from __future__ import annotations
import asyncio
import atexit
import contextlib
import copy
import logging
import math
import weakref
from collections.abc import Awaitable, Generator
from contextlib import suppress
from inspect import isawaitable
from typing import TYPE_CHECKING, Any, ClassVar, TypeVar
from tornado import gen
from tornado.ioloop import IOLoop
import dask
from dask.utils import parse_bytes, parse_timedelta
from dask.widgets import get_template
from distributed.core import Status, rpc
from distributed.deploy.adaptive import Adaptive
from distributed.deploy.cluster import Cluster
from distributed.scheduler import Scheduler
from distributed.security import Security
from distributed.utils import (
NoOpAwaitable,
TimeoutError,
import_term,
silence_logging_cmgr,
)
if TYPE_CHECKING:
# TODO import from typing (requires Python >=3.11)
from typing_extensions import Self
# Circular imports
from distributed import Nanny, Worker
logger = logging.getLogger(__name__)
class ProcessInterface:
"""
An interface for Scheduler and Worker processes for use in SpecCluster
This interface is responsible to submit a worker or scheduler process to a
resource manager like Kubernetes, Yarn, or SLURM/PBS/SGE/...
It should implement the methods below, like ``start`` and ``close``
"""
@property
def status(self):
return self._status
@status.setter
def status(self, new_status):
if not isinstance(new_status, Status):
raise TypeError(f"Expected Status; got {new_status!r}")
self._status = new_status
def __init__(self, scheduler=None, name=None):
self.address = getattr(self, "address", None)
self.external_address = None
self.lock = asyncio.Lock()
self.status = Status.created
self._event_finished = asyncio.Event()
def __await__(self):
async def _():
async with self.lock:
if self.status == Status.created:
await self.start()
assert self.status == Status.running
return self
return _().__await__()
async def start(self):
"""Submit the process to the resource manager
For workers this doesn't have to wait until the process actually starts,
but can return once the resource manager has the request, and will work
to make the job exist in the future
For the scheduler we will expect the scheduler's ``.address`` attribute
to be available after this completes.
"""
self.status = Status.running
async def close(self):
"""Close the process
This will be called by the Cluster object when we scale down a node,
but only after we ask the Scheduler to close the worker gracefully.
This method should kill the process a bit more forcefully and does not
need to worry about shutting down gracefully
"""
self.status = Status.closed
self._event_finished.set()
async def finished(self):
"""Wait until the server has finished"""
await self._event_finished.wait()
def __repr__(self):
return f"<{dask.utils.typename(type(self))}: status={self.status.name}>"
def _repr_html_(self):
return get_template("process_interface.html.j2").render(process_interface=self)
async def __aenter__(self):
await self
return self
async def __aexit__(self, exc_type, exc_value, traceback):
await self.close()
_T = TypeVar("_T")
async def _wrap_awaitable(aw: Awaitable[_T]) -> _T:
return await aw
[docs]class SpecCluster(Cluster):
"""Cluster that requires a full specification of workers
The SpecCluster class expects a full specification of the Scheduler and
Workers to use. It removes any handling of user inputs (like threads vs
processes, number of cores, and so on) and any handling of cluster resource
managers (like pods, jobs, and so on). Instead, it expects this
information to be passed in scheduler and worker specifications. This
class does handle all of the logic around asynchronously cleanly setting up
and tearing things down at the right times. Hopefully it can form a base
for other more user-centric classes.
Parameters
----------
workers: dict
A dictionary mapping names to worker classes and their specifications
See example below
scheduler: dict, optional
A similar mapping for a scheduler
worker: dict
A specification of a single worker.
This is used for any new workers that are created.
asynchronous: bool
If this is intended to be used directly within an event loop with
async/await
silence_logs: bool
Whether or not we should silence logging when setting up the cluster.
name: str, optional
A name to use when printing out the cluster, defaults to type name
Examples
--------
To create a SpecCluster you specify how to set up a Scheduler and Workers
>>> from dask.distributed import Scheduler, Worker, Nanny
>>> scheduler = {'cls': Scheduler, 'options': {"dashboard_address": ':8787'}}
>>> workers = {
... 'my-worker': {"cls": Worker, "options": {"nthreads": 1}},
... 'my-nanny': {"cls": Nanny, "options": {"nthreads": 2}},
... }
>>> cluster = SpecCluster(scheduler=scheduler, workers=workers)
The worker spec is stored as the ``.worker_spec`` attribute
>>> cluster.worker_spec
{
'my-worker': {"cls": Worker, "options": {"nthreads": 1}},
'my-nanny': {"cls": Nanny, "options": {"nthreads": 2}},
}
While the instantiation of this spec is stored in the ``.workers``
attribute
>>> cluster.workers
{
'my-worker': <Worker ...>
'my-nanny': <Nanny ...>
}
Should the spec change, we can await the cluster or call the
``._correct_state`` method to align the actual state to the specified
state.
