sdk_worker.py

#
# Licensed to the Apache Software Foundation (ASF) under one or more
# contributor license agreements.  See the NOTICE file distributed with
# this work for additional information regarding copyright ownership.
# The ASF licenses this file to You under the Apache License, Version 2.0
# (the "License"); you may not use this file except in compliance with
# the License.  You may obtain a copy of the License at
#
#    http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#

"""SDK harness for executing Python Fns via the Fn API."""

# pytype: skip-file
# mypy: disallow-untyped-defs

import abc
import collections
import contextlib
import functools
import json
import logging
import queue
import sys
import threading
import time
import traceback
from concurrent import futures
from typing import TYPE_CHECKING
from typing import Any
from typing import Callable
from typing import DefaultDict
from typing import Dict
from typing import FrozenSet
from typing import Generic
from typing import Iterable
from typing import Iterator
from typing import List
from typing import MutableMapping
from typing import Optional
from typing import Tuple
from typing import TypeVar
from typing import Union

import grpc

from apache_beam.coders import coder_impl
from apache_beam.metrics import monitoring_infos
from apache_beam.metrics.execution import MetricsEnvironment
from apache_beam.portability.api import beam_fn_api_pb2
from apache_beam.portability.api import beam_fn_api_pb2_grpc
from apache_beam.portability.api import metrics_pb2
from apache_beam.runners.worker import bundle_processor
from apache_beam.runners.worker import data_plane
from apache_beam.runners.worker import data_sampler
from apache_beam.runners.worker import statesampler
from apache_beam.runners.worker.channel_factory import GRPCChannelFactory
from apache_beam.runners.worker.data_plane import PeriodicThread
from apache_beam.runners.worker.statecache import CacheAware
from apache_beam.runners.worker.statecache import StateCache
from apache_beam.runners.worker.worker_id_interceptor import WorkerIdInterceptor
from apache_beam.runners.worker.worker_status import FnApiWorkerStatusHandler
from apache_beam.utils import thread_pool_executor
from apache_beam.utils.sentinel import Sentinel
from apache_beam.version import __version__ as beam_version

if TYPE_CHECKING:
  from apache_beam.portability.api import endpoints_pb2
  from apache_beam.utils.profiler import Profile

T = TypeVar('T')
_KT = TypeVar('_KT')
_VT = TypeVar('_VT')

_LOGGER = logging.getLogger(__name__)

DEFAULT_BUNDLE_PROCESSOR_CACHE_SHUTDOWN_THRESHOLD_S = 60
# The number of ProcessBundleRequest instruction ids the BundleProcessorCache
# will remember for not running instructions.
MAX_KNOWN_NOT_RUNNING_INSTRUCTIONS = 1000
# The number of ProcessBundleRequest instruction ids that BundleProcessorCache
# will remember for failed instructions.
MAX_FAILED_INSTRUCTIONS = 10000

# retry on transient UNAVAILABLE grpc error from state channels.
_GRPC_SERVICE_CONFIG = json.dumps({
    "methodConfig": [{
        "name": [{
            "service": "org.apache.beam.model.fn_execution.v1.BeamFnState"
        }],
        "retryPolicy": {
            "maxAttempts": 5,
            "initialBackoff": "0.1s",
            "maxBackoff": "5s",
            "backoffMultiplier": 2,
            "retryableStatusCodes": ["UNAVAILABLE"],
        },
    }]
})


class ShortIdCache(object):
  """ Cache for MonitoringInfo "short ids"
  """
  def __init__(self):
    # type: () -> None
    self._lock = threading.Lock()
    self._last_short_id = 0
    self._info_key_to_short_id = {}  # type: Dict[FrozenSet, str]
    self._short_id_to_info = {}  # type: Dict[str, metrics_pb2.MonitoringInfo]

  def get_short_id(self, monitoring_info):
    # type: (metrics_pb2.MonitoringInfo) -> str

    """ Returns the assigned shortId for a given MonitoringInfo, assigns one if
    not assigned already.
    """
    key = monitoring_infos.to_key(monitoring_info)
    with self._lock:
      try:
        return self._info_key_to_short_id[key]
      except KeyError:
        self._last_short_id += 1

        # Convert to a hex string (and drop the '0x') for some compression
        shortId = hex(self._last_short_id)[2:]

        payload_cleared = metrics_pb2.MonitoringInfo()
        payload_cleared.CopyFrom(monitoring_info)
        payload_cleared.ClearField('payload')

        self._info_key_to_short_id[key] = shortId
        self._short_id_to_info[shortId] = payload_cleared
        return shortId

  def get_infos(self, short_ids):
    #type: (Iterable[str]) -> Dict[str, metrics_pb2.MonitoringInfo]

    """ Gets the base MonitoringInfo (with payload cleared) for each short ID.

    Throws KeyError if an unassigned short ID is encountered.
    """
    return {
        short_id: self._short_id_to_info[short_id]
        for short_id in short_ids
    }


SHORT_ID_CACHE = ShortIdCache()


class SdkHarness(object):
  REQUEST_METHOD_PREFIX = '_request_'

  def __init__(
      self,
      control_address,  # type: str
      credentials=None,  # type: Optional[grpc.ChannelCredentials]
      worker_id=None,  # type: Optional[str]
      # Caching is disabled by default
      state_cache_size=0,  # type: int
      # time-based data buffering is disabled by default
      data_buffer_time_limit_ms=0,  # type: int
      profiler_factory=None,  # type: Optional[Callable[..., Profile]]
      status_address=None,  # type: Optional[str]
      # Heap dump through status api is disabled by default
      enable_heap_dump=False,  # type: bool
      data_sampler=None,  # type: Optional[data_sampler.DataSampler]
      # Unrecoverable SDK harness initialization error (if any)
      # that should be reported to the runner when proocessing the first bundle.
      deferred_exception=None,  # type: Optional[Exception]
      runner_capabilities=frozenset(),  # type: FrozenSet[str]
      element_processing_timeout_minutes=None,  # type: Optional[int]
  ):
    # type: (...) -> None
    self._alive = True
    self._worker_index = 0
    self._worker_id = worker_id
    self._state_cache = StateCache(state_cache_size)
    self._deferred_exception = deferred_exception
    options = [('grpc.max_receive_message_length', -1),
               ('grpc.max_send_message_length', -1)]
    if credentials is None:
      _LOGGER.info('Creating insecure control channel for %s.', control_address)
      self._control_channel = GRPCChannelFactory.insecure_channel(
          control_address, options=options)
    else:
      _LOGGER.info('Creating secure control channel for %s.', control_address)
      self._control_channel = GRPCChannelFactory.secure_channel(
          control_address, credentials, options=options)
    grpc.channel_ready_future(self._control_channel).result(timeout=60)
    _LOGGER.info('Control channel established.')

