from __future__ import annotations import asyncio import contextlib import dataclasses import heapq import inspect import itertools import json import logging import math import operator import os import pickle import random import sys import uuid import warnings import weakref from collections import defaultdict, deque from collections.abc import ( Callable, Collection, Container, Hashable, Iterable, Iterator, Mapping, Sequence, Set, ) from contextlib import suppress from functools import partial from numbers import Number from typing import TYPE_CHECKING, Any, ClassVar, Literal, NamedTuple, cast, overload import psutil from sortedcontainers import SortedDict, SortedSet from tlz import ( first, groupby, merge, merge_sorted, merge_with, partition, pluck, second, valmap, ) from tornado.ioloop import IOLoop import dask from dask.highlevelgraph import HighLevelGraph from dask.utils import ( format_bytes, format_time, key_split, parse_bytes, parse_timedelta, tmpfile, ) from dask.widgets import get_template from distributed import cluster_dump, preloading, profile from distributed import versions as version_module from distributed._stories import scheduler_story from distributed.active_memory_manager import ActiveMemoryManagerExtension, RetireWorker from distributed.batched import BatchedSend from distributed.collections import HeapSet from distributed.comm import ( Comm, CommClosedError, get_address_host, normalize_address, resolve_address, unparse_host_port, ) from distributed.comm.addressing import addresses_from_user_args from distributed.compatibility import PeriodicCallback from distributed.core import Status, clean_exception, rpc, send_recv from distributed.diagnostics.memory_sampler import MemorySamplerExtension from distributed.diagnostics.plugin import SchedulerPlugin, _get_plugin_name from distributed.event import EventExtension from distributed.http import get_handlers from distributed.lock import LockExtension from distributed.metrics import monotonic, time from distributed.multi_lock import MultiLockExtension from distributed.node import ServerNode from distributed.proctitle import setproctitle from distributed.protocol.pickle import dumps, loads from distributed.protocol.serialize import Serialized, serialize from distributed.publish import PublishExtension from distributed.pubsub import PubSubSchedulerExtension from distributed.queues import QueueExtension from distributed.recreate_tasks import ReplayTaskScheduler from distributed.security import Security from distributed.semaphore import SemaphoreExtension from distributed.shuffle import ShuffleSchedulerExtension from distributed.stealing import WorkStealing from distributed.utils import ( All, TimeoutError, empty_context, get_fileno_limit, key_split_group, log_errors, no_default, recursive_to_dict, validate_key, ) from distributed.utils_comm import ( gather_from_workers, retry_operation, scatter_to_workers, ) from distributed.utils_perf import disable_gc_diagnosis, enable_gc_diagnosis from distributed.variable import VariableExtension if TYPE_CHECKING: # TODO import from typing (requires Python >=3.10) from typing_extensions import TypeAlias # Not to be confused with distributed.worker_state_machine.TaskStateState TaskStateState: TypeAlias = Literal[ "released", "waiting", "no-worker", "queued", "processing", "memory", "erred", "forgotten", ] ALL_TASK_STATES: Set[TaskStateState] = set(TaskStateState.__args__) # type: ignore # TODO remove quotes (requires Python >=3.9) # {task key -> finish state} # Not to be confused with distributed.worker_state_machine.Recs Recs: TypeAlias = "dict[str, TaskStateState]" # {client or worker address: [{op: , ...}, ...]} Msgs: TypeAlias = "dict[str, list[dict[str, Any]]]" # (recommendations, client messages, worker messages) RecsMsgs: TypeAlias = "tuple[Recs, Msgs, Msgs]" logger = logging.getLogger(__name__) LOG_PDB = dask.config.get("distributed.admin.pdb-on-err") DEFAULT_DATA_SIZE = parse_bytes( dask.config.get("distributed.scheduler.default-data-size") ) STIMULUS_ID_UNSET = "" DEFAULT_EXTENSIONS = { "locks": LockExtension, "multi_locks": MultiLockExtension, "publish": PublishExtension, "replay-tasks": ReplayTaskScheduler, "queues": QueueExtension, "variables": VariableExtension, "pubsub": PubSubSchedulerExtension, "semaphores": SemaphoreExtension, "events": EventExtension, "amm": ActiveMemoryManagerExtension, "memory_sampler": MemorySamplerExtension, "shuffle": ShuffleSchedulerExtension, "stealing": WorkStealing, } class ClientState: """A simple object holding information about a client.""" #: A unique identifier for this client. This is generally an opaque #: string generated by the client itself. client_key: str #: Cached hash of :attr:`~ClientState.client_key` _hash: int #: A set of tasks this client wants to be kept in memory, so that it can download #: its result when desired. This is the reverse mapping of #: :class:`TaskState.who_wants`. Tasks are typically removed from this set when the #: corresponding object in the client's space (for example a ``Future`` or a Dask #: collection) gets garbage-collected. wants_what: set[TaskState] #: The last time we received a heartbeat from this client, in local scheduler time. last_seen: float #: Output of :func:`distributed.versions.get_versions` on the client versions: dict[str, Any] __slots__ = tuple(__annotations__) def __init__(self, client: str, *, versions: dict[str, Any] | None = None): self.client_key = client self._hash = hash(client) self.wants_what = set() self.last_seen = time() self.versions = versions or {} def __hash__(self) -> int: return self._hash def __eq__(self, other: object) -> bool: if not isinstance(other, ClientState): return False return self.client_key == other.client_key def __repr__(self) -> str: return f"" def __str__(self) -> str: return self.client_key def _to_dict_no_nest(self, *, exclude: Container[str] = ()) -> dict: """Dictionary representation for debugging purposes. Not type stable and not intended for roundtrips. See also -------- Client.dump_cluster_state distributed.utils.recursive_to_dict TaskState._to_dict """ return recursive_to_dict( self, exclude=set(exclude) | {"versions"}, # type: ignore members=True, ) class MemoryState: """Memory readings on a worker or on the whole cluster. See :doc:`worker-memory`. Attributes / properties: managed_total Sum of the output of sizeof() for all dask keys held by the worker in memory, plus number of bytes spilled to disk managed Sum of the output of sizeof() for the dask keys held in RAM. Note that this may be inaccurate, which may cause inaccurate unmanaged memory (see below). spilled Number of bytes for the dask keys spilled to the hard drive. Note that this is the size on disk; size in memory may be different due to compression and inaccuracies in sizeof(). In other words, given the same keys, 'managed' will change depending on the keys being in memory or spilled. process Total RSS memory measured by the OS on the worker process. This is always exactly equal to managed + unmanaged. unmanaged process - managed. This is the sum of - Python interpreter and modules - global variables - memory temporarily allocated by the dask tasks that are currently running - memory fragmentation - memory leaks - memory not yet garbage collected - memory not yet free()'d by the Python memory manager to the OS unmanaged_old Minimum of the 'unmanaged' measures over the last ``distributed.memory.recent-to-old-time`` seconds unmanaged_recent unmanaged - unmanaged_old; in other words process memory that has been recently allocated but is not accounted for by dask; hopefully it's mostly a temporary spike. optimistic managed + unmanaged_old; in other words the memory held long-term by the process under the hopeful assumption that all unmanaged_recent memory is a temporary spike """ process: int unmanaged_old: int managed: int spilled: int __slots__ = tuple(__annotations__) def __init__( self, *, process: int, unmanaged_old: int, managed: int, spilled: int, ): # Some data arrives with the heartbeat, some other arrives in realtime as the # tasks progress. Also, sizeof() is not guaranteed to return correct results. # This can cause glitches where a partial measure is larger than the whole, so # we need to force all numbers to add up exactly by definition. self.process = process self.managed = min(self.process, managed) self.spilled = spilled # Subtractions between unsigned ints guaranteed by construction to be >= 0 self.unmanaged_old = min(unmanaged_old, process - self.managed) @staticmethod def sum(*infos: MemoryState) -> MemoryState: process = 0 unmanaged_old = 0 managed = 0 spilled = 0 for ms in infos: process += ms.process unmanaged_old += ms.unmanaged_old spilled += ms.spilled managed += ms.managed return MemoryState( process=process, unmanaged_old=unmanaged_old, managed=managed, spilled=spilled, ) @property def managed_total(self) -> int: return self.managed + self.spilled @property def unmanaged(self) -> int: # This is never negative thanks to __init__ return self.process - self.managed @property def unmanaged_recent(self) -> int: # This is never negative thanks to __init__ return self.process - self.managed - self.unmanaged_old @property def optimistic(self) -> int: return self.managed + self.unmanaged_old @property def managed_in_memory(self) -> int: warnings.warn("managed_in_memory has been renamed to managed", FutureWarning) return self.managed @property def managed_spilled(self) -> int: warnings.warn("managed_spilled has been renamed to spilled", FutureWarning) return self.spilled def __repr__(self) -> str: return ( f"Process memory (RSS) : {format_bytes(self.process)}\n" f" - managed by Dask : {format_bytes(self.managed)}\n" f" - unmanaged (old) : {format_bytes(self.unmanaged_old)}\n" f" - unmanaged (recent): {format_bytes(self.unmanaged_recent)}\n" f"Spilled to disk : {format_bytes(self.spilled)}\n" ) def _to_dict(self, *, exclude: Container[str] = ()) -> dict: """Dictionary representation for debugging purposes. See also -------- Client.dump_cluster_state distributed.utils.recursive_to_dict """ return { k: getattr(self, k) for k in dir(self) if not k.startswith("_") and k not in {"sum", "managed_in_memory", "managed_spilled"} } class WorkerState: """A simple object holding information about a worker. Not to be confused with :class:`distributed.worker_state_machine.WorkerState`. """ #: This worker's unique key. This can be its connected address #: (such as ``"tcp://127.0.0.1:8891"``) or an alias (such as ``"alice"``). address: str pid: int name: Hashable #: The number of CPU threads made available on this worker nthreads: int #: Memory available to the worker, in bytes memory_limit: int local_directory: str services: dict[str, int] #: Output of :meth:`distributed.versions.get_versions` on the worker versions: dict[str, Any] #: Address of the associated :class:`~distributed.nanny.Nanny`, if present nanny: str #: Read-only worker status, synced one way from the remote Worker object status: Status #: Cached hash of :attr:`~WorkerState.address` _hash: int #: The total memory size, in bytes, used by the tasks this worker holds in memory #: (i.e. the tasks in this worker's :attr:`~WorkerState.has_what`). nbytes: int #: Worker memory unknown to the worker, in bytes, which has been there for more than #: 30 seconds. See :class:`MemoryState`. _memory_unmanaged_old: int #: History of the last 30 seconds' worth of unmanaged memory. Used to differentiate #: between "old" and "new" unmanaged memory. #: Format: ``[(timestamp, bytes), (timestamp, bytes), ...]`` _memory_unmanaged_history: deque[tuple[float, int]] metrics: dict[str, Any] #: The last time we received a heartbeat from this worker, in local scheduler time. last_seen: float time_delay: float bandwidth: float #: A set of all TaskStates on this worker that are actors. This only includes those #: actors whose state actually lives on this worker, not actors to which this worker #: has a reference. actors: set[TaskState] #: Underlying data of :meth:`WorkerState.has_what` _has_what: dict[TaskState, None] #: A set of tasks that have been submitted to this worker. Multiple tasks may be # submitted to a worker in advance and the worker will run them eventually, # depending on its execution resources (but see :doc:`work-stealing`). #: #: All the tasks here are in the "processing" state. #: This attribute is kept in sync with :attr:`TaskState.processing_on`. processing: set[TaskState] #: Running tasks that invoked :func:`distributed.secede` long_running: set[TaskState] #: A dictionary of tasks that are currently being run on this worker. #: Each task state is associated with the duration in seconds which the task has #: been running. executing: dict[TaskState, float] #: The available resources on this worker, e.g. ``{"GPU": 2}``. #: These are abstract quantities that constrain certain tasks from running at the #: same time on this worker. resources: dict[str, float] #: The sum of each resource used by all tasks allocated to this worker. #: The numbers in this dictionary can only be less or equal than those in this #: worker's :attr:`~WorkerState.resources`. used_resources: dict[str, float] #: Arbitrary additional metadata to be added to :meth:`~WorkerState.identity` extra: dict[str, Any] # The unique server ID this WorkerState is referencing server_id: str # Reference to scheduler task_groups scheduler_ref: weakref.ref[SchedulerState] | None task_prefix_count: defaultdict[str, int] _network_occ: float _occupancy_cache: float | None #: Keys that may need to be fetched to this worker, and the number of tasks that need them. #: All tasks are currently in `memory` on a worker other than this one. #: Much like `processing`, this does not exactly reflect worker state: #: keys here may be queued to fetch, in flight, or already in memory #: on the worker. needs_what: dict[TaskState, int] __slots__ = tuple(__annotations__) def __init__( self, *, address: str, status: Status, pid: int, name: object, nthreads: int = 0, memory_limit: int, local_directory: str, nanny: str, server_id: str, services: dict[str, int] | None = None, versions: dict[str, Any] | None = None, extra: dict[str, Any] | None = None, scheduler: SchedulerState | None = None, ): self.server_id = server_id self.address = address self.pid = pid self.name = name self.nthreads = nthreads self.memory_limit = memory_limit self.local_directory = local_directory self.services = services or {} self.versions = versions or {} self.nanny = nanny self.status = status self._hash = hash(self.server_id) self.nbytes = 0 self._memory_unmanaged_old = 0 self._memory_unmanaged_history = deque() self.metrics = {} self.last_seen = 0 self.time_delay = 0 self.bandwidth = parse_bytes(dask.config.get("distributed.scheduler.bandwidth")) self.actors = set() self._has_what = {} self.processing = set() self.long_running = set() self.executing = {} self.resources = {} self.used_resources = {} self.extra = extra or {} self.scheduler_ref = weakref.ref(scheduler) if scheduler else None self.task_prefix_count = defaultdict(int) self.needs_what = {} self._network_occ = 0 self._occupancy_cache = None def __hash__(self) -> int: return self._hash def __eq__(self, other: object) -> bool: return isinstance(other, WorkerState) and other.server_id == self.server_id @property def has_what(self) -> Set[TaskState]: """An insertion-sorted set-like of tasks which currently reside on this worker. All the tasks here are in the "memory" state. This is the reverse mapping of :attr:`TaskState.who_has`. This is a read-only public accessor. The data is implemented as a dict without values, because rebalance() relies on dicts being insertion-sorted. """ return self._has_what.keys() @property def host(self) -> str: return get_address_host(self.address) @property def memory(self) -> MemoryState: """Polished memory metrics for the worker. **Design note on managed memory** There are two measures available for managed memory: - ``self.nbytes`` - ``self.metrics["managed_bytes"]`` At rest, the two numbers must be identical. However, ``self.nbytes`` is immediately updated through the batched comms as soon as each task lands in memory on the worker; ``self.metrics["managed_bytes"]`` instead is updated by the heartbeat, which can lag several seconds behind. Below we are mixing likely newer managed memory info from ``self.nbytes`` with process and spilled memory from the heartbeat. This is deliberate, so that managed memory total is updated more frequently. Managed memory directly and immediately contributes to optimistic memory, which is in turn used in Active Memory Manager heuristics (at the moment of writing; more uses will likely be added in the future). So it's important to have it up to date; much more than it is for process memory. Having up-to-date managed memory info as soon as the scheduler learns about task completion also substantially simplifies unit tests. The flip side of this design is that it may cause some noise in the unmanaged_recent measure. e.g.: 1. Delete 100MB of managed data 2. The updated managed memory reaches the scheduler faster than the updated process memory 3. There's a blip where the scheduler thinks that there's a sudden 100MB increase in unmanaged_recent, since process memory hasn't changed but managed memory has decreased by 100MB 4. When the heartbeat arrives, process memory goes down and so does the unmanaged_recent. This is OK - one of the main reasons for the unmanaged_recent / unmanaged_old split is exactly to concentrate all the noise in unmanaged_recent and exclude it from optimistic memory, which is used for heuristics. Something that is less OK, but also less frequent, is that the sudden deletion of spilled keys will cause a negative blip in managed memory: 1. Delete 100MB of spilled data 2. The updated managed memory *total* reaches the scheduler faster than the updated spilled portion 3. This causes the managed memory to temporarily plummet and be replaced by unmanaged_recent, while spilled memory remains unaltered 4. When the heartbeat arrives, managed goes back up, unmanaged_recent goes back down, and spilled goes down by 100MB as it should have to begin with. https://github.com/dask/distributed/issues/6002 will let us solve this. """ return MemoryState( process=self.metrics["memory"], managed=max(0, self.nbytes - self.metrics["spilled_bytes"]["memory"]), spilled=self.metrics["spilled_bytes"]["disk"], unmanaged_old=self._memory_unmanaged_old, ) def clean(self) -> WorkerState: """Return a version of this object that is appropriate for serialization""" ws = WorkerState( address=self.address, status=self.status, pid=self.pid, name=self.name, nthreads=self.nthreads, memory_limit=self.memory_limit, local_directory=self.local_directory, services=self.services, nanny=self.nanny, extra=self.extra, server_id=self.server_id, ) ws._occupancy_cache = self.occupancy ws.executing = { ts.key: duration for ts, duration in self.executing.items() # type: ignore } return ws def __repr__(self) -> str: name = f", name: {self.name}" if self.name != self.address else "" return ( f"" ) def _repr_html_(self) -> str: return get_template("worker_state.html.j2").render( address=self.address, name=self.name, status=self.status.name, has_what=self.has_what, processing=self.processing, ) def identity(self) -> dict[str, Any]: return { "type": "Worker", "id": self.name, "host": self.host, "resources": self.resources, "local_directory": self.local_directory, "name": self.name, "nthreads": self.nthreads, "memory_limit": self.memory_limit, "last_seen": self.last_seen, "services": self.services, "metrics": self.metrics, "status": self.status.name, "nanny": self.nanny, **self.extra, } def _to_dict_no_nest(self, *, exclude: Container[str] = ()) -> dict[str, Any]: """Dictionary representation for debugging purposes. Not type stable and not intended for roundtrips. See also -------- Client.dump_cluster_state distributed.utils.recursive_to_dict TaskState._to_dict """ return recursive_to_dict( self, exclude=set(exclude) | {"versions"}, # type: ignore members=True, ) @property def scheduler(self) -> SchedulerState: assert self.scheduler_ref s = self.scheduler_ref() assert s return s def add_to_processing(self, ts: TaskState) -> None: """Assign a task to this worker for compute.""" if self.scheduler.validate: assert ts not in self.processing tp = ts.prefix self.task_prefix_count[tp.name] += 1 self.scheduler._task_prefix_count_global[tp.name] += 1 self.processing.add(ts) for dts in ts.dependencies: if self not in dts.who_has: self._inc_needs_replica(dts) def add_to_long_running(self, ts: TaskState) -> None: if self.scheduler.validate: assert ts in self.processing assert ts not in self.long_running self._remove_from_task_prefix_count(ts) # Cannot remove from processing since we're using this for things like # idleness detection. Idle workers are typically targeted for # downscaling but we should not downscale workers with long running # tasks self.long_running.add(ts) def remove_from_processing(self, ts: TaskState) -> None: """Remove a task from a workers processing""" if self.scheduler.validate: assert ts in self.processing if ts in self.long_running: self.long_running.discard(ts) else: self._remove_from_task_prefix_count(ts) self.processing.remove(ts) for dts in ts.dependencies: if dts in self.needs_what: self._dec_needs_replica(dts) def _remove_from_task_prefix_count(self, ts: TaskState) -> None: count = self.task_prefix_count[ts.prefix.name] - 1 if count: self.task_prefix_count[ts.prefix.name] = count else: del self.task_prefix_count[ts.prefix.name] count = self.scheduler._task_prefix_count_global[ts.prefix.name] - 1 if count: self.scheduler._task_prefix_count_global[ts.prefix.name] = count else: del self.scheduler._task_prefix_count_global[ts.prefix.name] def remove_replica(self, ts: TaskState) -> None: """The worker no longer has a task in memory""" if self.scheduler.validate: assert self in ts.who_has assert ts in self.has_what assert ts not in self.needs_what self.nbytes -= ts.get_nbytes() del self._has_what[ts] ts.who_has.remove(self) def _inc_needs_replica(self, ts: TaskState) -> None: """Assign a task fetch to this worker and update network occupancies""" if self.scheduler.validate: assert self not in ts.who_has assert ts not in self.has_what if ts not in self.needs_what: self.needs_what[ts] = 1 nbytes = ts.get_nbytes() self._network_occ += nbytes self.scheduler._network_occ_global += nbytes else: self.needs_what[ts] += 1 def _dec_needs_replica(self, ts: TaskState) -> None: if self.scheduler.validate: assert ts in self.needs_what self.needs_what[ts] -= 1 if self.needs_what[ts] == 0: del self.needs_what[ts] nbytes = ts.get_nbytes() self._network_occ -= nbytes self.scheduler._network_occ_global -= nbytes def add_replica(self, ts: TaskState) -> None: """The worker acquired a replica of task""" if self.scheduler.validate: assert self not in ts.who_has assert ts not in self.has_what nbytes = ts.get_nbytes() if ts in self.needs_what: del self.needs_what[ts] self._network_occ -= nbytes self.scheduler._network_occ_global -= nbytes ts.who_has.add(self) self.nbytes += nbytes self._has_what[ts] = None @property def occupancy(self) -> float: return self._occupancy_cache or self.scheduler._calc_occupancy( self.task_prefix_count, self._network_occ ) @dataclasses.dataclass class ErredTask: """Lightweight representation of an erred task without any dependency information or runspec. See also -------- TaskState """ key: Hashable timestamp: float erred_on: set[str] exception_text: str traceback_text: str class Computation: """Collection tracking a single compute or persist call See also -------- TaskPrefix TaskGroup TaskState """ start: float groups: set[TaskGroup] code: SortedSet id: uuid.UUID __slots__ = tuple(__annotations__) def __init__(self): self.start = time() self.groups = set() self.code = SortedSet() self.id = uuid.uuid4() @property def stop(self) -> float: if self.groups: return max(tg.stop for tg in self.groups) else: return -1 @property def states(self) -> dict[TaskStateState, int]: return merge_with(sum, (tg.states for tg in self.groups)) def __repr__(self) -> str: return ( f"" ) def _repr_html_(self) -> str: return get_template("computation.html.j2").render( id=self.id, start=self.start, stop=self.stop, groups=self.groups, states=self.states, code=self.code, ) class TaskPrefix: """Collection tracking all tasks within a group Keys often have a structure like ``("x-123", 0)`` A group takes the first section, like ``"x"`` See Also -------- TaskGroup """ #: The name of a group of tasks. #: For a task like ``("x-123", 0)`` this is the text ``"x"`` name: str #: An exponentially weighted moving average duration of all tasks with this prefix duration_average: float #: Numbers of times a task was marked as suspicious with this prefix suspicious: int #: Store timings for each prefix-action all_durations: defaultdict[str, float] #: This measures the maximum recorded live execution time and can be used to #: detect outliers max_exec_time: float #: Task groups associated to this prefix groups: list[TaskGroup] #: Accumulate count of number of tasks in each state state_counts: defaultdict[TaskStateState, int] __slots__ = tuple(__annotations__) def __init__(self, name: str): self.name = name self.groups = [] self.all_durations = defaultdict(float) self.state_counts = defaultdict(int) task_durations = dask.config.get("distributed.scheduler.default-task-durations") if self.name in task_durations: self.duration_average = parse_timedelta(task_durations[self.name]) else: self.duration_average = -1 self.max_exec_time = -1 self.suspicious = 0 def add_exec_time(self, duration: float) -> None: self.max_exec_time = max(duration, self.max_exec_time) if duration > 2 * self.duration_average: self.duration_average = -1 def add_duration(self, action: str, start: float, stop: float) -> None: duration = stop - start self.all_durations[action] += duration if action == "compute": old = self.duration_average if old < 0: self.duration_average = duration else: self.duration_average = 0.5 * duration + 0.5 * old @property def states(self) -> dict[str, int]: """The number of tasks in each state, like ``{"memory": 10, "processing": 3, "released": 4, ...}`` """ return merge_with(sum, [tg.states for tg in self.groups]) @property def active(self) -> list[TaskGroup]: return [ tg for tg in self.groups if any(k != "forgotten" and v != 0 for k, v in tg.states.items()) ] @property def active_states(self) -> dict[str, int]: return merge_with(sum, [tg.states for tg in self.active]) def __repr__(self) -> str: return ( "<" + self.name + ": " + ", ".join( "%s: %d" % (k, v) for (k, v) in sorted(self.states.items()) if v ) + ">" ) @property def nbytes_total(self) -> int: return sum(tg.nbytes_total for tg in self.groups) def __len__(self) -> int: return sum(map(len, self.groups)) @property def duration(self) -> float: return sum(tg.duration for tg in self.groups) @property def types(self) -> set[str]: return {typ for tg in self.groups for typ in tg.types} class TaskGroup: """Collection tracking all tasks within a group Keys often have a structure like ``("x-123", 0)`` A group takes the first section, like ``"x-123"`` See also -------- TaskPrefix """ #: The name of a group of tasks. #: For a task like ``("x-123", 0)`` this is the text ``"x-123"`` name: str #: The number of tasks in each state, #: like ``{"memory": 10, "processing": 3, "released": 4, ...