We can also ``.scale(...)`` the cluster, which adds new workers of a given
form.
>>> worker = {'cls': Worker, 'options': {}}
>>> cluster = SpecCluster(scheduler=scheduler, worker=worker)
>>> cluster.worker_spec
{}
>>> cluster.scale(3)
>>> cluster.worker_spec
{
0: {'cls': Worker, 'options': {}},
1: {'cls': Worker, 'options': {}},
2: {'cls': Worker, 'options': {}},
}
Note that above we are using the standard ``Worker`` and ``Nanny`` classes,
however in practice other classes could be used that handle resource
management like ``KubernetesPod`` or ``SLURMJob``. The spec does not need
to conform to the expectations of the standard Dask Worker class. It just
needs to be called with the provided options, support ``__await__`` and
``close`` methods and the ``worker_address`` property..
Also note that uniformity of the specification is not required. Other API
could be added externally (in subclasses) that adds workers of different
specifications into the same dictionary.
If a single entry in the spec will generate multiple dask workers then
please provide a `"group"` element to the spec, that includes the suffixes
that will be added to each name (this should be handled by your worker
class).
>>> cluster.worker_spec
{
0: {"cls": MultiWorker, "options": {"processes": 3}, "group": ["-0", "-1", -2"]}
1: {"cls": MultiWorker, "options": {"processes": 2}, "group": ["-0", "-1"]}
}
These suffixes should correspond to the names used by the workers when
they deploy.
>>> [ws.name for ws in cluster.scheduler.workers.values()]
["0-0", "0-1", "0-2", "1-0", "1-1"]
"""
_instances: ClassVar[weakref.WeakSet[SpecCluster]] = weakref.WeakSet()
def __init__(
self,
workers=None,
scheduler=None,
worker=None,
asynchronous=False,
loop=None,
security=None,
silence_logs=False,
name=None,
shutdown_on_close=True,
scheduler_sync_interval=1,
):
if loop is None and asynchronous:
loop = IOLoop.current()
self.__exit_stack = stack = contextlib.ExitStack()
self._created = weakref.WeakSet()
self.scheduler_spec = copy.copy(scheduler)
self.worker_spec = copy.copy(workers) or {}
self.new_spec = copy.copy(worker)
self.scheduler = None
self.workers = {}
self._i = 0
self.security = security or Security()
self._futures = set()
if silence_logs:
stack.enter_context(silence_logging_cmgr(level=silence_logs))
stack.enter_context(silence_logging_cmgr(level=silence_logs, root="bokeh"))
self._instances.add(self)
self._correct_state_waiting = None
self._name = name or type(self).__name__
self.shutdown_on_close = shutdown_on_close
super().__init__(
asynchronous=asynchronous,
loop=loop,
name=name,
scheduler_sync_interval=scheduler_sync_interval,
)
if not self.called_from_running_loop:
self._loop_runner.start()
self.sync(self._start)
try:
self.sync(self._correct_state)
except Exception:
self.sync(self.close)
self._loop_runner.stop()
raise
def close(self, timeout: float | None = None) -> Awaitable[None] | None:
aw = super().close(timeout)
if not self.asynchronous:
self._loop_runner.stop()
return aw
async def _start(self):
while self.status == Status.starting:
await asyncio.sleep(0.01)
if self.status == Status.running:
return
if self.status == Status.closed:
raise ValueError("Cluster is closed")
self._lock = asyncio.Lock()
self.status = Status.starting
if self.scheduler_spec is None:
try:
import distributed.dashboard # noqa: F401
except ImportError:
pass
else:
options = {"dashboard": True}
self.scheduler_spec = {"cls": Scheduler, "options": options}
try:
# Check if scheduler has already been created by a subclass
if self.scheduler is None:
cls = self.scheduler_spec["cls"]
if isinstance(cls, str):
cls = import_term(cls)
self.scheduler = cls(**self.scheduler_spec.get("options", {}))
self.scheduler = await self.scheduler
self.scheduler_comm = rpc(
getattr(self.scheduler, "external_address", None)
or self.scheduler.address,
connection_args=self.security.get_connection_args("client"),
)
await super()._start()
except Exception as e: # pragma: no cover
self.status = Status.failed
await self._close()
raise RuntimeError(f"Cluster failed to start: {e}") from e
def _correct_state(self):
if self._correct_state_waiting:
# If people call this frequently, we only want to run it once
return self._correct_state_waiting
else:
task = asyncio.ensure_future(self._correct_state_internal())
self._correct_state_waiting = task
return task
async def _correct_state_internal(self) -> None:
async with self._lock:
self._correct_state_waiting = None
to_close = set(self.workers) - set(self.worker_spec)
if to_close:
if self.scheduler.status == Status.running:
await self.scheduler_comm.retire_workers(workers=list(to_close))
tasks = [
asyncio.create_task(self.workers[w].close())
for w in to_close
if w in self.workers
]
await asyncio.gather(*tasks)
for name in to_close:
if name in self.workers:
del self.workers[name]
to_open = set(self.worker_spec) - set(self.workers)
workers = []
for name in to_open:
d = self.worker_spec[name]
cls, opts = d["cls"], d.get("options", {})
if "name" not in opts:
opts = opts.copy()
opts["name"] = name
if isinstance(cls, str):
cls = import_term(cls)
worker = cls(
getattr(self.scheduler, "contact_address", None)
or self.scheduler.address,
**opts,
)
self._created.add(worker)
workers.append(worker)
if workers:
worker_futs = [asyncio.ensure_future(w) for w in workers]
await asyncio.wait(worker_futs)
self.workers.update(dict(zip(to_open, workers)))
for w in workers:
w._cluster = weakref.ref(self)