    self._control_channel = grpc.intercept_channel(
        self._control_channel, WorkerIdInterceptor(self._worker_id))
    self._data_channel_factory = data_plane.GrpcClientDataChannelFactory(
        credentials, self._worker_id, data_buffer_time_limit_ms)
    self._state_handler_factory = GrpcStateHandlerFactory(
        state_cache=self._state_cache,
        credentials=credentials,
        worker_id=self._worker_id)
    self._profiler_factory = profiler_factory
    self.data_sampler = data_sampler
    self.runner_capabilities = runner_capabilities
    self._element_processing_timeout_minutes = element_processing_timeout_minutes

    def default_factory(id):
      # type: (str) -> beam_fn_api_pb2.ProcessBundleDescriptor
      return self._control_stub.GetProcessBundleDescriptor(
          beam_fn_api_pb2.GetProcessBundleDescriptorRequest(
              process_bundle_descriptor_id=id))

    self._fns = KeyedDefaultDict(default_factory)
    # BundleProcessor cache across all workers.
    self._bundle_processor_cache = BundleProcessorCache(
        self.runner_capabilities,
        state_handler_factory=self._state_handler_factory,
        data_channel_factory=self._data_channel_factory,
        fns=self._fns,
        data_sampler=self.data_sampler,
    )
    self._status_handler = None  # type: Optional[FnApiWorkerStatusHandler]
    if status_address:
      try:
        self._status_handler = FnApiWorkerStatusHandler(
            status_address,
            self._bundle_processor_cache,
            self._state_cache,
            enable_heap_dump,
            element_processing_timeout_minutes=self.
            _element_processing_timeout_minutes)
      except Exception:
        traceback_string = traceback.format_exc()
        _LOGGER.warning(
            'Error creating worker status request handler, '
            'skipping status report. Trace back: %s' % traceback_string)

    # TODO(BEAM-8998) use common
    # thread_pool_executor.shared_unbounded_instance() to process bundle
    # progress once dataflow runner's excessive progress polling is removed.
    self._report_progress_executor = futures.ThreadPoolExecutor(max_workers=1)
    self._worker_thread_pool = thread_pool_executor.shared_unbounded_instance()
    self._responses = queue.Queue(
    )  # type: queue.Queue[Union[beam_fn_api_pb2.InstructionResponse, Sentinel]]
    _LOGGER.info('Initializing SDKHarness with unbounded number of workers.')

  def run(self):
    # type: () -> None
    self._control_stub = beam_fn_api_pb2_grpc.BeamFnControlStub(
        self._control_channel)
    no_more_work = Sentinel.sentinel

    def get_responses():
      # type: () -> Iterator[beam_fn_api_pb2.InstructionResponse]
      while True:
        response = self._responses.get()
        if response is no_more_work:
          return
        yield response

    self._alive = True

    try:
      for work_request in self._control_stub.Control(get_responses()):
        _LOGGER.debug('Got work %s', work_request.instruction_id)
        request_type = work_request.WhichOneof('request')

        if request_type is None:
          raise RuntimeError(
              "Cannot interpret a request received over control channel. "
              "This is not expected. "
              "Verify that SDK was not accidentally downgraded at runtime. "
              f"SDK version: {beam_version}, "
              f"instruction id: {work_request.instruction_id}, "
              f"raw request: {str(work_request.SerializeToString())}")

        # Name spacing the request method with 'request_'. The called method
        # will be like self.request_register(request)
        getattr(self, SdkHarness.REQUEST_METHOD_PREFIX + request_type)(
            work_request)
    finally:
      self._alive = False
      if self.data_sampler:
        self.data_sampler.stop()

    _LOGGER.info('No more requests from control plane')
    _LOGGER.info('SDK Harness waiting for in-flight requests to complete')
    # Wait until existing requests are processed.
    self._worker_thread_pool.shutdown()
    # get_responses may be blocked on responses.get(), but we need to return
    # control to its caller.
    self._responses.put(no_more_work)
    # Stop all the workers and clean all the associated resources
    self._data_channel_factory.close()
    self._state_handler_factory.close()
    self._bundle_processor_cache.shutdown()
    if self._status_handler:
      self._status_handler.close()
    _LOGGER.info('Done consuming work.')

  def _execute(
      self,
      task,  # type: Callable[[], beam_fn_api_pb2.InstructionResponse]
      request  # type:  beam_fn_api_pb2.InstructionRequest
  ):
    # type: (...) -> None
    with statesampler.instruction_id(request.instruction_id):
      try:
        response = task()
      except:  # pylint: disable=bare-except
        traceback_string = traceback.format_exc()
        print(traceback_string, file=sys.stderr)
        _LOGGER.error(
            'Error processing instruction %s. Original traceback is\n%s\n',
            request.instruction_id,
            traceback_string)
        response = beam_fn_api_pb2.InstructionResponse(
            instruction_id=request.instruction_id, error=traceback_string)
      self._responses.put(response)

  def _request_register(self, request):
    # type: (beam_fn_api_pb2.InstructionRequest) -> None
    # registration request is handled synchronously
    self._execute(lambda: self.create_worker().do_instruction(request), request)

  def _request_process_bundle(self, request):
    # type: (beam_fn_api_pb2.InstructionRequest) -> None
    if self._deferred_exception:
      raise self._deferred_exception
    self._bundle_processor_cache.activate(request.instruction_id)
    self._request_execute(request)

  def _request_process_bundle_split(self, request):
    # type: (beam_fn_api_pb2.InstructionRequest) -> None
    self._request_process_bundle_action(request)

  def _request_process_bundle_progress(self, request):
    # type: (beam_fn_api_pb2.InstructionRequest) -> None
    self._request_process_bundle_action(request)

  def _request_process_bundle_action(self, request):
    # type: (beam_fn_api_pb2.InstructionRequest) -> None
    def task():
      # type: () -> None
      self._execute(
          lambda: self.create_worker().do_instruction(request), request)

    self._report_progress_executor.submit(task)

  def _request_finalize_bundle(self, request):
    # type: (beam_fn_api_pb2.InstructionRequest) -> None
    self._request_execute(request)

