failed. This is not tested! logger.exception("Nanny failed to start process", exc_info=True) # NOTE: doesn't wait for process to terminate, just for terminate signal to be sent await self.process.terminate() self.status = Status.failed try: msg = await self._wait_until_connected(uid) except Exception: # NOTE: doesn't wait for process to terminate, just for terminate signal to be sent await self.process.terminate() self.status = Status.failed raise finally: self.running.set() if not msg: return self.status self.worker_address = msg["address"] self.worker_dir = msg["dir"] assert self.worker_address self.status = Status.running return self.status def _on_exit(self, proc): if proc is not self.process: # Ignore exit of old process instance return self.mark_stopped() def _death_message(self, pid, exitcode): assert exitcode is not None if exitcode == 255: return "Worker process %d was killed by unknown signal" % (pid,) elif exitcode >= 0: return "Worker process %d exited with status %d" % (pid, exitcode) else: return "Worker process %d was killed by signal %d" % (pid, -exitcode) def is_alive(self): return self.process is not None and self.process.is_alive() @property def pid(self): return self.process.pid if self.process and self.process.is_alive() else None def mark_stopped(self): if self.status != Status.stopped: assert self.process is not None r = self.process.exitcode assert r is not None if r != 0: msg = self._death_message(self.process.pid, r) logger.info(msg) self.status = Status.stopped self.running.clear() self.stopped.set() # Release resources self.process.close() self.init_result_q = None self.child_stop_q = None self.process = None # Best effort to clean up worker directory if self.worker_dir and os.path.exists(self.worker_dir): shutil.rmtree(self.worker_dir, ignore_errors=True) self.worker_dir = None # User hook if self.on_exit is not None: self.on_exit(r) async def kill( self, timeout: float = 5, executor_wait: bool = True, reason: str = "workerprocess-kill", ) -> None: """ Ensure the worker process is stopped, waiting at most ``timeout * 0.8`` seconds before killing it abruptly. When `kill` returns, the worker process has been joined. If the worker process does not terminate within ``timeout`` seconds, even after being killed, `asyncio.TimeoutError` is raised. """ deadline = time() + timeout # If the process is not properly up it will not watch the closing queue # and we may end up leaking this process # Therefore wait for it to be properly started before killing it if self.status == Status.starting: await self.running.wait() assert self.status in ( Status.stopping, Status.stopped, Status.running, Status.failed, # process failed to start, but hasn't been joined yet Status.closing_gracefully, ), self.status if self.status == Status.stopped: return if self.status == Status.stopping: await self.stopped.wait() return self.status = Status.stopping logger.info("Nanny asking worker to close. Reason: %s", reason) process = self.process wait_timeout = timeout * 0.8 self.child_stop_q.put( { "op": "stop", "timeout": wait_timeout, "executor_wait": executor_wait, "reason": reason, } ) self.child_stop_q.close() assert process is not None try: try: await process.join(wait_timeout) return except asyncio.TimeoutError: pass logger.warning( f"Worker process still alive after {wait_timeout:.1f} seconds, killing" ) await process.kill() await process.join(max(0, deadline - time())) except ValueError as e: if "invalid operation on closed AsyncProcess" in str(e): return raise async def _wait_until_connected(self, uid): while True: if self.status != Status.starting: return # This is a multiprocessing queue and we'd block the event loop if # we simply called get try: msg = self.init_result_q.get_nowait() except Empty: await asyncio.sleep(self._init_msg_interval) continue if msg["uid"] != uid: # ensure that we didn't cross queues continue if "exception" in msg: raise msg["exception"] else: return msg @classmethod def _run( cls, silence_logs: bool, init_result_q: multiprocessing.Queue, child_stop_q: multiprocessing.Queue, uid: str, env: dict, config: dict, worker_factory: Callable[[], Worker], ) -> None: # pragma: no cover async def do_stop( *, worker: Worker, timeout: float = 5, executor_wait: bool = True, reason: str = "workerprocess-stop", ) -> None: await worker.close( nanny=False, executor_wait=executor_wait, timeout=timeout, reason=reason, ) def watch_stop_q(loop: IOLoop, worker: Worker) -> None: """ Wait for an incoming stop message and then stop the worker cleanly. """ try: msg = child_stop_q.get() except (TypeError, OSError, EOFError): logger.error("Worker process died unexpectedly") msg = {"op": "stop"} finally: child_stop_q.close() assert msg["op"] == "stop", msg del msg["op"] loop.add_callback(do_stop, worker=worker, **msg) async def run() -> None: """ Try to start worker and inform parent of outcome. """ failure_type: str | None = "initialize" try: worker = worker_factory() failure_type = "start" thread = threading.Thread( target=functools.partial( watch_stop_q, worker=worker, loop=IOLoop.current(), ), name="Nanny stop queue watch", daemon=True, ) thread.start() stack.callback(thread.join, timeout=2) async with worker: failure_type = None try: assert worker.address except ValueError: pass else: init_result_q.put( { "address": worker.address, "dir": worker.local_directory, "uid": uid, } ) init_result_q.close() await worker.finished() logger.info("Worker closed") except Exception as e: if failure_type is None: raise logger.exception(f"Failed to {failure_type} worker") init_result_q.put({"uid": uid, "exception": e}) init_result_q.close() finally: init_result_q.join_thread() with contextlib.ExitStack() as stack: @stack.callback def close_stop_q() -> None: try: child_stop_q.put({"op": "stop"}) # usually redundant except ValueError: pass try: child_stop_q.close() # usually redundant except ValueError: pass child_stop_q.join_thread() os.environ.update(env) dask.config.refresh() dask.config.set(config) from dask.multiprocessing import default_initializer default_initializer() if silence_logs: logger.setLevel(silence_logs) asyncio_run(run(), loop_factory=get_loop_factory()) def _get_env_variables(config_key: str) -> dict[str, str]: cfg = dask.config.get(config_key) if not isinstance(cfg, dict): raise TypeError( # pragma: nocover f"{config_key} configuration must be of type dict. Instead got {type(cfg)}" ) # Override dask config with explicitly defined env variables from the OS # Allow unsetting a variable in a config override by setting its value to None. cfg = {k: os.environ.get(k, str(v)) for k, v in cfg.items() if v is not None} return cfg