Source code for distributed.core

import asyncio
from collections import defaultdict, deque
from functools import partial
from inspect import isawaitable
import logging
import threading
import traceback
import uuid
import weakref

import dask
import tblib
from toolz import merge
from tornado import gen
from tornado.ioloop import IOLoop

from .comm import (
    connect,
    listen,
    CommClosedError,
    normalize_address,
    unparse_host_port,
    get_address_host_port,
)
from .metrics import time
from . import profile
from .system_monitor import SystemMonitor
from .utils import (
    is_coroutine_function,
    get_traceback,
    truncate_exception,
    ignoring,
    shutting_down,
    PeriodicCallback,
    parse_timedelta,
    has_keyword,
    CancelledError,
)
from . import protocol


class RPCClosed(IOError):
    pass


logger = logging.getLogger(__name__)


def raise_later(exc):
    def _raise(*args, **kwargs):
        raise exc

    return _raise


tick_maximum_delay = parse_timedelta(
    dask.config.get("distributed.admin.tick.limit"), default="ms"
)

LOG_PDB = dask.config.get("distributed.admin.pdb-on-err")


[docs]class Server: """ Dask Distributed Server Superclass for endpoints in a distributed cluster, such as Worker and Scheduler objects. **Handlers** Servers define operations with a ``handlers`` dict mapping operation names to functions. The first argument of a handler function will be a ``Comm`` for the communication established with the client. Other arguments will receive inputs from the keys of the incoming message which will always be a dictionary. >>> def pingpong(comm): ... return b'pong' >>> def add(comm, x, y): ... return x + y >>> handlers = {'ping': pingpong, 'add': add} >>> server = Server(handlers) # doctest: +SKIP >>> server.listen('tcp://0.0.0.0:8000') # doctest: +SKIP **Message Format** The server expects messages to be dictionaries with a special key, `'op'` that corresponds to the name of the operation, and other key-value pairs as required by the function. So in the example above the following would be good messages. * ``{'op': 'ping'}`` * ``{'op': 'add', 'x': 10, 'y': 20}`` """ default_ip = "" default_port = 0 def __init__( self, handlers, blocked_handlers=None, stream_handlers=None, connection_limit=512, deserialize=True, io_loop=None, ): self.handlers = { "identity": self.identity, "connection_stream": self.handle_stream, } self.handlers.update(handlers) if blocked_handlers is None: blocked_handlers = dask.config.get( "distributed.%s.blocked-handlers" % type(self).__name__.lower(), [] ) self.blocked_handlers = blocked_handlers self.stream_handlers = {} self.stream_handlers.update(stream_handlers or {}) self.id = type(self).__name__ + "-" + str(uuid.uuid4()) self._address = None self._listen_address = None self._port = None self._comms = {} self.deserialize = deserialize self.monitor = SystemMonitor() self.counters = None self.digests = None self.events = None self.event_counts = None self._ongoing_coroutines = weakref.WeakSet() self._event_finished = asyncio.Event() self.listeners = [] self.io_loop = io_loop or IOLoop.current() self.loop = self.io_loop if not hasattr(self.io_loop, "profile"): ref = weakref.ref(self.io_loop) if hasattr(self.io_loop, "asyncio_loop"): def stop(): loop = ref() return loop is None or loop.asyncio_loop.is_closed() else: def stop(): loop = ref() return loop is None or loop._closing self.io_loop.profile = profile.watch( omit=("profile.py", "selectors.py"), interval=dask.config.get("distributed.worker.profile.interval"), cycle=dask.config.get("distributed.worker.profile.cycle"), stop=stop, ) # Statistics counters for various events with ignoring(ImportError): from .counter import Digest self.digests = defaultdict(partial(Digest, loop=self.io_loop)) from .counter import Counter self.counters = defaultdict(partial(Counter, loop=self.io_loop)) self.events = defaultdict(lambda: deque(maxlen=10000)) self.event_counts = defaultdict(lambda: 0) self.periodic_callbacks = dict() pc = PeriodicCallback(self.monitor.update, 500, io_loop=self.io_loop) self.periodic_callbacks["monitor"] = pc self._last_tick = time() pc = PeriodicCallback( self._measure_tick, parse_timedelta( dask.config.get("distributed.admin.tick.interval"), default="ms" ) * 1000, io_loop=self.io_loop, ) self.periodic_callbacks["tick"] = pc self.thread_id = 0 def set_thread_ident(): self.thread_id = threading.get_ident() self.io_loop.add_callback(set_thread_ident) self.