  def _request_harness_monitoring_infos(self, request):
    # type: (beam_fn_api_pb2.InstructionRequest) -> None
    process_wide_monitoring_infos = MetricsEnvironment.process_wide_container(
    ).to_runner_api_monitoring_infos(None).values()
    self._execute(
        lambda: beam_fn_api_pb2.InstructionResponse(
            instruction_id=request.instruction_id, harness_monitoring_infos=(
                beam_fn_api_pb2.HarnessMonitoringInfosResponse(
                    monitoring_data={
                        SHORT_ID_CACHE.get_short_id(info): info.payload
                        for info in process_wide_monitoring_infos
                    }))),
        request)

  def _request_monitoring_infos(self, request):
    # type: (beam_fn_api_pb2.InstructionRequest) -> None
    self._execute(
        lambda: beam_fn_api_pb2.InstructionResponse(
            instruction_id=request.instruction_id, monitoring_infos=
            beam_fn_api_pb2.MonitoringInfosMetadataResponse(
                monitoring_info=SHORT_ID_CACHE.get_infos(
                    request.monitoring_infos.monitoring_info_id))),
        request)

  def _request_execute(self, request):
    # type: (beam_fn_api_pb2.InstructionRequest) -> None
    def task():
      # type: () -> None
      self._execute(
          lambda: self.create_worker().do_instruction(request), request)

    self._worker_thread_pool.submit(task)
    _LOGGER.debug(
        "Currently using %s threads." % len(self._worker_thread_pool._workers))

  def _request_sample_data(self, request):
    # type: (beam_fn_api_pb2.InstructionRequest) -> None

    def get_samples(request):
      # type: (beam_fn_api_pb2.InstructionRequest) -> beam_fn_api_pb2.InstructionResponse
      samples = beam_fn_api_pb2.SampleDataResponse()
      if self.data_sampler is not None:
        samples = self.data_sampler.samples(request.sample_data.pcollection_ids)

      return beam_fn_api_pb2.InstructionResponse(
          instruction_id=request.instruction_id, sample_data=samples)

    self._execute(lambda: get_samples(request), request)

  def create_worker(self):
    # type: () -> SdkWorker
    return SdkWorker(
        self._bundle_processor_cache, profiler_factory=self._profiler_factory)


class BundleProcessorCache(object):
  """A cache for ``BundleProcessor``s.

  ``BundleProcessor`` objects are cached by the id of their
  ``beam_fn_api_pb2.ProcessBundleDescriptor``.

  Attributes:
    fns (dict): A dictionary that maps bundle descriptor IDs to instances of
      ``beam_fn_api_pb2.ProcessBundleDescriptor``.
    state_handler_factory (``StateHandlerFactory``): Used to create state
      handlers to be used by a ``bundle_processor.BundleProcessor`` during
      processing.
    data_channel_factory (``data_plane.DataChannelFactory``)
    active_bundle_processors (dict): A dictionary, indexed by instruction IDs,
      containing ``bundle_processor.BundleProcessor`` objects that are currently
      active processing the corresponding instruction.
    cached_bundle_processors (dict): A dictionary, indexed by bundle processor
      id, of cached ``bundle_processor.BundleProcessor`` that are not currently
      performing processing.
  """
  periodic_shutdown = None  # type: Optional[PeriodicThread]

  def __init__(
      self,
      runner_capabilities,  # type: FrozenSet[str]
      state_handler_factory,  # type: StateHandlerFactory
      data_channel_factory,  # type: data_plane.DataChannelFactory
      fns,  # type: MutableMapping[str, beam_fn_api_pb2.ProcessBundleDescriptor]
      data_sampler=None,  # type: Optional[data_sampler.DataSampler]
  ):
    # type: (...) -> None
    self.runner_capabilities = runner_capabilities
    self.fns = fns
    self.state_handler_factory = state_handler_factory
    self.data_channel_factory = data_channel_factory
    self.known_not_running_instruction_ids = collections.OrderedDict(
    )  # type: collections.OrderedDict[str, bool]
    self.failed_instruction_ids = collections.OrderedDict(
    )  # type: collections.OrderedDict[str, Exception]
    self.active_bundle_processors = {
    }  # type: Dict[str, Tuple[str, bundle_processor.BundleProcessor]]
    self.cached_bundle_processors = collections.defaultdict(
        list)  # type: DefaultDict[str, List[bundle_processor.BundleProcessor]]
    self.last_access_times = collections.defaultdict(
        float)  # type: DefaultDict[str, float]
    self._schedule_periodic_shutdown()
    self._lock = threading.Lock()
    self.data_sampler = data_sampler

  def register(self, bundle_descriptor):
    # type: (beam_fn_api_pb2.ProcessBundleDescriptor) -> None

    """Register a ``beam_fn_api_pb2.ProcessBundleDescriptor`` by its id."""
    self.fns[bundle_descriptor.id] = bundle_descriptor

  def activate(self, instruction_id):
    # type: (str) -> None

    """Makes the ``instruction_id`` known to the bundle processor.

    Allows ``lookup`` to return ``None``. Necessary if ``lookup`` can occur
    before ``get``.
    """
    with self._lock:
      self.known_not_running_instruction_ids[instruction_id] = True

  def get(self, instruction_id, bundle_descriptor_id):
    # type: (str, str) -> bundle_processor.BundleProcessor

    """
    Return the requested ``BundleProcessor``, creating it if necessary.

    Moves the ``BundleProcessor`` from the inactive to the active cache.
    """
    with self._lock:
      try:
        # pop() is threadsafe
        processor = self.cached_bundle_processors[bundle_descriptor_id].pop()
        self.active_bundle_processors[
          instruction_id] = bundle_descriptor_id, processor
        try:
          del self.known_not_running_instruction_ids[instruction_id]
        except KeyError:
          # The instruction may have not been pre-registered before execution
          # since activate() may have never been invoked
          pass
        return processor
      except IndexError:
        pass

    # Make sure we instantiate the processor while not holding the lock.

    # Reduce risks of concurrent modifications of the same protos
    # captured in bundle descriptor when the same bundle descriptor is used
    # in different instructions.
    pbd = beam_fn_api_pb2.ProcessBundleDescriptor()
    pbd.MergeFrom(self.fns[bundle_descriptor_id])

    processor = bundle_processor.BundleProcessor(
        self.runner_capabilities,
        pbd,
        self.state_handler_factory.create_state_handler(
            pbd.state_api_service_descriptor),
        self.data_channel_factory,
        self.data_sampler)
    with self._lock:
      self.active_bundle_processors[
        instruction_id] = bundle_descriptor_id, processor
      try:
        del self.known_not_running_instruction_ids[instruction_id]
      except KeyError:
        # The instruction may have not been pre-registered before execution
        # since activate() may have never been invoked
        pass
    return processor

  def lookup(self, instruction_id):
    # type: (str) -> Optional[bundle_processor.BundleProcessor]

    """
    Return the requested ``BundleProcessor`` from the cache.

