_dask.py 12.3 KB
from __future__ import print_function, division, absolute_import

import asyncio
import concurrent.futures
import contextlib

import time
from uuid import uuid4
import weakref

from .parallel import AutoBatchingMixin, ParallelBackendBase, BatchedCalls
from .parallel import parallel_backend

try:
    import distributed
except ImportError:
    distributed = None

if distributed is not None:
    from dask.utils import funcname, itemgetter
    from dask.sizeof import sizeof
    from dask.distributed import (
        Client,
        as_completed,
        get_client,
        secede,
        rejoin
    )
    from distributed.utils import thread_state

    try:
        # asyncio.TimeoutError, Python3-only error thrown by recent versions of
        # distributed
        from distributed.utils import TimeoutError as _TimeoutError
    except ImportError:
        from tornado.gen import TimeoutError as _TimeoutError


def is_weakrefable(obj):
    try:
        weakref.ref(obj)
        return True
    except TypeError:
        return False


class _WeakKeyDictionary:
    """A variant of weakref.WeakKeyDictionary for unhashable objects.

    This datastructure is used to store futures for broadcasted data objects
    such as large numpy arrays or pandas dataframes that are not hashable and
    therefore cannot be used as keys of traditional python dicts.

    Futhermore using a dict with id(array) as key is not safe because the
    Python is likely to reuse id of recently collected arrays.
    """

    def __init__(self):
        self._data = {}

    def __getitem__(self, obj):
        ref, val = self._data[id(obj)]
        if ref() is not obj:
            # In case of a race condition with on_destroy.
            raise KeyError(obj)
        return val

    def __setitem__(self, obj, value):
        key = id(obj)
        try:
            ref, _ = self._data[key]
            if ref() is not obj:
                # In case of race condition with on_destroy.
                raise KeyError(obj)
        except KeyError:
            # Insert the new entry in the mapping along with a weakref
            # callback to automatically delete the entry from the mapping
            # as soon as the object used as key is garbage collected.
            def on_destroy(_):
                del self._data[key]
            ref = weakref.ref(obj, on_destroy)
        self._data[key] = ref, value

    def __len__(self):
        return len(self._data)

    def clear(self):
        self._data.clear()


def _funcname(x):
    try:
        if isinstance(x, list):
            x = x[0][0]
    except Exception:
        pass
    return funcname(x)


def _make_tasks_summary(tasks):
    """Summarize of list of (func, args, kwargs) function calls"""
    unique_funcs = {func for func, args, kwargs in tasks}

    if len(unique_funcs) == 1:
        mixed = False
    else:
        mixed = True
    return len(tasks), mixed, _funcname(tasks)


class Batch:
    """dask-compatible wrapper that executes a batch of tasks"""
    def __init__(self, tasks):
        # collect some metadata from the tasks to ease Batch calls
        # introspection when debugging
        self._num_tasks, self._mixed, self._funcname = _make_tasks_summary(
            tasks
        )

    def __call__(self, tasks=None):
        results = []
        with parallel_backend('dask'):
            for func, args, kwargs in tasks:
                results.append(func(*args, **kwargs))
        return results

    def __repr__(self):
        descr = f"batch_of_{self._funcname}_{self._num_tasks}_calls"
        if self._mixed:
            descr = "mixed_" + descr
        return descr


def _joblib_probe_task():
    # Noop used by the joblib connector to probe when workers are ready.
    pass


class DaskDistributedBackend(AutoBatchingMixin, ParallelBackendBase):
    MIN_IDEAL_BATCH_DURATION = 0.2
    MAX_IDEAL_BATCH_DURATION = 1.0
    supports_timeout = True

    def __init__(self, scheduler_host=None, scatter=None,
                 client=None, loop=None, wait_for_workers_timeout=10,
                 **submit_kwargs):
        super().__init__()

        if distributed is None:
            msg = ("You are trying to use 'dask' as a joblib parallel backend "
                   "but dask is not installed. Please install dask "
                   "to fix this error.")
            raise ValueError(msg)

        if client is None:
            if scheduler_host:
                client = Client(scheduler_host, loop=loop,
                                set_as_default=False)
            else:
                try:
                    client = get_client()
                except ValueError as e:
                    msg = ("To use Joblib with Dask first create a Dask Client"
                           "\n\n"
                           "    from dask.distributed import Client\n"
                           "    client = Client()\n"
                           "or\n"
                           "    client = Client('scheduler-address:8786')")
                    raise ValueError(msg) from e

        self.client = client

        if scatter is not None and not isinstance(scatter, (list, tuple)):
            raise TypeError("scatter must be a list/tuple, got "
                            "`%s`" % type(scatter).__name__)

        if scatter is not None and len(scatter) > 0:
            # Keep a reference to the scattered data to keep the ids the same
            self._scatter = list(scatter)
            scattered = self.client.scatter(scatter, broadcast=True)
            self.data_futures = {id(x): f for x, f in zip(scatter, scattered)}
        else:
            self._scatter = []
            self.data_futures = {}
        self.wait_for_workers_timeout = wait_for_workers_timeout
        self.submit_kwargs = submit_kwargs
        self.waiting_futures = as_completed(
            [],
            loop=client.loop,
            with_results=True,
            raise_errors=False
        )
        self._results = {}
        self._callbacks = {}