__stopped = False async def finished(self): """ Wait until the server has finished """ await self._event_finished.wait() def start_periodic_callbacks(self): """ Start Periodic Callbacks consistently This starts all PeriodicCallbacks stored in self.periodic_callbacks if they are not yet running. It does this safely on the IOLoop. """ self._last_tick = time() def start_pcs(): for pc in self.periodic_callbacks.values(): if not pc.is_running(): pc.start() self.io_loop.add_callback(start_pcs) def stop(self): if not self.__stopped: self.__stopped = True for listener in self.listeners: # Delay closing the server socket until the next IO loop tick. # Otherwise race conditions can appear if an event handler # for an accept() call is already scheduled by the IO loop, # raising EBADF. # The demonstrator for this is Worker.terminate(), which # closes the server socket in response to an incoming message. # See https://github.com/tornadoweb/tornado/issues/2069 self.io_loop.add_callback(listener.stop) @property def listener(self): if self.listeners: return self.listeners[0] else: return None def _measure_tick(self): now = time() diff = now - self._last_tick self._last_tick = now if diff > tick_maximum_delay: logger.info( "Event loop was unresponsive in %s for %.2fs. " "This is often caused by long-running GIL-holding " "functions or moving large chunks of data. " "This can cause timeouts and instability.", type(self).__name__, diff, ) if self.digests is not None: self.digests["tick-duration"].add(diff) def log_event(self, name, msg): msg["time"] = time() if isinstance(name, list): for n in name: self.events[n].append(msg) self.event_counts[n] += 1 else: self.events[name].append(msg) self.event_counts[name] += 1 @property def address(self): """ The address this Server can be contacted on. """ if not self._address: if self.listener is None: raise ValueError("cannot get address of non-running Server") self._address = self.listener.contact_address return self._address @property def listen_address(self): """ The address this Server is listening on. This may be a wildcard address such as `tcp://0.0.0.0:1234`. """ if not self._listen_address: if self.listener is None: raise ValueError("cannot get listen address of non-running Server") self._listen_address = self.listener.listen_address return self._listen_address @property def port(self): """ The port number this Server is listening on. This will raise ValueError if the Server is listening on a non-IP based protocol. """ if not self._port: _, self._port = get_address_host_port(self.address) return self._port def identity(self, comm=None): return {"type": type(self).__name__, "id": self.id} async def listen(self, port_or_addr=None, listen_args=None): if port_or_addr is None: port_or_addr = self.default_port if isinstance(port_or_addr, int): addr = unparse_host_port(self.default_ip, port_or_addr) elif isinstance(port_or_addr, tuple): addr = unparse_host_port(*port_or_addr) else: addr = port_or_addr assert isinstance(addr, str) listener = listen( addr, self.handle_comm, deserialize=self.deserialize, connection_args=listen_args, ) await listener.start() self.listeners.append(listener) async def handle_comm(self, comm, shutting_down=shutting_down): """ Dispatch new communications to coroutine-handlers Handlers is a dictionary mapping operation names to functions or coroutines. {'get_data': get_data, 'ping': pingpong} Coroutines should expect a single Comm object. """ if self.__stopped: comm.abort() return address = comm.peer_address op = None logger.debug("Connection from %r to %s", address, type(self).