    Will return ``None`` if the BundleProcessor is known but not yet ready. Will
    raise an error if the ``instruction_id`` is not known or has been discarded.
    """
    with self._lock:
      if instruction_id in self.failed_instruction_ids:
        e = self.failed_instruction_ids[instruction_id]
        raise RuntimeError(
            'Bundle processing associated with %s has failed. '
            'Check prior failing response and attached exception for details.' %
            instruction_id) from e
      processor = self.active_bundle_processors.get(
          instruction_id, (None, None))[-1]
      if processor:
        return processor
      if instruction_id in self.known_not_running_instruction_ids:
        return None
      raise RuntimeError('Unknown process bundle id %s.' % instruction_id)

  def discard(self, instruction_id, exception):
    # type: (str, Exception) -> None

    """
    Marks the instruction id as failed shutting down the ``BundleProcessor``.
    """
    with self._lock:
      self.failed_instruction_ids[instruction_id] = exception
      while len(self.failed_instruction_ids) > MAX_FAILED_INSTRUCTIONS:
        self.failed_instruction_ids.popitem(last=False)
      processor = self.active_bundle_processors[instruction_id][1]
      del self.active_bundle_processors[instruction_id]

    # Perform the shutdown while not holding the lock.
    processor.shutdown()

  def release(self, instruction_id):
    # type: (str) -> None

    """
    Release the requested ``BundleProcessor``.

    Resets the ``BundleProcessor`` and moves it from the active to the
    inactive cache.
    """
    with self._lock:
      self.known_not_running_instruction_ids[instruction_id] = True
      while len(self.known_not_running_instruction_ids
                ) > MAX_KNOWN_NOT_RUNNING_INSTRUCTIONS:
        self.known_not_running_instruction_ids.popitem(last=False)
      descriptor_id, processor = (
          self.active_bundle_processors.pop(instruction_id))

    # Make sure that we reset the processor while not holding the lock.
    processor.reset()
    with self._lock:
      self.last_access_times[descriptor_id] = time.time()
      self.cached_bundle_processors[descriptor_id].append(processor)

  def shutdown(self):
    # type: () -> None

    """
    Shutdown all ``BundleProcessor``s in the cache.
    """
    if self.periodic_shutdown:
      self.periodic_shutdown.cancel()
      self.periodic_shutdown.join()
      self.periodic_shutdown = None

    for instruction_id in list(self.active_bundle_processors.keys()):
      self.discard(instruction_id, RuntimeError('Shutdown invoked'))
    for cached_bundle_processors in self.cached_bundle_processors.values():
      BundleProcessorCache._shutdown_cached_bundle_processors(
          cached_bundle_processors)

  def _schedule_periodic_shutdown(self):
    # type: () -> None
    def shutdown_inactive_bundle_processors():
      # type: () -> None
      inactive_descriptor_ids = []
      inactive_time = time.time(
      ) - DEFAULT_BUNDLE_PROCESSOR_CACHE_SHUTDOWN_THRESHOLD_S
      with self._lock:
        for descriptor_id, last_access_time in self.last_access_times.items():
          if (inactive_time > last_access_time):
            inactive_descriptor_ids.append(descriptor_id)

      # Shutdown can be expensive, keep out of lock
      for descriptor_id in inactive_descriptor_ids:
        BundleProcessorCache._shutdown_cached_bundle_processors(
            self.cached_bundle_processors[descriptor_id])

    self.periodic_shutdown = PeriodicThread(
        DEFAULT_BUNDLE_PROCESSOR_CACHE_SHUTDOWN_THRESHOLD_S,
        shutdown_inactive_bundle_processors)
    self.periodic_shutdown.daemon = True
    self.periodic_shutdown.start()

  @staticmethod
  def _shutdown_cached_bundle_processors(cached_bundle_processors):
    # type: (List[bundle_processor.BundleProcessor]) -> None
    try:
      while True:
        # pop() is threadsafe
        bundle_processor = cached_bundle_processors.pop()
        bundle_processor.shutdown()
    except IndexError:
      pass


class SdkWorker(object):
  def __init__(
      self,
      bundle_processor_cache,  # type: BundleProcessorCache
      profiler_factory=None,  # type: Optional[Callable[..., Profile]]
  ):
    # type: (...) -> None
    self.bundle_processor_cache = bundle_processor_cache
    self.profiler_factory = profiler_factory

  def do_instruction(self, request):
    # type: (beam_fn_api_pb2.InstructionRequest) -> beam_fn_api_pb2.InstructionResponse
    request_type = request.WhichOneof('request')
    if request_type:
      # E.g. if register is set, this will call self.register(request.register))
      return getattr(self, request_type)(
          getattr(request, request_type), request.instruction_id)
    else:
      raise NotImplementedError

  def register(
      self,
      request,  # type: beam_fn_api_pb2.RegisterRequest
      instruction_id  # type: str
  ):
    # type: (...) -> beam_fn_api_pb2.InstructionResponse

    """Registers a set of ``beam_fn_api_pb2.ProcessBundleDescriptor``s.

    This set of ``beam_fn_api_pb2.ProcessBundleDescriptor`` come as part of a
    ``beam_fn_api_pb2.RegisterRequest``, which the runner sends to the SDK
    worker before starting processing to register stages.
    """

    for process_bundle_descriptor in request.process_bundle_descriptor:
      self.bundle_processor_cache.register(process_bundle_descriptor)
    return beam_fn_api_pb2.InstructionResponse(
        instruction_id=instruction_id,
        register=beam_fn_api_pb2.RegisterResponse())

  def process_bundle(
      self,
      request,  # type: beam_fn_api_pb2.ProcessBundleRequest
      instruction_id  # type: str
  ):
    # type: (...) -> beam_fn_api_pb2.InstructionResponse
    bundle_processor = self.bundle_processor_cache.get(
        instruction_id, request.process_bundle_descriptor_id)
    try:
      with bundle_processor.state_handler.process_instruction_id(
          instruction_id, request.cache_tokens):
        with self.maybe_profile(instruction_id):
          delayed_applications, requests_finalization = (
              bundle_processor.process_bundle(instruction_id))
          monitoring_infos = bundle_processor.monitoring_infos()
          response = beam_fn_api_pb2.InstructionResponse(
              instruction_id=instruction_id,
              process_bundle=beam_fn_api_pb2.ProcessBundleResponse(
                  residual_roots=delayed_applications,
                  monitoring_infos=monitoring_infos,
                  monitoring_data={
                      SHORT_ID_CACHE.get_short_id(info): info.payload
                      for info in monitoring_infos
                  },
                  requires_finalization=requests_finalization))
      # Don't release here if finalize is needed.
      if not requests_finalization:
        self.bundle_processor_cache.release(instruction_id)
      return response
    except Exception as e:  # pylint: disable=bare-except
      # Don't re-use bundle processors on failure.
      self.bundle_processor_cache.discard(instruction_id, e)
      raise

