sdks/python/apache_beam/runners/worker/operations.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.
#

"""Worker operations executor."""

# pytype: skip-file
# pylint: disable=super-with-arguments

# ruff: noqa: UP006
import collections
import logging
import threading
import warnings
from typing import TYPE_CHECKING
from typing import Any
from typing import DefaultDict
from typing import Dict
from typing import FrozenSet
from typing import Iterable
from typing import Iterator
from typing import List
from typing import Mapping
from typing import NamedTuple
from typing import Optional
from typing import Tuple

from apache_beam import coders
from apache_beam.internal import pickler
from apache_beam.io import iobase
from apache_beam.metrics import monitoring_infos
from apache_beam.metrics.cells import DistributionData
from apache_beam.metrics.execution import MetricsContainer
from apache_beam.portability.api import metrics_pb2
from apache_beam.runners import common
from apache_beam.runners.common import Receiver
from apache_beam.runners.worker import opcounters
from apache_beam.runners.worker import operation_specs
from apache_beam.runners.worker import sideinputs
from apache_beam.runners.worker.data_sampler import DataSampler
from apache_beam.transforms import combiners
from apache_beam.transforms import core
from apache_beam.transforms import sideinputs as apache_sideinputs
from apache_beam.transforms import userstate
from apache_beam.transforms import window
from apache_beam.transforms.combiners import PhasedCombineFnExecutor
from apache_beam.transforms.combiners import curry_combine_fn
from apache_beam.transforms.window import GlobalWindows
from apache_beam.typehints.batch import BatchConverter
from apache_beam.utils.windowed_value import WindowedBatch
from apache_beam.utils.windowed_value import WindowedValue

if TYPE_CHECKING:
  from apache_beam.runners.sdf_utils import SplitResultPrimary
  from apache_beam.runners.sdf_utils import SplitResultResidual
  from apache_beam.runners.worker.bundle_processor import ExecutionContext
  from apache_beam.runners.worker.data_sampler import OutputSampler
  from apache_beam.runners.worker.statesampler import StateSampler
  from apache_beam.transforms.userstate import TimerSpec

# Allow some "pure mode" declarations.
try:
  import cython
except ImportError:

  class FakeCython(object):
    compiled = False

  globals()['cython'] = FakeCython()

_globally_windowed_value = GlobalWindows.windowed_value(None)
_global_window_type = type(_globally_windowed_value.windows[0])

_LOGGER = logging.getLogger(__name__)

SdfSplitResultsPrimary = Tuple['DoOperation', 'SplitResultPrimary']
SdfSplitResultsResidual = Tuple['DoOperation', 'SplitResultResidual']


# TODO(BEAM-9324) Remove these workarounds once upgraded to Cython 3
def _cast_to_operation(value):
  if cython.compiled:
    return cython.cast(Operation, value)
  else:
    return value


# TODO(BEAM-9324) Remove these workarounds once upgraded to Cython 3
def _cast_to_receiver(value):
  if cython.compiled:
    return cython.cast(Receiver, value)
  else:
    return value


class ConsumerSet(Receiver):
  """A ConsumerSet represents a graph edge between two Operation nodes.

  The ConsumerSet object collects information from the output of the
  Operation at one end of its edge and the input of the Operation at
  the other edge.
  ConsumerSet are attached to the outputting Operation.
  """
  @staticmethod
  def create(
      counter_factory,
      step_name,  # type: str
      output_index,
      consumers,  # type: List[Operation]
      coder,
      producer_type_hints,
      producer_batch_converter,  # type: Optional[BatchConverter]
      output_sampler=None,  # type: Optional[OutputSampler]
  ):
    # type: (...) -> ConsumerSet
    if len(consumers) == 1:
      consumer = consumers[0]

      consumer_batch_preference = consumer.get_batching_preference()
      consumer_batch_converter = consumer.get_input_batch_converter()
      if (not consumer_batch_preference.supports_batches and
          producer_batch_converter is None and
          consumer_batch_converter is None):
        return SingletonElementConsumerSet(
            counter_factory,
            step_name,
            output_index,
            consumer,
            coder,
            producer_type_hints,
            output_sampler)

