"""Encapsulated manager for in-memory tasks on a worker. This module covers: - spill/unspill data depending on the 'distributed.worker.memory.target' threshold - spill/unspill data depending on the 'distributed.worker.memory.spill' threshold - pause/unpause the worker depending on the 'distributed.worker.memory.pause' threshold - kill the worker depending on the 'distributed.worker.memory.terminate' threshold This module does *not* cover: - Changes in behaviour in Worker, Scheduler, task stealing, Active Memory Manager, etc. caused by the Worker being in paused status - Worker restart after it's been killed - Scheduler-side heuristics regarding memory usage, e.g. the Active Memory Manager See also: - :mod:`distributed.spill`, which implements the spill-to-disk mechanism and is wrapped by this module. Unlike this module, :mod:`distributed.spill` is agnostic to the Worker. - :mod:`distributed.active_memory_manager`, which runs on the scheduler side """ from __future__ import annotations import asyncio import logging import os import sys import warnings from collections.abc import Callable, Container, Hashable, MutableMapping from contextlib import suppress from functools import partial from typing import TYPE_CHECKING, Any, Literal, Union, cast import psutil import dask.config from dask.system import CPU_COUNT from dask.typing import Key from dask.utils import format_bytes, parse_bytes, parse_timedelta from distributed import system from distributed.compatibility import WINDOWS, PeriodicCallback from distributed.core import Status from distributed.gc import ThrottledGC from distributed.metrics import context_meter, monotonic from distributed.spill import ManualEvictProto, SpillBuffer from distributed.utils import RateLimiterFilter, has_arg, log_errors if TYPE_CHECKING: # TODO import from typing (requires Python >=3.10) from typing_extensions import TypeAlias # Circular imports from distributed.nanny import Nanny from distributed.worker import Worker WorkerDataParameter: TypeAlias = Union[ # pre-initialized MutableMapping[Key, object], # constructor Callable[[], MutableMapping[Key, object]], # constructor, passed worker.local_directory Callable[[str], MutableMapping[Key, object]], # (constructor, kwargs to constructor) tuple[Callable[..., MutableMapping[Key, object]], dict[str, Any]], # initialize internally None, ] worker_logger = logging.getLogger("distributed.worker.memory") worker_logger.addFilter(RateLimiterFilter(r"Unmanaged memory use is high", rate="300s")) nanny_logger = logging.getLogger("distributed.nanny.memory") class WorkerMemoryManager: """Management of worker memory usage Parameters ---------- worker Worker to manage For meaning of the remaining parameters, see the matching parameter names in :class:`~.distributed.worker.Worker`. Notes ----- If data is a callable and has the argument ``worker_local_directory`` in its signature, it will be filled with the worker's attr:``local_directory``. """ data: MutableMapping[Key, object] # {task key: task payload} memory_limit: int | None memory_target_fraction: float | Literal[False] memory_spill_fraction: float | Literal[False] memory_pause_fraction: float | Literal[False] max_spill: int | Literal[False] memory_monitor_interval: float _throttled_gc: ThrottledGC def __init__( self, worker: Worker, *, nthreads: int, memory_limit: str | float = "auto", # This should be None most of the times, short of a power user replacing the # SpillBuffer with their own custom dict-like data: WorkerDataParameter = None, # Deprecated parameters; use dask.config instead memory_target_fraction: float | Literal[False] | None = None, memory_spill_fraction: float | Literal[False] | None = None, memory_pause_fraction: float | Literal[False] | None = None, ): self.memory_limit = parse_memory_limit( memory_limit, nthreads, logger=worker_logger ) self.memory_target_fraction = _parse_threshold( "distributed.worker.memory.target", "memory_target_fraction", memory_target_fraction, ) self.memory_spill_fraction = _parse_threshold( "distributed.worker.memory.spill", "memory_spill_fraction", memory_spill_fraction, ) self.memory_pause_fraction = _parse_threshold( "distributed.worker.memory.pause", "memory_pause_fraction", memory_pause_fraction, ) max_spill = dask.config.get("distributed.worker.memory.max-spill") self.max_spill = False if max_spill is False else parse_bytes(max_spill) if isinstance(data, MutableMapping): self.data = data elif callable(data): if has_arg(data, "worker_local_directory"): data = cast("Callable[[str], MutableMapping[Key, object]]", data) self.data = data(worker.local_directory) else: data = cast("Callable[[], MutableMapping[Key, object]]", data) self.data = data() elif isinstance(data, tuple): func, kwargs = data if not callable(func): raise ValueError("Expecting a callable") if has_arg(func, "worker_local_directory"): self.data = func( worker_local_directory=worker.local_directory, **kwargs ) else: self.data = func(**kwargs) elif self.memory_limit and ( self.memory_target_fraction or self.memory_spill_fraction ): if self.memory_target_fraction: target = int( self.memory_limit * (self.memory_target_fraction or self.memory_spill_fraction) ) else: target = sys.maxsize self.data = SpillBuffer( os.path.join(worker.local_directory, "storage"), target=target, max_spill=self.max_spill, ) else: self.data = {} if not isinstance(self.data, MutableMapping): raise TypeError(f"Worker.data must be a MutableMapping; got {self.data}") if self.data: raise ValueError("Worker.data must be empty at initialization time") self.memory_monitor_interval = parse_timedelta( dask.config.get("distributed.worker.memory.monitor-interval"), default=False, ) assert isinstance(self.memory_monitor_interval, (int, float)) if self.memory_limit and ( self.memory_spill_fraction is not False or self.memory_pause_fraction is not False ): assert self.memory_monitor_interval is not None pc = PeriodicCallback( # Don't store worker as self.worker to avoid creating a circular # dependency. We could have alternatively used a weakref. partial(self.memory_monitor, worker), self.memory_monitor_interval * 1000, ) worker.periodic_callbacks["memory_monitor"] = pc self._throttled_gc = ThrottledGC(logger=worker_logger) @log_errors async def memory_monitor(self, worker: Worker) -> None: """Track this process's memory usage and act accordingly. If process memory rises above the spill threshold (70%), start dumping data to disk until it goes below the target threshold (60%). If process memory rises above the pause threshold (80%), stop execution of new tasks. """ # Don't use psutil directly; instead read from the same API that is used # to send info to the Scheduler (e.g. for the benefit of Active Memory # Manager) and which can be easily mocked in unit tests. memory = worker.monitor.get_process_memory() self._maybe_pause_or_unpause(worker, memory) await self._maybe_spill(worker, memory) def _maybe_pause_or_unpause(self, worker: Worker, memory: int) -> None: if self.memory_pause_fraction is False: return assert self.memory_limit frac = memory / self.memory_limit # Pause worker threads if above 80% memory use if frac > self.memory_pause_fraction: # Try to free some memory while in paused state self._throttled_gc.collect() if worker.status == Status.running: worker_logger.warning( "Worker is at %d%% memory usage. Pausing worker. " "Process memory: %s -- Worker memory limit: %s", int(frac * 100), format_bytes(memory), ( format_bytes(self.memory_limit) if self.memory_limit is not None else "None" ), ) worker.status = Status.paused elif worker.status == Status.paused: worker_logger.warning( "Worker is at %d%% memory usage. Resuming worker. " "Process memory: %s -- Worker memory limit: %s", int(frac * 100), format_bytes(memory), ( format_bytes(self.memory_limit) if self.memory_limit is not None else "None" ), ) worker.status = Status.running async def _maybe_spill(self, worker: Worker, memory: int) -> None: """If process memory is above the ``spill`` threshold, evict keys until it goes below the ``target`` threshold """ if self.memory_spill_fraction is False: return # SpillBuffer or a duct-type compatible MutableMapping which offers the # fast property and evict() methods. Dask-CUDA uses this. if not hasattr(self.data, "fast") or not hasattr(self.data, "evict"): return assert self.memory_limit frac = memory / self.memory_limit if frac <= self.memory_spill_fraction: return worker_logger.debug( "Worker is at %.0f%% memory usage. Start spilling data to disk.", frac * 100, ) def metrics_callback(label: Hashable, value: float, unit: str) -> None: if not