  def process_bundle_split(
      self,
      request,  # type: beam_fn_api_pb2.ProcessBundleSplitRequest
      instruction_id  # type: str
  ):
    # type: (...) -> beam_fn_api_pb2.InstructionResponse
    try:
      processor = self.bundle_processor_cache.lookup(request.instruction_id)
    except RuntimeError:
      return beam_fn_api_pb2.InstructionResponse(
          instruction_id=instruction_id, error=traceback.format_exc())
    # Return an empty response if we aren't running. This can happen
    # if the ProcessBundleRequest has not started or already finished.
    process_bundle_split = (
        processor.try_split(request)
        if processor else beam_fn_api_pb2.ProcessBundleSplitResponse())
    return beam_fn_api_pb2.InstructionResponse(
        instruction_id=instruction_id,
        process_bundle_split=process_bundle_split)

  def process_bundle_progress(
      self,
      request,  # type: beam_fn_api_pb2.ProcessBundleProgressRequest
      instruction_id  # type: str
  ):
    # type: (...) -> beam_fn_api_pb2.InstructionResponse
    try:
      processor = self.bundle_processor_cache.lookup(request.instruction_id)
    except RuntimeError:
      return beam_fn_api_pb2.InstructionResponse(
          instruction_id=instruction_id, error=traceback.format_exc())
    if processor:
      monitoring_infos = processor.monitoring_infos()
      consuming_received_data = processor.consuming_received_data
    else:
      # Return an empty response if we aren't running. This can happen
      # if the ProcessBundleRequest has not started or already finished.
      monitoring_infos = []
      consuming_received_data = False
    return beam_fn_api_pb2.InstructionResponse(
        instruction_id=instruction_id,
        process_bundle_progress=beam_fn_api_pb2.ProcessBundleProgressResponse(
            monitoring_infos=monitoring_infos,
            monitoring_data={
                SHORT_ID_CACHE.get_short_id(info): info.payload
                for info in monitoring_infos
            },
            consuming_received_data=consuming_received_data))

  def finalize_bundle(
      self,
      request,  # type: beam_fn_api_pb2.FinalizeBundleRequest
      instruction_id  # type: str
  ):
    # type: (...) -> beam_fn_api_pb2.InstructionResponse
    try:
      processor = self.bundle_processor_cache.lookup(request.instruction_id)
    except RuntimeError:
      return beam_fn_api_pb2.InstructionResponse(
          instruction_id=instruction_id, error=traceback.format_exc())
    if processor:
      try:
        finalize_response = processor.finalize_bundle()
        self.bundle_processor_cache.release(request.instruction_id)
        return beam_fn_api_pb2.InstructionResponse(
            instruction_id=instruction_id, finalize_bundle=finalize_response)
      except Exception as e:
        self.bundle_processor_cache.discard(request.instruction_id, e)
        raise
    # We can reach this state if there was an erroneous request to finalize
    # the bundle while it is being initialized or has already been finalized
    # and released.
    raise RuntimeError(
        'Bundle is not in a finalizable state for %s' % instruction_id)

  @contextlib.contextmanager
  def maybe_profile(self, instruction_id):
    # type: (str) -> Iterator[None]
    if self.profiler_factory:
      profiler = self.profiler_factory(instruction_id)
      if profiler:
        with profiler:
          yield
      else:
        yield
    else:
      yield


class StateHandler(metaclass=abc.ABCMeta):
  """An abstract object representing a ``StateHandler``."""
  @abc.abstractmethod
  def get_raw(
      self,
      state_key,  # type: beam_fn_api_pb2.StateKey
      continuation_token=None  # type: Optional[bytes]
  ):
    # type: (...) -> Tuple[bytes, Optional[bytes]]

    """Gets the contents of state for the given state key.

    State is associated to a state key, AND an instruction_id, which is set
    when calling process_instruction_id.

    Returns a tuple with the contents in state, and an optional continuation
    token, which is used to page the API.
    """
    raise NotImplementedError(type(self))

  @abc.abstractmethod
  def append_raw(
      self,
      state_key,  # type: beam_fn_api_pb2.StateKey
      data  # type: bytes
  ):
    # type: (...) -> _Future

    """Append the input data into the state key.

    Returns a future that allows one to wait for the completion of the call.

    State is associated to a state key, AND an instruction_id, which is set
    when calling process_instruction_id.
    """
    raise NotImplementedError(type(self))

  @abc.abstractmethod
  def clear(self, state_key):
    # type: (beam_fn_api_pb2.StateKey) -> _Future

    """Clears the contents of a cell for the input state key.

    Returns a future that allows one to wait for the completion of the call.

    State is associated to a state key, AND an instruction_id, which is set
    when calling process_instruction_id.
    """
    raise NotImplementedError(type(self))

  @abc.abstractmethod
  @contextlib.contextmanager
  def process_instruction_id(self, bundle_id):
    # type: (str) -> Iterator[None]

    """Switch the context of the state handler to a specific instruction.

    This must be called before performing any write or read operations on the
    existing state.
    """
    raise NotImplementedError(type(self))

  @abc.abstractmethod
  def done(self):
    # type: () -> None

    """Mark the state handler as done, and potentially delete all context."""
    raise NotImplementedError(type(self))


class StateHandlerFactory(metaclass=abc.ABCMeta):
  """An abstract factory for creating ``DataChannel``."""
  @abc.abstractmethod
  def create_state_handler(self, api_service_descriptor):
    # type: (endpoints_pb2.ApiServiceDescriptor) -> CachingStateHandler

    """Returns a ``StateHandler`` from the given ApiServiceDescriptor."""
    raise NotImplementedError(type(self))

  @abc.abstractmethod
  def close(self):
    # type: () -> None

    """Close all channels that this factory owns."""
    raise NotImplementedError(type(self))


class GrpcStateHandlerFactory(StateHandlerFactory):
  """A factory for ``GrpcStateHandler``.