    return GeneralPurposeConsumerSet(
        counter_factory,
        step_name,
        output_index,
        coder,
        producer_type_hints,
        consumers,
        producer_batch_converter,
        output_sampler)

  def __init__(
      self,
      counter_factory,
      step_name,  # type: str
      output_index,
      consumers,
      coder,
      producer_type_hints,
      producer_batch_converter,
      output_sampler):
    self.opcounter = opcounters.OperationCounters(
        counter_factory,
        step_name,
        coder,
        output_index,
        producer_type_hints=producer_type_hints,
        producer_batch_converter=producer_batch_converter)
    # Used in repr.
    self.step_name = step_name
    self.output_index = output_index
    self.coder = coder
    self.consumers = consumers
    self.output_sampler = output_sampler
    self.element_sampler = (
        output_sampler.element_sampler if output_sampler else None)
    self.execution_context = None  # type: Optional[ExecutionContext]

  def try_split(self, fraction_of_remainder):
    # type: (...) -> Optional[Any]
    # TODO(SDF): Consider supporting splitting each consumer individually.
    # This would never come up in the existing SDF expansion, but might
    # be useful to support fused SDF nodes.
    # This would require dedicated delivery of the split results to each
    # of the consumers separately.
    return None

  def current_element_progress(self):
    # type: () -> Optional[iobase.RestrictionProgress]

    """Returns the progress of the current element.

    This progress should be an instance of
    apache_beam.io.iobase.RestrictionProgress, or None if progress is unknown.
    """
    # TODO(SDF): Could implement this as a weighted average, if it becomes
    # useful to split on.
    return None

  def update_counters_start(self, windowed_value):
    # type: (WindowedValue) -> None
    self.opcounter.update_from(windowed_value)

    if self.execution_context is not None:
      self.execution_context.output_sampler = self.output_sampler

    # The following code is optimized by inlining a function call. Because this
    # is called for every element, a function call is too expensive (order of
    # 100s of nanoseconds). Furthermore, a lock was purposefully not used
    # between here and the DataSampler as an additional operation. The tradeoff
    # is that some samples might be dropped, but it is better than the
    # alternative which is double sampling the same element.
    if self.element_sampler is not None:
      if not self.element_sampler.has_element:
        self.element_sampler.el = windowed_value
        self.element_sampler.has_element = True

  def update_counters_finish(self):
    # type: () -> None
    self.opcounter.update_collect()

  def update_counters_batch(self, windowed_batch):
    # type: (WindowedBatch) -> None
    self.opcounter.update_from_batch(windowed_batch)

  def __repr__(self):
    return '%s[%s.out%s, coder=%s, len(consumers)=%s]' % (
        self.__class__.__name__,
        self.step_name,
        self.output_index,
        self.coder,
        len(self.consumers))


class SingletonElementConsumerSet(ConsumerSet):
  """ConsumerSet representing a single consumer that can only process elements
  (not batches)."""
  def __init__(
      self,
      counter_factory,
      step_name,
      output_index,
      consumer,  # type: Operation
      coder,
      producer_type_hints,
      output_sampler):
    super().__init__(
        counter_factory,
        step_name,
        output_index, [consumer],
        coder,
        producer_type_hints,
        None,
        output_sampler)
    self.consumer = consumer

  def receive(self, windowed_value):
    # type: (WindowedValue) -> None
    self.update_counters_start(windowed_value)
    self.consumer.process(windowed_value)
    self.update_counters_finish()

  def receive_batch(self, windowed_batch):
    raise AssertionError(
        "SingletonElementConsumerSet.receive_batch is not implemented")

  def flush(self):
    # SingletonElementConsumerSet has no buffer to flush
    pass

  def try_split(self, fraction_of_remainder):
    # type: (...) -> Optional[Any]
    return self.consumer.try_split(fraction_of_remainder)

  def current_element_progress(self):
    return self.consumer.current_element_progress()