isinstance(label, tuple): label = (label,) worker.digest_metric(("memory-monitor", *label, unit), value) # Work around bug with Tornado 6.2 PeriodicCallback, which does not properly # insulate contextvars. Without this hack, you would see metrics that are # clearly emitted by Worker.execute labelled with 'memory-monitor'.So we're # wrapping our change in contextvars (inside add_callback) inside create_task(), # which copies and insulates the context. async def _() -> None: with context_meter.add_callback(metrics_callback): # Measure delta between the measures from the SpillBuffer and the total # end-to-end duration of _spill await self._spill(worker, memory) await asyncio.create_task(_(), name="memory-monitor-spill") # End work around async def _spill(self, worker: Worker, memory: int) -> None: """Evict keys until the process memory goes below the ``target`` threshold""" assert self.memory_limit total_spilled = 0 # Implement hysteresis cycle where spilling starts at the spill threshold and # stops at the target threshold. Normally that here the target threshold defines # process memory, whereas normally it defines reported managed memory (e.g. # output of sizeof() ). If target=False, disable hysteresis. target = self.memory_limit * ( self.memory_target_fraction or self.memory_spill_fraction ) count = 0 need = memory - target last_checked_for_pause = last_yielded = monotonic() data = cast(ManualEvictProto, self.data) while memory > target: if not data.fast: worker_logger.warning( "Unmanaged memory use is high. This may indicate a memory leak " "or the memory may not be released to the OS; see " "https://distributed.dask.org/en/latest/worker-memory.html#memory-not-released-back-to-the-os " "for more information. " "-- Unmanaged memory: %s -- Worker memory limit: %s", format_bytes(memory), format_bytes(self.memory_limit), ) break weight = data.evict() if weight == -1: # Failed to evict: # disk full, spill size limit exceeded, or pickle error break total_spilled += weight count += 1 memory = worker.monitor.get_process_memory() if total_spilled > need and memory > target: # Issue a GC to ensure that the evicted data is actually # freed from memory and taken into account by the monitor # before trying to evict even more data. self._throttled_gc.collect() memory = worker.monitor.get_process_memory() now = monotonic() # Spilling may potentially take multiple seconds; we may pass the pause # threshold in the meantime. if now - last_checked_for_pause > self.memory_monitor_interval: self._maybe_pause_or_unpause(worker, memory) last_checked_for_pause = now # Increase spilling aggressiveness when the fast buffer is filled with a lot # of small values. This artificially chokes the rest of the event loop - # namely, the reception of new data from other workers. While this is # somewhat of an ugly hack, DO NOT tweak this without a thorough cycle of # stress testing. See: https://github.com/dask/distributed/issues/6110. if now - last_yielded > 0.5: await asyncio.sleep(0) last_yielded = monotonic() if count: worker_logger.debug( "Moved %d tasks worth %s to disk", count, format_bytes(total_spilled), ) def _to_dict(self, *, exclude: Container[str] = ()) -> dict: info = {k: v for k, v in self.__dict__.items() if not k.startswith("_")} info["data"] = dict.fromkeys(self.data) return info class NannyMemoryManager: memory_limit: int | None memory_terminate_fraction: float | Literal[False] memory_monitor_interval: float | None _last_terminated_pid: int def __init__( self, nanny: Nanny, *, memory_limit: str | float = "auto", ): self.memory_limit = parse_memory_limit( memory_limit, nanny.nthreads, logger=nanny_logger ) self.memory_terminate_fraction = dask.config.get( "distributed.worker.memory.terminate" ) self.memory_monitor_interval = parse_timedelta( dask.config.get("distributed.worker.memory.monitor-interval"), default=False, ) assert isinstance(self.memory_monitor_interval, (int, float)) self._last_terminated_pid = -1 if self.memory_limit and self.memory_terminate_fraction is not False: pc = PeriodicCallback( partial(self.memory_monitor, nanny), self.memory_monitor_interval * 1000, ) nanny.periodic_callbacks["memory_monitor"] = pc def memory_monitor(self, nanny: Nanny) -> None: """Track worker's memory. Restart if it goes above terminate