}`` states: dict[TaskStateState, int] #: The other TaskGroups on which this one depends dependencies: set[TaskGroup] #: The total number of bytes that this task group has produced nbytes_total: int #: The total amount of time spent on all tasks in this TaskGroup duration: float #: The result types of this TaskGroup types: set[str] #: The worker most recently assigned a task from this group, or None when the group #: is not identified to be root-like by `SchedulerState.decide_worker`. last_worker: WorkerState | None #: If `last_worker` is not None, the number of times that worker should be assigned #: subsequent tasks until a new worker is chosen. last_worker_tasks_left: int prefix: TaskPrefix | None start: float stop: float all_durations: defaultdict[str, float] __slots__ = tuple(__annotations__) def __init__(self, name: str): self.name = name self.prefix = None self.states = dict.fromkeys(ALL_TASK_STATES, 0) self.dependencies = set() self.nbytes_total = 0 self.duration = 0 self.types = set() self.start = 0.0 self.stop = 0.0 self.all_durations = defaultdict(float) self.last_worker = None self.last_worker_tasks_left = 0 def add_duration(self, action: str, start: float, stop: float) -> None: duration = stop - start self.all_durations[action] += duration if action == "compute": if self.stop < stop: self.stop = stop self.start = self.start or start self.duration += duration assert self.prefix is not None self.prefix.add_duration(action, start, stop) def add(self, other: TaskState) -> None: self.states[other.state] += 1 other.group = self def __repr__(self) -> str: return ( "<" + (self.name or "no-group") + ": " + ", ".join( "%s: %d" % (k, v) for (k, v) in sorted(self.states.items()) if v ) + ">" ) def __len__(self) -> int: return sum(self.states.values()) def _to_dict_no_nest(self, *, exclude: Container[str] = ()) -> dict[str, Any]: """Dictionary representation for debugging purposes. Not type stable and not intended for roundtrips. See also -------- Client.dump_cluster_state distributed.utils.recursive_to_dict TaskState._to_dict """ return recursive_to_dict(self, exclude=exclude, members=True) class TaskState: """A simple object holding information about a task. Not to be confused with :class:`distributed.worker_state_machine.TaskState`, which holds similar information on the Worker side. """ #: The key is the unique identifier of a task, generally formed from the name of the #: function, followed by a hash of the function and arguments, like #: ``'inc-ab31c010444977004d656610d2d421ec'``. key: str #: The broad class of tasks to which this task belongs like "inc" or "read_csv" prefix: TaskPrefix #: A specification of how to run the task. The type and meaning of this value is #: opaque to the scheduler, as it is only interpreted by the worker to which the #: task is sent for executing. #: #: As a special case, this attribute may also be ``None``, in which case the task is #: "pure data" (such as, for example, a piece of data loaded in the scheduler using #: :meth:`Client.scatter`). A "pure data" task cannot be computed again if its #: value is lost. run_spec: object #: The priority provides each task with a relative ranking which is used to break #: ties when many tasks are being considered for execution. #: #: This ranking is generally a 2-item tuple. The first (and dominant) item #: corresponds to when it was submitted. Generally, earlier tasks take precedence. #: The second item is determined by the client, and is a way to prioritize tasks #: within a large graph that may be important, such as if they are on the critical #: path, or good to run in order to release many dependencies. This is explained #: further in :doc:`Scheduling Policy `. priority: tuple[int, ...] # Attribute underlying the state property _state: TaskStateState #: The set of tasks this task depends on for proper execution. Only tasks still #: alive are listed in this set. If, for whatever reason, this task also depends on #: a forgotten task, the :attr:`has_lost_dependencies` flag is set. #: #: A task can only be executed once all its dependencies have already been #: successfully executed and have their result stored on at least one worker. This #: is tracked by progressively draining the :attr:`waiting_on` set. dependencies: set[TaskState] #: The set of tasks which depend on this task. Only tasks still alive are listed in #: this set. This is the reverse mapping of :attr:`dependencies`. dependents: set[TaskState] #: Whether any of the dependencies of this task has been forgotten. For memory #: consumption reasons, forgotten tasks are not kept in memory even though they may #: have dependent tasks. When a task is forgotten, therefore, each of its #: dependents has their :attr:`has_lost_dependencies` attribute set to ``True``. #: #: If :attr:`has_lost_dependencies` is true, this task cannot go into the #: "processing" state anymore. has_lost_dependencies: bool #: The set of tasks this task is waiting on *before* it can be executed. This is #: always a subset of :attr:`dependencies`. Each time one of the dependencies has #: finished processing, it is removed from the :attr:`waiting_on` set. #: #: Once :attr:`waiting_on` becomes empty, this task can move from the "waiting" #: state to the "processing" state (unless one of the dependencies errored out, in #: which case this task is instead marked "erred"). waiting_on: set[TaskState] #: The set of tasks which need this task to remain alive. This is always a subset #: of :attr:`dependents`. Each time one of the dependents has finished processing, #: it is removed from the :attr:`waiters` set. #: #: Once both :attr:`waiters` and :attr:`who_wants` become empty, this task can be #: released (if it has a non-empty :attr:`run_spec`) or forgotten (otherwise) by the #: scheduler, and by any workers in :attr:`who_has`. #: #: .. note:: #: Counter-intuitively, :attr:`waiting_on` and :attr:`waiters` are not reverse #: mappings of each other. waiters: set[TaskState] #: The set of clients who want the result of this task to remain alive. #: This is the reverse mapping of :attr:`ClientState.wants_what`. #: #: When a client submits a graph to the scheduler it also specifies which output #: tasks it desires, such that their results are not released from memory. #: #: Once a task has finished executing (i.e. moves into the "memory" or "erred" #: state), the clients in :attr:`who_wants` are notified. #: #: Once both :attr:`waiters` and :attr:`who_wants` become empty, this task can be #: released (if it has a non-empty :attr:`run_spec`) or forgotten (otherwise) by the #: scheduler, and by any workers in :attr:`who_has`. who_wants: set[ClientState] #: The set of workers who have this task's result in memory. It is non-empty iff the #: task is in the "memory" state. There can be more than one worker in this set if, #: for example, :meth:`Client.scatter` or :meth:`Client.replicate` was used. #: #: This is the reverse mapping of :attr:`WorkerState.has_what`. who_has: set[WorkerState] #: If this task is in the "processing" state, which worker is currently processing #: it. This attribute is kept in sync with :attr:`WorkerState.processing`. processing_on: WorkerState | None #: The number of times this task can automatically be retried in case of failure. #: If a task fails executing (the worker returns with an error), its :attr:`retries` #: attribute is checked. If it is equal to 0, the task is marked "erred". If it is #: greater than 0, the :attr:`retries` attribute is decremented and execution is #: attempted again. retries: int #: The number of bytes, as determined by ``sizeof``, of the result of a finished #: task. This number is used for diagnostics and to help prioritize work. #: Set to -1 for unfinished tasks. nbytes: int #: The type of the object as a string. Only present for tasks that have been #: computed. type: str #: If this task failed executing, the exception object is stored here. exception: Serialized | None #: If this task failed executing, the traceback object is stored here. traceback: Serialized | None #: string representation of exception exception_text: str #: string representation of traceback traceback_text: str #: If this task or one of its dependencies failed executing, the failed task is #: stored here (possibly itself). exception_blame: TaskState | None #: Worker addresses on which errors appeared, causing this task to be in an error #: state. erred_on: set[str] #: The number of times this task has been involved in a worker death. #: #: Some tasks may cause workers to die (such as calling ``os._exit(0)``). When a #: worker dies, all of the tasks on that worker are reassigned to others. This #: combination of behaviors can cause a bad task to catastrophically destroy all #: workers on the cluster, one after another. Whenever a worker dies, we mark each #: task currently processing on that worker (as recorded by #: :attr:`WorkerState.processing`) as suspicious. If a task is involved in three #: deaths (or some other fixed constant) then we mark the task as ``erred``. suspicious: int #: A set of hostnames where this task can be run (or ``None`` if empty). Usually #: this is empty unless the task has been specifically restricted to only run on #: certain hosts. A hostname may correspond to one or several connected workers. host_restrictions: set[str] #: A set of complete worker addresses where this can be run (or ``None`` if empty). #: Usually this is empty unless the task has been specifically restricted to only #: run on certain workers. #: Note this is tracking worker addresses, not worker states, since the specific #: workers may not be connected at this time. worker_restrictions: set[str] #: Resources required by this task, such as ``{'gpu': 1}`` or ``{'memory': 1e9}`` #: These are user-defined names and are matched against the : contents of each #: :attr:`WorkerState.resources` dictionary. resource_restrictions: dict[str, float] #: False #: Each of :attr:`host_restrictions`, :attr:`worker_restrictions` and #: :attr:`resource_restrictions` is a hard constraint: if no worker is available #: satisfying those restrictions, the task cannot go into the "processing" state #: and will instead go into the "no-worker" state. #: True #: The above restrictions are mere preferences: if no worker is available #: satisfying those restrictions, the task can still go into the #: "processing" state and be sent for execution to another connected worker. loose_restrictions: bool #: Whether this task is an Actor actor: bool #: The group of tasks to which this one belongs group: TaskGroup #: Same as of group.name group_key: str #: Metadata related to task metadata: dict[str, Any] #: Task annotations annotations: dict[str, Any] #: The unique identifier of a specific execution of a task. This identifier #: is used to sign a task such that the assigned worker is expected to return #: the same identifier in the task-finished message. This is used to correlate #: responses. #: Only the most recently assigned worker is trusted. All other results will #: be rejected. run_id: int | None #: Cached hash of :attr:`~TaskState.client_key` _hash: int # Support for weakrefs to a class with __slots__ __weakref__: Any = None __slots__ = tuple(__annotations__) #: Global iterator used to create unique task run IDs _run_id_iterator: ClassVar[itertools.count] = itertools.count() # Instances not part of slots since class variable _instances: ClassVar[weakref.WeakSet[TaskState]] = weakref.WeakSet() def __init__( self, key: str, run_spec: object, state: TaskStateState, ): self.key = key self._hash = hash(key) self.run_spec = run_spec self._state = state self.exception = None self.exception_blame = None self.traceback = None self.exception_text = "" self.traceback_text = "" self.suspicious = 0 self.retries = 0 self.nbytes = -1 self.priority = None # type: ignore self.who_wants = set() self.dependencies = set() self.dependents = set() self.waiting_on = set() self.waiters = set() self.who_has = set() self.processing_on = None self.has_lost_dependencies = False self.host_restrictions = None # type: ignore self.worker_restrictions = None # type: ignore self.resource_restrictions = {} self.loose_restrictions = False self.actor = False self.prefix = None # type: ignore self.type = None # type: ignore self.group_key = key_split_group(key) self.group = None # type: ignore self.metadata = {} self.annotations = {} self.erred_on = set() self.run_id = None TaskState._instances.add(self) def __hash__(self) -> int: return self._hash def __eq__(self, other: object) -> bool: return isinstance(other, TaskState) and self.key == other.key @property def state(self) -> TaskStateState: """This task's current state. Valid states are ``released``, ``waiting``, ``no-worker``, ``processing``, ``memory``, ``erred`` and ``forgotten``. If it is ``forgotten``, the task isn't stored in the ``tasks`` dictionary anymore and will probably disappear soon from memory. """ return self._state @state.setter def state(self, value: TaskStateState) -> None: self.group.states[self._state] -= 1 self.group.states[value] += 1 self._state = value self.prefix.state_counts[value] += 1 def add_dependency(self, other: TaskState) -> None: """Add another task as a dependency of this task""" self.dependencies.add(other) self.group.dependencies.add(other.group) other.dependents.add(self) def get_nbytes(self) -> int: return self.nbytes if self.nbytes >= 0 else DEFAULT_DATA_SIZE def set_nbytes(self, nbytes: int) -> None: diff = nbytes old_nbytes = self.nbytes if old_nbytes >= 0: diff -= old_nbytes self.group.nbytes_total += diff for ws in self.who_has: ws.nbytes += diff self.nbytes = nbytes def __repr__(self) -> str: return f"" def _repr_html_(self) -> str: return get_template("task_state.html.j2").render( state=self.state, nbytes=self.nbytes, key=self.key, ) def validate(self) -> None: try: for cs in self.who_wants: assert isinstance(cs, ClientState), (repr(cs), self.who_wants) for ws in self.who_has: assert isinstance(ws, WorkerState), (repr(ws), self.who_has) for ts in self.dependencies: assert isinstance(ts, TaskState), (repr(ts), self.dependencies) for ts in self.dependents: assert isinstance(ts, TaskState), (repr(ts), self.dependents) validate_task_state(self) except Exception as e: logger.exception(e) if LOG_PDB: import pdb pdb.set_trace() def get_nbytes_deps(self) -> int: return sum(ts.get_nbytes() for ts in self.dependencies) def _to_dict_no_nest(self, *, exclude: Container[str] = ()) -> dict[str, Any]: """Dictionary representation for debugging purposes. Not type stable and not intended for roundtrips. See also -------- Client.dump_cluster_state distributed.utils.recursive_to_dict Notes ----- This class uses ``_to_dict_no_nest`` instead of ``_to_dict``. When a task references another task, or when a WorkerState.tasks contains tasks, this method is not executed for the inner task, even if the inner task was never seen before; you get a repr instead. All tasks should neatly appear under Scheduler.tasks. This also prevents a RecursionError during particularly heavy loads, which have been observed to happen whenever there's an acyclic dependency chain of ~200+ tasks. """ return recursive_to_dict(self, exclude=exclude, members=True) class Transition(NamedTuple): """An entry in :attr:`SchedulerState.transition_log`""" key: str start: TaskStateState finish: TaskStateState recommendations: Recs stimulus_id: str timestamp: float class SchedulerState: """Underlying task state of dynamic scheduler Tracks the current state of workers, data, and computations. Handles transitions between different task states. Notifies the Scheduler of changes by messaging passing through Queues, which the Scheduler listens to responds accordingly. All events are handled quickly, in linear time with respect to their input (which is often of constant size) and generally within a millisecond. Users typically do not interact with ``Transitions`` directly. Instead users interact with the ``Client``, which in turn engages the ``Scheduler`` affecting different transitions here under-the-hood. In the background ``Worker``s also engage with the ``Scheduler`` affecting these state transitions as well. """ bandwidth: int #: Clients currently connected to the scheduler clients: dict[str, ClientState] extensions: dict[str, Any] # TODO write a scheduler extension Protocol plugins: dict[str, SchedulerPlugin] host_info: dict[str, dict[str, Any]] #: If True, enable expensive internal consistency check. #: Typically disabled in production. validate: bool ####################### # Workers-related state ####################### #: Workers currently connected to the scheduler #: (actually a SortedDict, but the sortedcontainers package isn't annotated) workers: dict[str, WorkerState] #: Worker {name: address} aliases: dict[Hashable, str] #: Workers that are currently in running state running: set[WorkerState] #: Workers that are currently in running state and not fully utilized #: Definition based on occupancy #: (actually a SortedDict, but the sortedcontainers package isn't annotated) idle: dict[str, WorkerState] #: Similar to `idle` #: Definition based on assigned tasks idle_task_count: set[WorkerState] #: Workers that are fully utilized. May include non-running workers. saturated: set[WorkerState] total_nthreads: int #: Cluster-wide resources. {resource name: {worker address: amount}} resources: dict[str, dict[str, float]] ##################### # Tasks-related state ##################### #: Total number of tasks ever processed n_tasks: int #: All tasks currently known to the scheduler tasks: dict[str, TaskState] #: Tasks in the "queued" state, ordered by priority queued: HeapSet[TaskState] #: Tasks in the "no-worker" state unrunnable: set[TaskState] #: Subset of tasks that exist in memory on more than one worker replicated_tasks: set[TaskState] unknown_durations: dict[str, set[TaskState]] task_groups: dict[str, TaskGroup] task_prefixes: dict[str, TaskPrefix] task_metadata: dict[str, Any] ######### # History ######### #: History of computations. #: The length can be tweaked through #: distributed.diagnostics.computations.max-history computations: deque[Computation] #: History of erred tasks. #: The length can be tweaked through #: distributed.diagnostics.erred-tasks.max-history erred_tasks: deque[ErredTask] #: History of task state transitions. #: The length can be tweaked through #: distributed.scheduler.transition-log-length transition_log: deque[Transition] #: Total number of transitions since the cluster was started transition_counter: int #: Total number of transitions as of the previous call to check_idle() _idle_transition_counter: int #: Raise an error if the :attr:`transition_counter` ever reaches this value. #: This is meant for debugging only, to catch infinite recursion loops. #: In production, it should always be set to False. transition_counter_max: int | Literal[False] _task_prefix_count_global: defaultdict[str, int] _network_occ_global: float ###################### # Cached configuration ###################### #: distributed.scheduler.unknown-task-duration UNKNOWN_TASK_DURATION: float #: distributed.worker.memory.recent-to-old-time MEMORY_RECENT_TO_OLD_TIME: float #: distributed.worker.memory.rebalance.measure MEMORY_REBALANCE_MEASURE: str #: distributed.worker.memory.rebalance.sender-min MEMORY_REBALANCE_SENDER_MIN: float #: distributed.worker.memory.rebalance.recipient-max MEMORY_REBALANCE_RECIPIENT_MAX: float #: distributed.worker.memory.rebalance.sender-recipient-gap / 2 MEMORY_REBALANCE_HALF_GAP: float #: distributed.scheduler.worker-saturation WORKER_SATURATION: float __slots__ = tuple(__annotations__) def __init__( self, aliases: dict[Hashable, str], clients: dict[str, ClientState], workers: SortedDict[str, WorkerState], host_info: dict[str, dict[str, Any]], resources: dict[str, dict[str, float]], tasks: dict[str, TaskState], unrunnable: set[TaskState], queued: HeapSet[TaskState], validate: bool, plugins: Iterable[SchedulerPlugin] = (), transition_counter_max: int | Literal[False] = False, **kwargs: Any, # Passed verbatim to Server.__init__() ): logger.info("State start") self.aliases = aliases self.bandwidth = parse_bytes(dask.config.get("distributed.scheduler.bandwidth")) self.clients = clients self.clients["fire-and-forget"] = ClientState("fire-and-forget") self.extensions = {} self.host_info = host_info self.idle = SortedDict() self.idle_task_count = set() self.n_tasks = 0 self.resources = resources self.saturated = set() self.tasks = tasks self.replicated_tasks = { ts for ts in self.tasks.values() if len(ts.who_has) > 1 } self.computations = deque( maxlen=dask.config.get("distributed.diagnostics.computations.max-history") ) self.erred_tasks = deque( maxlen=dask.config.get("distributed.diagnostics.erred-tasks.max-history") ) self.task_groups = {} self.task_prefixes = {} self.task_metadata = {} self.total_nthreads = 0 self.unknown_durations = {} self.queued = queued self.unrunnable = unrunnable self.validate = validate self.workers = workers self._task_prefix_count_global = defaultdict(int) self._network_occ_global = 0.0 self.running = { ws for ws in self.workers.values() if ws.status == Status.running } self.plugins = {} if not plugins else {_get_plugin_name(p): p for p in plugins} self.transition_log = deque( maxlen=dask.config.get("distributed.scheduler.transition-log-length") ) self.transition_counter = 0 self._idle_transition_counter = 0 self.transition_counter_max = transition_counter_max # Variables from dask.config, cached by __init__ for performance self.UNKNOWN_TASK_DURATION = parse_timedelta( dask.config.get("distributed.scheduler.unknown-task-duration") ) self.MEMORY_RECENT_TO_OLD_TIME = parse_timedelta( dask.config.get("distributed.worker.memory.recent-to-old-time") ) self.MEMORY_REBALANCE_MEASURE = dask.config.get( "distributed.worker.memory.rebalance.measure" ) self.MEMORY_REBALANCE_SENDER_MIN = dask.config.get( "distributed.worker.memory.rebalance.sender-min" ) self.MEMORY_REBALANCE_RECIPIENT_MAX = dask.config.get( "distributed.worker.memory.rebalance.recipient-max" ) self.MEMORY_REBALANCE_HALF_GAP = ( dask.config.get("distributed.worker.memory.rebalance.sender-recipient-gap") / 2.0 ) self.WORKER_SATURATION = dask.config.get( "distributed.scheduler.worker-saturation" ) if self.WORKER_SATURATION == "inf": # Special case necessary because there's no way to parse a float infinity # from a DASK_* environment variable self.WORKER_SATURATION = math.inf if ( not isinstance(self.WORKER_SATURATION, (int, float)) or self.WORKER_SATURATION <= 0 ): raise ValueError( # pragma: nocover "`distributed.scheduler.worker-saturation` must be a float > 0; got " + repr(self.WORKER_SATURATION) ) @property def memory(self) -> MemoryState: return MemoryState.sum(*(w.memory for w in self.workers.values())) @property def __pdict__(self) -> dict[str, Any]: return { "bandwidth": self.bandwidth, "resources": self.resources, "saturated": self.saturated, "unrunnable": self.unrunnable, "queued": self.queued, "n_tasks": self.n_tasks, "unknown_durations": self.unknown_durations, "validate": self.validate, "tasks": self.tasks, "task_groups": self.task_groups, "task_prefixes": self.task_prefixes, "total_nthreads": self.total_nthreads, "total_occupancy": self.total_occupancy, "erred_tasks": self.erred_tasks, "extensions": self.extensions, "clients": self.clients, "workers": self.workers, "idle": self.idle, "host_info": self.host_info, } def new_task( self, key: str, spec: object, state: TaskStateState, computation: Computation | None = None, ) -> TaskState: """Create a new task, and associated states""" ts = TaskState(key, spec, state) prefix_key = key_split(key) tp = self.task_prefixes.get(prefix_key) if tp is None: self.task_prefixes[prefix_key] = tp = TaskPrefix(prefix_key) ts.prefix = tp group_key = ts.group_key tg = self.task_groups.get(group_key) if tg is None: self.task_groups[group_key] = tg = TaskGroup(group_key) if computation: computation.groups.add(tg) tg.prefix = tp tp.groups.append(tg) tg.add(ts) self.tasks[key] = ts return ts def _clear_task_state(self) -> None: logger.debug("Clear task state") for collection in ( self.unrunnable, self.erred_tasks, self.computations, self.task_prefixes, self.task_groups, self.task_metadata, self.unknown_durations, self.replicated_tasks, ): collection.clear() # type: ignore @property def total_occupancy(self) -> float: return self._calc_occupancy( self._task_prefix_count_global, self._network_occ_global, ) def _calc_occupancy( self, task_prefix_count: dict[str, int], network_occ: float, ) -> float: res = 0.0 for prefix_name, count in task_prefix_count.items(): # TODO: Deal with unknown tasks better prefix = self.task_prefixes[prefix_name] assert prefix is not None duration = prefix.duration_average if duration < 0: if prefix.max_exec_time > 0: duration = 2 * prefix.max_exec_time else: duration = self.UNKNOWN_TASK_DURATION res += duration * count occ = res + network_occ / self.bandwidth assert occ >= 0, occ return occ ##################### # State Transitions # ##################### def _transition( self, key: str, finish: TaskStateState, stimulus_id: str, **kwargs: Any ) -> RecsMsgs: """Transition a key from its current state to the finish state Examples -------- >>> self._transition('x', 'waiting') {'x': 'processing'}, {}, {} Returns ------- Tuple of: - Dictionary of recommendations for future transitions {key: new state} - Messages to clients {client address: [msg, msg, ...]} - Messages to workers {worker address: [msg, msg, ...]