__name__) self._comms[comm] = op try: while True: try: msg = await comm.read() logger.debug("Message from %r: %s", address, msg) except EnvironmentError as e: if not shutting_down(): logger.debug( "Lost connection to %r while reading message: %s." " Last operation: %s", address, e, op, ) break except Exception as e: logger.exception(e) await comm.write(error_message(e, status="uncaught-error")) continue if not isinstance(msg, dict): raise TypeError( "Bad message type. Expected dict, got\n " + str(msg) ) try: op = msg.pop("op") except KeyError: raise ValueError( "Received unexpected message without 'op' key: " + str(msg) ) if self.counters is not None: self.counters["op"].add(op) self._comms[comm] = op serializers = msg.pop("serializers", None) close_desired = msg.pop("close", False) reply = msg.pop("reply", True) if op == "close": if reply: await comm.write("OK") break result = None try: if op in self.blocked_handlers: _msg = ( "The '{op}' handler has been explicitly disallowed " "in {obj}, possibly due to security concerns." ) exc = ValueError(_msg.format(op=op, obj=type(self).__name__)) handler = raise_later(exc) else: handler = self.handlers[op] except KeyError: logger.warning( "No handler %s found in %s", op, type(self).__name__, exc_info=True, ) else: if serializers is not None and has_keyword(handler, "serializers"): msg["serializers"] = serializers # add back in logger.debug("Calling into handler %s", handler.__name__) try: result = handler(comm, **msg) if isawaitable(result): result = asyncio.ensure_future(result) self._ongoing_coroutines.add(result) result = await result except (CommClosedError, CancelledError) as e: if self.status == "running": logger.info("Lost connection to %r: %s", address, e) break except Exception as e: logger.exception(e) result = error_message(e, status="uncaught-error") if reply and result != "dont-reply": try: await comm.write(result, serializers=serializers) except (EnvironmentError, TypeError) as e: logger.debug( "Lost connection to %r while sending result for op %r: %s", address, op, e, ) break msg = result = None if close_desired: await comm.close() if comm.closed(): break finally: del self._comms[comm] if not shutting_down() and not comm.closed(): try: comm.abort() except Exception as e: logger.error( "Failed while closing connection to %r: %s", address, e ) async def handle_stream(self, comm, extra=None, every_cycle=[]): extra = extra or {} logger.info("Starting established connection") io_error = None closed = False try: while not closed: msgs = await comm.read() if not isinstance(msgs, (tuple, list)): msgs = (msgs,) if not comm.closed(): for msg in msgs: if msg == "OK": # from close break op = msg.pop("op") if op: if op == "close-stream": closed = True break handler = self.stream_handlers[op] if is_coroutine_function(handler): self.loop.add_callback(handler, **merge(extra, msg)) else: handler(**merge(extra, msg)) else: logger.error("odd message %s", msg) await asyncio.sleep(0) for func in every_cycle: func() except (CommClosedError, EnvironmentError) as e: io_error = e except Exception as e: logger.exception(e) if LOG_PDB: import pdb pdb.set_trace() raise finally: await comm.close() assert comm.closed() @gen.coroutine def close(self): for pc in self.periodic_callbacks.values(): pc.stop() for listener in self.listeners: self.listener.stop() for i in range(20): # let comms close naturally for a second if not self._comms: break else: yield asyncio.sleep(0.05) yield [comm.close() for comm in self._comms] # then forcefully close for cb in self._ongoing_coroutines: cb.cancel() for i in range(10): if all(cb.cancelled() for c in self._ongoing_coroutines): break else: yield asyncio.sleep(0.01) self._event_finished.set()
def pingpong(comm): return b"pong" async def send_recv(comm, reply=True, serializers=None, deserializers=None, **kwargs): """ Send and recv with a Comm. Keyword arguments turn into the message response = yield send_recv(comm, op='ping', reply=True) """ msg = kwargs msg["reply"] = reply please_close = kwargs.get("close") force_close = False if deserializers is None: deserializers = serializers if deserializers is not None: msg["serializers"] = deserializers try: await comm.write(msg, serializers=serializers, on_error="raise") if reply: response = await comm.read(deserializers=deserializers) else: response = None