class GeneralPurposeConsumerSet(ConsumerSet):
  """ConsumerSet implementation that handles all combinations of possible edges.
  """
  MAX_BATCH_SIZE = 4096

  def __init__(
      self,
      counter_factory,
      step_name,  # type: str
      output_index,
      coder,
      producer_type_hints,
      consumers,  # type: List[Operation]
      producer_batch_converter,
      output_sampler):
    super().__init__(
        counter_factory,
        step_name,
        output_index,
        consumers,
        coder,
        producer_type_hints,
        producer_batch_converter,
        output_sampler)

    self.producer_batch_converter = producer_batch_converter

    # Partition consumers into three groups:
    # - consumers that will be passed elements
    # - consumers that will be passed batches (where their input batch type
    #   matches the output of the producer)
    # - consumers that will be passed converted batches
    self.element_consumers: List[Operation] = []
    self.passthrough_batch_consumers: List[Operation] = []
    other_batch_consumers: DefaultDict[
        BatchConverter, List[Operation]] = collections.defaultdict(lambda: [])

    for consumer in consumers:
      if not consumer.get_batching_preference().supports_batches:
        self.element_consumers.append(consumer)
      elif (consumer.get_input_batch_converter() ==
            self.producer_batch_converter):
        self.passthrough_batch_consumers.append(consumer)
      else:
        # Batch consumer with a mismatched batch type
        if consumer.get_batching_preference().supports_elements:
          # Pass it elements if we can
          self.element_consumers.append(consumer)
        else:
          # As a last resort, explode and rebatch
          consumer_batch_converter = consumer.get_input_batch_converter()
          # This consumer supports batches, it must have a batch converter
          assert consumer_batch_converter is not None
          other_batch_consumers[consumer_batch_converter].append(consumer)

    self.other_batch_consumers: Dict[BatchConverter, List[Operation]] = dict(
        other_batch_consumers)

    self.has_batch_consumers = (
        self.passthrough_batch_consumers or self.other_batch_consumers)
    self._batched_elements: List[Any] = []

  def receive(self, windowed_value):
    # type: (WindowedValue) -> None

    self.update_counters_start(windowed_value)

    for consumer in self.element_consumers:
      _cast_to_operation(consumer).process(windowed_value)

    # TODO: Do this branching when contstructing ConsumerSet
    if self.has_batch_consumers:
      self._batched_elements.append(windowed_value)
      if len(self._batched_elements) >= self.MAX_BATCH_SIZE:
        self.flush()

    # TODO(https://github.com/apache/beam/issues/21655): Properly estimate
    # sizes in the batch-consumer only case, this undercounts large iterables
    self.update_counters_finish()

  def receive_batch(self, windowed_batch):
    if self.element_consumers:
      for wv in windowed_batch.as_windowed_values(
          self.producer_batch_converter.explode_batch):
        for consumer in self.element_consumers:
          _cast_to_operation(consumer).process(wv)

    for consumer in self.passthrough_batch_consumers:
      _cast_to_operation(consumer).process_batch(windowed_batch)

    for (consumer_batch_converter,
         consumers) in self.other_batch_consumers.items():
      # Explode and rebatch into the new batch type (ouch!)
      # TODO: Register direct conversions for equivalent batch types

      for consumer in consumers:
        warnings.warn(
            f"Input to operation {consumer} must be rebatched from type "
            f"{self.producer_batch_converter.batch_type!r} to "
            f"{consumer_batch_converter.batch_type!r}.\n"
            "This is very inefficient, consider re-structuring your pipeline "
            "or adding a DoFn to directly convert between these types.",
            InefficientExecutionWarning)
        _cast_to_operation(consumer).process_batch(
            windowed_batch.with_values(
                consumer_batch_converter.produce_batch(
                    self.producer_batch_converter.explode_batch(
                        windowed_batch.values))))

    self.update_counters_batch(windowed_batch)

  def flush(self):
    if not self.has_batch_consumers or not self._batched_elements:
      return

    for batch_converter, consumers in self.other_batch_consumers.items():
      for windowed_batch in WindowedBatch.from_windowed_values(
          self._batched_elements, produce_fn=batch_converter.produce_batch):
        for consumer in consumers:
          _cast_to_operation(consumer).process_batch(windowed_batch)

    for consumer in self.passthrough_batch_consumers:
      for windowed_batch in WindowedBatch.from_windowed_values(
          self._batched_elements,
          produce_fn=self.producer_batch_converter.produce_batch):
        _cast_to_operation(consumer).process_batch(windowed_batch)

    self._batched_elements = []


class Operation(object):
  """An operation representing the live version of a work item specification.