} See Also -------- Scheduler.transitions : transitive version of this function """ try: ts = self.tasks.get(key) if ts is None: return {}, {}, {} start = ts._state if start == finish: return {}, {}, {} # Notes: # - in case of transition through released, this counter is incremented by 2 # - this increase happens before the actual transitions, so that it can # catch potential infinite recursions self.transition_counter += 1 if self.transition_counter_max: assert self.transition_counter < self.transition_counter_max recommendations: dict = {} worker_msgs: dict = {} client_msgs: dict = {} if self.plugins: dependents = set(ts.dependents) dependencies = set(ts.dependencies) func = self._TRANSITIONS_TABLE.get((start, finish)) if func is not None: recommendations, client_msgs, worker_msgs = func( self, key, stimulus_id, **kwargs ) elif "released" not in (start, finish): assert not kwargs, (kwargs, start, finish) a_recs, a_cmsgs, a_wmsgs = self._transition( key, "released", stimulus_id ) v = a_recs.get(key, finish) func = self._TRANSITIONS_TABLE["released", v] b_recs, b_cmsgs, b_wmsgs = func(self, key, stimulus_id) recommendations.update(a_recs) for c, new_msgs in a_cmsgs.items(): client_msgs.setdefault(c, []).extend(new_msgs) for w, new_msgs in a_wmsgs.items(): worker_msgs.setdefault(w, []).extend(new_msgs) recommendations.update(b_recs) for c, new_msgs in b_cmsgs.items(): client_msgs.setdefault(c, []).extend(new_msgs) for w, new_msgs in b_wmsgs.items(): worker_msgs.setdefault(w, []).extend(new_msgs) start = "released" else: raise RuntimeError( f"Impossible transition from {start} to {finish} for {key!r}: " f"{stimulus_id=}, {kwargs=}, story={self.story(ts)}" ) if not stimulus_id: stimulus_id = STIMULUS_ID_UNSET actual_finish = ts._state self.transition_log.append( Transition( key, start, actual_finish, recommendations, stimulus_id, time() ) ) if self.validate: if stimulus_id == STIMULUS_ID_UNSET: raise RuntimeError( "stimulus_id not set during Scheduler transition" ) logger.debug( "Transitioned %r %s->%s (actual: %s). Consequence: %s", key, start, finish, actual_finish, dict(recommendations), ) if self.plugins: # Temporarily put back forgotten key for plugin to retrieve it if ts._state == "forgotten": ts.dependents = dependents ts.dependencies = dependencies self.tasks[ts.key] = ts for plugin in list(self.plugins.values()): try: plugin.transition(key, start, actual_finish, **kwargs) except Exception: logger.info("Plugin failed with exception", exc_info=True) if ts.state == "forgotten": del self.tasks[ts.key] tg = ts.group if ts.state == "forgotten" and tg.name in self.task_groups: # Remove TaskGroup if all tasks are in the forgotten state if all(v == 0 or k == "forgotten" for k, v in tg.states.items()): ts.prefix.groups.remove(tg) del self.task_groups[tg.name] return recommendations, client_msgs, worker_msgs except Exception: logger.exception("Error transitioning %r from %r to %r", key, start, finish) if LOG_PDB: import pdb pdb.set_trace() raise def _transitions( self, recommendations: Recs, client_msgs: Msgs, worker_msgs: Msgs, stimulus_id: str, ) -> None: """Process transitions until none are left This includes feedback from previous transitions and continues until we reach a steady state """ keys: set[str] = set() recommendations = recommendations.copy() while recommendations: key, finish = recommendations.popitem() keys.add(key) new_recs, new_cmsgs, new_wmsgs = self._transition(key, finish, stimulus_id) recommendations.update(new_recs) for c, new_msgs in new_cmsgs.items(): client_msgs.setdefault(c, []).extend(new_msgs) for w, new_msgs in new_wmsgs.items(): worker_msgs.setdefault(w, []).extend(new_msgs) if self.validate: # FIXME downcast antipattern scheduler = cast(Scheduler, self) for key in keys: scheduler.validate_key(key) def transition_released_waiting(self, key: str, stimulus_id: str) -> RecsMsgs: ts = self.tasks[key] if self.validate: assert ts.run_spec assert not ts.waiting_on assert not ts.who_has assert not ts.processing_on for dts in ts.dependencies: assert dts.state not in {"forgotten", "erred"} if ts.has_lost_dependencies: return {key: "forgotten"}, {}, {} ts.state = "waiting" recommendations: Recs = {} for dts in ts.dependencies: if not dts.who_has: ts.waiting_on.add(dts) if dts.state == "released": recommendations[dts.key] = "waiting" else: dts.waiters.add(ts) ts.waiters = {dts for dts in ts.dependents if dts.state == "waiting"} if not ts.waiting_on: # NOTE: waiting->processing will send tasks to queued or no-worker as # necessary recommendations[key] = "processing" return recommendations, {}, {} def transition_no_worker_processing(self, key: str, stimulus_id: str) -> RecsMsgs: ts = self.tasks[key] worker_msgs: Msgs = {} if self.validate: assert not ts.actor, f"Actors can't be in `no-worker`: {ts}" assert ts in self.unrunnable if ws := self.decide_worker_non_rootish(ts): self.unrunnable.discard(ts) worker_msgs = self._add_to_processing(ts, ws) # If no worker, task just stays in `no-worker` return {}, {}, worker_msgs def decide_worker_rootish_queuing_disabled( self, ts: TaskState ) -> WorkerState | None: """Pick a worker for a runnable root-ish task, without queuing. This attempts to schedule sibling tasks on the same worker, reducing future data transfer. It does not consider the location of dependencies, since they'll end up on every worker anyway. It assumes it's being called on a batch of tasks in priority order, and maintains state in `SchedulerState.last_root_worker` and `SchedulerState.last_root_worker_tasks_left` to achieve this. This will send every runnable task to a worker, often causing root task overproduction. Returns ------- ws: WorkerState | None The worker to assign the task to. If there are no workers in the cluster, returns None, in which case the task should be transitioned to ``no-worker``. """ if self.validate: # See root-ish-ness note below in `decide_worker_rootish_queuing_enabled` assert math.isinf(self.WORKER_SATURATION) pool = self.idle.values() if self.idle else self.running if not pool: return None tg = ts.group lws = tg.last_worker if ( lws and tg.last_worker_tasks_left and lws.status == Status.running and self.workers.get(lws.address) is lws ): ws = lws else: # Last-used worker is full, unknown, retiring, or paused; # pick a new worker for the next few tasks ws = min(pool, key=partial(self.worker_objective, ts)) tg.last_worker_tasks_left = math.floor( (len(tg) / self.total_nthreads) * ws.nthreads ) # Record `last_worker`, or clear it on the final task tg.last_worker = ( ws if tg.states["released"] + tg.states["waiting"] > 1 else None ) tg.last_worker_tasks_left -= 1 if self.validate and ws is not None: assert self.workers.get(ws.address) is ws assert ws in self.running, (ws, self.running) return ws def decide_worker_rootish_queuing_enabled(self) -> WorkerState | None: """Pick a worker for a runnable root-ish task, if not all are busy. Picks the least-busy worker out of the ``idle`` workers (idle workers have fewer tasks running than threads, as set by ``distributed.scheduler.worker-saturation``). It does not consider the location of dependencies, since they'll end up on every worker anyway. If all workers are full, returns None, meaning the task should transition to ``queued``. The scheduler will wait to send it to a worker until a thread opens up. This ensures that downstream tasks always run before new root tasks are started. This does not try to schedule sibling tasks on the same worker; in fact, it usually does the opposite. Even though this increases subsequent data transfer, it typically reduces overall memory use by eliminating root task overproduction. Returns ------- ws: WorkerState | None The worker to assign the task to. If there are no idle workers, returns None, in which case the task should be transitioned to ``queued``. """ if self.validate: # We don't `assert self.is_rootish(ts)` here, because that check is # dependent on cluster size. It's possible a task looked root-ish when it # was queued, but the cluster has since scaled up and it no longer does when # coming out of the queue. If `is_rootish` changes to a static definition, # then add that assertion here (and actually pass in the task). assert not math.isinf(self.WORKER_SATURATION) if not self.idle_task_count: # All workers busy? Task gets/stays queued. return None # Just pick the least busy worker. # NOTE: this will lead to worst-case scheduling with regards to co-assignment. ws = min( self.idle_task_count, key=lambda ws: len(ws.processing) / ws.nthreads, ) if self.validate: assert not _worker_full(ws, self.WORKER_SATURATION), ( ws, _task_slots_available(ws, self.WORKER_SATURATION), ) assert ws in self.running, (ws, self.running) if self.validate and ws is not None: assert self.workers.get(ws.address) is ws assert ws in self.running, (ws, self.running) return ws def decide_worker_non_rootish(self, ts: TaskState) -> WorkerState | None: """Pick a worker for a runnable non-root task, considering dependencies and restrictions. Out of eligible workers holding dependencies of ``ts``, selects the worker where, considering worker backlong and data-transfer costs, the task is estimated to start running the soonest. Returns ------- ws: WorkerState | None The worker to assign the task to. If no workers satisfy the restrictions of ``ts`` or there are no running workers, returns None, in which case the task should be transitioned to ``no-worker``. """ if not self.running: return None valid_workers = self.valid_workers(ts) if valid_workers is None and len(self.running) < len(self.workers): if not self.running: return None # If there were no restrictions, `valid_workers()` didn't subset by # `running`. valid_workers = self.running if ts.dependencies or valid_workers is not None: ws = decide_worker( ts, self.running, valid_workers, partial(self.worker_objective, ts), ) else: # TODO if `is_rootish` would always return True for tasks without # dependencies, we could remove all this logic. The rootish assignment logic # would behave more or less the same as this, maybe without guaranteed # round-robin though? This path is only reachable when `ts` doesn't have # dependencies, but its group is also smaller than the cluster. # Fastpath when there are no related tasks or restrictions worker_pool = self.idle or self.workers # FIXME idle and workers are SortedDict's declared as dicts # because sortedcontainers is not annotated wp_vals = cast("Sequence[WorkerState]", worker_pool.values()) n_workers = len(wp_vals) if n_workers < 20: # smart but linear in small case ws = min(wp_vals, key=operator.attrgetter("occupancy")) assert ws if ws.occupancy == 0: # special case to use round-robin; linear search # for next worker with zero occupancy (or just # land back where we started). start = self.n_tasks % n_workers for i in range(n_workers): wp_i = wp_vals[(i + start) % n_workers] if wp_i.occupancy == 0: ws = wp_i break else: # dumb but fast in large case ws = wp_vals[self.n_tasks % n_workers] if self.validate and ws is not None: assert self.workers.get(ws.address) is ws assert ws in self.running, (ws, self.running) return ws def transition_waiting_processing(self, key: str, stimulus_id: str) -> RecsMsgs: """Possibly schedule a ready task. This is the primary dispatch for ready tasks. If there's no appropriate worker for the task (but the task is otherwise runnable), it will be recommended to ``no-worker`` or ``queued``. """ ts = self.tasks[key] if self.is_rootish(ts): # NOTE: having two root-ish methods is temporary. When the feature flag is # removed, there should only be one, which combines co-assignment and # queuing. Eventually, special-casing root tasks might be removed entirely, # with better heuristics. if math.isinf(self.WORKER_SATURATION): if not (ws := self.decide_worker_rootish_queuing_disabled(ts)): return {ts.key: "no-worker"}, {}, {} else: if not (ws := self.decide_worker_rootish_queuing_enabled()): return {ts.key: "queued"}, {}, {} else: if not (ws := self.decide_worker_non_rootish(ts)): return {ts.key: "no-worker"}, {}, {} worker_msgs = self._add_to_processing(ts, ws) return {}, {}, worker_msgs def transition_waiting_memory( self, key: str, stimulus_id: str, *, nbytes: int | None = None, type: bytes | None = None, typename: str | None = None, worker: str, **kwargs: Any, ) -> RecsMsgs: """This transition exclusively happens in a race condition where the scheduler believes that the only copy of a dependency task has just been lost, so it transitions all dependents back to waiting, but actually a replica has already been acquired by a worker computing the dependency - the scheduler just doesn't know yet - and the execution finishes before the cancellation message from the scheduler has a chance to reach the worker. Shortly, the cancellation request will reach the worker, thus deleting the data from memory. """ ts = self.tasks[key] if self.validate: assert not ts.processing_on assert ts.waiting_on assert ts.state == "waiting" return {}, {}, {} def transition_processing_memory( self, key: str, stimulus_id: str, *, nbytes: int | None = None, type: bytes | None = None, typename: str | None = None, worker: str, startstops: list[dict] | None = None, **kwargs: Any, ) -> RecsMsgs: ts = self.tasks[key] assert worker assert isinstance(worker, str) if self.validate: assert ts.processing_on wss = ts.processing_on assert wss assert ts in wss.processing del wss assert not ts.waiting_on assert not ts.who_has, (ts, ts.who_has) assert not ts.exception_blame assert ts.state == "processing" ws = self.workers.get(worker) if ws is None: return {key: "released"}, {}, {} if ws != ts.processing_on: # pragma: nocover assert ts.processing_on raise RuntimeError( f"Task {ts.key!r} transitioned from processing to memory on worker " f"{ws}, while it was expected from {ts.processing_on}. This should " f"be impossible. {stimulus_id=}, story={self.story(ts)}" ) ############################# # Update Timing Information # ############################# if startstops: for startstop in startstops: ts.group.add_duration( stop=startstop["stop"], start=startstop["start"], action=startstop["action"], ) s = self.unknown_durations.pop(ts.prefix.name, set()) steal = self.extensions.get("stealing") if steal: for tts in s: if tts.processing_on: steal.recalculate_cost(tts) ############################ # Update State Information # ############################ if nbytes is not None: ts.set_nbytes(nbytes) self._exit_processing_common(ts) recommendations: Recs = {} client_msgs: Msgs = {} self._add_to_memory( ts, ws, recommendations, client_msgs, type=type, typename=typename ) if self.validate: assert not ts.processing_on assert not ts.waiting_on return recommendations, client_msgs, {} def transition_memory_released( self, key: str, stimulus_id: str, *, safe: bool = False ) -> RecsMsgs: ts = self.tasks[key] if self.validate: assert not ts.waiting_on assert not ts.processing_on if safe: assert not ts.waiters if ts.actor: for ws in ts.who_has: ws.actors.discard(ts) if ts.who_wants: ts.exception_blame = ts ts.exception = Serialized( *serialize(ValueError("Worker holding Actor was lost")) ) return {ts.key: "erred"}, {}, {} # don't try to recreate recommendations: Recs = {} client_msgs: Msgs = {} worker_msgs: Msgs = {} # XXX factor this out? worker_msg = { "op": "free-keys", "keys": [key], "stimulus_id": stimulus_id, } for ws in ts.who_has: worker_msgs[ws.address] = [worker_msg] self.remove_all_replicas(ts) ts.state = "released" report_msg = {"op": "lost-data", "key": key} for cs in ts.who_wants: client_msgs[cs.client_key] = [report_msg] if not ts.run_spec: # pure data recommendations[key] = "forgotten" elif ts.has_lost_dependencies: recommendations[key] = "forgotten" elif ts.who_wants or ts.waiters: recommendations[key] = "waiting" for dts in ts.waiters: if dts.state in ("no-worker", "processing"): recommendations[dts.key] = "waiting" elif dts.state == "waiting": dts.waiting_on.add(ts) if self.validate: assert not ts.waiting_on return recommendations, client_msgs, worker_msgs def transition_released_erred(self, key: str, stimulus_id: str) -> RecsMsgs: ts = self.tasks[key] recommendations: Recs = {} client_msgs: Msgs = {} if self.validate: with log_errors(pdb=LOG_PDB): assert ts.exception_blame assert not ts.who_has assert not ts.waiting_on assert not ts.waiters failing_ts = ts.exception_blame assert failing_ts for dts in ts.dependents: dts.exception_blame = failing_ts if not dts.who_has: recommendations[dts.key] = "erred" report_msg = { "op": "task-erred", "key": key, "exception": failing_ts.exception, "traceback": failing_ts.traceback, } for cs in ts.who_wants: client_msgs[cs.client_key] = [report_msg] ts.state = "erred" # TODO: waiting data? return recommendations, client_msgs, {} def transition_erred_released(self, key: str, stimulus_id: str) -> RecsMsgs: ts = self.tasks[key] recommendations: Recs = {} client_msgs: Msgs = {} worker_msgs: Msgs = {} if self.validate: with log_errors(pdb=LOG_PDB): assert ts.exception_blame assert not ts.who_has assert not ts.waiting_on assert not ts.waiters ts.exception = None ts.exception_blame = None ts.traceback = None for dts in ts.dependents: if dts.state == "erred": recommendations[dts.key] = "waiting" w_msg = { "op": "free-keys", "keys": [key], "stimulus_id": stimulus_id, } for ws_addr in ts.erred_on: worker_msgs[ws_addr] = [w_msg] ts.erred_on.clear() report_msg = {"op": "task-retried", "key": key} for cs in ts.who_wants: client_msgs[cs.client_key] = [report_msg] ts.state = "released" return recommendations, client_msgs, worker_msgs def transition_waiting_released(self, key: str, stimulus_id: str) -> RecsMsgs: ts = self.tasks[key] recommendations: Recs = {} if self.validate: assert not ts.who_has assert not ts.processing_on for dts in ts.dependencies: if ts in dts.waiters: dts.waiters.discard(ts) if not dts.waiters and not dts.who_wants: recommendations[dts.key] = "released" ts.waiting_on.clear() ts.state = "released" if ts.has_lost_dependencies: recommendations[key] = "forgotten" elif not ts.exception_blame and (ts.who_wants or ts.waiters): recommendations[key] = "waiting" else: ts.waiters.clear() return recommendations, {}, {} def transition_processing_released(self, key: str, stimulus_id: str) -> RecsMsgs: ts = self.tasks[key] recommendations: Recs = {} worker_msgs: Msgs = {} if self.validate: assert ts.processing_on assert not ts.who_has assert not ts.waiting_on assert ts.state == "processing" ws = self._exit_processing_common(ts) if ws: worker_msgs[ws.address] = [ { "op": "free-keys", "keys": [key], "stimulus_id": stimulus_id, } ] self._propagate_released(ts, recommendations) return recommendations, {}, worker_msgs def transition_processing_erred( self, key: str, stimulus_id: str, *, worker: str, cause: str | None = None, exception: Serialized | None = None, traceback: Serialized | None = None, exception_text: str | None = None, traceback_text: str | None = None, **kwargs: Any, ) -> RecsMsgs: """Processed a recommended transition processing -> erred. Parameters ---------- key Key of the task to transition stimulus_id ID of the stimulus causing the transition worker Address of the worker where the task erred. Not necessarily ``ts.processing_on``. cause Address of the task that caused this task to be transitioned to erred exception Exception caused by the task traceback Traceback caused by the task exception_text String representation of the exception traceback_text String representation of the traceback Returns ------- Recommendations, client messages and worker messages to process """ ts = self.tasks[key] recommendations: Recs = {} client_msgs: Msgs = {} if self.validate: assert cause or ts.exception_blame assert ts.processing_on assert not ts.who_has assert not ts.waiting_on if ts.actor: ws = ts.processing_on assert ws ws.actors.remove(ts) self._exit_processing_common(ts) ts.erred_on.add(worker) if exception is not None: ts.exception = exception ts.exception_text = exception_text # type: ignore if traceback is not None: ts.traceback = traceback ts.traceback_text = traceback_text # type: ignore if cause is not None: failing_ts = self.tasks[cause] ts.exception_blame = failing_ts else: failing_ts = ts.exception_blame # type: ignore self.erred_tasks.appendleft( ErredTask( ts.key, time(), ts.erred_on.copy(), exception_text or "", traceback_text or "", ) ) for dts in ts.dependents: dts.exception_blame = failing_ts recommendations[dts.key] = "erred" for dts in ts.dependencies: dts.waiters.discard(ts) if not dts.waiters and not dts.who_wants: recommendations[dts.key] = "released" ts.waiters.clear() # do anything with this? ts.state = "erred" report_msg = { "op": "task-erred", "key": key, "exception": failing_ts.exception, "traceback": failing_ts.traceback, } for cs in ts.who_wants: client_msgs[cs.client_key] = [report_msg] cs = self.clients["fire-and-forget"] if ts in cs.wants_what: self._client_releases_keys( cs=cs, keys=[key], recommendations=recommendations, ) if self.validate: assert not ts.processing_on return recommendations, client_msgs, {} def transition_no_worker_released(self, key: str, stimulus_id: str) -> RecsMsgs: ts = self.tasks[key] if self.validate: assert self.tasks[key].state == "no-worker" assert not ts.who_has assert not ts.waiting_on self.unrunnable.remove(ts) ts.state = "released" for dts in ts.dependencies: dts.waiters.discard(ts) ts.waiters.clear() return {}, {}, {} def transition_waiting_queued(self, key: str, stimulus_id: str) -> RecsMsgs: ts = self.tasks[key] if self.validate: assert not self.idle_task_count, (ts, self.idle_task_count) self._validate_ready(ts) ts.state = "queued" self.queued.add(ts) return {}, {}, {} def transition_waiting_no_worker(self, key: str, stimulus_id: str) -> RecsMsgs: ts = self.tasks[key] if self.validate: self._validate_ready(ts) ts.state = "no-worker" self.unrunnable.add(ts) return {}, {}, {} def transition_queued_released(self, key: str, stimulus_id: str) -> RecsMsgs: ts = self.tasks[key] if self.validate: assert ts in self.queued assert not ts.processing_on self.queued.remove(ts) recommendations: Recs = {} self._propagate_released(ts, recommendations) return recommendations, {}, {} def transition_queued_processing(self, key: str, stimulus_id: str) -> RecsMsgs: ts = self.tasks[key] recommendations: Recs = {} worker_msgs: Msgs = {} if self.validate: assert not ts.actor, f"Actors can't be queued: {ts}" assert ts in self.queued if ws := self.decide_worker_rootish_queuing_enabled(): self.queued.discard(ts) worker_msgs = self._add_to_processing(ts, ws) # If no worker, task just stays `queued` return recommendations, {}, worker_msgs def _remove_key(self, key: str) -> None: ts = self.tasks.pop(key) assert ts.state == "forgotten" self.unrunnable.discard(ts) for cs in ts.who_wants: cs.wants_what.remove(ts) ts.who_wants.clear() ts.processing_on = None ts.exception_blame = ts.exception = ts.traceback = None self.task_metadata.pop(key, None) def transition_memory_forgotten(self, key: str, stimulus_id: str) -> RecsMsgs: ts = self.tasks[key] if self.validate: assert ts.state == "memory" assert not ts.processing_on assert not ts.waiting_on if not ts.run_spec: # It's ok to forget a pure data task pass elif ts.has_lost_dependencies: # It's ok to forget a task with forgotten dependencies pass elif not ts.who_wants and not ts.waiters and not ts.dependents: # It's ok to forget a task that nobody needs pass else: raise AssertionError("Unreachable", ts) # pragma: nocover if ts.actor: for ws in ts.who_has: ws.actors.discard(ts) recommendations: Recs = {} worker_msgs: Msgs = {} self._propagate_forgotten(ts, recommendations, worker_msgs, stimulus_id) client_msgs = _task_to_client_msgs(ts) self._remove_key(key) return recommendations, client_msgs, worker_msgs def transition_released_forgotten(self, key: str, stimulus_id: str) -> RecsMsgs: ts = self.tasks[key] if self.validate: assert ts.state in ("released", "erred") assert not ts.who_has assert not ts.processing_on assert ts not in self.queued assert not ts.waiting_on, (ts, ts.waiting_on) if not ts.run_spec: # It's ok to forget a pure data task pass elif ts.has_lost_dependencies: # It's ok to forget a task with forgotten dependencies pass elif not ts.who_wants and not ts.waiters and not ts.dependents: # It's ok to forget a task that nobody needs pass else: raise AssertionError("Unreachable", str(ts)) # pragma: nocover recommendations: Recs = {} worker_msgs: Msgs = {} self._propagate_forgotten(ts, recommendations, worker_msgs, stimulus_id) client_msgs = _task_to_client_msgs(ts) self._remove_key(key) return recommendations, client_msgs, worker_msgs # { # (start, finish): # transition__( # self, key: str, stimulus_id: str, **kwargs # ) -> (recommendations, client_msgs, worker_msgs) # } _TRANSITIONS_TABLE: ClassVar[ Mapping[ tuple[TaskStateState, TaskStateState], Callable[..., RecsMsgs], ] ] = { ("released", "waiting"): transition_released_waiting, ("waiting", "released"): transition_waiting_released, ("waiting", "processing"): transition_waiting_processing, ("waiting", "no-worker"): transition_waiting_no_worker, ("waiting", "queued"): transition_waiting_queued, ("waiting", "memory"): transition_waiting_memory, ("queued", "released"): transition_queued_released, ("queued", "processing"): transition_queued_processing, ("processing", "released"): transition_processing_released, ("processing", "memory"): transition_processing_memory, ("processing", "erred"): transition_processing_erred, ("no-worker", "released"): transition_no_worker_released, ("no-worker", "processing"): transition_no_worker_processing, ("released", "forgotten"): transition_released_forgotten, ("memory", "forgotten"): transition_memory_forgotten, ("erred", "released"): transition_erred_released, ("memory", "released"): transition_memory_released, ("released", "erred"): transition_released_erred, } def story(self, *keys_or_tasks_or_stimuli: str | TaskState) -> list[Transition]: """Get all transitions that touch one of the input keys or stimulus_id's""" keys_or_stimuli = { key.key if isinstance(key, TaskState) else key for key in keys_or_tasks_or_stimuli } return scheduler_story(keys_or_stimuli, self.transition_log) ############################## # Assigning Tasks to Workers # ############################## def is_rootish(self, ts: TaskState) -> bool: """ Whether ``ts`` is a root or root-like task. Root-ish tasks are part of a group that's much larger than the cluster, and have few or no dependencies. """ if ts.resource_restrictions or ts.worker_restrictions or ts.host_restrictions: return False tg = ts.group # TODO short-circuit to True if `not ts.dependencies`? return ( len(tg) > self.total_nthreads * 2 and len(tg.dependencies) < 5 and sum(map(len, tg.dependencies)) < 5 ) def check_idle_saturated(self, ws: WorkerState, occ: float = -1.0) -> None: """Update the status of the idle and saturated state The scheduler keeps track of workers that are .. - Saturated: have enough work to stay busy - Idle: do not have enough work to stay busy They are considered saturated if they both have enough tasks to occupy all of their threads, and if the expected runtime of those tasks is large enough. If ``distributed.scheduler.worker-saturation`` is not ``inf`` (scheduler-side queuing is enabled), they are considered idle if they have fewer tasks processing than the ``worker-saturation`` threshold dictates. Otherwise, they are considered idle if they have fewer tasks processing than threads, or if their tasks' total expected runtime is less than half the expected runtime of the same number of average tasks. This is useful for load balancing and adaptivity. """ if self.total_nthreads == 0 or ws.status == Status.closed: return if occ < 0: occ = ws.occupancy p = len(ws.processing) self.saturated.discard(ws) if ws.status != Status.running: self.idle.pop(ws.address, None) elif self.is_unoccupied(ws, occ, p): self.idle[ws.address] = ws else: self.idle.pop(ws.address, None) nc = ws.nthreads if p > nc: pending = occ * (p - nc) / (p * nc) if 0.4 < pending > 1.9 * (self.total_occupancy / self.total_nthreads): self.saturated.add(ws) if not _worker_full(ws, self.WORKER_SATURATION) and ws.status == Status.running: self.idle_task_count.add(ws) else: self.idle_task_count.discard(ws) def is_unoccupied( self, ws: WorkerState, occupancy: float, nprocessing: int ) -> bool: nthreads = ws.nthreads return ( nprocessing < nthreads or occupancy < nthreads * (self.total_occupancy / self.total_nthreads) / 2 ) def get_comm_cost(self, ts: TaskState, ws: WorkerState) -> float: """ Get the estimated communication cost (in s.) to compute the task on the given worker. """ if 10 * len(ts.dependencies) < len(ws.has_what): # In the common case where the number of dependencies is # much less than the number of tasks that we have, # construct the set of deps that require communication in # O(len(dependencies)) rather than O(len(has_what)) time. # Factor of 10 is a guess at the overhead of explicit # iteration as opposed to just calling set.difference deps = {dep for dep in ts.dependencies if dep not in ws.has_what} else: deps = ts.dependencies.difference(ws.has_what) nbytes = sum(dts.nbytes for dts in deps) return nbytes / self.bandwidth def get_task_duration(self, ts: TaskState) -> float: """Get the estimated computation cost of the given task (not including any communication cost). If no data has been observed, value of `distributed.scheduler.default-task-durations` are used. If none is set for this task, `distributed.scheduler.unknown-task-duration` is used instead. """ duration: float = ts.prefix.duration_average if duration >= 0: return duration s = self.unknown_durations.get(ts.prefix.name) if s is None: self.unknown_durations[ts.prefix.name] = s = set() s.add(ts) return self.UNKNOWN_TASK_DURATION def valid_workers(self, ts: TaskState) -> set[WorkerState] | None: """Return set of currently valid workers for key If all workers are valid then this returns ``None``, in which case any *running* worker can be used. Otherwise, the subset of running workers valid for this task is returned. This checks tracks the following state: * worker_restrictions * host_restrictions * resource_restrictions """ s: set[str] | None = None if ts.worker_restrictions: s = {addr for addr in ts.worker_restrictions if addr in self.workers} if ts.host_restrictions: # Resolve the alias here rather than early, for the worker # may not be connected when host_restrictions is populated hr = [self.coerce_hostname(h) for h in ts.host_restrictions] # XXX need HostState? sl = [] for h in hr: dh = self.host_info.get(h) if dh is not None: sl.append(dh["addresses"]) ss = set.union(*sl) if sl else set() if s is None: s = ss else: s |= ss if ts.resource_restrictions: dw = {} for resource, required in ts.resource_restrictions.items(): dr = self.resources.get(resource) if dr is None: self.resources[resource] = dr = {} sw = set() for addr, supplied in dr.items(): if supplied >= required: sw.add(addr) dw[resource] = sw ww = set.intersection(*dw.values()) if s is None: s = ww else: s &= ww if s is None: return None # All workers are valid if not s: return set() # No workers are valid # Some workers are valid s_ws = {self.workers[addr] for addr in s} if len(self.running) < len(self.workers): s_ws &= self.running return s_ws def acquire_resources(self, ts: TaskState, ws: WorkerState) -> None: for r, required in ts.resource_restrictions.items(): ws.used_resources[r] += required def release_resources(self, ts: TaskState, ws: WorkerState) -> None: for r, required in ts.resource_restrictions.items(): ws.used_resources[r] -= required def coerce_hostname(self, host: Hashable) -> str: """ Coerce the hostname of a worker. """ alias = self.aliases.get(host) if alias is not None: ws = self.workers[alias] return ws.host else: assert isinstance(host, str) return host def worker_objective(self, ts: TaskState, ws: WorkerState) -> tuple: """Objective function to determine which worker should get the task Minimize expected start time. If a tie then break with data storage. """ comm_bytes = sum( dts.get_nbytes() for dts in ts.dependencies if ws not in dts.who_has ) stack_time = ws.occupancy / ws.nthreads start_time = stack_time + comm_bytes / self.bandwidth if ts.actor: return (len(ws.actors), start_time, ws.nbytes) else: return (start_time, ws.nbytes) def add_replica(self, ts: TaskState, ws: WorkerState) -> None: """Note that a worker holds a replica of a task with state='memory'""" ws.add_replica(ts) if len(ts.who_has) == 2: self.replicated_tasks.add(ts) def remove_replica(self, ts: TaskState, ws: WorkerState) -> None: """Note that a worker no longer holds a replica of a task""" ws.remove_replica(ts) if len(ts.who_has) == 1: self.replicated_tasks.remove(ts) def remove_all_replicas(self, ts: TaskState) -> None: """Remove all replicas of a task from all workers""" nbytes = ts.get_nbytes() for ws in ts.who_has: ws.nbytes -= nbytes del ws._has_what[ts] if len(ts.who_has) > 1: self.replicated_tasks.remove(ts) ts.who_has.clear() def bulk_schedule_unrunnable_after_adding_worker(self, ws: WorkerState) -> Recs: """Send ``no-worker`` tasks to ``processing`` that this worker can handle. Returns priority-ordered recommendations. """ runnable: list[TaskState] = [] for ts in self.unrunnable: valid = self.valid_workers(ts) if valid is None or ws in valid: runnable.append(ts) # Recommendations are processed LIFO, hence the reversed order runnable.sort(key=operator.attrgetter("priority"), reverse=True) return {ts.key: "processing" for ts in runnable} def _validate_ready(self, ts: TaskState) -> None: """Validation for ready states (processing, queued, no-worker)""" assert not ts.waiting_on assert not ts.who_has assert not ts.exception_blame assert not ts.processing_on assert not ts.has_lost_dependencies assert ts not in self.unrunnable assert ts not in self.queued assert all(dts.who_has for dts in ts.dependencies) def _add_to_processing(self, ts: TaskState, ws: WorkerState) -> Msgs: """Set a task as processing on a worker and return the worker messages to send""" if self.validate: self._validate_ready(ts) assert ws in self.running, self.running assert (o := self.workers.get(ws.address)) is ws, (ws, o) ws.add_to_processing(ts) ts.processing_on = ws ts.state = "processing" self.acquire_resources(ts, ws) self.check_idle_saturated(ws) self.n_tasks += 1 if ts.actor: ws.actors.add(ts) return {ws.address: [self._task_to_msg(ts)]} def _exit_processing_common(self, ts: TaskState) -> WorkerState | None: """Remove *ts* from the set of processing tasks. Returns ------- Worker state of the worker that processed *ts* if the worker is current, None if the worker is stale. See also -------- Scheduler._set_duration_estimate """ ws = ts.processing_on assert ws ts.processing_on = None ws.remove_from_processing(ts) if self.workers.get(ws.address) is not ws: # may have been removed return None self.check_idle_saturated(ws) self.release_resources(ts, ws) return ws def _add_to_memory( self, ts: TaskState, ws: WorkerState, recommendations: Recs, client_msgs: Msgs, type: bytes | None = None, typename: str | None = None, ) -> None: """Add ts to the set of in-memory tasks""" if self.validate: assert ts not in ws.has_what self.add_replica(ts, ws) deps = list(ts.dependents) if len(deps) > 1: deps.sort(key=operator.attrgetter("priority"), reverse=True) for dts in deps: s = dts.waiting_on if ts in s: s.discard(ts) if not s: # new task ready to run recommendations[dts.key] = "processing" for dts in ts.dependencies: s = dts.waiters s.discard(ts) if not s and not dts.who_wants: recommendations[dts.key] = "released" if not ts.waiters and not ts.who_wants: recommendations[ts.key] = "released" else: report_msg: dict[str, Any] = {"op": "key-in-memory", "key": ts.key} if type is not None: report_msg["type"] = type for cs in ts.who_wants: client_msgs[cs.client_key] = [report_msg] ts.state = "memory" ts.type = typename # type: ignore ts.group.types.add(typename) # type: ignore cs = self.clients["fire-and-forget"] if ts in cs.wants_what: self._client_releases_keys( cs=cs, keys=[ts.key], recommendations=recommendations, ) def _propagate_released(self, ts: TaskState, recommendations: Recs) -> None: ts.state = "released" key = ts.key if ts.has_lost_dependencies: recommendations[key] = "forgotten" elif ts.waiters or ts.who_wants: recommendations[key] = "waiting" if recommendations.get(key) != "waiting": for dts in ts.dependencies: if dts.state != "released": dts.waiters.discard(ts) if not dts.waiters and not dts.who_wants: recommendations[dts.key] = "released" ts.waiters.clear() if self.validate: assert not ts.processing_on assert ts not in self.queued def _propagate_forgotten( self, ts: TaskState, recommendations: Recs, worker_msgs: Msgs, stimulus_id: str, ) -> None: ts.state = "forgotten" for dts in ts.dependents: dts.has_lost_dependencies = True dts.dependencies.remove(ts) dts.waiting_on.discard(ts) if dts.state not in ("memory", "erred"): # Cannot compute task anymore recommendations[dts.key] = "forgotten" ts.dependents.clear() ts.waiters.clear() for dts in ts.dependencies: dts.dependents.remove(ts) dts.waiters.discard(ts) if not dts.dependents and not dts.who_wants: # Task not needed anymore assert dts is not ts recommendations[dts.key] = "forgotten" ts.dependencies.clear() ts.waiting_on.clear() for ws in ts.who_has: if ws.address in self.workers: # in case worker has died worker_msgs[ws.address] = [ { "op": "free-keys", "keys": [ts.key], "stimulus_id": stimulus_id, } ] self.remove_all_replicas(ts) def _client_releases_keys( self, keys: Collection[str], cs: ClientState, recommendations: Recs, ) -> None: """Remove keys from client desired list""" logger.debug("Client %s releases keys: %s", cs.client_key, keys) for key in keys: ts = self.tasks.get(key) if ts is not None and ts in cs.wants_what: cs.wants_what.remove(ts) ts.who_wants.remove(cs) if not ts.who_wants: if not ts.dependents: # No live dependents, can forget recommendations[ts.key] = "forgotten" elif ts.state != "erred" and not ts.waiters: recommendations[ts.key] = "released" def _task_to_msg(self, ts: TaskState, duration: float = -1) -> dict[str, Any]: """Convert a single computational task to a message""" # FIXME: The duration attribute is not used on worker. We could save ourselves the # time to compute and submit this if duration < 0: duration = self.get_task_duration(ts) ts.run_id = next(TaskState._run_id_iterator) msg: dict[str, Any] = { "op": "compute-task", "key": ts.key, "run_id": ts.run_id, "priority": ts.priority, "duration": duration, "stimulus_id": f"compute-task-{time()}", "who_has": { dts.key: [ws.address for ws in dts.who_has] for dts in ts.dependencies }, "nbytes": {dts.key: dts.nbytes for dts in ts.dependencies}, "run_spec": None, "function": None, "args": None, "kwargs": None, "resource_restrictions": ts.resource_restrictions, "actor": ts.actor, "annotations": ts.annotations, } if self.validate: assert all(msg["who_has"].values()) if isinstance(ts.run_spec, dict): msg.update(ts.run_spec) else: msg["run_spec"] = ts.run_spec return msg class Scheduler(SchedulerState, ServerNode): """Dynamic distributed task scheduler The scheduler tracks the current state of workers, data, and computations. The scheduler listens for events and responds by controlling workers appropriately. It continuously tries to use the workers to execute an ever growing dask graph. All events are handled quickly, in linear time with respect to their input (which is often of constant size) and generally within a millisecond. To accomplish this the scheduler tracks a lot of state. Every operation maintains the consistency of this state. The scheduler communicates with the outside world through Comm objects. It maintains a consistent and valid view of the world even when listening to several clients at once. A Scheduler is typically started either with the ``dask scheduler`` executable:: $ dask scheduler Scheduler started at 127.0.0.1:8786 Or within a LocalCluster a Client starts up without connection information:: >>> c = Client() # doctest: +SKIP >>> c.cluster.scheduler # doctest: +SKIP Scheduler(...) Users typically do not interact with the scheduler directly but rather with the client object ``Client``. The ``contact_address`` parameter allows to advertise a specific address to the workers for communication with the scheduler, which is different than the address the scheduler binds to. This is useful when the scheduler listens on a private address, which therefore cannot be used by the workers to contact it. **State** The scheduler contains the following state variables. Each variable is listed along with what it stores and a brief description. * **tasks:** ``{task key: TaskState}`` Tasks currently known to the scheduler * **unrunnable:** ``{TaskState}`` Tasks in the "no-worker" state * **workers:** ``{worker key: WorkerState}`` Workers currently connected to the scheduler * **idle:** ``{WorkerState}``: Set of workers that are not fully utilized * **saturated:** ``{WorkerState}``: Set of workers that are not over-utilized * **host_info:** ``{hostname: dict}``: Information about each worker host * **clients:** ``{client key: ClientState}`` Clients currently connected to the scheduler * **services:** ``{str: port}``: Other services running on this scheduler, like Bokeh * **loop:** ``IOLoop``: The running Tornado IOLoop * **client_comms:** ``{client key: Comm}`` For each client, a Comm object used to receive task requests and report task status updates. * **stream_comms:** ``{worker key: Comm}`` For each worker, a Comm object from which we both accept stimuli and report results * **task_duration:** ``{key-prefix: time}`` Time we expect certain functions to take, e.g. ``{'sum': 0.25}`` """ default_port = 8786 _instances: ClassVar[weakref.WeakSet[Scheduler]] = weakref.WeakSet() def __init__( self, loop=None, delete_interval="500ms", synchronize_worker_interval="60s", services=None, service_kwargs=None, allowed_failures=None, extensions=None, validate=None, scheduler_file=None, security=None, worker_ttl=None, idle_timeout=None, interface=None, host=None, port=0, protocol=None, dashboard_address=None, dashboard=None, http_prefix="/", preload=None, preload_argv=(), plugins=(), contact_address=None, transition_counter_max=False, jupyter=False, **kwargs, ): if loop is not None: warnings.warn( "the loop kwarg to Scheduler is deprecated", DeprecationWarning, stacklevel=2, ) self.loop = self.io_loop = IOLoop.current() self._setup_logging(logger) # Attributes if contact_address is None: contact_address = dask.config.get("distributed.scheduler.contact-address") self.contact_address = contact_address if allowed_failures is None: allowed_failures = dask.config.get("distributed.scheduler.allowed-failures") self.allowed_failures = allowed_failures if validate is None: validate = dask.config.get("distributed.scheduler.validate") self.proc = psutil.Process() self.delete_interval = parse_timedelta(delete_interval, default="ms") self.synchronize_worker_interval = parse_timedelta( synchronize_worker_interval, default="ms" ) self.service_specs = services or {} self.service_kwargs = service_kwargs or {} self.services = {} self.scheduler_file = scheduler_file worker_ttl = worker_ttl or dask.config.get("distributed.scheduler.worker-ttl") self.worker_ttl = parse_timedelta(worker_ttl) if worker_ttl else None idle_timeout = idle_timeout or dask.config.get( "distributed.scheduler.idle-timeout" ) if idle_timeout: self.idle_timeout = parse_timedelta(idle_timeout) else: self.idle_timeout = None self.idle_since = time() self.time_started = self.idle_since # compatibility for dask-gateway self._lock = asyncio.Lock() self.bandwidth_workers = defaultdict(float) self.bandwidth_types = defaultdict(float) if not preload: preload = dask.config.get("distributed.scheduler.preload") if not preload_argv: preload_argv = dask.config.get("distributed.scheduler.preload-argv") self.preloads = preloading.process_preloads(self, preload, preload_argv) 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("scheduler") self.connection_args["handshake_overrides"] = { # common denominator "pickle-protocol": 4 } self._start_address = addresses_from_user_args( host=host, port=port, interface=interface, protocol=protocol, security=security, default_port=self.default_port, ) http_server_modules = dask.config.get("distributed.scheduler.http.routes") show_dashboard = dashboard or (dashboard is None and dashboard_address) # install vanilla route if show_dashboard but bokeh is not installed if show_dashboard: try: import distributed.dashboard.scheduler except ImportError: show_dashboard = False http_server_modules.append("distributed.http.scheduler.missing_bokeh") routes = get_handlers( server=self, modules=http_server_modules, prefix=http_prefix ) self.start_http_server(routes, dashboard_address, default_port=8787) if show_dashboard: distributed.dashboard.scheduler.connect( self.http_application, self.http_server, self, prefix=http_prefix ) self.jupyter = jupyter if self.jupyter: try: from jupyter_server.serverapp import ServerApp except ImportError: raise ImportError( "In order to use the Dask jupyter option you " "need to have jupyterlab installed" ) from traitlets.config import Config j = ServerApp.instance( config=Config( { "ServerApp": { "base_url": "jupyter", # SECURITY: We usually expect the dashboard to be a read-only view into # the scheduler activity. However, by adding an open Jupyter application # we are allowing arbitrary remote code execution on the scheduler via the # dashboard server. This option should only be used when the dashboard is # protected via other means, or when you don't care about cluster security. "token": "", "allow_remote_access": True, } } ) ) j.initialize( new_httpserver=False, argv=[], ) self._jupyter_server_application = j self.http_application.add_application(j.web_app) # Communication state self.client_comms = {} self.stream_comms = {} # Task state tasks = {} self.generation = 0 self._last_client = None self._last_time = 0 unrunnable = set() queued: HeapSet[TaskState] = HeapSet(key=operator.attrgetter("priority")) self.datasets = {} # Prefix-keyed containers # Client state clients = {} # Worker state workers = SortedDict() host_info = {} resources = {} aliases = {} self._worker_collections = [ workers, host_info, resources, aliases, ] self.events = defaultdict( partial( deque, maxlen=dask.config.get("distributed.scheduler.events-log-length") ) ) self.event_counts = defaultdict(int) self.event_subscriber = defaultdict(set) self.worker_plugins = {} self.nanny_plugins = {} worker_handlers = { "task-finished": self.handle_task_finished, "task-erred": self.handle_task_erred, "release-worker-data": self.release_worker_data, "add-keys": self.add_keys, "long-running": self.handle_long_running, "reschedule": self._reschedule, "keep-alive": lambda *args, **kwargs: None, "log-event": self.log_worker_event, "worker-status-change": self.handle_worker_status_change, "request-refresh-who-has": self.handle_request_refresh_who_has, } client_handlers = { "update-graph": self.update_graph, "update-graph-hlg": self.update_graph_hlg, "client-desires-keys": self.client_desires_keys, "update-data": self.update_data, "report-key": self.report_on_key, "client-releases-keys": self.client_releases_keys, "heartbeat-client": self.client_heartbeat, "close-client": self.remove_client, "subscribe-topic": self.subscribe_topic, "unsubscribe-topic": self.unsubscribe_topic, } self.handlers = { "register-client": self.add_client, "scatter": self.scatter, "register-worker": self.add_worker, "register_nanny": self.add_nanny, "unregister": self.remove_worker, "gather": self.gather, "cancel": self.stimulus_cancel, "retry": self.stimulus_retry, "feed": self.feed, "terminate": self.close, "broadcast": self.broadcast, "proxy": self.proxy, "ncores": self.get_ncores, "ncores_running": self.get_ncores_running, "has_what": self.get_has_what, "who_has": self.get_who_has, "processing": self.get_processing, "call_stack": self.get_call_stack, "profile": self.get_profile, "performance_report": self.performance_report, "get_logs": self.get_logs, "logs": self.get_logs, "worker_logs": self.get_worker_logs, "log_event": self.log_event, "events": self.get_events, "nbytes": self.get_nbytes, "versions": self.versions, "add_keys": self.add_keys, "rebalance": self.rebalance, "replicate": self.replicate, "run_function": self.run_function, "restart": self.restart, "update_data": self.update_data, "set_resources": self.add_resources, "retire_workers": self.retire_workers, "get_metadata": self.get_metadata, "set_metadata": self.set_metadata, "set_restrictions": self.set_restrictions, "heartbeat_worker": self.heartbeat_worker, "get_task_status": self.get_task_status, "get_task_stream": self.get_task_stream, "get_task_prefix_states": self.get_task_prefix_states, "register_scheduler_plugin": self.register_scheduler_plugin, "register_worker_plugin": self.register_worker_plugin, "unregister_worker_plugin": self.unregister_worker_plugin, "register_nanny_plugin": self.register_nanny_plugin, "unregister_nanny_plugin": self.unregister_nanny_plugin, "adaptive_target": self.adaptive_target, "workers_to_close": self.workers_to_close, "subscribe_worker_status": self.subscribe_worker_status, "start_task_metadata": self.start_task_metadata, "stop_task_metadata": self.stop_task_metadata, "get_cluster_state": self.get_cluster_state, "dump_cluster_state_to_url": self.dump_cluster_state_to_url, "benchmark_hardware": self.benchmark_hardware, "get_story": self.get_story, } connection_limit = get_fileno_limit() / 2 SchedulerState.__init__( self, aliases=aliases, clients=clients, workers=workers, host_info=host_info, resources=resources, tasks=tasks, unrunnable=unrunnable, queued=queued, validate=validate, plugins=plugins, transition_counter_max=transition_counter_max, ) ServerNode.__init__( self, handlers=self.handlers, stream_handlers=merge(worker_handlers, client_handlers), connection_limit=connection_limit, deserialize=False, connection_args=self.connection_args, **kwargs, ) if self.worker_ttl: pc = PeriodicCallback(self.check_worker_ttl, self.worker_ttl * 1000) self.periodic_callbacks["worker-ttl"] = pc if self.idle_timeout: pc = PeriodicCallback(self.check_idle, self.idle_timeout * 1000 / 4) self.periodic_callbacks["idle-timeout"] = pc if extensions is None: extensions = DEFAULT_EXTENSIONS.copy() if not dask.config.get("distributed.scheduler.work-stealing"): if "stealing" in extensions: del extensions["stealing"] for name, extension in extensions.items(): self.extensions[name] = extension(self) setproctitle("dask scheduler [not started]") Scheduler._instances.add(self) self.rpc.allow_offload = False ################## # Administration # ################## def __repr__(self): return ( f"" ) def _repr_html_(self): return get_template("scheduler.html.j2").render( address=self.address, workers=self.workers, threads=self.total_nthreads, tasks=self.tasks, ) def identity(self): """Basic information about ourselves and our cluster""" d = { "type": type(self).__name__, "id": str(self.id), "address": self.address, "services": {key: v.port for (key, v) in self.services.items()}, "started": self.time_started, "workers": { worker.address: worker.identity() for worker in self.workers.values() }, } return d def _to_dict(self, *, exclude: Container[str] = ()) -> dict: """Dictionary representation for debugging purposes. Not type stable and not intended for roundtrips. See also -------- Server.identity Client.dump_cluster_state distributed.utils.recursive_to_dict """ info = super()._to_dict(exclude=exclude) extra = { "transition_log": self.transition_log, "transition_counter": self.transition_counter, "tasks": self.tasks, "task_groups": self.task_groups, # Overwrite dict of WorkerState.identity from info "workers": self.workers, "clients": self.clients, "memory": self.memory, "events": self.events, "extensions": self.extensions, } extra = {k: v for k, v in extra.items() if k not in exclude} info.update(recursive_to_dict(extra, exclude=exclude)) return info async def get_cluster_state( self, exclude: "Collection[str]", ) -> dict: "Produce the state dict used in a cluster state dump" # Kick off state-dumping on workers before we block the event loop in `self._to_dict`. workers_future = asyncio.gather( self.broadcast( msg={"op": "dump_state", "exclude": exclude}, on_error="return", ), self.broadcast( msg={"op": "versions"}, on_error="ignore", ), ) try: scheduler_state = self._to_dict(exclude=exclude) worker_states, worker_versions = await workers_future finally: # Ensure the tasks aren't left running if anything fails. # Someday (py3.11), use a trio-style TaskGroup for this. workers_future.cancel() # Convert any RPC errors to strings worker_states = { k: repr(v) if isinstance(v, Exception) else v for k, v in worker_states.items() } return { "scheduler": scheduler_state, "workers": worker_states, "versions": {"scheduler": self.versions(), "workers": worker_versions}, } async def dump_cluster_state_to_url( self, url: str, exclude: "Collection[str]", format: Literal["msgpack", "yaml"], **storage_options: dict[str, Any], ) -> None: "Write a cluster state dump to an fsspec-compatible URL." await cluster_dump.write_state( partial(self.get_cluster_state, exclude), url, format, **storage_options ) def get_worker_service_addr( self, worker: str, service_name: str, protocol: bool = False ) -> tuple[str, int] | str | None: """ Get the (host, port) address of the named service on the *worker*. Returns None if the service doesn't exist. Parameters ---------- worker : address service_name : str Common services include 'bokeh' and 'nanny' protocol : boolean Whether or not to include a full address with protocol (True) or just a (host, port) pair """ ws = self.workers[worker] port = ws.services.get(service_name) if port is None: return None elif protocol: return "%(protocol)s://%(host)s:%(port)d" % { "protocol": ws.address.split("://")[0], "host": ws.host, "port": port, } else: return ws.host, port async def start_unsafe(self): """Clear out old state and restart all running coroutines""" await super().start_unsafe() enable_gc_diagnosis() self._clear_task_state() for addr in self._start_address: await self.listen( addr, allow_offload=False, handshake_overrides={"pickle-protocol": 4, "compression": None}, **self.security.get_listen_args("scheduler"), ) self.ip = get_address_host(self.listen_address) listen_ip = self.ip if listen_ip == "0.0.0.0": listen_ip = "" if self.address.startswith("inproc://"): listen_ip = "localhost" # Services listen on all addresses self.start_services(listen_ip) for listener in self.listeners: logger.info(" Scheduler at: %25s", listener.contact_address) for k, v in self.services.items(): logger.info("%11s at: %25s", k, "%s:%d" % (listen_ip, v.port)) if self.scheduler_file: with open(self.scheduler_file, "w") as f: json.dump(self.identity(), f, indent=2) fn = self.scheduler_file # remove file when we close the process def del_scheduler_file(): if os.path.exists(fn): os.remove(fn) weakref.finalize(self, del_scheduler_file) for preload in self.preloads: try: await preload.start() except Exception: logger.exception("Failed to start preload") if self.jupyter: # Allow insecure communications from local users if self.address.startswith("tls://"): await self.listen("tcp://localhost:0") os.environ["DASK_SCHEDULER_ADDRESS"] = self.listeners[-1].contact_address await asyncio.gather( *[plugin.start(self) for plugin in list(self.plugins.values())] ) self.start_periodic_callbacks() setproctitle(f"dask scheduler [{self.address}]") return self async def close(self, fast=None, close_workers=None): """Send cleanup signal to all coroutines then wait until finished See Also -------- Scheduler.cleanup """ if fast is not None or close_workers is not None: warnings.warn( "The 'fast' and 'close_workers' parameters in Scheduler.close have no effect and will be removed in a future version of distributed.", FutureWarning, ) if self.status in (Status.closing, Status.closed): await self.finished() return async def log_errors(func): try: await func() except Exception: logger.exception("Plugin call failed during scheduler.close") await asyncio.gather( *[log_errors(plugin.before_close) for plugin in list(self.plugins.values())] ) self.status = Status.closing logger.info("Scheduler closing...") setproctitle("dask scheduler [closing]") for preload in self.preloads: try: await preload.teardown() except Exception: logger.exception("Failed to tear down preload") await asyncio.gather( *[log_errors(plugin.close) for plugin in list(self.plugins.values())] ) for pc in self.periodic_callbacks.values(): pc.stop() self.periodic_callbacks.clear() self.stop_services() for ext in self.extensions.values(): with suppress(AttributeError): ext.teardown() logger.info("Scheduler closing all comms") futures = [] for _, comm in list(self.stream_comms.items()): # FIXME use `self.remove_worker()` instead after https://github.com/dask/distributed/issues/6390 if not comm.closed(): # This closes the Worker and ensures that if a Nanny is around, # it is closed as well comm.send({"op": "close", "reason": "scheduler-close"}) comm.send({"op": "close-stream"}) # ^ TODO remove? `Worker.close` will close the stream anyway. with suppress(AttributeError): futures.append(comm.close()) await asyncio.gather(*futures) if self.jupyter: await self._jupyter_server_application._cleanup() for comm in self.client_comms.values(): comm.abort() await self.rpc.close() self.status = Status.closed self.stop() await super().close() setproctitle("dask scheduler [closed]") disable_gc_diagnosis() ########### # Stimuli # ########### def heartbeat_worker( self, comm=None, *, address, resolve_address: bool = True, now: float | None = None, resources: dict[str, float] | None = None, host_info: dict | None = None, metrics: dict, executing: dict[str, float] | None = None, extensions: dict | None = None, ) -> dict[str, Any]: address = self.coerce_address(address, resolve_address) address = normalize_address(address) ws = self.workers.get(address) if ws is None: logger.warning(f"Received heartbeat from unregistered worker {address!r}.") return {"status": "missing"} host = get_address_host(address) local_now = time() host_info = host_info or {} dh: dict = self.host_info.setdefault(host, {}) dh["last-seen"] = local_now frac = 1 / len(self.workers) self.bandwidth = ( self.bandwidth * (1 - frac) + metrics["bandwidth"]["total"] * frac ) for other, (bw, count) in metrics["bandwidth"]["workers"].items(): if (address, other) not in self.bandwidth_workers: self.bandwidth_workers[address, other] = bw / count else: alpha = (1 - frac) ** count self.bandwidth_workers[address, other] = self.bandwidth_workers[ address, other ] * alpha + bw * (1 - alpha) for typ, (bw, count) in metrics["bandwidth"]["types"].items(): if typ not in self.bandwidth_types: self.bandwidth_types[typ] = bw / count else: alpha = (1 - frac) ** count self.bandwidth_types[typ] = self.bandwidth_types[typ] * alpha + bw * ( 1 - alpha ) ws.last_seen = local_now if executing is not None: # NOTE: the executing dict is unused ws.executing = {} for key, duration in executing.items(): if key in self.tasks: ts = self.tasks[key] ws.executing[ts] = duration ts.prefix.add_exec_time(duration) ws.metrics = metrics # Calculate RSS - dask keys, separating "old" and "new" usage # See MemoryState for details max_memory_unmanaged_old_hist_age = local_now - self.MEMORY_RECENT_TO_OLD_TIME memory_unmanaged_old = ws._memory_unmanaged_old while ws._memory_unmanaged_history: timestamp, size = ws._memory_unmanaged_history[0] if timestamp >= max_memory_unmanaged_old_hist_age: break ws._memory_unmanaged_history.popleft() if size == memory_unmanaged_old: memory_unmanaged_old = 0 # recalculate min() # ws._nbytes is updated at a different time and sizeof() may not be accurate, # so size may be (temporarily) negative; floor it to zero. size = max( 0, metrics["memory"] - ws.nbytes + metrics["spilled_bytes"]["memory"] ) ws._memory_unmanaged_history.append((local_now, size)) if not memory_unmanaged_old: # The worker has just been started or the previous minimum has been expunged # because too old. # Note: this algorithm is capped to 200 * MEMORY_RECENT_TO_OLD_TIME elements # cluster-wide by heartbeat_interval(), regardless of the number of workers ws._memory_unmanaged_old = min(map(second, ws._memory_unmanaged_history)) elif size < memory_unmanaged_old: ws._memory_unmanaged_old = size if host_info: dh = self.host_info.setdefault(host, {}) dh.update(host_info) if now: ws.time_delay = local_now - now if resources: self.add_resources(worker=address, resources=resources) if extensions: for name, data in extensions.items(): self.extensions[name].heartbeat(ws, data) return { "status": "OK", "time": local_now, "heartbeat-interval": heartbeat_interval(len(self.workers)), } @log_errors async def add_worker( self, comm: Comm, *, address: str, status: str, server_id: str, keys=(), nthreads=None, name=None, resolve_address=True, nbytes=None, types=None, now=None, resources=None, host_info=None, memory_limit=None, metrics=None, pid=0, services=None, local_directory=None, versions: dict[str, Any] | None = None, nanny=None, extra=None, stimulus_id: str, ) -> None: """Add a new worker to the cluster""" address = self.coerce_address(address, resolve_address) address = normalize_address(address) host = get_address_host(address) if address in self.workers: raise ValueError("Worker already exists %s" % address) if nbytes: err = ( f"Worker {address!r} connected with {len(nbytes)} key(s) in memory! Worker reconnection is not supported. " f"Keys: {list(nbytes)}" ) logger.error(err) await comm.write({"status": "error", "message": err, "time": time()}) return if name in self.aliases: logger.warning("Worker tried to connect with a duplicate name: %s", name) msg = { "status": "error", "message": "name taken, %s" % name, "time": time(), } await comm.write(msg) return self.log_event(address, {"action": "add-worker"}) self.log_event("all", {"action": "add-worker", "worker": address}) self.workers[address] = ws = WorkerState( address=address, status=Status.lookup[status], # type: ignore pid=pid, nthreads=nthreads, memory_limit=memory_limit or 0, name=name, local_directory=local_directory, services=services, versions=versions, nanny=nanny, extra=extra, server_id=server_id, scheduler=self, ) if ws.status == Status.running: self.running.add(ws) dh = self.host_info.get(host) if dh is None: self.host_info[host] = dh = {} dh_addresses = dh.get("addresses") if dh_addresses is None: dh["addresses"] = dh_addresses = set() dh["nthreads"] = 0 dh_addresses.add(address) dh["nthreads"] += nthreads self.total_nthreads += nthreads self.aliases[name] = address self.heartbeat_worker( address=address, resolve_address=resolve_address, now=now, resources=resources, host_info=host_info, metrics=metrics, ) # Do not need to adjust self.total_occupancy as self.occupancy[ws] cannot # exist before this. self.check_idle_saturated(ws) # for key in keys: # TODO # self.mark_key_in_memory(key, [address]) self.stream_comms[address] = BatchedSend(interval="5ms", loop=self.loop) for plugin in list(self.plugins.values()): try: result = plugin.add_worker(scheduler=self, worker=address) if result is not None and inspect.isawaitable(result): await result except Exception as e: logger.exception(e) if ws.status == Status.running: self.transitions( self.bulk_schedule_unrunnable_after_adding_worker(ws), stimulus_id ) self.stimulus_queue_slots_maybe_opened(stimulus_id=stimulus_id) logger.info("Register worker %s", ws) msg = { "status": "OK", "time": time(), "heartbeat-interval": heartbeat_interval(len(self.workers)), "worker-plugins": self.worker_plugins, } version_warning = version_module.error_message( version_module.get_versions(), {w: ws.versions for w, ws in self.workers.items()}, versions, source_name=str(ws.server_id), ) msg.update(version_warning) await comm.write(msg) # This will keep running until the worker is removed await self.handle_worker(comm, address) async def add_nanny(self) -> dict[str, Any]: msg = { "status": "OK", "nanny-plugins": self.nanny_plugins, } return msg def update_graph_hlg( self, client=None, hlg=None, keys=None, dependencies=None, restrictions=None, priority=None, loose_restrictions=None, resources=None, submitting_task=None, retries=None, user_priority=0, actors=None, fifo_timeout=0, code=None, ): unpacked_graph = HighLevelGraph.__dask_distributed_unpack__(hlg) dsk = unpacked_graph["dsk"] dependencies = unpacked_graph["deps"] annotations = unpacked_graph["annotations"] # Remove any self-dependencies (happens on test_publish_bag() and others) for k, v in dependencies.items(): deps = set(v) if k in deps: deps.remove(k) dependencies[k] = deps if priority is None: # Removing all non-local keys before calling order() dsk_keys = set(dsk) # intersection() of sets is much faster than dict_keys stripped_deps = { k: v.intersection(dsk_keys) for k, v in dependencies.items() if k in dsk_keys } priority = dask.order.order(dsk, dependencies=stripped_deps) return self.update_graph( client, dsk, keys, dependencies, restrictions, priority, loose_restrictions, resources, submitting_task, retries, user_priority, actors, fifo_timeout, annotations, code=code, stimulus_id=f"update-graph-{time()}", ) def update_graph( self, client=None, tasks=None, keys=None, dependencies=None, restrictions=None, priority=None, loose_restrictions=None, resources=None, submitting_task=None, retries=None, user_priority=0, actors=None, fifo_timeout=0, annotations=None, code=None, stimulus_id=None, ): """ Add new computations to the internal dask graph This happens whenever the Client calls submit, map, get, or compute. """ stimulus_id = stimulus_id or f"update-graph-{time()}" start = time() fifo_timeout = parse_timedelta(fifo_timeout) keys = set(keys) if len(tasks) > 1: self.log_event( ["all", client], {"action": "update_graph", "count": len(tasks)} ) # Remove aliases for k in list(tasks): if tasks[k] is k: del tasks[k] dependencies = dependencies or {} if self.total_occupancy > 1e-9 and self.computations: # Still working on something. Assign new tasks to same computation computation = self.computations[-1] else: computation = Computation() self.computations.append(computation) if code and code not in computation.code: # add new code blocks computation.code.add(code) n = 0 while len(tasks) != n: # walk through new tasks, cancel any bad deps n = len(tasks) for k, deps in list(dependencies.items()): if any( dep not in self.tasks and dep not in tasks for dep in deps ): # bad key logger.info("User asked for computation on lost data, %s", k) del tasks[k] del dependencies[k] if k in keys: keys.remove(k) self.report({"op": "cancelled-key", "key": k}, client=client) self.client_releases_keys( keys=[k], client=client, stimulus_id=stimulus_id ) # Avoid computation that is already finished already_in_memory = set() # tasks that are already done for k, v in dependencies.items(): if v and k in self.tasks: ts = self.tasks[k] if ts.state in ("memory", "erred"): already_in_memory.add(k) if already_in_memory: dependents = dask.core.reverse_dict(dependencies) stack = list(already_in_memory) done = set(already_in_memory) while stack: # remove unnecessary dependencies key = stack.pop() try: deps = dependencies[key] except KeyError: deps = self.tasks[key].dependencies for dep in deps: if dep in dependents: child_deps = dependents[dep] elif dep in self.tasks: child_deps = self.tasks[dep].dependencies else: child_deps = set() if all(d in done for d in child_deps): if dep in self.tasks and dep not in done: done.add(dep) stack.append(dep) for d in done: tasks.pop(d, None) dependencies.pop(d, None) # Get or create task states stack = list(keys) touched_keys = set() touched_tasks = [] while stack: k = stack.pop() if k in touched_keys: continue # XXX Have a method get_task_state(self, k) ? ts = self.tasks.get(k) if ts is None: ts = self.new_task(k, tasks.get(k), "released", computation=computation) elif not ts.run_spec: ts.run_spec = tasks.get(k) touched_keys.add(k) touched_tasks.append(ts) stack.extend(dependencies.get(k, ())) self.client_desires_keys(keys=keys, client=client) # Add dependencies for key, deps in dependencies.items(): ts = self.tasks.get(key) if ts is None or ts.dependencies: continue for dep in deps: dts = self.tasks[dep] ts.add_dependency(dts) # Compute priorities if isinstance(user_priority, Number): user_priority = {k: user_priority for k in tasks} annotations = annotations or {} restrictions = restrictions or {} loose_restrictions = loose_restrictions or [] resources = resources or {} retries = retries or {} # Override existing taxonomy with per task annotations if annotations: if "priority" in annotations: user_priority.update(annotations["priority"]) if "workers" in annotations: restrictions.update(annotations["workers"]) if "allow_other_workers" in annotations: loose_restrictions.extend( k for k, v in annotations["allow_other_workers"].items() if v ) if "retries" in annotations: retries.update(annotations["retries"]) if "resources" in annotations: resources.update(annotations["resources"]) for a, kv in annotations.items(): for k, v in kv.items(): # Tasks might have been culled, in which case # we have nothing to annotate. ts = self.tasks.get(k) if ts is not None: ts.annotations[a] = v # Add actors if actors is True: actors = list(keys) for actor in actors or []: ts = self.tasks[actor] ts.actor = True priority = priority or dask.order.order( tasks ) # TODO: define order wrt old graph if submitting_task: # sub-tasks get better priority than parent tasks ts = self.tasks.get(submitting_task) if ts is not None: generation = ts.priority[0] - 0.01 else: # super-task already cleaned up generation = self.generation elif self._last_time + fifo_timeout < start: self.generation += 1 # older graph generations take precedence generation = self.generation self._last_time = start else: generation = self.generation for key in set(priority) & touched_keys: ts = self.tasks[key] if ts.priority is None: ts.priority = (-(user_priority.get(key, 0)), generation, priority[key]) # Ensure all runnables have a priority runnables = [ts for ts in touched_tasks if ts.run_spec] for ts in runnables: if ts.priority is None and ts.run_spec: ts.priority = (self.generation, 0) if restrictions: # *restrictions* is a dict keying task ids to lists of # restriction specifications (either worker names or addresses) for k, v in restrictions.items(): if v is None: continue ts = self.tasks.get(k) if ts is None: continue ts.host_restrictions = set() ts.worker_restrictions = set() # Make sure `v` is a collection and not a single worker name / address if not isinstance(v, (list, tuple, set)): v = [v] for w in v: try: w = self.coerce_address(w) except ValueError: # Not a valid address, but perhaps it's a hostname ts.host_restrictions.add(w) else: ts.worker_restrictions.add(w) if loose_restrictions: for k in loose_restrictions: ts = self.tasks[k] ts.loose_restrictions = True if resources: for k, v in resources.items(): if v is None: continue assert isinstance(v, dict) ts = self.tasks.get(k) if ts is None: continue ts.resource_restrictions = v if retries: for k, v in retries.items(): assert isinstance(v, int) ts = self.tasks.get(k) if ts is None: continue ts.retries = v # Compute recommendations recommendations: Recs = {} for ts in sorted(runnables, key=operator.attrgetter("priority"), reverse=True): if ts.state == "released" and ts.run_spec: recommendations[ts.key] = "waiting" for ts in touched_tasks: for dts in ts.dependencies: if dts.exception_blame: ts.exception_blame = dts.exception_blame recommendations[ts.key] = "erred" break for plugin in list(self.plugins.values()): try: plugin.update_graph( self, client=client, tasks=tasks, keys=keys, restrictions=restrictions or {}, dependencies=dependencies, priority=priority, loose_restrictions=loose_restrictions, resources=resources, annotations=annotations, ) except Exception as e: logger.exception(e) self.transitions(recommendations, stimulus_id) for ts in touched_tasks: if ts.state in ("memory", "erred"): self.report_on_key(ts=ts, client=client) end = time() self.digest_metric("update-graph-duration", end - start) # TODO: balance workers def stimulus_queue_slots_maybe_opened(self, *, stimulus_id: str) -> None: """Respond to an event which may have opened spots on worker threadpools Selects the appropriate number of tasks from the front of the queue according to the total number of task slots available on workers (potentially 0), and transitions them to ``processing``. Notes ----- Other transitions related to this stimulus should be fully processed beforehand, so any tasks that became runnable are already in ``processing``. Otherwise, overproduction can occur if queued tasks get scheduled before downstream tasks. Must be called after `check_idle_saturated`; i.e. `idle_task_count` must be up to date. """ if not self.queued: return slots_available = sum( _task_slots_available(ws, self.WORKER_SATURATION) for ws in self.idle_task_count ) if slots_available == 0: return recommendations: Recs = {} for qts in self.queued.peekn(slots_available): if self.validate: assert qts.state == "queued", qts.state assert not qts.processing_on, (qts, qts.processing_on) assert not qts.waiting_on, (qts, qts.processing_on) assert qts.who_wants or qts.waiters, qts recommendations[qts.key] = "processing" self.transitions(recommendations, stimulus_id) def stimulus_task_finished(self, key, worker, stimulus_id, run_id, **kwargs): """Mark that a task has finished execution on a particular worker""" logger.debug("Stimulus task finished %s[%d] %s", key, run_id, worker) recommendations: Recs = {} client_msgs: Msgs = {} worker_msgs: Msgs = {} ws: WorkerState = self.workers[worker] ts: TaskState = self.tasks.get(key) if ts is None or ts.state in ("released", "queued"): logger.debug( "Received already computed task, worker: %s, state: %s" ", key: %s, who_has: %s", worker, ts.state if ts else "forgotten", key, ts.who_has if ts else {}, ) worker_msgs[worker] = [ { "op": "free-keys", "keys": [key], "stimulus_id": stimulus_id, } ] elif ts.run_id != run_id: if not ts.processing_on or ts.processing_on.address != worker: logger.debug( "Received stale task run, worker: %s, key: %s, run_id: %d (%d)", worker, key, run_id, ts.run_id, ) worker_msgs[worker] = [ { "op": "free-keys", "keys": [key], "stimulus_id": stimulus_id, } ] else: recommendations[ts.key] = "released" elif ts.state == "memory": self.add_keys(worker=worker, keys=[key]) else: ts.metadata.update(kwargs["metadata"]) r: tuple = self._transition( key, "memory", stimulus_id, worker=worker, **kwargs ) recommendations, client_msgs, worker_msgs = r if ts.state == "memory": assert ws in ts.who_has return recommendations, client_msgs, worker_msgs def stimulus_task_erred( self, key=None, worker=None, exception=None, stimulus_id=None, traceback=None, **kwargs, ): """Mark that a task has erred on a particular worker""" logger.debug("Stimulus task erred %s, %s", key, worker) ts: TaskState = self.tasks.get(key) if ts is None or ts.state != "processing": return {}, {}, {} if ts.retries > 0: ts.retries -= 1 return self._transition(key, "waiting", stimulus_id) else: return self._transition( key, "erred", stimulus_id, cause=key, exception=exception, traceback=traceback, worker=worker, **kwargs, ) def stimulus_retry(self, keys, client=None): logger.info("Client %s requests to retry %d keys", client, len(keys)) if client: self.log_event(client, {"action": "retry", "count": len(keys)}) stack = list(keys) seen = set() roots = [] while stack: key = stack.pop() seen.add(key) ts = self.tasks[key] erred_deps = [dts.key for dts in ts.dependencies if dts.state == "erred"] if erred_deps: stack.extend(erred_deps) else: roots.append(key) recommendations: Recs = {key: "waiting" for key in roots} self.transitions(recommendations, f"stimulus-retry-{time()}") if self.validate: for key in seen: assert not self.tasks[key].exception_blame return tuple(seen) def close_worker(self, worker: str) -> None: """Ask a worker to shut itself down. Do not wait for it to take effect. Note that there is no guarantee that the worker will actually accept the command. Note that :meth:`remove_worker` sends the same command internally if close=True. See also -------- retire_workers remove_worker """ if worker not in self.workers: return logger.info("Closing worker %s", worker) self.log_event(worker, {"action": "close-worker"}) self.worker_send(worker, {"op": "close", "reason": "scheduler-close-worker"}) @log_errors async def remove_worker( self, address: str, *, stimulus_id: str, safe: bool = False, close: bool = True ) -> Literal["OK", "already-removed"]: """Remove worker from cluster. We do this when a worker reports that it plans to leave or when it appears to be unresponsive. This may send its tasks back to a released state. See also -------- retire_workers close_worker """ if self.status == Status.closed: return "already-removed" address = self.coerce_address(address) if address not in self.workers: return "already-removed" host = get_address_host(address) ws: WorkerState = self.workers[address] event_msg = { "action": "remove-worker", "processing-tasks": {ts.key for ts in ws.processing}, } self.log_event(address, event_msg.copy()) event_msg["worker"] = address self.log_event("all", event_msg) logger.info("Remove worker %s", ws) if close: with suppress(AttributeError, CommClosedError): self.stream_comms[address].send( {"op": "close", "reason": "scheduler-remove-worker"} ) self.remove_resources(address) dh: dict = self.host_info[host] dh_addresses: set = dh["addresses"] dh_addresses.remove(address) dh["nthreads"] -= ws.nthreads self.total_nthreads -= ws.nthreads if not dh_addresses: del self.host_info[host] self.rpc.remove(address) del self.stream_comms[address] del self.aliases[ws.name] self.idle.pop(ws.address, None) self.idle_task_count.discard(ws) self.saturated.discard(ws) del self.workers[address] ws.status = Status.closed self.running.discard(ws) recommendations: Recs = {} ts: TaskState for ts in list(ws.processing): k = ts.key recommendations[k] = "released" if not safe: ts.suspicious += 1 ts.prefix.suspicious += 1 if ts.suspicious > self.allowed_failures: del recommendations[k] e = pickle.dumps( KilledWorker( task=k, last_worker=ws.clean(), allowed_failures=self.allowed_failures, ), ) r = self.transition( k, "erred", exception=e, cause=k, stimulus_id=stimulus_id, worker=address, ) recommendations.update(r) logger.info( "Task %s marked as failed because %d workers died" " while trying to run it", ts.key, self.allowed_failures, ) for ts in list(ws.has_what): self.remove_replica(ts, ws) if not ts.who_has: if ts.run_spec: recommendations[ts.key] = "released" else: # pure data recommendations[ts.key] = "forgotten" self.transitions(recommendations, stimulus_id=stimulus_id) for plugin in list(self.plugins.values()): try: result = plugin.remove_worker(scheduler=self, worker=address) if inspect.isawaitable(result): await result except Exception as e: logger.exception(e) if not self.workers: logger.info("Lost all workers") for w in self.workers: self.bandwidth_workers.pop((address, w), None) self.bandwidth_workers.pop((w, address), None) async def remove_worker_from_events(): # If the worker isn't registered anymore after the delay, remove from events if address not in self.workers and address in self.events: del self.events[address] cleanup_delay = parse_timedelta( dask.config.get("distributed.scheduler.events-cleanup-delay") ) self._ongoing_background_tasks.call_later( cleanup_delay, remove_worker_from_events ) logger.debug("Removed worker %s", ws) return "OK" async def stimulus_cancel(self, keys, client, force=False): """Stop execution on a list of keys""" logger.info("Client %s requests to cancel %d keys", client, len(keys)) if client: self.log_event( client, {"action": "cancel", "count": len(keys), "force": force} ) await asyncio.gather( *[self._cancel_key(key, client, force=force) for key in keys] ) async def _cancel_key(self, key, client, force=False): """Cancel a particular key and all dependents""" # TODO: this should be converted to use the transition mechanism ts = self.tasks.get(key) try: cs = self.clients[client] except KeyError: return # no key yet, lets try again in a moment start = time() while ts is None or not ts.who_wants: await asyncio.sleep(0.1) ts = self.tasks.get(key) if time() - start >= 1: return if force or ts.who_wants == {cs}: # no one else wants this key await asyncio.gather( *[ self._cancel_key(dts.key, client, force=force) for dts in ts.dependents ] ) logger.info("Scheduler cancels key %s. Force=%s", key, force) self.report({"op": "cancelled-key", "key": key}) clients = list(ts.who_wants) if force else [cs] for cs in clients: self.client_releases_keys( keys=[key], client=cs.client_key, stimulus_id=f"cancel-key-{time()}" ) def client_desires_keys(self, keys=None, client=None): cs: ClientState = self.clients.get(client) if cs is None: # For publish, queues etc. self.clients[client] = cs = ClientState(client) for k in keys: ts = self.tasks.get(k) if ts is None: # For publish, queues etc. ts = self.new_task(k, None, "released") ts.who_wants.add(cs) cs.wants_what.add(ts) if ts.state in ("memory", "erred"): self.report_on_key(ts=ts, client=client) def client_releases_keys(self, keys=None, client=None, stimulus_id=None): """Remove keys from client desired list""" stimulus_id = stimulus_id or f"client-releases-keys-{time()}" if not isinstance(keys, list): keys = list(keys) cs = self.clients[client] recommendations: Recs = {} self._client_releases_keys(keys=keys, cs=cs, recommendations=recommendations) self.transitions(recommendations, stimulus_id) self.stimulus_queue_slots_maybe_opened(stimulus_id=stimulus_id) def client_heartbeat(self, client=None): """Handle heartbeats from Client""" cs: ClientState = self.clients[client] cs.last_seen = time() ################### # Task Validation # ################### def validate_released(self, key): ts: TaskState = self.tasks[key] assert ts.state == "released" assert not ts.waiters assert not ts.waiting_on assert not ts.who_has assert not ts.processing_on assert not any([ts in dts.waiters for dts in ts.dependencies]) assert ts not in self.unrunnable assert ts not in self.queued def validate_waiting(self, key): ts: TaskState = self.tasks[key] assert ts.waiting_on assert not ts.who_has assert not ts.processing_on assert ts not in self.unrunnable assert ts not in self.queued for dts in ts.dependencies: # We are waiting on a dependency iff it's not stored assert bool(dts.who_has) != (dts in ts.waiting_on) assert ts in dts.waiters # XXX even if dts._who_has? def validate_queued(self, key): ts: TaskState = self.tasks[key] dts: TaskState assert ts in self.queued assert not ts.waiting_on assert not ts.who_has assert not ts.processing_on assert not ( ts.worker_restrictions or ts.host_restrictions or ts.resource_restrictions ) for dts in ts.dependencies: assert dts.who_has assert ts in dts.waiters def validate_processing(self, key): ts: TaskState = self.tasks[key] dts: TaskState assert not ts.waiting_on ws = ts.processing_on assert ws assert ts in ws.processing assert not ts.who_has assert ts not in self.queued for dts in ts.dependencies: assert dts.who_has assert ts in dts.waiters def validate_memory(self, key): ts: TaskState = self.tasks[key] dts: TaskState assert ts.who_has assert bool(ts in self.replicated_tasks) == (len(ts.who_has) > 1) assert not ts.processing_on assert not ts.waiting_on assert ts not in self.unrunnable assert ts not in self.queued for dts in ts.dependents: assert (dts in ts.waiters) == ( dts.state in ("waiting", "queued", "processing", "no-worker") ) assert ts not in dts.waiting_on def validate_no_worker(self, key): ts: TaskState = self.tasks[key] assert ts in self.unrunnable assert not ts.waiting_on assert ts in self.unrunnable assert not ts.processing_on assert not ts.who_has assert ts not in self.queued for dts in ts.dependencies: assert dts.who_has def validate_erred(self, key): ts: TaskState = self.tasks[key] assert ts.exception_blame assert not ts.who_has assert ts not in self.queued def validate_key(self, key, ts: TaskState | None = None): try: if ts is None: ts = self.tasks.get(key) if ts is None: logger.debug("Key lost: %s", key) else: ts.validate() try: func = getattr(self, "validate_" + ts.state.replace("-", "_")) except AttributeError: logger.error( "self.validate_%s not found", ts.state.replace("-", "_") ) else: func(key) except Exception as e: logger.exception(e) if LOG_PDB: import pdb pdb.set_trace() raise def validate_state(self, allow_overlap: bool = False) -> None: validate_state(self.tasks, self.workers, self.clients) if not (set(self.workers) == set(self.stream_comms)): raise ValueError("Workers not the same in all collections") assert self.running.issuperset(self.idle.values()), ( self.running.copy(), set(self.idle.values()), ) assert self.running.issuperset(self.idle_task_count), ( self.running.copy(), self.idle_task_count.copy(), ) assert self.running.issuperset(self.saturated), ( self.running.copy(), self.saturated.copy(), ) assert self.saturated.isdisjoint(self.idle.values()), ( self.saturated.copy(), set(self.idle.values()), ) task_prefix_counts: defaultdict[str, int] = defaultdict(int) for w, ws in self.workers.items(): assert isinstance(w, str), (type(w), w) assert isinstance(ws, WorkerState), (type(ws), ws) assert ws.address == w if ws.status == Status.running: assert ws in self.running else: assert ws not in self.running assert ws.address not in self.idle assert ws not in