except EnvironmentError: # On communication errors, we should simply close the communication force_close = True raise finally: if please_close: await comm.close() elif force_close: comm.abort() if isinstance(response, dict) and response.get("status") == "uncaught-error": if comm.deserialize: typ, exc, tb = clean_exception(**response) raise exc.with_traceback(tb) else: raise Exception(response["text"]) return response def addr_from_args(addr=None, ip=None, port=None): if addr is None: addr = (ip, port) else: assert ip is None and port is None if isinstance(addr, tuple): addr = unparse_host_port(*addr) return normalize_address(addr)
[docs]class rpc: """ Conveniently interact with a remote server >>> remote = rpc(address) # doctest: +SKIP >>> response = yield remote.add(x=10, y=20) # doctest: +SKIP One rpc object can be reused for several interactions. Additionally, this object creates and destroys many comms as necessary and so is safe to use in multiple overlapping communications. When done, close comms explicitly. >>> remote.close_comms() # doctest: +SKIP """ active = weakref.WeakSet() comms = () address = None def __init__( self, arg=None, comm=None, deserialize=True, timeout=None, connection_args=None, serializers=None, deserializers=None, ): self.comms = {} self.address = coerce_to_address(arg) self.timeout = timeout self.status = "running" self.deserialize = deserialize self.serializers = serializers self.deserializers = deserializers if deserializers is not None else serializers self.connection_args = connection_args self._created = weakref.WeakSet() rpc.active.add(self) async def live_comm(self): """ Get an open communication Some comms to the ip/port target may be in current use by other coroutines. We track this with the `comms` dict :: {comm: True/False if open and ready for use} This function produces an open communication, either by taking one that we've already made or making a new one if they are all taken. This also removes comms that have been closed. When the caller is done with the stream they should set self.comms[comm] = True As is done in __getattr__ below. """ if self.status == "closed": raise RPCClosed("RPC Closed") to_clear = set() open = False for comm, open in self.comms.items(): if comm.closed(): to_clear.add(comm) if open: break for s in to_clear: del self.comms[s] if not open or comm.closed(): comm = await connect( self.address, self.timeout, deserialize=self.deserialize, connection_args=self.connection_args, ) comm.name = "rpc" self.comms[comm] = False # mark as taken return comm def close_comms(self): async def _close_comm(comm): # Make sure we tell the peer to close try: if not comm.closed(): await comm.write({"op": "close", "reply": False}) await comm.close() except EnvironmentError: comm.abort() tasks = [] for comm in list(self.comms): if comm and not comm.closed(): # IOLoop.current().add_callback(_close_comm, comm) task = asyncio.ensure_future(_close_comm(comm)) tasks.append(task) for comm in list(self._created): if comm and not comm.closed(): # IOLoop.current().add_callback(_close_comm, comm) task = asyncio.ensure_future(_close_comm(comm)) tasks.append(task) self.comms.clear() return tasks def __getattr__(self, key): async def send_recv_from_rpc(**kwargs): if self.serializers is not None and kwargs.get("serializers") is None: kwargs["serializers"] = self.serializers if self.deserializers is not None and kwargs.get("deserializers") is None: kwargs["deserializers"] = self.deserializers try: comm = await self.live_comm() comm.name = "rpc." + key result = await send_recv(comm=comm, op=key, **kwargs) except (RPCClosed, CommClosedError) as e: raise e.__class__( "%s: while trying to call remote method %r" % (e, key) ) self.comms[comm] = True # mark as open return result return send_recv_from_rpc def close_rpc(self): if self.status != "closed": rpc.active.discard(self) self.status = "closed" return asyncio.gather(*self.close_comms()) def __enter__(self): return self def __exit__(self, *args): asyncio.ensure_future(self.close_rpc()) async def __aenter__(self): return self async def __aexit__(self, *args): await self.close_rpc() def __del__(self): if self.status != "closed": rpc.active.discard(self) self.status = "closed" still_open = [comm for comm in self.comms if not comm.closed()] if still_open: logger.warning( "rpc object %s deleted with %d open comms", self, len(still_open) ) for comm in still_open: comm.abort() def __repr__(self): return "<rpc to %r, %d comms>" % (self.address, len(self.comms))