  An operation can have one or more outputs and for each output it can have
  one or more receiver operations that will take that as input.
  """
  def __init__(
      self,
      name_context,  # type: common.NameContext
      spec,
      counter_factory,
      state_sampler  # type: StateSampler
  ):
    """Initializes a worker operation instance.

    Args:
      name_context: A NameContext instance, with the name information for this
        operation.
      spec: A operation_specs.Worker* instance.
      counter_factory: The CounterFactory to use for our counters.
      state_sampler: The StateSampler for the current operation.
    """
    assert isinstance(name_context, common.NameContext)
    self.name_context = name_context

    self.spec = spec
    self.counter_factory = counter_factory
    self.execution_context = None  # type: Optional[ExecutionContext]
    self.consumers = collections.defaultdict(
        list)  # type: DefaultDict[int, List[Operation]]

    # These are overwritten in the legacy harness.
    self.metrics_container = MetricsContainer(self.name_context.metrics_name())

    self.state_sampler = state_sampler
    self.scoped_start_state = self.state_sampler.scoped_state(
        self.name_context, 'start', metrics_container=self.metrics_container)
    self.scoped_process_state = self.state_sampler.scoped_state(
        self.name_context, 'process', metrics_container=self.metrics_container)
    self.scoped_finish_state = self.state_sampler.scoped_state(
        self.name_context, 'finish', metrics_container=self.metrics_container)
    # TODO(ccy): the '-abort' state can be added when the abort is supported in
    # Operations.
    self.receivers = []  # type: List[ConsumerSet]
    # Legacy workers cannot call setup() until after setting additional state
    # on the operation.
    self.setup_done = False
    self.step_name = None  # type: Optional[str]
    self.data_sampler: Optional[DataSampler] = None

  def setup(self, data_sampler=None):
    # type: (Optional[DataSampler]) -> None

    """Set up operation.

    This must be called before any other methods of the operation."""
    with self.scoped_start_state:
      self.data_sampler = data_sampler
      self.debug_logging_enabled = logging.getLogger().isEnabledFor(
          logging.DEBUG)
      transform_id = self.name_context.transform_id

      # Everything except WorkerSideInputSource, which is not a
      # top-level operation, should have output_coders
      #TODO(pabloem): Define better what step name is used here.
      if getattr(self.spec, 'output_coders', None):

        def get_output_sampler(output_num):
          if data_sampler is None:
            return None
          return data_sampler.sampler_for_output(transform_id, output_num)

        self.receivers = [
            ConsumerSet.create(
                self.counter_factory,
                self.name_context.logging_name(),
                i,
                self.consumers[i],
                coder,
                self._get_runtime_performance_hints(),
                self.get_output_batch_converter(),
                get_output_sampler(i))
            for i, coder in enumerate(self.spec.output_coders)
        ]
    self.setup_done = True

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

    """Start operation."""
    if not self.setup_done:
      # For legacy workers.
      self.setup(self.data_sampler)

    # The ExecutionContext is per instruction and so cannot be set at
    # initialization time.
    if self.data_sampler is not None:
      for receiver in self.receivers:
        receiver.execution_context = self.execution_context

  def get_batching_preference(self):
    # By default operations don't support batching, require Receiver to unbatch
    return common.BatchingPreference.BATCH_FORBIDDEN

  def get_input_batch_converter(self) -> Optional[BatchConverter]:
    """Returns a batch type converter if this operation can accept a batch,
    otherwise None."""
    return None

  def get_output_batch_converter(self) -> Optional[BatchConverter]:
    """Returns a batch type converter if this operation can produce a batch,
    otherwise None."""
    return None

  def process(self, o):
    # type: (WindowedValue) -> None

    """Process element in operation."""
    pass

  def process_batch(self, batch: WindowedBatch):
    pass

  def finalize_bundle(self):
    # type: () -> None
    pass

  def needs_finalization(self):
    return False

  def try_split(self, fraction_of_remainder):
    # type: (...) -> Optional[Any]
    return None

  def current_element_progress(self):
    return None

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

    """Finish operation."""
    for receiver in self.receivers:
      _cast_to_receiver(receiver).flush()

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

    """Tear down operation.