fraction.""" if ( nanny.status != Status.running or nanny.process is None or nanny.process.process is None or nanny.process.process.pid is None ): return # pragma: nocover process = nanny.process.process try: memory = psutil.Process(process.pid).memory_info().rss except (ProcessLookupError, psutil.NoSuchProcess, psutil.AccessDenied): return # pragma: nocover if memory / self.memory_limit <= self.memory_terminate_fraction: return if self._last_terminated_pid != process.pid: msg = ( f"Worker {nanny.worker_address} (pid={process.pid}) exceeded " f"{self.memory_terminate_fraction * 100:.0f}% memory budget. " f"Restarting..." ) nanny_logger.warning(msg) self._last_terminated_pid = process.pid event = { "worker": nanny.worker_address, "pid": process.pid, "rss": memory, } nanny.log_event("worker-restart-memory", event) process.terminate() else: # We already sent SIGTERM to the worker, but the process is still alive # since the previous iteration of the memory_monitor - for example, some # user code may have tampered with signal handlers. # Send SIGKILL for immediate termination. # # Note that this should not be a disk-related issue. Unlike in a regular # worker shutdown, where the worker cleans up its own spill directory, in # case of SIGTERM no atexit or weakref.finalize callback is triggered # whatsoever; instead, the nanny cleans up the spill directory *after* the # worker has been shut down and before starting a new one. # This is important, as spill directory cleanup may potentially take tens of # seconds and, if the worker did it, any task that was running and leaking # would continue to do so for the whole duration of the cleanup, increasing # the risk of going beyond 100%. nanny_logger.warning( f"Worker {nanny.worker_address} (pid={process.pid}) is slow to %s", # On Windows, kill() is an alias to terminate() ( "terminate; trying again" if WINDOWS else "accept SIGTERM; sending SIGKILL" ), ) process.kill() def parse_memory_limit( memory_limit: str | float | None, nthreads: int, total_cores: int = CPU_COUNT, *, logger: logging.Logger, ) -> int | None: if memory_limit is None: return None orig = memory_limit if memory_limit == "auto": memory_limit = int(system.MEMORY_LIMIT * min(1, nthreads / total_cores)) with suppress(ValueError, TypeError): memory_limit = float(memory_limit) if isinstance(memory_limit, float) and memory_limit <= 1: memory_limit = int(memory_limit * system.MEMORY_LIMIT) if isinstance(memory_limit, str): memory_limit = parse_bytes(memory_limit) else: memory_limit = int(memory_limit) assert isinstance(memory_limit, int) if memory_limit == 0: return None if system.MEMORY_LIMIT < memory_limit: logger.warning( "Ignoring provided memory limit %s due to system memory limit of %s", orig, format_bytes(system.MEMORY_LIMIT), ) return system.MEMORY_LIMIT else: return memory_limit def _parse_threshold( config_key: str, deprecated_param_name: str, deprecated_param_value: float | Literal[False] | None, ) -> float | Literal[False]: if deprecated_param_value is not None: warnings.warn( f"Parameter {deprecated_param_name} has been deprecated and will be " f"removed in a future version; please use dask config key {config_key} " "instead", FutureWarning, ) return deprecated_param_value return dask.config.get(config_key) def _warn_deprecated(w: Nanny | Worker, name: str) -> None: warnings.warn( f"The `{type(w).__name__}.{name}` attribute has been moved to " f"`{type(w).__name__}.memory_manager.{name}", FutureWarning, ) class DeprecatedMemoryManagerAttribute: name: str def __set_name__(self, owner: type, name: str) -> None: self.name = name def __get__(self, instance: Nanny | Worker | None, owner: type) -> Any: if instance is None: # This is triggered by Sphinx return None # pragma: nocover _warn_deprecated(instance, self.name) return getattr(instance.memory_manager, self.name) def __set__(self, instance: Nanny | Worker, value: Any) -> None: _warn_deprecated(instance, self.name) setattr(instance.memory_manager, self.name, value) class DeprecatedMemoryMonitor: def __get__(self, instance: Nanny | Worker | None, owner: type) -> Any: if instance is None: # This is triggered by Sphinx return None # pragma: nocover _warn_deprecated(instance, "memory_monitor") return partial(instance.memory_manager.memory_monitor, instance) # type: ignore