class PooledRPCCall: """ The result of ConnectionPool()('host:port') See Also: ConnectionPool """ def __init__(self, addr, pool, serializers=None, deserializers=None): self.addr = addr self.pool = pool self.serializers = serializers self.deserializers = deserializers if deserializers is not None else serializers @property def address(self): return self.addr def __getattr__(self, key): async def send_recv_from_rpc(**kwargs): if self.serializers is not None and kwargs.get("serializers") is None: kwargs["serializers"] = self.serializers if self.deserializers is not None and kwargs.get("deserializers") is None: kwargs["deserializers"] = self.deserializers comm = await self.pool.connect(self.addr) name, comm.name = comm.name, "ConnectionPool." + key try: result = await send_recv(comm=comm, op=key, **kwargs) finally: self.pool.reuse(self.addr, comm) comm.name = name return result return send_recv_from_rpc async def close_rpc(self): pass # For compatibility with rpc() def __enter__(self): return self def __exit__(self, *args): pass def __repr__(self): return "<pooled rpc to %r>" % (self.addr,) class ConnectionPool: """ A maximum sized pool of Comm objects. This provides a connect method that mirrors the normal distributed.connect method, but provides connection sharing and tracks connection limits. This object provides an ``rpc`` like interface:: >>> rpc = ConnectionPool(limit=512) >>> scheduler = rpc('127.0.0.1:8786') >>> workers = [rpc(address) for address ...] >>> info = yield scheduler.identity() It creates enough comms to satisfy concurrent connections to any particular address:: >>> a, b = yield [scheduler.who_has(), scheduler.has_what()] It reuses existing comms so that we don't have to continuously reconnect. It also maintains a comm limit to avoid "too many open file handle" issues. Whenever this maximum is reached we clear out all idling comms. If that doesn't do the trick then we wait until one of the occupied comms closes. Parameters ---------- limit: int The number of open comms to maintain at once deserialize: bool Whether or not to deserialize data by default or pass it through """ _instances = weakref.WeakSet() def __init__( self, limit=512, deserialize=True, serializers=None, deserializers=None, connection_args=None, timeout=None, server=None, ): self.limit = limit # Max number of open comms # Invariant: len(available) == open - active self.available = defaultdict(set) # Invariant: len(occupied) == active self.occupied = defaultdict(set) self.deserialize = deserialize self.serializers = serializers self.deserializers = deserializers if deserializers is not None else serializers self.connection_args = connection_args self.timeout = timeout self._n_connecting = 0 # Invariant: semaphore._value == limit - open - _n_connecting self.semaphore = asyncio.Semaphore(self.limit) self.server = weakref.ref(server) if server else None self._created = weakref.WeakSet() self._instances.add(self) def _validate(self): """ Validate important invariants of this class Used only for testing / debugging """ assert self.semaphore._value == self.limit - self.open - self._n_connecting @property def active(self): return sum(map(len, self.occupied.values())) @property def open(self): return self.active + sum(map(len, self.available.values())) def __repr__(self): return "<ConnectionPool: open=%d, active=%d, connecting=%d>" % ( self.open, self.active, self._n_connecting, ) def __call__(self, addr=None, ip=None, port=None): """ Cached rpc objects """ addr = addr_from_args(addr=addr, ip=ip, port=port) return PooledRPCCall( addr, self, serializers=self.serializers, deserializers=self.deserializers ) async def connect(self, addr, timeout=None): """ Get a Comm to the given address. For internal use. """ available = self.available[addr] occupied = self.occupied[addr] while available: comm = available.pop() if comm.closed(): self.semaphore.release() else: occupied.add(comm) return comm if self.semaphore.locked(): self.collect() self._n_connecting += 1 await self.semaphore.acquire() try: comm = await connect( addr, timeout=timeout or self.timeout, deserialize=self.deserialize, connection_args=self.connection_args, ) comm.name = "ConnectionPool" comm._pool = weakref.ref(self) self._created.add(comm) except Exception: self.semaphore.release() raise finally: self._n_connecting -= 1 occupied.add(comm) return comm def reuse(self, addr, comm): """ Reuse an open communication to the given address. For internal use. """ # if the pool is asked to re-use a comm it does not know about, ignore # this comm: just close it. if comm not in self.occupied[addr]: IOLoop.current().add_callback(comm.close) else: self.occupied[addr].remove(comm) if comm.closed(): self.semaphore.release() else: self.available[addr].add(comm) if self.semaphore.locked() and self._n_connecting > 0: self.collect() def collect(self): """ Collect open but unused communications, to allow opening other ones. """ logger.info( "Collecting unused comms. open: %d, active: %d, connecting: %d", self.open, self.active, self._n_connecting, ) for addr, comms in self.available.items(): for comm in comms: IOLoop.current().add_callback(comm.close) self.semaphore.release() comms.clear() def remove(self, addr): """ Remove all Comms to a given address. """ logger.info("Removing comms to %s", addr) if addr in self.available: comms = self.available.pop(addr) for comm in comms: IOLoop.current().add_callback(comm.close) self.semaphore.release() if addr in self.occupied: comms = self.occupied.pop(addr) for comm in comms: IOLoop.current().add_callback(comm.close) self.semaphore.release() async def close(self): """ Close all communications """ for d in [self.available, self.occupied]: comms = [comm for comms in d.values() for comm in comms] await asyncio.gather( *[comm.close() for comm in comms], return_exceptions=True ) for _ in comms: self.semaphore.release() for comm in self._created: IOLoop.current().add_callback(comm.abort) def coerce_to_address(o): if isinstance(o, (list, tuple)): o = unparse_host_port(*o) return normalize_address(o) def collect_causes(e): causes = [] while e.__cause__ is not None: causes.append(e.__cause__) e = e.__cause__ return causes def error_message(e, status="error"): """ Produce message to send back given an exception has occurred This does the following: 1. Gets the traceback 2. Truncates the exception and the traceback 3. Serializes the exception and traceback or 4. If they can't be serialized send string versions 5. Format a message and return See Also -------- clean_exception: deserialize and unpack message into exception/traceback """ MAX_ERROR_LEN = dask.config.get("distributed.admin.max-error-length") tblib.pickling_support.install(e, *collect_causes(e)) tb = get_traceback() e2 = truncate_exception(e, MAX_ERROR_LEN) try: e3 = protocol.pickle.dumps(e2) protocol.pickle.loads(e3) except Exception: e2 = Exception(str(e2)) e4 = protocol.to_serialize(e2) try: tb2 = protocol.pickle.dumps(tb) except Exception: tb = tb2 = "".join(traceback.format_tb(tb)) if len(tb2) > MAX_ERROR_LEN: tb_result = None else: tb_result = protocol.to_serialize(tb) return {"status": status, "exception": e4, "traceback": tb_result, "text": str(e2)} def clean_exception(exception, traceback, **kwargs): """ Reraise exception and traceback. Deserialize if necessary See Also -------- error_message: create and serialize errors into message """ if isinstance(exception, bytes) or isinstance(exception, bytearray): try: exception = protocol.pickle.loads(exception) except Exception: exception = Exception(exception) elif isinstance(exception, str): exception = Exception(exception) if isinstance(traceback, bytes): try: traceback = protocol.pickle.loads(traceback) except (TypeError, AttributeError): traceback = None elif isinstance(traceback, str): traceback = None # happens if the traceback failed serializing return type(exception), exception, traceback