    No other methods of this operation should be called after this."""
    pass

  def reset(self):
    # type: () -> None
    self.metrics_container.reset()

  def output(self, windowed_value, output_index=0):
    # type: (WindowedValue, int) -> None
    _cast_to_receiver(self.receivers[output_index]).receive(windowed_value)

  def add_receiver(self, operation, output_index=0):
    # type: (Operation, int) -> None

    """Adds a receiver operation for the specified output."""
    self.consumers[output_index].append(operation)

  def monitoring_infos(self, transform_id, tag_to_pcollection_id):
    # type: (str, Dict[str, str]) -> Dict[FrozenSet, metrics_pb2.MonitoringInfo]

    """Returns the list of MonitoringInfos collected by this operation."""
    all_monitoring_infos = self.execution_time_monitoring_infos(transform_id)
    all_monitoring_infos.update(
        self.pcollection_count_monitoring_infos(tag_to_pcollection_id))
    all_monitoring_infos.update(self.user_monitoring_infos(transform_id))
    return all_monitoring_infos

  def pcollection_count_monitoring_infos(self, tag_to_pcollection_id):
    # type: (Dict[str, str]) -> Dict[FrozenSet, metrics_pb2.MonitoringInfo]

    """Returns the element count MonitoringInfo collected by this operation."""

    # Skip producing monitoring infos if there is more then one receiver
    # since there is no way to provide a mapping from tag to pcollection id
    # within Operation.
    if len(self.receivers) != 1 or len(tag_to_pcollection_id) != 1:
      return {}

    all_monitoring_infos = {}
    pcollection_id = next(iter(tag_to_pcollection_id.values()))
    receiver = self.receivers[0]
    elem_count_mi = monitoring_infos.int64_counter(
        monitoring_infos.ELEMENT_COUNT_URN,
        receiver.opcounter.element_counter.value(),
        pcollection=pcollection_id,
    )

    (unused_mean, sum, count, min, max) = (
        receiver.opcounter.mean_byte_counter.value())

    sampled_byte_count = monitoring_infos.int64_distribution(
        monitoring_infos.SAMPLED_BYTE_SIZE_URN,
        DistributionData(sum, count, min, max),
        pcollection=pcollection_id,
    )
    all_monitoring_infos[monitoring_infos.to_key(elem_count_mi)] = elem_count_mi
    all_monitoring_infos[monitoring_infos.to_key(
        sampled_byte_count)] = sampled_byte_count

    return all_monitoring_infos

  def user_monitoring_infos(self, transform_id):
    # type: (str) -> Dict[FrozenSet, metrics_pb2.MonitoringInfo]

    """Returns the user MonitoringInfos collected by this operation."""
    return self.metrics_container.to_runner_api_monitoring_infos(transform_id)

  def execution_time_monitoring_infos(self, transform_id):
    # type: (str) -> Dict[FrozenSet, metrics_pb2.MonitoringInfo]
    total_time_spent_msecs = (
        self.scoped_start_state.sampled_msecs_int() +
        self.scoped_process_state.sampled_msecs_int() +
        self.scoped_finish_state.sampled_msecs_int())
    mis = [
        monitoring_infos.int64_counter(
            monitoring_infos.START_BUNDLE_MSECS_URN,
            self.scoped_start_state.sampled_msecs_int(),
            ptransform=transform_id),
        monitoring_infos.int64_counter(
            monitoring_infos.PROCESS_BUNDLE_MSECS_URN,
            self.scoped_process_state.sampled_msecs_int(),
            ptransform=transform_id),
        monitoring_infos.int64_counter(
            monitoring_infos.FINISH_BUNDLE_MSECS_URN,
            self.scoped_finish_state.sampled_msecs_int(),
            ptransform=transform_id),
        monitoring_infos.int64_counter(
            monitoring_infos.TOTAL_MSECS_URN,
            total_time_spent_msecs,
            ptransform=transform_id),
    ]
    return {monitoring_infos.to_key(mi): mi for mi in mis}

  def __str__(self):
    """Generates a useful string for this object.