# Collect exceptions from failed workers. This must happen after all
# *other* workers have finished initialising, so that we can have a
# proper teardown.
await asyncio.gather(*worker_futs)
def _update_worker_status(self, op, msg):
if op == "remove":
name = self.scheduler_info["workers"][msg]["name"]
def f():
if (
name in self.workers
and msg not in self.scheduler_info["workers"]
and not any(
d["name"] == name
for d in self.scheduler_info["workers"].values()
)
):
self._futures.add(asyncio.ensure_future(self.workers[name].close()))
del self.workers[name]
delay = parse_timedelta(
dask.config.get("distributed.deploy.lost-worker-timeout")
)
asyncio.get_running_loop().call_later(delay, f)
super()._update_worker_status(op, msg)
def __await__(self: Self) -> Generator[Any, Any, Self]:
async def _() -> Self:
if self.status == Status.created:
await self._start()
await self.scheduler
await self._correct_state()
if self.workers:
await asyncio.wait(
[
asyncio.create_task(_wrap_awaitable(w))
for w in self.workers.values()
]
) # maybe there are more
return self
return _().__await__()
async def _close(self):
while self.status == Status.closing:
await asyncio.sleep(0.1)
if self.status == Status.closed:
return
if self.status == Status.running or self.status == Status.failed:
self.status = Status.closing
# Need to call stop here before we close all servers to avoid having
# dangling tasks in the ioloop
with suppress(AttributeError):
self._adaptive.stop()
f = self.scale(0)
if isawaitable(f):
await f
await self._correct_state()
await asyncio.gather(*self._futures)
if self.scheduler_comm:
async with self._lock:
with suppress(OSError):
await self.scheduler_comm.terminate()
await self.scheduler_comm.close_rpc()
else:
logger.warning("Cluster closed without starting up")
if self.scheduler:
await self.scheduler.close()
for w in self._created:
assert w.status in {
Status.closing,
Status.closed,
Status.failed,
}, w.status
self.__exit_stack.__exit__(None, None, None)
await super()._close()
async def __aenter__(self):
try:
await self
await self._correct_state()
assert self.status == Status.running
return self
except Exception:
await self._close()
raise
def _threads_per_worker(self) -> int:
"""Return the number of threads per worker for new workers"""
if not self.new_spec: # pragma: no cover
raise ValueError("To scale by cores= you must specify cores per worker")
for name in ["nthreads", "ncores", "threads", "cores"]:
with suppress(KeyError):
return self.new_spec["options"][name]
raise RuntimeError("unreachable")
def _memory_per_worker(self) -> int:
"""Return the memory limit per worker for new workers"""
if not self.new_spec: # pragma: no cover
raise ValueError(
"to scale by memory= your worker definition must include a "
"memory_limit definition"
)
for name in ["memory_limit", "memory"]:
with suppress(KeyError):
return parse_bytes(self.new_spec["options"][name])
raise ValueError(
"to use scale(memory=...) your worker definition must include a "
"memory_limit definition"
)
[docs] def scale(self, n=0, memory=None, cores=None):
if memory is not None:
n = max(n, int(math.ceil(parse_bytes(memory) / self._memory_per_worker())))
if cores is not None:
n = max(n, int(math.ceil(cores / self._threads_per_worker())))
if len(self.worker_spec) > n:
not_yet_launched = set(self.worker_spec) - {
v["name"] for v in self.scheduler_info["workers"].values()
}
while len(self.worker_spec) > n and not_yet_launched:
del self.worker_spec[not_yet_launched.pop()]
while len(self.worker_spec) > n:
self.worker_spec.popitem()
if self.status not in (Status.closing, Status.closed):
while len(self.worker_spec) < n:
self.worker_spec.update(self.new_worker_spec())
self.loop.add_callback(self._correct_state)
if self.asynchronous:
return NoOpAwaitable()
def _new_worker_name(self, worker_number):
"""Returns new worker name.