    async def _collect(self):
        while self._continue:
            async for future, result in self.waiting_futures:
                cf_future = self._results.pop(future)
                callback = self._callbacks.pop(future)
                if future.status == "error":
                    typ, exc, tb = result
                    cf_future.set_exception(exc)
                else:
                    cf_future.set_result(result)
                    callback(result)
            await asyncio.sleep(0.01)

    def __reduce__(self):
        return (DaskDistributedBackend, ())

    def get_nested_backend(self):
        return DaskDistributedBackend(client=self.client), -1

    def configure(self, n_jobs=1, parallel=None, **backend_args):
        self.parallel = parallel
        return self.effective_n_jobs(n_jobs)

    def start_call(self):
        self._continue = True
        self.client.loop.add_callback(self._collect)
        self.call_data_futures = _WeakKeyDictionary()

    def stop_call(self):
        # The explicit call to clear is required to break a cycling reference
        # to the futures.
        self._continue = False
        # wait for the future collection routine (self._backend._collect) to
        # finish in order to limit asyncio warnings due to aborting _collect
        # during a following backend termination call
        time.sleep(0.01)
        self.call_data_futures.clear()

    def effective_n_jobs(self, n_jobs):
        effective_n_jobs = sum(self.client.ncores().values())
        if effective_n_jobs != 0 or not self.wait_for_workers_timeout:
            return effective_n_jobs

        # If there is no worker, schedule a probe task to wait for the workers
        # to come up and be available. If the dask cluster is in adaptive mode
        # task might cause the cluster to provision some workers.
        try:
            self.client.submit(_joblib_probe_task).result(
                timeout=self.wait_for_workers_timeout)
        except _TimeoutError as e:
            error_msg = (
                "DaskDistributedBackend has no worker after {} seconds. "
                "Make sure that workers are started and can properly connect "
                "to the scheduler and increase the joblib/dask connection "
                "timeout with:\n\n"
                "parallel_backend('dask', wait_for_workers_timeout={})"
            ).format(self.wait_for_workers_timeout,
                     max(10, 2 * self.wait_for_workers_timeout))
            raise TimeoutError(error_msg) from e
        return sum(self.client.ncores().values())

    async def _to_func_args(self, func):
        itemgetters = dict()

        # Futures that are dynamically generated during a single call to
        # Parallel.__call__.
        call_data_futures = getattr(self, 'call_data_futures', None)

        async def maybe_to_futures(args):
            out = []
            for arg in args:
                arg_id = id(arg)
                if arg_id in itemgetters:
                    out.append(itemgetters[arg_id])
                    continue

                f = self.data_futures.get(arg_id, None)
                if f is None and call_data_futures is not None:
                    try:
                        f = call_data_futures[arg]
                    except KeyError:
                        pass
                    if f is None:
                        if is_weakrefable(arg) and sizeof(arg) > 1e3:
                            # Automatically scatter large objects to some of
                            # the workers to avoid duplicated data transfers.
                            # Rely on automated inter-worker data stealing if
                            # more workers need to reuse this data
                            # concurrently.
                            # set hash=False - nested scatter calls (i.e
                            # calling client.scatter inside a dask worker)
                            # using hash=True often raise CancelledError,
                            # see dask/distributed#3703
                            [f] = await self.client.scatter(
                                [arg],
                                asynchronous=True,
                                hash=False
                            )
                            call_data_futures[arg] = f

                if f is not None:
                    out.append(f)
                else:
                    out.append(arg)
            return out

        tasks = []
        for f, args, kwargs in func.items:
            args = list(await maybe_to_futures(args))
            kwargs = dict(zip(kwargs.keys(),
                              await maybe_to_futures(kwargs.values())))
            tasks.append((f, args, kwargs))

        return (Batch(tasks), tasks)

    def apply_async(self, func, callback=None):

        cf_future = concurrent.futures.Future()
        cf_future.get = cf_future.result  # achieve AsyncResult API

        async def f(func, callback):
            batch, tasks = await self._to_func_args(func)
            key = f'{repr(batch)}-{uuid4().hex}'

            dask_future = self.client.submit(
                batch, tasks=tasks, key=key, **self.submit_kwargs
            )
            self.waiting_futures.add(dask_future)
            self._callbacks[dask_future] = callback
            self._results[dask_future] = cf_future

        self.client.loop.add_callback(f, func, callback)

        return cf_future

    def abort_everything(self, ensure_ready=True):
        """ Tell the client to cancel any task submitted via this instance

        joblib.Parallel will never access those results
        """
        with self.waiting_futures.lock:
            self.waiting_futures.futures.clear()
            while not self.waiting_futures.queue.empty():
                self.waiting_futures.queue.get()

    @contextlib.contextmanager
    def retrieval_context(self):
        """Override ParallelBackendBase.retrieval_context to avoid deadlocks.

        This removes thread from the worker's thread pool (using 'secede').
        Seceding avoids deadlock in nested parallelism settings.
        """
        # See 'joblib.Parallel.__call__' and 'joblib.Parallel.retrieve' for how
        # this is used.
        if hasattr(thread_state, 'execution_state'):
            # we are in a worker. Secede to avoid deadlock.
            secede()

        yield

        if hasattr(thread_state, 'execution_state'):
            rejoin()