  Caches the created channels by ``state descriptor url``.
  """
  def __init__(self, state_cache, credentials=None, worker_id=None):
    # type: (StateCache, Optional[grpc.ChannelCredentials], Optional[str]) -> None
    self._state_handler_cache = {}  # type: Dict[str, CachingStateHandler]
    self._lock = threading.Lock()
    self._throwing_state_handler = ThrowingStateHandler()
    self._credentials = credentials
    self._state_cache = state_cache
    self._worker_id = worker_id

  def create_state_handler(self, api_service_descriptor):
    # type: (endpoints_pb2.ApiServiceDescriptor) -> CachingStateHandler
    if not api_service_descriptor:
      return self._throwing_state_handler
    url = api_service_descriptor.url
    if url not in self._state_handler_cache:
      with self._lock:
        if url not in self._state_handler_cache:
          # Options to have no limits (-1) on the size of the messages
          # received or sent over the data plane. The actual buffer size is
          # controlled in a layer above.
          options = [('grpc.max_receive_message_length', -1),
                     ('grpc.max_send_message_length', -1),
                     ('grpc.service_config', _GRPC_SERVICE_CONFIG)]
          if self._credentials is None:
            _LOGGER.info('Creating insecure state channel for %s.', url)
            grpc_channel = GRPCChannelFactory.insecure_channel(
                url, options=options)
          else:
            _LOGGER.info('Creating secure state channel for %s.', url)
            grpc_channel = GRPCChannelFactory.secure_channel(
                url, self._credentials, options=options)
          _LOGGER.info('State channel established.')
          # Add workerId to the grpc channel
          grpc_channel = grpc.intercept_channel(
              grpc_channel, WorkerIdInterceptor(self._worker_id))
          self._state_handler_cache[url] = GlobalCachingStateHandler(
              self._state_cache,
              GrpcStateHandler(
                  beam_fn_api_pb2_grpc.BeamFnStateStub(grpc_channel)))
    return self._state_handler_cache[url]

  def close(self):
    # type: () -> None
    _LOGGER.info('Closing all cached gRPC state handlers.')
    for _, state_handler in self._state_handler_cache.items():
      state_handler.done()
    self._state_handler_cache.clear()
    self._state_cache.invalidate_all()


class CachingStateHandler(metaclass=abc.ABCMeta):
  @abc.abstractmethod
  @contextlib.contextmanager
  def process_instruction_id(self, bundle_id, cache_tokens):
    # type: (str, Iterable[beam_fn_api_pb2.ProcessBundleRequest.CacheToken]) -> Iterator[None]
    raise NotImplementedError(type(self))

  @abc.abstractmethod
  def blocking_get(
      self,
      state_key,  # type: beam_fn_api_pb2.StateKey
      coder,  # type: coder_impl.CoderImpl
  ):
    # type: (...) -> Iterable[Any]
    raise NotImplementedError(type(self))

  @abc.abstractmethod
  def extend(
      self,
      state_key,  # type: beam_fn_api_pb2.StateKey
      coder,  # type: coder_impl.CoderImpl
      elements,  # type: Iterable[Any]
  ):
    # type: (...) -> _Future
    raise NotImplementedError(type(self))

  @abc.abstractmethod
  def clear(self, state_key):
    # type: (beam_fn_api_pb2.StateKey) -> _Future
    raise NotImplementedError(type(self))

  @abc.abstractmethod
  def done(self):
    # type: () -> None
    raise NotImplementedError(type(self))


class ThrowingStateHandler(CachingStateHandler):
  """A caching state handler that errors on any requests."""
  @contextlib.contextmanager
  def process_instruction_id(self, bundle_id, cache_tokens):
    # type: (str, Iterable[beam_fn_api_pb2.ProcessBundleRequest.CacheToken]) -> Iterator[None]
    raise RuntimeError(
        'Unable to handle state requests for ProcessBundleDescriptor '
        'for bundle id %s.' % bundle_id)

  def blocking_get(
      self,
      state_key,  # type: beam_fn_api_pb2.StateKey
      coder,  # type: coder_impl.CoderImpl
  ):
    # type: (...) -> Iterable[Any]
    raise RuntimeError(
        'Unable to handle state requests for ProcessBundleDescriptor without '
        'state ApiServiceDescriptor for state key %s.' % state_key)

  def extend(
      self,
      state_key,  # type: beam_fn_api_pb2.StateKey
      coder,  # type: coder_impl.CoderImpl
      elements,  # type: Iterable[Any]
  ):
    # type: (...) -> _Future
    raise RuntimeError(
        'Unable to handle state requests for ProcessBundleDescriptor without '
        'state ApiServiceDescriptor for state key %s.' % state_key)

  def clear(self, state_key):
    # type: (beam_fn_api_pb2.StateKey) -> _Future
    raise RuntimeError(
        'Unable to handle state requests for ProcessBundleDescriptor without '
        'state ApiServiceDescriptor for state key %s.' % state_key)

  def done(self):
    # type: () -> None
    raise RuntimeError(
        'Unable to handle state requests for ProcessBundleDescriptor.')


class GrpcStateHandler(StateHandler):

  _DONE = Sentinel.sentinel

  def __init__(self, state_stub):
    # type: (beam_fn_api_pb2_grpc.BeamFnStateStub) -> None
    self._lock = threading.Lock()
    self._state_stub = state_stub
    self._requests = queue.Queue(
    )  # type: queue.Queue[Union[beam_fn_api_pb2.StateRequest, Sentinel]]
    self._responses_by_id = {}  # type: Dict[str, _Future]
    self._last_id = 0
    self._exception = None  # type: Optional[Exception]
    self._context = threading.local()
    self.start()