self.saturated assert ws.long_running.issubset(ws.processing) if not ws.processing: assert not ws.occupancy if ws.status == Status.running: assert ws.address in self.idle assert not ws.needs_what.keys() & ws.has_what actual_needs_what: defaultdict[TaskState, int] = defaultdict(int) for ts in ws.processing: for tss in ts.dependencies: if tss not in ws.has_what: actual_needs_what[tss] += 1 assert actual_needs_what == ws.needs_what assert (ws.status == Status.running) == (ws in self.running) for name, count in ws.task_prefix_count.items(): task_prefix_counts[name] += count assert task_prefix_counts.keys() == self._task_prefix_count_global.keys() for name, global_count in self._task_prefix_count_global.items(): assert ( task_prefix_counts[name] == global_count ), f"{name}: {task_prefix_counts[name]} (wss), {global_count} (global)" for ws in self.running: assert ws.status == Status.running assert ws.address in self.workers for k, ts in self.tasks.items(): assert isinstance(ts, TaskState), (type(ts), ts) assert ts.key == k assert bool(ts in self.replicated_tasks) == (len(ts.who_has) > 1) self.validate_key(k, ts) for ts in self.replicated_tasks: assert ts.state == "memory" assert ts.key in self.tasks for c, cs in self.clients.items(): # client=None is often used in tests... assert c is None or type(c) == str, (type(c), c) assert type(cs) == ClientState, (type(cs), cs) assert cs.client_key == c a = {w: ws.nbytes for w, ws in self.workers.items()} b = { w: sum(ts.get_nbytes() for ts in ws.has_what) for w, ws in self.workers.items() } assert a == b, (a, b) if self.transition_counter_max: assert self.transition_counter < self.transition_counter_max ################### # Manage Messages # ################### def report(self, msg: dict, ts: TaskState | None = None, client: str | None = None): """ Publish updates to all listening Queues and Comms If the message contains a key then we only send the message to those comms that care about the key. """ if ts is None: msg_key = msg.get("key") if msg_key is not None: tasks: dict = self.tasks ts = tasks.get(msg_key) client_comms: dict = self.client_comms if ts is None: # Notify all clients client_keys = list(client_comms) elif client is None: # Notify clients interested in key client_keys = [cs.client_key for cs in ts.who_wants] else: # Notify clients interested in key (including `client`) client_keys = [ cs.client_key for cs in ts.who_wants if cs.client_key != client ] client_keys.append(client) k: str for k in client_keys: c = client_comms.get(k) if c is None: continue try: c.send(msg) # logger.debug("Scheduler sends message to client %s", msg) except CommClosedError: if self.status == Status.running: logger.critical( "Closed comm %r while trying to write %s", c, msg, exc_info=True ) async def add_client( self, comm: Comm, client: str, versions: dict[str, Any] ) -> None: """Add client to network We listen to all future messages from this Comm. """ assert client is not None comm.name = "Scheduler->Client" logger.info("Receive client connection: %s", client) self.log_event(["all", client], {"action": "add-client", "client": client}) self.clients[client] = ClientState(client, versions=versions) for plugin in list(self.plugins.values()): try: plugin.add_client(scheduler=self, client=client) except Exception as e: logger.exception(e) try: bcomm = BatchedSend(interval="2ms", loop=self.loop) bcomm.start(comm) self.client_comms[client] = bcomm msg = {"op": "stream-start"} version_warning = version_module.error_message( version_module.get_versions(), {w: ws.versions for w, ws in self.workers.items()}, versions, ) msg.update(version_warning) bcomm.send(msg) try: await self.handle_stream(comm=comm, extra={"client": client}) finally: self.remove_client(client=client, stimulus_id=f"remove-client-{time()}") logger.debug("Finished handling client %s", client) finally: if not comm.closed(): self.client_comms[client].send({"op": "stream-closed"}) try: if not sys.is_finalizing(): await self.client_comms[client].close() del self.client_comms[client] if self.status == Status.running: logger.info("Close client connection: %s", client) except TypeError: # comm becomes None during GC pass def remove_client(self, client: str, stimulus_id: str | None = None) -> None: """Remove client from network""" stimulus_id = stimulus_id or f"remove-client-{time()}" if self.status == Status.running: logger.info("Remove client %s", client) self.log_event(["all", client], {"action": "remove-client", "client": client}) try: cs: ClientState = self.clients[client] except KeyError: # XXX is this a legitimate condition? pass else: self.client_releases_keys( keys=[ts.key for ts in cs.wants_what], client=cs.client_key, stimulus_id=stimulus_id, ) del self.clients[client] for plugin in list(self.plugins.values()): try: plugin.remove_client(scheduler=self, client=client) except Exception as e: logger.exception(e) async def remove_client_from_events(): # If the client isn't registered anymore after the delay, remove from events if client not in self.clients and client in self.events: del self.events[client] cleanup_delay = parse_timedelta( dask.config.get("distributed.scheduler.events-cleanup-delay") ) if not self._ongoing_background_tasks.closed: self._ongoing_background_tasks.call_later( cleanup_delay, remove_client_from_events ) def send_task_to_worker(self, worker, ts: TaskState, duration: float = -1): """Send a single computational task to a worker""" try: msg: dict = self._task_to_msg(ts, duration) self.worker_send(worker, msg) except Exception as e: logger.exception(e) if LOG_PDB: import pdb pdb.set_trace() raise def handle_uncaught_error(self, **msg): logger.exception(clean_exception(**msg)[1]) def handle_task_finished( self, key: str, worker: str, stimulus_id: str, **msg ) -> None: if worker not in self.workers: return validate_key(key) r: tuple = self.stimulus_task_finished( key=key, worker=worker, stimulus_id=stimulus_id, **msg ) recommendations, client_msgs, worker_msgs = r self._transitions(recommendations, client_msgs, worker_msgs, stimulus_id) self.send_all(client_msgs, worker_msgs) self.stimulus_queue_slots_maybe_opened(stimulus_id=stimulus_id) def handle_task_erred(self, key: str, stimulus_id: str, **msg) -> None: r: tuple = self.stimulus_task_erred(key=key, stimulus_id=stimulus_id, **msg) recommendations, client_msgs, worker_msgs = r self._transitions(recommendations, client_msgs, worker_msgs, stimulus_id) self.send_all(client_msgs, worker_msgs) self.stimulus_queue_slots_maybe_opened(stimulus_id=stimulus_id) def release_worker_data(self, key: str, worker: str, stimulus_id: str) -> None: ts = self.tasks.get(key) ws = self.workers.get(worker) if not ts or not ws or ws not in ts.who_has: return self.remove_replica(ts, ws) if not ts.who_has: self.transitions({key: "released"}, stimulus_id) def handle_long_running( self, key: str, worker: str, compute_duration: float | None, stimulus_id: str ) -> None: """A task has seceded from the thread pool We stop the task from being stolen in the future, and change task duration accounting as if the task has stopped. """ if key not in self.tasks: logger.debug("Skipping long_running since key %s was already released", key) return ts = self.tasks[key] steal = self.extensions.get("stealing") if steal is not None: steal.remove_key_from_stealable(ts) ws = ts.processing_on if ws is None: logger.debug("Received long-running signal from duplicate task. Ignoring.") return if compute_duration is not None: old_duration = ts.prefix.duration_average if old_duration < 0: ts.prefix.duration_average = compute_duration else: ts.prefix.duration_average = (old_duration + compute_duration) / 2 ws.add_to_long_running(ts) self.check_idle_saturated(ws) self.stimulus_queue_slots_maybe_opened(stimulus_id=stimulus_id) def handle_worker_status_change( self, status: str | Status, worker: str | WorkerState, stimulus_id: str ) -> None: ws = self.workers.get(worker) if isinstance(worker, str) else worker if not ws: return prev_status = ws.status ws.status = Status[status] if isinstance(status, str) else status if ws.status == prev_status: return self.log_event( ws.address, { "action": "worker-status-change", "prev-status": prev_status.name, "status": ws.status.name, }, ) logger.debug(f"Worker status {prev_status.name} -> {status} - {ws}") if ws.status == Status.running: self.running.add(ws) self.check_idle_saturated(ws) self.transitions( self.bulk_schedule_unrunnable_after_adding_worker(ws), stimulus_id ) self.stimulus_queue_slots_maybe_opened(stimulus_id=stimulus_id) else: self.running.discard(ws) self.idle.pop(ws.address, None) self.idle_task_count.discard(ws) self.saturated.discard(ws) async def handle_request_refresh_who_has( self, keys: Iterable[str], worker: str, stimulus_id: str ) -> None: """Asynchronous request (through bulk comms) from a Worker to refresh the who_has for some keys. Not to be confused with scheduler.who_has, which is a synchronous RPC request from a Client. """ who_has = {} free_keys = [] for key in keys: if key in self.tasks: who_has[key] = [ws.address for ws in self.tasks[key].who_has] else: free_keys.append(key) if who_has: self.stream_comms[worker].send( { "op": "refresh-who-has", "who_has": who_has, "stimulus_id": stimulus_id, } ) if free_keys: self.stream_comms[worker].send( { "op": "free-keys", "keys": free_keys, "stimulus_id": stimulus_id, } ) async def handle_worker(self, comm: Comm, worker: str) -> None: """ Listen to responses from a single worker This is the main loop for scheduler-worker interaction See Also -------- Scheduler.handle_client: Equivalent coroutine for clients """ comm.name = "Scheduler connection to worker" worker_comm = self.stream_comms[worker] worker_comm.start(comm) logger.info("Starting worker compute stream, %s", worker) try: await self.handle_stream(comm=comm, extra={"worker": worker}) finally: if worker in self.stream_comms: worker_comm.abort() await self.remove_worker( worker, stimulus_id=f"handle-worker-cleanup-{time()}" ) def add_plugin( self, plugin: SchedulerPlugin, *, idempotent: bool = False, name: str | None = None, **kwargs, ): """Add external plugin to scheduler. See https://distributed.readthedocs.io/en/latest/plugins.html Parameters ---------- plugin : SchedulerPlugin SchedulerPlugin instance to add idempotent : bool If true, the plugin is assumed to already exist and no action is taken. name : str A name for the plugin, if None, the name attribute is checked on the Plugin instance and generated if not discovered. """ if name is None: name = _get_plugin_name(plugin) if name in self.plugins: if idempotent: return warnings.warn( f"Scheduler already contains a plugin with name {name}; overwriting.", category=UserWarning, ) self.plugins[name] = plugin def remove_plugin( self, name: str | None = None, plugin: SchedulerPlugin | None = None, ) -> None: """Remove external plugin from scheduler Parameters ---------- name : str Name of the plugin to remove """ assert name is not None try: del self.plugins[name] except KeyError: raise ValueError( f"Could not find plugin {name!r} among the current scheduler plugins" ) async def register_scheduler_plugin(self, plugin, name=None, idempotent=None): """Register a plugin on the scheduler.""" if not dask.config.get("distributed.scheduler.pickle"): raise ValueError( "Cannot register a scheduler plugin as the scheduler " "has been explicitly disallowed from deserializing " "arbitrary bytestrings using pickle via the " "'distributed.scheduler.pickle' configuration setting." ) if not isinstance(plugin, SchedulerPlugin): plugin = loads(plugin) if name is None: name = _get_plugin_name(plugin) if name in self.plugins and idempotent: return if hasattr(plugin, "start"): result = plugin.start(self) if inspect.isawaitable(result): await result self.add_plugin(plugin, name=name, idempotent=idempotent) def worker_send(self, worker: str, msg: dict[str, Any]) -> None: """Send message to worker This also handles connection failures by adding a callback to remove the worker on the next cycle. """ stream_comms: dict = self.stream_comms try: stream_comms[worker].send(msg) except (CommClosedError, AttributeError): self._ongoing_background_tasks.call_soon( self.remove_worker, address=worker, stimulus_id=f"worker-send-comm-fail-{time()}", ) def client_send(self, client, msg): """Send message to client""" client_comms: dict = self.client_comms c = client_comms.get(client) if c is None: return try: c.send(msg) except CommClosedError: if self.status == Status.running: logger.critical( "Closed comm %r while trying to write %s", c, msg, exc_info=True ) def send_all(self, client_msgs: Msgs, worker_msgs: Msgs) -> None: """Send messages to client and workers""" for client, msgs in client_msgs.items(): c = self.client_comms.get(client) if c is None: continue try: c.send(*msgs) except CommClosedError: if self.status == Status.running: logger.critical( "Closed comm %r while trying to write %s", c, msgs, exc_info=True, ) for worker, msgs in worker_msgs.items(): try: w = self.stream_comms[worker] w.send(*msgs) except KeyError: # worker already gone pass except (CommClosedError, AttributeError): self._ongoing_background_tasks.call_soon( self.remove_worker, address=worker, stimulus_id=f"send-all-comm-fail-{time()}", ) ############################ # Less common interactions # ############################ async def scatter( self, comm=None, data=None, workers=None, client=None, broadcast=False, timeout=2, ): """Send data out to workers See also -------- Scheduler.broadcast: """ start = time() while True: if workers is None: wss = self.running else: workers = [self.coerce_address(w) for w in workers] wss = {self.workers[w] for w in workers} wss = {ws for ws in wss if ws.status == Status.running} if wss: break if time() > start + timeout: raise TimeoutError("No valid workers found") await asyncio.sleep(0.1) nthreads = {ws.address: ws.nthreads for ws in wss} assert isinstance(data, dict) keys, who_has, nbytes = await scatter_to_workers(nthreads, data, rpc=self.rpc) self.update_data(who_has=who_has, nbytes=nbytes, client=client) if broadcast: n = len(nthreads) if broadcast is True else broadcast await self.replicate(keys=keys, workers=workers, n=n) self.log_event( [client, "all"], {"action": "scatter", "client": client, "count": len(data)} ) return keys async def gather(self, keys, serializers=None): """Collect data from workers to the scheduler""" stimulus_id = f"gather-{time()}" keys = list(keys) who_has = {} for key in keys: ts: TaskState = self.tasks.get(key) if ts is not None: who_has[key] = [ws.address for ws in ts.who_has] else: who_has[key] = [] data, missing_keys, missing_workers = await gather_from_workers( who_has, rpc=self.rpc, close=False, serializers=serializers ) if not missing_keys: result = {"status": "OK", "data": data} else: missing_states = [ (self.tasks[key].state if key in self.tasks else None) for key in missing_keys ] logger.exception( "Couldn't gather keys %s state: %s workers: %s", missing_keys, missing_states, missing_workers, ) result = {"status": "error", "keys": missing_keys} with log_errors(): # Remove suspicious workers from the scheduler and shut them down. await asyncio.gather( *( self.remove_worker( address=worker, close=True, stimulus_id=stimulus_id ) for worker in missing_workers ) ) for key, workers in missing_keys.items(): logger.exception( "Shut down workers that don't have promised key: %s, %s", str(workers), str(key), ) self.log_event("all", {"action": "gather", "count": len(keys)}) return result @log_errors async def restart(self, client=None, timeout=30, wait_for_workers=True): """ Restart all workers. Reset local state. Optionally wait for workers to return. Workers without nannies are shut down, hoping an external deployment system will restart them. Therefore, if not using nannies and your deployment system does not automatically restart workers, ``restart`` will just shut down all workers, then time out! After ``restart``, all connected workers are new, regardless of whether ``TimeoutError`` was raised. Any workers that failed to shut down in time are removed, and may or may not shut down on their own in the future. Parameters ---------- timeout: How long to wait for workers to shut down and come back, if ``wait_for_workers`` is True, otherwise just how long to wait for workers to shut down. Raises ``asyncio.TimeoutError`` if this is exceeded. wait_for_workers: Whether to wait for all workers to reconnect, or just for them to shut down (default True). Use ``restart(wait_for_workers=False)`` combined with :meth:`Client.wait_for_workers` for granular control over how many workers to wait for. See also -------- Client.restart Client.restart_workers """ stimulus_id = f"restart-{time()}" logger.info("Restarting workers and releasing all keys.") for cs in self.clients.values(): self.client_releases_keys( keys=[ts.key for ts in cs.wants_what], client=cs.client_key, stimulus_id=stimulus_id, ) self._clear_task_state() assert not self.tasks self.report({"op": "restart"}) for plugin in list(self.plugins.values()): try: plugin.restart(self) except Exception as e: logger.exception(e) n_workers = len(self.workers) nanny_workers = { addr: ws.nanny for addr, ws in self.workers.items() if ws.nanny } # Close non-Nanny workers. We have no way to restart them, so we just let them go, # and assume a deployment system is going to restart them for us. await asyncio.gather( *( self.remove_worker(address=addr, stimulus_id=stimulus_id) for addr in self.workers if addr not in nanny_workers ) ) logger.debug("Send kill signal to nannies: %s", nanny_workers) async with contextlib.AsyncExitStack() as stack: nannies = await asyncio.gather( *( stack.enter_async_context( rpc(nanny_address, connection_args=self.connection_args) ) for nanny_address in nanny_workers.values() ) ) start = monotonic() resps = await asyncio.gather( *( asyncio.wait_for( # FIXME does not raise if the process fails to shut down, # see https://github.com/dask/distributed/pull/6427/files#r894917424 # NOTE: Nanny will automatically restart worker process when it's killed nanny.kill( reason="scheduler-restart", timeout=timeout, ), timeout, ) for nanny in nannies ), return_exceptions=True, ) # NOTE: the `WorkerState` entries for these workers will be removed # naturally when they disconnect from the scheduler. # Remove any workers that failed to shut down, so we can guarantee # that after `restart`, there are no old workers around. bad_nannies = [ addr for addr, resp in zip(nanny_workers, resps) if resp is not None ] if bad_nannies: await asyncio.gather( *( self.remove_worker(addr, stimulus_id=stimulus_id) for addr in bad_nannies ) ) raise TimeoutError( f"{len(bad_nannies)}/{len(nannies)} nanny worker(s) did not shut down within {timeout}s" ) self.log_event([client, "all"], {"action": "restart", "client": client}) if wait_for_workers: while len(self.workers) < n_workers: # NOTE: if new (unrelated) workers join while we're waiting, we may return before # our shut-down workers have come back up. That's fine; workers are interchangeable. if monotonic() < start + timeout: await asyncio.sleep(0.2) else: msg = ( f"Waited for {n_workers} worker(s) to reconnect after restarting, " f"but after {timeout}s, only {len(self.workers)} have returned. " "Consider a longer timeout, or `wait_for_workers=False`." ) if (n_nanny := len(nanny_workers)) < n_workers: msg += ( f" The {n_workers - n_nanny} worker(s) not using Nannies were just shut " "down instead of restarted (restart is only possible with Nannies). If " "your deployment system does not automatically re-launch terminated " "processes, then those workers will never come back, and `Client.restart` " "will always time out. Do not use `Client.restart` in that case." ) raise TimeoutError(msg) from None logger.info("Restarting finished.") async def broadcast( self, comm=None, *, msg: dict, workers: "list[str] | None" = None, hosts: "list[str] | None" = None, nanny: bool = False, serializers=None, on_error: "Literal['raise', 'return', 'return_pickle', 'ignore']" = "raise", ) -> dict: # dict[str, Any] """Broadcast message to workers, return all results""" if workers is None: if hosts is None: workers = list(self.workers) else: workers = [] if hosts is not None: for host in hosts: dh: dict = self.host_info.get(host) # type: ignore if dh is not None: workers.extend(dh["addresses"]) # TODO replace with worker_list if nanny: addresses = [self.workers[w].nanny for w in workers] else: addresses = workers ERROR = object() async def send_message(addr): try: comm = await self.rpc.connect(addr) comm.name = "Scheduler Broadcast" try: resp = await send_recv( comm, close=True, serializers=serializers, **msg ) finally: self.rpc.reuse(addr, comm) return resp except Exception as e: logger.error(f"broadcast to {addr} failed: {e.__class__.__name__}: {e}") if on_error == "raise": raise elif on_error == "return": return e elif on_error == "return_pickle": return dumps(e) elif on_error == "ignore": return ERROR else: raise ValueError( "on_error must be 'raise', 'return', 'return_pickle', " f"or 'ignore'; got {on_error!r}" ) results = await All( [send_message(address) for address in addresses if address is not None] ) return {k: v for k, v in zip(workers, results) if v is not ERROR} async def proxy(self, comm=None, msg=None, worker=None, serializers=None): """Proxy a communication through the scheduler to some other worker""" d = await self.broadcast( comm=comm, msg=msg, workers=[worker], serializers=serializers ) return d[worker] async def gather_on_worker( self, worker_address: str, who_has: "dict[str, list[str]]" ) -> set: """Peer-to-peer copy of keys from multiple workers to a single worker Parameters ---------- worker_address: str Recipient worker address to copy keys to who_has: dict[Hashable, list[str]] {key: [sender address, sender address, ...], key: ...} Returns ------- returns: set of keys that failed to be copied """ try: result = await retry_operation( self.rpc(addr=worker_address).gather, who_has=who_has ) except OSError as e: # This can happen e.g. if the worker is going through controlled shutdown; # it doesn't necessarily mean that it went unexpectedly missing logger.warning( f"Communication with worker {worker_address} failed during " f"replication: {e.__class__.__name__}: {e}" ) return set(who_has) ws = self.workers.get(worker_address) if not ws: logger.warning(f"Worker {worker_address} lost during replication") return set(who_has) elif result["status"] == "OK": keys_failed = set() keys_ok: Set = who_has.keys() elif result["status"] == "partial-fail": keys_failed = set(result["keys"]) keys_ok = who_has.keys() - keys_failed logger.warning( f"Worker {worker_address} failed to acquire keys: {result['keys']}" ) else: # pragma: nocover raise ValueError(f"Unexpected message from {worker_address}: {result}") for key in keys_ok: ts: TaskState = self.tasks.get(key) # type: ignore if ts is None or ts.state != "memory": logger.warning(f"Key lost during replication: {key}") continue if ws not in ts.who_has: self.add_replica(ts, ws) return keys_failed async def delete_worker_data( self, worker_address: str, keys: "Collection[str]", stimulus_id: str ) -> None: """Delete data from a worker and update the corresponding worker/task states Parameters ---------- worker_address: str Worker address to delete keys from keys: list[str] List of keys to delete on the specified worker """ try: await retry_operation( self.rpc(addr=worker_address).free_keys, keys=list(keys), stimulus_id=f"delete-data-{time()}", ) except OSError as e: # This can happen e.g. if the worker is going through controlled shutdown; # it doesn't necessarily mean that it went unexpectedly missing logger.warning( f"Communication with worker {worker_address} failed during " f"replication: {e.__class__.