This can be overridden in SpecCluster derived classes to customise the
worker names.
"""
return worker_number
[docs] def new_worker_spec(self):
"""Return name and spec for the next worker
Returns
-------
d: dict mapping names to worker specs
See Also
--------
scale
"""
new_worker_name = self._new_worker_name(self._i)
while new_worker_name in self.worker_spec:
self._i += 1
new_worker_name = self._new_worker_name(self._i)
return {new_worker_name: self.new_spec}
@property
def _supports_scaling(self):
return bool(self.new_spec)
async def scale_down(self, workers):
# We may have groups, if so, map worker addresses to job names
if not all(w in self.worker_spec for w in workers):
mapping = {}
for name, spec in self.worker_spec.items():
if "group" in spec:
for suffix in spec["group"]:
mapping[str(name) + suffix] = name
else:
mapping[name] = name
workers = {mapping.get(w, w) for w in workers}
for w in workers:
if w in self.worker_spec:
del self.worker_spec[w]
await self
scale_up = scale # backwards compatibility
@property
def plan(self):
out = set()
for name, spec in self.worker_spec.items():
if "group" in spec:
out.update({str(name) + suffix for suffix in spec["group"]})
else:
out.add(name)
return out
@property
def requested(self):
out = set()
for name in self.workers:
try:
spec = self.worker_spec[name]
except KeyError:
continue
if "group" in spec:
out.update({str(name) + suffix for suffix in spec["group"]})
else:
out.add(name)
return out
[docs] def adapt(
self,
Adaptive: type[Adaptive] = Adaptive,
minimum: float = 0,
maximum: float = math.inf,
minimum_cores: int | None = None,
maximum_cores: int | None = None,
minimum_memory: str | None = None,
maximum_memory: str | None = None,
**kwargs: Any,
) -> Adaptive:
"""Turn on adaptivity
This scales Dask clusters automatically based on scheduler activity.
Parameters
----------
minimum : int
Minimum number of workers
maximum : int
Maximum number of workers
minimum_cores : int
Minimum number of cores/threads to keep around in the cluster
maximum_cores : int
Maximum number of cores/threads to keep around in the cluster
minimum_memory : str
Minimum amount of memory to keep around in the cluster
Expressed as a string like "100 GiB"
maximum_memory : str
Maximum amount of memory to keep around in the cluster
Expressed as a string like "100 GiB"
Examples
--------
>>> cluster.adapt(minimum=0, maximum_memory="100 GiB", interval='500ms')
See Also
--------
dask.distributed.Adaptive : for more keyword arguments
"""
if minimum_cores is not None:
minimum = max(
minimum or 0, math.ceil(minimum_cores / self._threads_per_worker())
)
if minimum_memory is not None:
minimum = max(
minimum or 0,
math.ceil(parse_bytes(minimum_memory) / self._memory_per_worker()),
)
if maximum_cores is not None:
maximum = min(
maximum, math.floor(maximum_cores / self._threads_per_worker())
)
if maximum_memory is not None:
maximum = min(
maximum,
math.floor(parse_bytes(maximum_memory) / self._memory_per_worker()),
)
return super().adapt(
Adaptive=Adaptive, minimum=minimum, maximum=maximum, **kwargs
)
[docs] @classmethod
def from_name(cls, name: str) -> ProcessInterface:
"""Create an instance of this class to represent an existing cluster by name."""
raise NotImplementedError()
def init_spec(spec: dict[str, Any], *args: Any) -> dict[str, Worker | Nanny]:
workers = {}
for k, d in spec.items():
cls = d["cls"]
if isinstance(cls, str):
cls = import_term(cls)
workers[k] = cls(*args, **d.get("opts", {}))
return workers
async def run_spec(spec: dict[str, Any], *args: Any) -> dict[str, Worker | Nanny]:
workers = init_spec(spec, *args)
if workers:
await asyncio.gather(*workers.values())
return workers
@atexit.register
def close_clusters():
for cluster in list(SpecCluster._instances):
if getattr(cluster, "shutdown_on_close", False):
with suppress(gen.TimeoutError, TimeoutError):
if getattr(cluster, "status", Status.closed) != Status.closed:
cluster.close(timeout=10)