  @contextlib.contextmanager
  def process_instruction_id(self, bundle_id):
    # type: (str) -> Iterator[None]
    if getattr(self._context, 'process_instruction_id', None) is not None:
      raise RuntimeError(
          'Already bound to %r' % self._context.process_instruction_id)
    self._context.process_instruction_id = bundle_id
    try:
      yield
    finally:
      self._context.process_instruction_id = None

  def start(self):
    # type: () -> None
    self._done = False

    def request_iter():
      # type: () -> Iterator[beam_fn_api_pb2.StateRequest]
      while True:
        request = self._requests.get()
        if request is self._DONE or self._done:
          break
        yield request

    responses = self._state_stub.State(request_iter())

    def pull_responses():
      # type: () -> None
      try:
        for response in responses:
          # Popping an item from a dictionary is atomic in cPython
          future = self._responses_by_id.pop(response.id)
          future.set(response)
          if self._done:
            break
      except Exception as e:
        self._exception = e
        raise

    reader = threading.Thread(target=pull_responses, name='read_state')
    reader.daemon = True
    reader.start()

  def done(self):
    # type: () -> None
    self._done = True
    self._requests.put(self._DONE)

  def get_raw(
      self,
      state_key,  # type: beam_fn_api_pb2.StateKey
      continuation_token=None  # type: Optional[bytes]
  ):
    # type: (...) -> Tuple[bytes, Optional[bytes]]
    response = self._blocking_request(
        beam_fn_api_pb2.StateRequest(
            state_key=state_key,
            get=beam_fn_api_pb2.StateGetRequest(
                continuation_token=continuation_token)))
    return response.get.data, response.get.continuation_token

  def append_raw(
      self,
      state_key,  # type: Optional[beam_fn_api_pb2.StateKey]
      data  # type: bytes
  ):
    # type: (...) -> _Future
    return self._request(
        beam_fn_api_pb2.StateRequest(
            state_key=state_key,
            append=beam_fn_api_pb2.StateAppendRequest(data=data)))

  def clear(self, state_key):
    # type: (Optional[beam_fn_api_pb2.StateKey]) -> _Future
    return self._request(
        beam_fn_api_pb2.StateRequest(
            state_key=state_key, clear=beam_fn_api_pb2.StateClearRequest()))

  def _request(self, request):
    # type: (beam_fn_api_pb2.StateRequest) -> _Future[beam_fn_api_pb2.StateResponse]
    request.id = self._next_id()
    request.instruction_id = self._context.process_instruction_id
    # Adding a new item to a dictionary is atomic in cPython
    self._responses_by_id[request.id] = future = _Future[
        beam_fn_api_pb2.StateResponse]()
    # Request queue is thread-safe
    self._requests.put(request)
    return future

  def _blocking_request(self, request):
    # type: (beam_fn_api_pb2.StateRequest) -> beam_fn_api_pb2.StateResponse
    req_future = self._request(request)
    while not req_future.wait(timeout=1):
      if self._exception:
        raise self._exception
      elif self._done:
        raise RuntimeError()
    response = req_future.get()
    if response.error:
      raise RuntimeError(response.error)
    else:
      return response

  def _next_id(self):
    # type: () -> str
    with self._lock:
      # Use a lock here because this GrpcStateHandler is shared across all
      # requests which have the same process bundle descriptor. State requests
      # can concurrently access this section if a Runner uses threads / workers
      # (aka "parallelism") to send data to this SdkHarness and its workers.
      self._last_id += 1
      request_id = self._last_id
    return str(request_id)


class GlobalCachingStateHandler(CachingStateHandler):
  """ A State handler which retrieves and caches state.
   If caching is activated, caches across bundles using a supplied cache token.
   If activated but no cache token is supplied, caching is done at the bundle
   level.
  """
  def __init__(
      self,
      global_state_cache,  # type: StateCache
      underlying_state  # type: StateHandler
  ):
    # type: (...) -> None
    self._underlying = underlying_state
    self._state_cache = global_state_cache
    self._context = threading.local()

    # state retrieval time statistics
    self._retrieval_time = 0.0
    self._get_raw_called = 0
    self._warn_interval = 60.0

  @contextlib.contextmanager
  def process_instruction_id(self, bundle_id, cache_tokens):
    # type: (str, Iterable[beam_fn_api_pb2.ProcessBundleRequest.CacheToken]) -> Iterator[None]
    if getattr(self._context, 'user_state_cache_token', None) is not None:
      raise RuntimeError(
          'Cache tokens already set to %s' %
          self._context.user_state_cache_token)
    self._context.side_input_cache_tokens = {}
    user_state_cache_token = None
    for cache_token_struct in cache_tokens:
      if cache_token_struct.HasField("user_state"):
        # There should only be one user state token present
        assert not user_state_cache_token
        user_state_cache_token = cache_token_struct.token
      elif cache_token_struct.HasField("side_input"):
        self._context.side_input_cache_tokens[
            cache_token_struct.side_input.transform_id,
            cache_token_struct.side_input.
            side_input_id] = cache_token_struct.token
    # TODO: Consider a two-level cache to avoid extra logic and locking
    # for items cached at the bundle level.
    self._context.bundle_cache_token = bundle_id
    try:
      self._context.user_state_cache_token = user_state_cache_token
      with self._underlying.process_instruction_id(bundle_id):
        yield
    finally:
      self._context.side_input_cache_tokens = {}
      self._context.user_state_cache_token = None
      self._context.bundle_cache_token = None

  def blocking_get(
      self,
      state_key,  # type: beam_fn_api_pb2.StateKey
      coder,  # type: coder_impl.CoderImpl
  ):
    # type: (...) -> Iterable[Any]
    cache_token = self._get_cache_token(state_key)
    if not cache_token:
      # Cache disabled / no cache token. Can't do a lookup/store in the cache.
      # Fall back to lazily materializing the state, one element at a time.
      return self._lazy_iterator(state_key, coder)
    # Cache lookup
    cache_state_key = self._convert_to_cache_key(state_key)
    return self._state_cache.get(
        (cache_state_key, cache_token),
        lambda key: self._partially_cached_iterable(state_key, coder))

  def extend(
      self,
      state_key,  # type: beam_fn_api_pb2.StateKey
      coder,  # type: coder_impl.CoderImpl
      elements,  # type: Iterable[Any]
  ):
    # type: (...) -> _Future
    cache_token = self._get_cache_token(state_key)
    if cache_token:
      # Update the cache if the value is already present and
      # can be updated.
      cache_key = self._convert_to_cache_key(state_key)
      cached_value = self._state_cache.peek((cache_key, cache_token))
      if isinstance(cached_value, list):
        # The state is fully cached and can be extended

        # Materialize provided iterable to ensure reproducible iterations,
        # here and when writing to the state handler below.
        elements = list(elements)
        cached_value.extend(elements)
        # Re-insert into the cache the updated value so the updated size is
        # reflected.
        self._state_cache.put((cache_key, cache_token), cached_value)