__name__}: {e}" ) return ws = self.workers.get(worker_address) if not ws: return for key in keys: ts: TaskState = self.tasks.get(key) # type: ignore if ts is not None and ws in ts.who_has: assert ts.state == "memory" self.remove_replica(ts, ws) if not ts.who_has: # Last copy deleted self.transitions({key: "released"}, stimulus_id) self.log_event(ws.address, {"action": "remove-worker-data", "keys": keys}) @log_errors async def rebalance( self, comm=None, keys: Iterable[str] | None = None, workers: Iterable[str] | None = None, stimulus_id: str | None = None, ) -> dict: """Rebalance keys so that each worker ends up with roughly the same process memory (managed+unmanaged). .. warning:: This operation is generally not well tested against normal operation of the scheduler. It is not recommended to use it while waiting on computations. **Algorithm** #. Find the mean occupancy of the cluster, defined as data managed by dask + unmanaged process memory that has been there for at least 30 seconds (``distributed.worker.memory.recent-to-old-time``). This lets us ignore temporary spikes caused by task heap usage. Alternatively, you may change how memory is measured both for the individual workers as well as to calculate the mean through ``distributed.worker.memory.rebalance.measure``. Namely, this can be useful to disregard inaccurate OS memory measurements. #. Discard workers whose occupancy is within 5% of the mean cluster occupancy (``distributed.worker.memory.rebalance.sender-recipient-gap`` / 2). This helps avoid data from bouncing around the cluster repeatedly. #. Workers above the mean are senders; those below are recipients. #. Discard senders whose absolute occupancy is below 30% (``distributed.worker.memory.rebalance.sender-min``). In other words, no data is moved regardless of imbalancing as long as all workers are below 30%. #. Discard recipients whose absolute occupancy is above 60% (``distributed.worker.memory.rebalance.recipient-max``). Note that this threshold by default is the same as ``distributed.worker.memory.target`` to prevent workers from accepting data and immediately spilling it out to disk. #. Iteratively pick the sender and recipient that are farthest from the mean and move the *least recently inserted* key between the two, until either all senders or all recipients fall within 5% of the mean. A recipient will be skipped if it already has a copy of the data. In other words, this method does not degrade replication. A key will be skipped if there are no recipients available with enough memory to accept the key and that don't already hold a copy. The least recently insertd (LRI) policy is a greedy choice with the advantage of being O(1), trivial to implement (it relies on python dict insertion-sorting) and hopefully good enough in most cases. Discarded alternative policies were: - Largest first. O(n*log(n)) save for non-trivial additional data structures and risks causing the largest chunks of data to repeatedly move around the cluster like pinballs. - Least recently used (LRU). This information is currently available on the workers only and not trivial to replicate on the scheduler; transmitting it over the network would be very expensive. Also, note that dask will go out of its way to minimise the amount of time intermediate keys are held in memory, so in such a case LRI is a close approximation of LRU. Parameters ---------- keys: optional allowlist of dask keys that should be considered for moving. All other keys will be ignored. Note that this offers no guarantee that a key will actually be moved (e.g. because it is unnecessary or because there are no viable recipient workers for it). workers: optional allowlist of workers addresses to be considered as senders or recipients. All other workers will be ignored. The mean cluster occupancy will be calculated only using the allowed workers. """ stimulus_id = stimulus_id or f"rebalance-{time()}" wss: Collection[WorkerState] if workers is not None: wss = [self.workers[w] for w in workers] else: wss = self.workers.values() if not wss: return {"status": "OK"} if keys is not None: if not isinstance(keys, Set): keys = set(keys) # unless already a set-like if not keys: return {"status": "OK"} missing_data = [ k for k in keys if k not in self.tasks or not self.tasks[k].who_has ] if missing_data: return {"status": "partial-fail", "keys": missing_data} msgs = self._rebalance_find_msgs(keys, wss) if not msgs: return {"status": "OK"} async with self._lock: result = await self._rebalance_move_data(msgs, stimulus_id) if result["status"] == "partial-fail" and keys is None: # Only return failed keys if the client explicitly asked for them result = {"status": "OK"} return result def _rebalance_find_msgs( self, keys: Set[Hashable] | None, workers: Iterable[WorkerState], ) -> list[tuple[WorkerState, WorkerState, TaskState]]: """Identify workers that need to lose keys and those that can receive them, together with how many bytes each needs to lose/receive. Then, pair a sender worker with a recipient worker for each key, until the cluster is rebalanced. This method only defines the work to be performed; it does not start any network transfers itself. The big-O complexity is O(wt + ke*log(we)), where - wt is the total number of workers on the cluster (or the number of allowed workers, if explicitly stated by the user) - we is the number of workers that are eligible to be senders or recipients - kt is the total number of keys on the cluster (or on the allowed workers) - ke is the number of keys that need to be moved in order to achieve a balanced cluster There is a degenerate edge case O(wt + kt*log(we)) when kt is much greater than the number of allowed keys, or when most keys are replicated or cannot be moved for some other reason. Returns list of tuples to feed into _rebalance_move_data: - sender worker - recipient worker - task to be transferred """ # Heaps of workers, managed by the heapq module, that need to send/receive data, # with how many bytes each needs to send/receive. # # Each element of the heap is a tuple constructed as follows: # - snd_bytes_max/rec_bytes_max: maximum number of bytes to send or receive. # This number is negative, so that the workers farthest from the cluster mean # are at the top of the smallest-first heaps. # - snd_bytes_min/rec_bytes_min: minimum number of bytes after sending/receiving # which the worker should not be considered anymore. This is also negative. # - arbitrary unique number, there just to to make sure that WorkerState objects # are never used for sorting in the unlikely event that two processes have # exactly the same number of bytes allocated. # - WorkerState # - iterator of all tasks in memory on the worker (senders only), insertion # sorted (least recently inserted first). # Note that this iterator will typically *not* be exhausted. It will only be # exhausted if, after moving away from the worker all keys that can be moved, # is insufficient to drop snd_bytes_min above 0. senders: list[tuple[int, int, int, WorkerState, Iterator[TaskState]]] = [] recipients: list[tuple[int, int, int, WorkerState]] = [] # Output: [(sender, recipient, task), ...] msgs: list[tuple[WorkerState, WorkerState, TaskState]] = [] # By default, this is the optimistic memory, meaning total process memory minus # unmanaged memory that appeared over the last 30 seconds # (distributed.worker.memory.recent-to-old-time). # This lets us ignore temporary spikes caused by task heap usage. memory_by_worker = [ (ws, getattr(ws.memory, self.MEMORY_REBALANCE_MEASURE)) for ws in workers ] mean_memory = sum(m for _, m in memory_by_worker) // len(memory_by_worker) for ws, ws_memory in memory_by_worker: if ws.memory_limit: half_gap = int(self.MEMORY_REBALANCE_HALF_GAP * ws.memory_limit) sender_min = self.MEMORY_REBALANCE_SENDER_MIN * ws.memory_limit recipient_max = self.MEMORY_REBALANCE_RECIPIENT_MAX * ws.memory_limit else: half_gap = 0 sender_min = 0.0 recipient_max = math.inf if ( ws._has_what and ws_memory >= mean_memory + half_gap and ws_memory >= sender_min ): # This may send the worker below sender_min (by design) snd_bytes_max = mean_memory - ws_memory # negative snd_bytes_min = snd_bytes_max + half_gap # negative # See definition of senders above senders.append( (snd_bytes_max, snd_bytes_min, id(ws), ws, iter(ws._has_what)) ) elif ws_memory < mean_memory - half_gap and ws_memory < recipient_max: # This may send the worker above recipient_max (by design) rec_bytes_max = ws_memory - mean_memory # negative rec_bytes_min = rec_bytes_max + half_gap # negative # See definition of recipients above recipients.append((rec_bytes_max, rec_bytes_min, id(ws), ws)) # Fast exit in case no transfers are necessary or possible if not senders or not recipients: self.log_event( "all", { "action": "rebalance", "senders": len(senders), "recipients": len(recipients), "moved_keys": 0, }, ) return [] heapq.heapify(senders) heapq.heapify(recipients) while senders and recipients: snd_bytes_max, snd_bytes_min, _, snd_ws, ts_iter = senders[0] # Iterate through tasks in memory, least recently inserted first for ts in ts_iter: if keys is not None and ts.key not in keys: continue nbytes = ts.nbytes if nbytes + snd_bytes_max > 0: # Moving this task would cause the sender to go below mean and # potentially risk becoming a recipient, which would cause tasks to # bounce around. Move on to the next task of the same sender. continue # Find the recipient, farthest from the mean, which # 1. has enough available RAM for this task, and # 2. doesn't hold a copy of this task already # There may not be any that satisfies these conditions; in this case # this task won't be moved. skipped_recipients = [] use_recipient = False while recipients and not use_recipient: rec_bytes_max, rec_bytes_min, _, rec_ws = recipients[0] if nbytes + rec_bytes_max > 0: # recipients are sorted by rec_bytes_max. # The next ones will be worse; no reason to continue iterating break use_recipient = ts not in rec_ws._has_what if not use_recipient: skipped_recipients.append(heapq.heappop(recipients)) for recipient in skipped_recipients: heapq.heappush(recipients, recipient) if not use_recipient: # This task has no recipients available. Leave it on the sender and # move on to the next task of the same sender. continue # Schedule task for transfer from sender to recipient msgs.append((snd_ws, rec_ws, ts)) # *_bytes_max/min are all negative for heap sorting snd_bytes_max += nbytes snd_bytes_min += nbytes rec_bytes_max += nbytes rec_bytes_min += nbytes # Stop iterating on the tasks of this sender for now and, if it still # has bytes to lose, push it back into the senders heap; it may or may # not come back on top again. if snd_bytes_min < 0: # See definition of senders above heapq.heapreplace( senders, (snd_bytes_max, snd_bytes_min, id(snd_ws), snd_ws, ts_iter), ) else: heapq.heappop(senders) # If recipient still has bytes to gain, push it back into the recipients # heap; it may or may not come back on top again. if rec_bytes_min < 0: # See definition of recipients above heapq.heapreplace( recipients, (rec_bytes_max, rec_bytes_min, id(rec_ws), rec_ws), ) else: heapq.heappop(recipients) # Move to next sender with the most data to lose. # It may or may not be the same sender again. break else: # for ts in ts_iter # Exhausted tasks on this sender heapq.heappop(senders) return msgs async def _rebalance_move_data( self, msgs: "list[tuple[WorkerState, WorkerState, TaskState]]", stimulus_id: str ) -> dict: """Perform the actual transfer of data across the network in rebalance(). Takes in input the output of _rebalance_find_msgs(), that is a list of tuples: - sender worker - recipient worker - task to be transferred FIXME this method is not robust when the cluster is not idle. """ to_recipients: defaultdict[str, defaultdict[str, list[str]]] = defaultdict( lambda: defaultdict(list) ) for snd_ws, rec_ws, ts in msgs: to_recipients[rec_ws.address][ts.key].append(snd_ws.address) failed_keys_by_recipient = dict( zip( to_recipients, await asyncio.gather( *( # Note: this never raises exceptions self.gather_on_worker(w, who_has) for w, who_has in to_recipients.items() ) ), ) ) to_senders = defaultdict(list) for snd_ws, rec_ws, ts in msgs: if ts.key not in failed_keys_by_recipient[rec_ws.address]: to_senders[snd_ws.address].append(ts.key) # Note: this never raises exceptions await asyncio.gather( *(self.delete_worker_data(r, v, stimulus_id) for r, v in to_senders.items()) ) for r, v in to_recipients.items(): self.log_event(r, {"action": "rebalance", "who_has": v}) self.log_event( "all", { "action": "rebalance", "senders": valmap(len, to_senders), "recipients": valmap(len, to_recipients), "moved_keys": len(msgs), }, ) missing_keys = {k for r in failed_keys_by_recipient.values() for k in r} if missing_keys: return {"status": "partial-fail", "keys": list(missing_keys)} else: return {"status": "OK"} async def replicate( self, comm=None, keys=None, n=None, workers=None, branching_factor=2, delete=True, lock=True, stimulus_id=None, ): """Replicate data throughout cluster This performs a tree copy of the data throughout the network individually on each piece of data. Parameters ---------- keys: Iterable list of keys to replicate n: int Number of replications we expect to see within the cluster branching_factor: int, optional The number of workers that can copy data in each generation. The larger the branching factor, the more data we copy in a single step, but the more a given worker risks being swamped by data requests. See also -------- Scheduler.rebalance """ stimulus_id = stimulus_id or f"replicate-{time()}" assert branching_factor > 0 async with self._lock if lock else empty_context: if workers is not None: workers = {self.workers[w] for w in self.workers_list(workers)} workers = {ws for ws in workers if ws.status == Status.running} else: workers = self.running if n is None: n = len(workers) else: n = min(n, len(workers)) if n == 0: raise ValueError("Can not use replicate to delete data") tasks = {self.tasks[k] for k in keys} missing_data = [ts.key for ts in tasks if not ts.who_has] if missing_data: return {"status": "partial-fail", "keys": missing_data} # Delete extraneous data if delete: del_worker_tasks = defaultdict(set) for ts in tasks: del_candidates = tuple(ts.who_has & workers) if len(del_candidates) > n: for ws in random.sample( del_candidates, len(del_candidates) - n ): del_worker_tasks[ws].add(ts) # Note: this never raises exceptions await asyncio.gather( *[ self.delete_worker_data( ws.address, [t.key for t in tasks], stimulus_id ) for ws, tasks in del_worker_tasks.items() ] ) # Copy not-yet-filled data while tasks: gathers = defaultdict(dict) for ts in list(tasks): if ts.state == "forgotten": # task is no longer needed by any client or dependent task tasks.remove(ts) continue n_missing = n - len(ts.who_has & workers) if n_missing <= 0: # Already replicated enough tasks.remove(ts) continue count = min(n_missing, branching_factor * len(ts.who_has)) assert count > 0 for ws in random.sample(tuple(workers - ts.who_has), count): gathers[ws.address][ts.key] = [ wws.address for wws in ts.who_has ] await asyncio.gather( *( # Note: this never raises exceptions self.gather_on_worker(w, who_has) for w, who_has in gathers.items() ) ) for r, v in gathers.items(): self.log_event(r, {"action": "replicate-add", "who_has": v}) self.log_event( "all", { "action": "replicate", "workers": list(workers), "key-count": len(keys), "branching-factor": branching_factor, }, ) def workers_to_close( self, comm=None, memory_ratio: int | float | None = None, n: int | None = None, key: Callable[[WorkerState], Hashable] | bytes | None = None, minimum: int | None = None, target: int | None = None, attribute: str = "address", ) -> list[str]: """ Find workers that we can close with low cost This returns a list of workers that are good candidates to retire. These workers are not running anything and are storing relatively little data relative to their peers. If all workers are idle then we still maintain enough workers to have enough RAM to store our data, with a comfortable buffer. This is for use with systems like ``distributed.deploy.adaptive``. Parameters ---------- memory_ratio : Number Amount of extra space we want to have for our stored data. Defaults to 2, or that we want to have twice as much memory as we currently have data. n : int Number of workers to close minimum : int Minimum number of workers to keep around key : Callable(WorkerState) An optional callable mapping a WorkerState object to a group affiliation. Groups will be closed together. This is useful when closing workers must be done collectively, such as by hostname. target : int Target number of workers to have after we close attribute : str The attribute of the WorkerState object to return, like "address" or "name". Defaults to "address". Examples -------- >>> scheduler.workers_to_close() ['tcp://192.168.0.1:1234', 'tcp://192.168.0.2:1234'] Group workers by hostname prior to closing >>> scheduler.workers_to_close(key=lambda ws: ws.host) ['tcp://192.168.0.1:1234', 'tcp://192.168.0.1:4567'] Remove two workers >>> scheduler.workers_to_close(n=2) Keep enough workers to have twice as much memory as we we need. >>> scheduler.workers_to_close(memory_ratio=2) Returns ------- to_close: list of worker addresses that are OK to close See Also -------- Scheduler.retire_workers """ if target is not None and n is None: n = len(self.workers) - target if n is not None: if n < 0: n = 0 target = len(self.workers) - n if n is None and memory_ratio is None: memory_ratio = 2 with log_errors(): if not n and all([ws.processing for ws in self.workers.values()]): return [] if key is None: key = operator.attrgetter("address") if isinstance(key, bytes) and dask.config.get( "distributed.scheduler.pickle" ): key = pickle.loads(key) groups = groupby(key, self.workers.values()) limit_bytes = { k: sum(ws.memory_limit for ws in v) for k, v in groups.items() } group_bytes = {k: sum(ws.nbytes for ws in v) for k, v in groups.items()} limit = sum(limit_bytes.values()) total = sum(group_bytes.values()) def _key(group): is_idle = not any([wws.processing for wws in groups[group]]) bytes = -group_bytes[group] return is_idle, bytes idle = sorted(groups, key=_key) to_close = [] n_remain = len(self.workers) while idle: group = idle.pop() if n is None and any([ws.processing for ws in groups[group]]): break if minimum and n_remain - len(groups[group]) < minimum: break limit -= limit_bytes[group] if ( n is not None and n_remain - len(groups[group]) >= (target or 0) ) or (memory_ratio is not None and limit >= memory_ratio * total): to_close.append(group) n_remain -= len(groups[group]) else: break result = [getattr(ws, attribute) for g in to_close for ws in groups[g]] if result: logger.debug("Suggest closing workers: %s", result) return result @log_errors async def retire_workers( self, workers: list[str] | None = None, *, names: list | None = None, close_workers: bool = False, remove: bool = True, stimulus_id: str | None = None, **kwargs: Any, ) -> dict[str, Any]: """Gracefully retire workers from cluster. Any key that is in memory exclusively on the retired workers is replicated somewhere else. Parameters ---------- workers: list[str] (optional) List of worker addresses to retire. names: list (optional) List of worker names to retire. Mutually exclusive with ``workers``. If neither ``workers`` nor ``names`` are provided, we call ``workers_to_close`` which finds a good set. close_workers: bool (defaults to False) Whether or not to actually close the worker explicitly from here. Otherwise we expect some external job scheduler to finish off the worker. remove: bool (defaults to True) Whether or not to remove the worker metadata immediately or else wait for the worker to contact us. If close_workers=False and remove=False, this method just flushes the tasks in memory out of the workers and then returns. If close_workers=True and remove=False, this method will return while the workers are still in the cluster, although they won't accept new tasks. If close_workers=False or for whatever reason a worker doesn't accept the close command, it will be left permanently unable to accept new tasks and it is expected to be closed in some other way. **kwargs: dict Extra options to pass to workers_to_close to determine which workers we should drop Returns ------- Dictionary mapping worker ID/address to dictionary of information about that worker for each retired worker. If there are keys that exist in memory only on the workers being retired and it was impossible to replicate them somewhere else (e.g. because there aren't any other running workers), the workers holding such keys won't be retired and won't appear in the returned dict. See Also -------- Scheduler.workers_to_close """ stimulus_id = stimulus_id or f"retire-workers-{time()}" # This lock makes retire_workers, rebalance, and replicate mutually # exclusive and will no longer be necessary once rebalance and replicate are # migrated to the Active Memory Manager. # Note that, incidentally, it also prevents multiple calls to retire_workers # from running in parallel - this is unnecessary. async with self._lock: if names is not None: if workers is not None: raise TypeError("names and workers are mutually exclusive") if names: logger.info("Retire worker names %s", names) # Support cases where names are passed through a CLI and become # strings names_set = {str(name) for name in names} wss = {ws for ws in self.workers.values() if str(ws.name) in names_set} elif workers is not None: wss = { self.workers[address] for address in workers if address in self.workers } else: wss = { self.workers[address] for address in self.workers_to_close(**kwargs) } if not wss: return {} stop_amm = False amm: ActiveMemoryManagerExtension = self.extensions["amm"] if not amm.running: amm = ActiveMemoryManagerExtension( self, policies=set(), register=False, start=True, interval=2.0 ) stop_amm = True try: coros = [] for ws in wss: logger.info("Retiring worker %s", ws.address) policy = RetireWorker(ws.address) amm.add_policy(policy) # Change Worker.status to closing_gracefully. Immediately set # the same on the scheduler to prevent race conditions. prev_status = ws.status self.handle_worker_status_change( Status.closing_gracefully, ws, stimulus_id ) # FIXME: We should send a message to the nanny first; # eventually workers won't be able to close their own nannies. self.stream_comms[ws.address].send( { "op": "worker-status-change", "status": ws.status.name, "stimulus_id": stimulus_id, } ) coros.append( self._track_retire_worker( ws, policy, prev_status=prev_status, close=close_workers, remove=remove, stimulus_id=stimulus_id, ) ) # Give the AMM a kick, in addition to its periodic running. This is # to avoid unnecessarily waiting for a potentially arbitrarily long # time (depending on interval settings) amm.run_once() workers_info = dict(await asyncio.gather(*coros)) workers_info.pop(None, None) finally: if stop_amm: amm.stop() self.log_event("all", {"action": "retire-workers", "workers": workers_info}) self.log_event(list(workers_info), {"action": "retired"}) return workers_info async def _track_retire_worker( self, ws: WorkerState, policy: RetireWorker, prev_status: Status, close: bool, remove: bool, stimulus_id: str, ) -> tuple: # tuple[str | None, dict] while not policy.done(): # Sleep 0.01s when there are 4 tasks or less # Sleep 0.5s when there are 200 or more poll_interval = max(0.01, min(0.5, len(ws.has_what) / 400)) await asyncio.sleep(poll_interval) if policy.no_recipients: # Abort retirement. This time we don't need to worry about race # conditions and we can wait for a scheduler->worker->scheduler # round-trip. self.stream_comms[ws.address].send( { "op": "worker-status-change", "status": prev_status.name, "stimulus_id": stimulus_id, } ) return None, {} logger.debug( "All unique keys on worker %s have been replicated elsewhere", ws.address ) if remove: await self.remove_worker( ws.address, safe=True, close=close, stimulus_id=stimulus_id ) elif close: self.close_worker(ws.address) logger.info("Retired worker %s", ws.address) return ws.address, ws.identity() def add_keys(self, worker=None, keys=(), stimulus_id=None): """ Learn that a worker has certain keys This should not be used in practice and is mostly here for legacy reasons. However, it is sent by workers from time to time. """ if worker not in self.workers: return "not found" ws: WorkerState = self.workers[worker] redundant_replicas = [] for key in keys: ts: TaskState = self.tasks.get(key) if ts is not None and ts.state == "memory": if ws not in ts.who_has: self.add_replica(ts, ws) else: redundant_replicas.append(key) if redundant_replicas: if not stimulus_id: stimulus_id = f"redundant-replicas-{time()}" self.worker_send( worker, { "op": "remove-replicas", "keys": redundant_replicas, "stimulus_id": stimulus_id, }, ) return "OK" @log_errors def update_data( self, *, who_has: dict, nbytes: dict, client=None, ): """ Learn that new data has entered the network from an external source See Also -------- Scheduler.mark_key_in_memory """ who_has = {k: [self.coerce_address(vv) for vv in v] for k, v in who_has.items()} logger.debug("Update data %s", who_has) for key, workers in who_has.items(): ts = self.tasks.get(key) if ts is None: ts = self.new_task(key, None, "memory") ts.state = "memory" ts_nbytes = nbytes.get(key, -1) if ts_nbytes >= 0: ts.set_nbytes(ts_nbytes) for w in workers: ws = self.workers[w] if ws not in ts.who_has: self.add_replica(ts, ws) self.report({"op": "key-in-memory", "key": key, "workers": list(workers)}) if client: self.client_desires_keys(keys=list(who_has), client=client) @overload def report_on_key(self, key: str, *, client: str | None = None) -> None: ... @overload def report_on_key(self, *, ts: TaskState, client: str | None = None) -> None: ... def report_on_key(self, key=None, *, ts=None, client=None): if (ts is None) == (key is None): raise ValueError( # pragma: nocover f"ts and key are mutually exclusive; received {key=!r}, {ts=!r}" ) if ts is None: assert key is not None ts = self.tasks.get(key) else: key = ts.key if ts is not None: report_msg = _task_to_report_msg(ts) else: report_msg = {"op": "cancelled-key", "key": key} if report_msg is not None: self.report(report_msg, ts=ts, client=client) @log_errors async def feed( self, comm, function=None, setup=None, teardown=None, interval="1s", **kwargs ): """ Provides a data Comm to external requester Caution: this runs arbitrary Python code on the scheduler. This should eventually be phased out. It is mostly used by diagnostics. """ if not dask.config.get("distributed.scheduler.pickle"): logger.warning( "Tried to call 'feed' route with custom functions, but " "pickle is disallowed. Set the 'distributed.scheduler.pickle'" "config value to True to use the 'feed' route (this is mostly " "commonly used with progress bars)" ) return interval = parse_timedelta(interval) if function: function = pickle.loads(function) if setup: setup = pickle.loads(setup) if teardown: teardown = pickle.loads(teardown) state = setup(self) if setup else None if inspect.isawaitable(state): state = await state try: while self.status == Status.running: if state is None: response = function(self) else: response = function(self, state) await comm.write(response) await asyncio.sleep(interval) except OSError: pass finally: if teardown: teardown(self, state) def log_worker_event( self, worker: str, topic: str | Collection[str], msg: Any ) -> None: if isinstance(msg, dict): msg["worker"] = worker self.log_event(topic, msg) def subscribe_worker_status(self, comm=None): WorkerStatusPlugin(self, comm) ident = self.identity() for v in ident["workers"].values(): del v["metrics"] del v["last_seen"] return ident def get_processing(self, workers=None): if workers is not None: workers = set(map(self.coerce_address, workers)) return {w: [ts.key for ts in self.workers[w].processing] for w in workers} else: return { w: [ts.key for ts in ws.processing] for w, ws in self.workers.items() } def get_who_has(self, keys: Iterable[str] | None = None) -> dict[str, list[str]]: if keys is not None: return { key: [ws.address for ws in self.tasks[key].who_has] if key in self.tasks else [] for key in keys } else: return { key: [ws.address for ws in ts.who_has] for key, ts in self.tasks.items() } def get_has_what(self, workers=None): if workers is not None: workers = map(self.coerce_address, workers) return { w: [ts.key for ts in self.workers[w].has_what] if w in self.workers else [] for w in workers } else: return {w: [ts.key for ts in ws.has_what] for w, ws in self.workers.items()} def get_ncores(self, workers=None): if workers is not None: workers = map(self.coerce_address, workers) return {w: self.workers[w].nthreads for w in workers if w in self.workers} else: return {w: ws.nthreads for w, ws in self.workers.items()} def get_ncores_running(self, workers=None): ncores = self.get_ncores(workers=workers) return { w: n for w, n in ncores.items() if self.workers[w].status == Status.running } async def get_call_stack(self, keys=None): if keys is not None: stack = list(keys) processing = set() while stack: key = stack.pop() ts = self.tasks[key] if ts.state == "waiting": stack.extend([dts.key for dts in ts.dependencies]) elif ts.state == "processing": processing.add(ts) workers = defaultdict(list) for ts in processing: if ts.processing_on: workers[ts.processing_on.address].append(ts.key) else: workers = {w: None for w in self.workers} if not workers: return {} results = await asyncio.gather( *(self.rpc(w).call_stack(keys=v) for w, v in workers.items()) ) response = {w: r for w, r in zip(workers, results) if r} return response async def benchmark_hardware(self) -> "dict[str, dict[str, float]]": """ Run a benchmark on the workers for memory, disk, and network bandwidths Returns ------- result: dict A dictionary mapping the names "disk", "memory", and "network" to dictionaries mapping sizes to bandwidths. These bandwidths are averaged over many workers running computations across the cluster. """ out: dict[str, defaultdict[str, list[float]]] = { name: defaultdict(list) for name in ["disk", "memory", "network"] } # disk result = await self.broadcast(msg={"op": "benchmark_disk"}) for d in result.values(): for size, duration in d.items(): out["disk"][size].append(duration) # memory result = await self.broadcast(msg={"op": "benchmark_memory"}) for d in result.values(): for size, duration in d.items(): out["memory"][size].append(duration) # network workers = list(self.workers) # On an adaptive cluster, if multiple workers are started on the same physical host, # they are more likely to connect to the Scheduler in sequence, ending up next to # each other in this list. # The transfer speed within such clusters of workers will be effectively that of # localhost. This could happen across different VMs and/or docker images, so # implementing logic based on IP addresses would not necessarily help. # Randomize the connections to even out the mean measures. random.shuffle(workers) futures = [ self.rpc(a).benchmark_network(address=b) for a, b in partition(2, workers) ] responses = await asyncio.gather(*futures) for d in responses: for size, duration in d.items(): out["network"][size].append(duration) result = {} for mode in out: result[mode] = { size: sum(durations) / len(durations) for size, durations in out[mode].items() } return result @log_errors def get_nbytes(self, keys=None, summary=True): if keys is not None: result = {k: self.tasks[k].nbytes for k in keys} else: result = {k: ts.nbytes for k, ts in self.tasks.items() if ts.nbytes >= 0} if