    # Write to state handler
    futures = []
    out = coder_impl.create_OutputStream()
    for element in elements:
      coder.encode_to_stream(element, out, True)
      if out.size() > data_plane._DEFAULT_SIZE_FLUSH_THRESHOLD:
        futures.append(self._underlying.append_raw(state_key, out.get()))
        out = coder_impl.create_OutputStream()
    if out.size():
      futures.append(self._underlying.append_raw(state_key, out.get()))
    return _DeferredCall(
        lambda *results: beam_fn_api_pb2.StateResponse(
            error='\n'.join(
                result.error for result in results if result and result.error),
            append=beam_fn_api_pb2.StateAppendResponse()),
        *futures)

  def clear(self, state_key):
    # type: (beam_fn_api_pb2.StateKey) -> _Future
    cache_token = self._get_cache_token(state_key)
    if cache_token:
      cache_key = self._convert_to_cache_key(state_key)
      self._state_cache.put((cache_key, cache_token), [])
    return self._underlying.clear(state_key)

  def done(self):
    # type: () -> None
    self._underlying.done()

  def _lazy_iterator(
      self,
      state_key,  # type: beam_fn_api_pb2.StateKey
      coder,  # type: coder_impl.CoderImpl
      continuation_token=None  # type: Optional[bytes]
  ):
    # type: (...) -> Iterator[Any]

    """Materializes the state lazily, one element at a time.
       :return A generator which returns the next element if advanced.
    """
    while True:
      input_stream, continuation_token = self._get_raw(
          state_key, continuation_token)

      while input_stream.size() > 0:
        yield coder.decode_from_stream(input_stream, True)
      if not continuation_token:
        break

  def _get_raw(
      self,
      state_key,  # type: beam_fn_api_pb2.StateKey
      continuation_token  # type: Optional[bytes]
  ):
    # type: (...) -> Tuple[coder_impl.create_InputStream, Optional[bytes]]

    """Call underlying get_raw with performance statistics and detection."""
    start_time = time.time()

    data, continuation_token = (
        self._underlying.get_raw(state_key, continuation_token))

    input_stream = coder_impl.create_InputStream(data)

    self._retrieval_time += time.time() - start_time
    self._get_raw_called += 1

    if self._retrieval_time > self._warn_interval:
      _LOGGER.warning(
          "Retrieving state %d times costed %.0f seconds. It may be due to "
          "insufficient state cache size and/or frequent direct access of "
          "states.\nConsider adding '--max_cache_memory_usage_mb' pipeline "
          "option to increase state cache size or switch to materialized "
          "(pvalue.AsList) side input if applicable." %
          (self._get_raw_called, self._retrieval_time))
      # reset counts
      self._retrieval_time = 0.0
      self._get_raw_called = 0
      self._warn_interval *= 2

    return input_stream, continuation_token

  def _get_cache_token(self, state_key):
    # type: (beam_fn_api_pb2.StateKey) -> Optional[bytes]
    if not self._state_cache.is_cache_enabled():
      return None
    elif state_key.HasField('bag_user_state'):
      if self._context.user_state_cache_token:
        return self._context.user_state_cache_token
      else:
        return self._context.bundle_cache_token
    elif state_key.WhichOneof('type').endswith('_side_input'):  # type: ignore[union-attr]
      side_input = getattr(state_key, state_key.WhichOneof('type'))  # type: ignore[arg-type]
      return self._context.side_input_cache_tokens.get(
          (side_input.transform_id, side_input.side_input_id),
          self._context.bundle_cache_token)
    return None

  def _partially_cached_iterable(
      self,
      state_key,  # type: beam_fn_api_pb2.StateKey
      coder  # type: coder_impl.CoderImpl
  ):
    # type: (...) -> Iterable[Any]

    """Materialized the first page of data, concatenated with a lazy iterable
    of the rest, if any.
    """
    input_stream, continuation_token = self._get_raw(state_key, None)
    head = []
    while input_stream.size() > 0:
      head.append(coder.decode_from_stream(input_stream, True))

    if not continuation_token:
      return head
    else:
      return self.ContinuationIterable(
          head,
          functools.partial(
              self._lazy_iterator, state_key, coder, continuation_token))

  class ContinuationIterable(Generic[T], CacheAware):
    def __init__(self, head, continue_iterator_fn):
      # type: (Iterable[T], Callable[[], Iterable[T]]) -> None
      self.head = head
      self.continue_iterator_fn = continue_iterator_fn

    def __iter__(self):
      # type: () -> Iterator[T]
      for item in self.head:
        yield item
      for item in self.continue_iterator_fn():
        yield item

    def get_referents_for_cache(self):
      # type: () -> List[Any]
      # Only capture the size of the elements and not the
      # continuation iterator since it references objects
      # we don't want to include in the cache measurement.
      return [self.head]

  @staticmethod
  def _convert_to_cache_key(state_key):
    # type: (beam_fn_api_pb2.StateKey) -> bytes
    return state_key.SerializeToString()


class _Future(Generic[T]):
  """A simple future object to implement blocking requests.
  """
  def __init__(self):
    # type: () -> None
    self._event = threading.Event()

  def wait(self, timeout=None):
    # type: (Optional[float]) -> bool
    return self._event.wait(timeout)

  def get(self, timeout=None):
    # type: (Optional[float]) -> T
    if self.wait(timeout):
      return self._value
    else:
      raise LookupError()

  def set(self, value):
    # type: (T) -> _Future[T]
    self._value = value
    self._event.set()
    return self

  @classmethod
  def done(cls):
    # type: () -> _Future[None]
    if not hasattr(cls, 'DONE'):
      done_future = _Future[None]()
      done_future.set(None)
      cls.DONE = done_future  # type: ignore[attr-defined]
    return cls.DONE  # type: ignore[attr-defined]


class _DeferredCall(_Future[T]):
  def __init__(self, func, *args):
    # type: (Callable[..., Any], *Any) -> None
    self._func = func
    self._args = [
        arg if isinstance(arg, _Future) else _Future().set(arg) for arg in args
    ]

  def wait(self, timeout=None):
    # type: (Optional[float]) -> bool
    return all(arg.wait(timeout) for arg in self._args)

  def get(self, timeout=None):
    # type: (Optional[float]) -> T
    return self._func(*(arg.get(timeout) for arg in self._args))

  def set(self, value):
    # type: (T) -> _Future[T]
    raise NotImplementedError()


class KeyedDefaultDict(DefaultDict[_KT, _VT]):
  if TYPE_CHECKING:
    # we promise to only use a subset of what DefaultDict can do
    def __init__(self, default_factory):
      # type: (Callable[[_KT], _VT]) -> None
      pass

  def __missing__(self, key):
    # type: (_KT) -> _VT
    # typing: default_factory takes an arg, but the base class does not
    self[key] = self.default_factory(key)  # type: ignore # pylint: disable=E1137
    return self[key]