summary: out = defaultdict(lambda: 0) for k, v in result.items(): out[key_split(k)] += v result = dict(out) return result def run_function(self, comm, function, args=(), kwargs=None, wait=True): """Run a function within this process See Also -------- Client.run_on_scheduler """ from distributed.worker import run if not dask.config.get("distributed.scheduler.pickle"): raise ValueError( "Cannot run function as the scheduler has been explicitly disallowed from " "deserializing arbitrary bytestrings using pickle via the " "'distributed.scheduler.pickle' configuration setting." ) kwargs = kwargs or {} self.log_event("all", {"action": "run-function", "function": function}) return run(self, comm, function=function, args=args, kwargs=kwargs, wait=wait) def set_metadata(self, keys: list[str], value: object = None) -> None: metadata = self.task_metadata for key in keys[:-1]: if key not in metadata or not isinstance(metadata[key], (dict, list)): metadata[key] = {} metadata = metadata[key] metadata[keys[-1]] = value def get_metadata(self, keys: list[str], default=no_default): metadata = self.task_metadata try: for key in keys: metadata = metadata[key] return metadata except KeyError: if default != no_default: return default else: raise def set_restrictions(self, worker: dict[str, Collection[str] | str]): for key, restrictions in worker.items(): ts = self.tasks[key] if isinstance(restrictions, str): restrictions = {restrictions} ts.worker_restrictions = set(restrictions) @log_errors def get_task_prefix_states(self): state = {} for tp in self.task_prefixes.values(): active_states = tp.active_states if any( active_states.get(s) for s in {"memory", "erred", "released", "processing", "waiting"} ): state[tp.name] = { "memory": active_states["memory"], "erred": active_states["erred"], "released": active_states["released"], "processing": active_states["processing"], "waiting": active_states["waiting"], } return state def get_task_status(self, keys=None): return { key: (self.tasks[key].state if key in self.tasks else None) for key in keys } def get_task_stream(self, start=None, stop=None, count=None): from distributed.diagnostics.task_stream import TaskStreamPlugin if TaskStreamPlugin.name not in self.plugins: self.add_plugin(TaskStreamPlugin(self)) plugin = self.plugins[TaskStreamPlugin.name] return plugin.collect(start=start, stop=stop, count=count) def start_task_metadata(self, name=None): plugin = CollectTaskMetaDataPlugin(scheduler=self, name=name) self.add_plugin(plugin) def stop_task_metadata(self, name=None): plugins = [ p for p in list(self.plugins.values()) if isinstance(p, CollectTaskMetaDataPlugin) and p.name == name ] if len(plugins) != 1: raise ValueError( "Expected to find exactly one CollectTaskMetaDataPlugin " f"with name {name} but found {len(plugins)}." ) plugin = plugins[0] self.remove_plugin(name=plugin.name) return {"metadata": plugin.metadata, "state": plugin.state} async def register_worker_plugin(self, comm, plugin, name=None): """Registers a worker plugin on all running and future workers""" self.worker_plugins[name] = plugin responses = await self.broadcast( msg=dict(op="plugin-add", plugin=plugin, name=name) ) return responses async def unregister_worker_plugin(self, comm, name): """Unregisters a worker plugin""" try: self.worker_plugins.pop(name) except KeyError: raise ValueError(f"The worker plugin {name} does not exist") responses = await self.broadcast(msg=dict(op="plugin-remove", name=name)) return responses async def register_nanny_plugin(self, comm, plugin, name=None): """Registers a setup function, and call it on every worker""" self.nanny_plugins[name] = plugin responses = await self.broadcast( msg=dict(op="plugin_add", plugin=plugin, name=name), nanny=True, ) return responses async def unregister_nanny_plugin(self, comm, name): """Unregisters a worker plugin""" try: self.nanny_plugins.pop(name) except KeyError: raise ValueError(f"The nanny plugin {name} does not exist") responses = await self.broadcast( msg=dict(op="plugin_remove", name=name), nanny=True ) return responses def transition( self, key: str, finish: TaskStateState, stimulus_id: str, **kwargs: Any, ) -> Recs: """Transition a key from its current state to the finish state Examples -------- >>> self.transition('x', 'waiting') {'x': 'processing'} Returns ------- Dictionary of recommendations for future transitions See Also -------- Scheduler.transitions: transitive version of this function """ recommendations, client_msgs, worker_msgs = self._transition( key, finish, stimulus_id, **kwargs ) self.send_all(client_msgs, worker_msgs) return recommendations def transitions(self, recommendations: Recs, stimulus_id: str) -> None: """Process transitions until none are left This includes feedback from previous transitions and continues until we reach a steady state """ client_msgs: Msgs = {} worker_msgs: Msgs = {} self._transitions(recommendations, client_msgs, worker_msgs, stimulus_id) self.send_all(client_msgs, worker_msgs) async def get_story(self, keys_or_stimuli: Iterable[str]) -> list[Transition]: """RPC hook for :meth:`SchedulerState.story`. Note that the msgpack serialization/deserialization round-trip will transform the :class:`Transition` namedtuples into regular tuples. """ return self.story(*keys_or_stimuli) def _reschedule( self, key: str, worker: str | None = None, *, stimulus_id: str ) -> None: """Reschedule a task. This function should only be used when the task has already been released in some way on the worker it's assigned to — either via cancellation or a Reschedule exception — and you are certain the worker will not send any further updates about the task to the scheduler. """ try: ts = self.tasks[key] except KeyError: logger.warning( f"Attempting to reschedule task {key}, which was not " "found on the scheduler. Aborting reschedule." ) return if ts.state != "processing": return if worker and ts.processing_on and ts.processing_on.address != worker: return # transition_processing_released will immediately suggest an additional # transition to waiting if the task has any waiters or clients holding a future. self.transitions({key: "released"}, stimulus_id=stimulus_id) ##################### # Utility functions # ##################### def add_resources(self, worker: str, resources=None): ws: WorkerState = self.workers[worker] if resources: ws.resources.update(resources) ws.used_resources = {} for resource, quantity in ws.resources.items(): ws.used_resources[resource] = 0 dr = self.resources.get(resource, None) if dr is None: self.resources[resource] = dr = {} dr[worker] = quantity return "OK" def remove_resources(self, worker): ws: WorkerState = self.workers[worker] for resource in ws.resources: dr: dict = self.resources.get(resource, None) if dr is None: self.resources[resource] = dr = {} del dr[worker] def coerce_address(self, addr, resolve=True): """ Coerce possible input addresses to canonical form. *resolve* can be disabled for testing with fake hostnames. Handles strings, tuples, or aliases. """ # XXX how many address-parsing routines do we have? if addr in self.aliases: addr = self.aliases[addr] if isinstance(addr, tuple): addr = unparse_host_port(*addr) if not isinstance(addr, str): raise TypeError(f"addresses should be strings or tuples, got {addr!r}") if resolve: addr = resolve_address(addr) else: addr = normalize_address(addr) return addr def workers_list(self, workers): """ List of qualifying workers Takes a list of worker addresses or hostnames. Returns a list of all worker addresses that match """ if workers is None: return list(self.workers) out = set() for w in workers: if ":" in w: out.add(w) else: out.update({ww for ww in self.workers if w in ww}) # TODO: quadratic return list(out) async def get_profile( self, comm=None, workers=None, scheduler=False, server=False, merge_workers=True, start=None, stop=None, key=None, ): if workers is None: workers = self.workers else: workers = set(self.workers) & set(workers) if scheduler: return profile.get_profile(self.io_loop.profile, start=start, stop=stop) results = await asyncio.gather( *( self.rpc(w).profile(start=start, stop=stop, key=key, server=server) for w in workers ), return_exceptions=True, ) results = [r for r in results if not isinstance(r, Exception)] if merge_workers: response = profile.merge(*results) else: response = dict(zip(workers, results)) return response async def get_profile_metadata( self, workers: "Iterable[str] | None" = None, start: float = 0, stop: "float | None" = None, profile_cycle_interval: "str | float | None" = None, ): dt = profile_cycle_interval or dask.config.get( "distributed.worker.profile.cycle" ) dt = parse_timedelta(dt, default="ms") if workers is None: workers = self.workers else: workers = set(self.workers) & set(workers) results = await asyncio.gather( *(self.rpc(w).profile_metadata(start=start, stop=stop) for w in workers), return_exceptions=True, ) results = [r for r in results if not isinstance(r, Exception)] counts = [ (time, sum(pluck(1, group))) for time, group in itertools.groupby( merge_sorted( *(v["counts"] for v in results), ), lambda t: t[0] // dt * dt, ) ] keys: dict[str, list[list]] = { k: [] for v in results for t, d in v["keys"] for k in d } groups1 = [v["keys"] for v in results] groups2 = list(merge_sorted(*groups1, key=first)) last = 0 for t, d in groups2: tt = t // dt * dt if tt > last: last = tt for v in keys.values(): v.append([tt, 0]) for k, v in d.items(): keys[k][-1][1] += v return {"counts": counts, "keys": keys} async def performance_report( self, start: float, last_count: int, code="", mode=None ): stop = time() # Profiles compute, scheduler, workers = await asyncio.gather( *[ self.get_profile(start=start), self.get_profile(scheduler=True, start=start), self.get_profile(server=True, start=start), ] ) from distributed import profile def profile_to_figure(state): data = profile.plot_data(state) figure, source = profile.plot_figure(data, sizing_mode="stretch_both") return figure compute, scheduler, workers = map( profile_to_figure, (compute, scheduler, workers) ) # Task stream task_stream = self.get_task_stream(start=start) total_tasks = len(task_stream) timespent: defaultdict[str, float] = defaultdict(float) for d in task_stream: for x in d["startstops"]: timespent[x["action"]] += x["stop"] - x["start"] tasks_timings = "" for k in sorted(timespent.keys()): tasks_timings += f"\n
  • {k} time: {format_time(timespent[k])}
    • " from distributed.dashboard.components.scheduler import task_stream_figure from distributed.diagnostics.task_stream import rectangles rects = rectangles(task_stream) source, task_stream = task_stream_figure(sizing_mode="stretch_both") source.data.update(rects) # Bandwidth from distributed.dashboard.components.scheduler import ( BandwidthTypes, BandwidthWorkers, ) bandwidth_workers = BandwidthWorkers(self, sizing_mode="stretch_both") bandwidth_workers.update() bandwidth_types = BandwidthTypes(self, sizing_mode="stretch_both") bandwidth_types.update() # System monitor from distributed.dashboard.components.shared import SystemMonitor sysmon = SystemMonitor(self, last_count=last_count, sizing_mode="stretch_both") sysmon.update() # Scheduler logs from distributed.dashboard.components.scheduler import ( _BOKEH_STYLES_KWARGS, SchedulerLogs, ) logs = SchedulerLogs(self, start=start) from bokeh.models import Div, Tabs import distributed from distributed.dashboard.core import TabPanel # HTML html = """

    Dask Performance Report

    Select different tabs on the top for additional information

    Duration: {time}

    Tasks Information

    • number of tasks: {ntasks}
      • {tasks_timings}

    Scheduler Information

    • Address: {address}
    • Workers: {nworkers}
    • Threads: {threads}
    • Memory: {memory}
    • Dask Version: {dask_version}
    • Dask.Distributed Version: {distributed_version}

    Calling Code 

    {code}
    """.format( time=format_time(stop - start), ntasks=total_tasks, tasks_timings=tasks_timings, address=self.address, nworkers=len(self.workers), threads=sum(ws.nthreads for ws in self.workers.values()), memory=format_bytes(sum(ws.memory_limit for ws in self.workers.values())), code=code, dask_version=dask.__version__, distributed_version=distributed.__version__, ) html = Div(text=html, **_BOKEH_STYLES_KWARGS) html = TabPanel(child=html, title="Summary") compute = TabPanel(child=compute, title="Worker Profile (compute)") workers = TabPanel(child=workers, title="Worker Profile (administrative)") scheduler = TabPanel( child=scheduler, title="Scheduler Profile (administrative)" ) task_stream = TabPanel(child=task_stream, title="Task Stream") bandwidth_workers = TabPanel( child=bandwidth_workers.root, title="Bandwidth (Workers)" ) bandwidth_types = TabPanel( child=bandwidth_types.root, title="Bandwidth (Types)" ) system = TabPanel(child=sysmon.root, title="System") logs = TabPanel(child=logs.root, title="Scheduler Logs") tabs = Tabs( tabs=[ html, task_stream, system, logs, compute, workers, scheduler, bandwidth_workers, bandwidth_types, ], sizing_mode="stretch_both", ) from bokeh.core.templates import get_env from bokeh.plotting import output_file, save with tmpfile(extension=".html") as fn: output_file(filename=fn, title="Dask Performance Report", mode=mode) template_directory = os.path.join( os.path.dirname(os.path.abspath(__file__)), "dashboard", "templates" ) template_environment = get_env() template_environment.loader.searchpath.append(template_directory) template = template_environment.get_template("performance_report.html") save(tabs, filename=fn, template=template) with open(fn) as f: data = f.read() return data async def get_worker_logs(self, n=None, workers=None, nanny=False): results = await self.broadcast( msg={"op": "get_logs", "n": n}, workers=workers, nanny=nanny ) return results def log_event(self, topic: str | Collection[str], msg: Any) -> None: event = (time(), msg) if not isinstance(topic, str): for t in topic: self.events[t].append(event) self.event_counts[t] += 1 self._report_event(t, event) else: self.events[topic].append(event) self.event_counts[topic] += 1 self._report_event(topic, event) for plugin in list(self.plugins.values()): try: plugin.log_event(topic, msg) except Exception: logger.info("Plugin failed with exception", exc_info=True) def _report_event(self, name, event): msg = { "op": "event", "topic": name, "event": event, } client_msgs = {client: [msg] for client in self.event_subscriber[name]} self.send_all(client_msgs, worker_msgs={}) def subscribe_topic(self, topic, client): self.event_subscriber[topic].add(client) def unsubscribe_topic(self, topic, client): self.event_subscriber[topic].discard(client) def get_events(self, topic=None): if topic is not None: return tuple(self.events[topic]) else: return valmap(tuple, self.events) async def get_worker_monitor_info(self, recent=False, starts=None): if starts is None: starts = {} results = await asyncio.gather( *( self.rpc(w).get_monitor_info(recent=recent, start=starts.get(w, 0)) for w in self.workers ) ) return dict(zip(self.workers, results)) ########### # Cleanup # ########### async def check_worker_ttl(self): now = time() stimulus_id = f"check-worker-ttl-{now}" for ws in self.workers.values(): if (ws.last_seen < now - self.worker_ttl) and ( ws.last_seen < now - 10 * heartbeat_interval(len(self.workers)) ): logger.warning( "Worker failed to heartbeat within %s seconds. Closing: %s", self.worker_ttl, ws, ) await self.remove_worker(address=ws.address, stimulus_id=stimulus_id) def check_idle(self): assert self.idle_timeout if self.status in (Status.closing, Status.closed): return if self.transition_counter != self._idle_transition_counter: self._idle_transition_counter = self.transition_counter self.idle_since = None return if ( self.queued or self.unrunnable or any([ws.processing for ws in self.workers.values()]) ): self.idle_since = None return if not self.idle_since: self.idle_since = time() if time() > self.idle_since + self.idle_timeout: assert self.idle_since logger.info( "Scheduler closing after being idle for %s", format_time(self.idle_timeout), ) self._ongoing_background_tasks.call_soon(self.close) def adaptive_target(self, target_duration=None): """Desired number of workers based on the current workload This looks at the current running tasks and memory use, and returns a number of desired workers. This is often used by adaptive scheduling. Parameters ---------- target_duration : str A desired duration of time for computations to take. This affects how rapidly the scheduler will ask to scale. See Also -------- distributed.deploy.Adaptive """ if target_duration is None: target_duration = dask.config.get("distributed.adaptive.target-duration") target_duration = parse_timedelta(target_duration) # CPU # TODO consider any user-specified default task durations for queued tasks queued_occupancy = len(self.queued) * self.UNKNOWN_TASK_DURATION cpu = math.ceil( (self.total_occupancy + queued_occupancy) / target_duration ) # TODO: threads per worker # Avoid a few long tasks from asking for many cores tasks_ready = len(self.queued) for ws in self.workers.values(): tasks_ready += len(ws.processing) if tasks_ready > cpu: break else: cpu = min(tasks_ready, cpu) if (self.unrunnable or self.queued) and not self.workers: cpu = max(1, cpu) # add more workers if more than 60% of memory is used limit = sum(ws.memory_limit for ws in self.workers.values()) used = sum(ws.nbytes for ws in self.workers.values()) memory = 0 if used > 0.6 * limit and limit > 0: memory = 2 * len(self.workers) target = max(memory, cpu) if target >= len(self.workers): return target else: # Scale down? to_close = self.workers_to_close() return len(self.workers) - len(to_close) def request_acquire_replicas( self, addr: str, keys: Iterable[str], *, stimulus_id: str ) -> None: """Asynchronously ask a worker to acquire a replica of the listed keys from other workers. This is a fire-and-forget operation which offers no feedback for success or failure, and is intended for housekeeping and not for computation. """ who_has = {} nbytes = {} for key in keys: ts = self.tasks[key] assert ts.who_has who_has[key] = [ws.address for ws in ts.who_has] nbytes[key] = ts.nbytes self.stream_comms[addr].send( { "op": "acquire-replicas", "who_has": who_has, "nbytes": nbytes, "stimulus_id": stimulus_id, }, ) def request_remove_replicas( self, addr: str, keys: list[str], *, stimulus_id: str ) -> None: """Asynchronously ask a worker to discard its replica of the listed keys. This must never be used to destroy the last replica of a key. This is a fire-and-forget operation, intended for housekeeping and not for computation. The replica disappears immediately from TaskState.who_has on the Scheduler side; if the worker refuses to delete, e.g. because the task is a dependency of another task running on it, it will (also asynchronously) inform the scheduler to re-add itself to who_has. If the worker agrees to discard the task, there is no feedback. """ ws = self.workers[addr] # The scheduler immediately forgets about the replica and suggests the worker to # drop it. The worker may refuse, at which point it will send back an add-keys # message to reinstate it. for key in keys: ts = self.tasks[key] if self.validate: # Do not destroy the last copy assert len(ts.who_has) > 1 self.remove_replica(ts, ws) self.stream_comms[addr].send( { "op": "remove-replicas", "keys": keys, "stimulus_id": stimulus_id, } ) def _task_to_report_msg(ts: TaskState) -> dict[str, Any] | None: if ts.state == "forgotten": return {"op": "cancelled-key", "key": ts.key} elif ts.state == "memory": return {"op": "key-in-memory", "key": ts.key} elif ts.state == "erred": failing_ts = ts.exception_blame assert failing_ts return { "op": "task-erred", "key": ts.key, "exception": failing_ts.exception, "traceback": failing_ts.traceback, } else: return None def _task_to_client_msgs(ts: TaskState) -> dict[str, list[dict[str, Any]]]: if ts.who_wants: report_msg = _task_to_report_msg(ts) if report_msg is not None: return {cs.client_key: [report_msg] for cs in ts.who_wants} return {} def decide_worker( ts: TaskState, all_workers: Iterable[WorkerState], valid_workers: set[WorkerState] | None, objective: Callable[[WorkerState], Any], ) -> WorkerState | None: """ Decide which worker should take task *ts*. We choose the worker that has the data on which *ts* depends. If several workers have dependencies then we choose the less-busy worker. Optionally provide *valid_workers* of where jobs are allowed to occur (if all workers are allowed to take the task, pass None instead). If the task requires data communication because no eligible worker has all the dependencies already, then we choose to minimize the number of bytes sent between workers. This is determined by calling the *objective* function. """ assert all(dts.who_has for dts in ts.dependencies) if ts.actor: candidates = set(all_workers) else: candidates = {wws for dts in ts.dependencies for wws in dts.who_has} if valid_workers is None: if not candidates: candidates = set(all_workers) else: candidates &= valid_workers if not candidates: candidates = valid_workers if not candidates: if ts.loose_restrictions: return decide_worker(ts, all_workers, None, objective) if not candidates: return None elif len(candidates) == 1: return next(iter(candidates)) else: return min(candidates, key=objective) def validate_task_state(ts: TaskState) -> None: """Validate the given TaskState""" assert ts.state in ALL_TASK_STATES, ts if ts.waiting_on: assert ts.waiting_on.issubset(ts.dependencies), ( "waiting not subset of dependencies", str(ts.waiting_on), str(ts.dependencies), ) if ts.waiters: assert ts.waiters.issubset(ts.dependents), ( "waiters not subset of dependents", str(ts.waiters), str(ts.dependents), ) for dts in ts.waiting_on: assert not dts.who_has, ("waiting on in-memory dep", str(ts), str(dts)) assert dts.state != "released", ("waiting on released dep", str(ts), str(dts)) for dts in ts.dependencies: assert ts in dts.dependents, ( "not in dependency's dependents", str(ts), str(dts), str(dts.dependents), ) if ts.state in ("waiting", "queued", "processing", "no-worker"): assert dts in ts.waiting_on or dts.who_has, ( "dep missing", str(ts), str(dts), ) assert dts.state != "forgotten" for dts in ts.waiters: assert dts.state in ("waiting", "queued", "processing", "no-worker"), ( "waiter not in play", str(ts), str(dts), ) for dts in ts.dependents: assert ts in dts.dependencies, ( "not in dependent's dependencies", str(ts), str(dts), str(dts.dependencies), ) assert dts.state != "forgotten" assert (ts.processing_on is not None) == (ts.state == "processing") assert bool(ts.who_has) == (ts.state == "memory"), (ts, ts.who_has, ts.state) if ts.state == "queued": assert not ts.processing_on assert not ts.who_has assert all(dts.who_has for dts in ts.dependencies), ( "task queued without all deps", str(ts), str(ts.dependencies), ) if ts.state == "processing": assert all(dts.who_has for dts in ts.dependencies), ( "task processing without all deps", str(ts), str(ts.dependencies), ) assert not ts.waiting_on if ts.who_has: assert ts.waiters or ts.who_wants, ( "unneeded task in memory", str(ts), str(ts.who_has), ) if ts.run_spec: # was computed assert ts.type assert isinstance(ts.type, str) assert not any([ts in dts.waiting_on for dts in ts.dependents]) for ws in ts.who_has: assert ts in ws.has_what, ( "not in who_has' has_what", str(ts), str(ws), str(ws.has_what), ) for cs in ts.who_wants: assert ts in cs.wants_what, ( "not in who_wants' wants_what", str(ts), str(cs), str(cs.wants_what), ) if ts.actor: if ts.state == "memory": assert sum(ts in ws.actors for ws in ts.who_has) == 1 if ts.state == "processing": assert ts.processing_on assert ts in ts.processing_on.actors assert ts.state != "queued" def validate_worker_state(ws: WorkerState) -> None: for ts in ws.has_what: assert ws in ts.who_has, ( "not in has_what' who_has", str(ws), str(ts), str(ts.who_has), ) for ts in ws.actors: assert ts.state in ("memory", "processing") def validate_state( tasks: dict[str, TaskState], workers: dict[str, WorkerState], clients: dict[str, ClientState], ) -> None: """Validate a current runtime state. This performs a sequence of checks on the entire graph, running in about linear time. This raises assert errors if anything doesn't check out. """ for ts in tasks.values(): validate_task_state(ts) for ws in workers.values(): validate_worker_state(ws) for cs in clients.values(): for ts in cs.wants_what: assert cs in ts.who_wants, ( "not in wants_what' who_wants", str(cs), str(ts), str(ts.who_wants), ) def heartbeat_interval(n: int) -> float: """Interval in seconds that we desire heartbeats based on number of workers""" if n <= 10: return 0.5 elif n < 50: return 1 elif n < 200: return 2 else: # No more than 200 heartbeats a second scaled by workers return n / 200 + 1 def _task_slots_available(ws: WorkerState, saturation_factor: float) -> int: """Number of tasks that can be sent to this worker without oversaturating it""" assert not math.isinf(saturation_factor) return max(math.ceil(saturation_factor * ws.nthreads), 1) - ( len(ws.processing) - len(ws.long_running) ) def _worker_full(ws: WorkerState, saturation_factor: float) -> bool: if math.isinf(saturation_factor): return False return _task_slots_available(ws, saturation_factor) <= 0 class KilledWorker(Exception): def __init__(self, task: str, last_worker: WorkerState, allowed_failures: int): super().__init__(task, last_worker, allowed_failures) @property def task(self) -> str: return self.args[0] @property def last_worker(self) -> WorkerState: return self.args[1] @property def allowed_failures(self) -> int: return self.args[2] def __str__(self) -> str: return ( f"Attempted to run task {self.task} on {self.allowed_failures} different " "workers, but all those workers died while running it. " f"The last worker that attempt to run the task was {self.last_worker.address}. " "Inspecting worker logs is often a good next step to diagnose what went wrong. " "For more information see https://distributed.dask.org/en/stable/killed.html." ) class WorkerStatusPlugin(SchedulerPlugin): """A plugin to share worker status with a remote observer This is used in cluster managers to keep updated about the status of the scheduler. """ name: ClassVar[str] = "worker-status" bcomm: BatchedSend def __init__(self, scheduler: Scheduler, comm: Comm): self.bcomm = BatchedSend(interval="5ms") self.bcomm.start(comm) scheduler.add_plugin(self) def add_worker(self, scheduler: Scheduler, worker: str) -> None: ident = scheduler.workers[worker].identity() del ident["metrics"] del ident["last_seen"] try: self.bcomm.send(["add", {"workers": {worker: ident}}]) except CommClosedError: scheduler.remove_plugin(name=self.name) def remove_worker(self, scheduler: Scheduler, worker: str) -> None: try: self.bcomm.send(["remove", worker]) except CommClosedError: scheduler.remove_plugin(name=self.name) def teardown(self) -> None: self.bcomm.close() class CollectTaskMetaDataPlugin(SchedulerPlugin): scheduler: Scheduler name: str keys: set[str] metadata: dict[str, Any] state: dict[str, str] def __init__(self, scheduler: Scheduler, name: str): self.scheduler = scheduler self.name = name self.keys = set() self.metadata = {} self.state = {} def update_graph( self, scheduler: Scheduler, keys: set[str], restrictions: dict[str, float], **kwargs: Any, ) -> None: self.keys.update(keys) def transition( self, key: str, start: TaskStateState, finish: TaskStateState, *args: Any, **kwargs: Any, ) -> None: if finish in ("memory", "erred"): ts = self.scheduler.tasks.get(key) if ts is not None and ts.key in self.keys: self.metadata[key] = ts.metadata self.state[key] = finish self.keys.discard(key)