    Compactly displays interesting fields.  In particular, pickled
    fields are not displayed.  Note that we collapse the fields of the
    contained Worker* object into this object, since there is a 1-1
    mapping between Operation and operation_specs.Worker*.

    Returns:
      Compact string representing this object.
    """
    return self.str_internal()

  def str_internal(self, is_recursive=False):
    """Internal helper for __str__ that supports recursion.

    When recursing on receivers, keep the output short.
    Args:
      is_recursive: whether to omit some details, particularly receivers.
    Returns:
      Compact string representing this object.
    """
    printable_name = self.__class__.__name__
    if hasattr(self, 'step_name'):
      printable_name += ' %s' % self.name_context.logging_name()
      if is_recursive:
        # If we have a step name, stop here, no more detail needed.
        return '<%s>' % printable_name

    if self.spec is None:
      printable_fields = []
    else:
      printable_fields = operation_specs.worker_printable_fields(self.spec)

    if not is_recursive and getattr(self, 'receivers', []):
      printable_fields.append(
          'receivers=[%s]' %
          ', '.join([str(receiver) for receiver in self.receivers]))

    return '<%s %s>' % (printable_name, ', '.join(printable_fields))

  def _get_runtime_performance_hints(self):
    # type: () -> Optional[Dict[Optional[str], Tuple[Optional[str], Any]]]

    """Returns any type hints required for performance runtime
    type-checking."""
    return None


class ReadOperation(Operation):
  def start(self):
    with self.scoped_start_state:
      super(ReadOperation, self).start()
      range_tracker = self.spec.source.source.get_range_tracker(
          self.spec.source.start_position, self.spec.source.stop_position)
      for value in self.spec.source.source.read(range_tracker):
        if isinstance(value, WindowedValue):
          windowed_value = value
        else:
          windowed_value = _globally_windowed_value.with_value(value)
        self.output(windowed_value)


class ImpulseReadOperation(Operation):
  def __init__(
      self,
      name_context,  # type: common.NameContext
      counter_factory,
      state_sampler,  # type: StateSampler
      consumers,  # type: Mapping[Any, List[Operation]]
      source,  # type: iobase.BoundedSource
      output_coder):
    super(ImpulseReadOperation,
          self).__init__(name_context, None, counter_factory, state_sampler)
    self.source = source

    self.receivers = [
        ConsumerSet.create(
            self.counter_factory,
            self.name_context.step_name,
            0,
            next(iter(consumers.values())),
            output_coder,
            self._get_runtime_performance_hints(),
            self.get_output_batch_converter())
    ]

  def process(self, unused_impulse):
    # type: (WindowedValue) -> None
    with self.scoped_process_state:
      range_tracker = self.source.get_range_tracker(None, None)
      for value in self.source.read(range_tracker):
        if isinstance(value, WindowedValue):
          windowed_value = value
        else:
          windowed_value = _globally_windowed_value.with_value(value)
        self.output(windowed_value)


class InMemoryWriteOperation(Operation):
  """A write operation that will write to an in-memory sink."""
  def process(self, o):
    # type: (WindowedValue) -> None
    with self.scoped_process_state:
      if self.debug_logging_enabled:
        _LOGGER.debug('Processing [%s] in %s', o, self)
      self.spec.output_buffer.append(
          o if self.spec.write_windowed_values else o.value)


class _TaggedReceivers(dict):
  def __init__(self, counter_factory, step_name):
    self._counter_factory = counter_factory
    self._step_name = step_name

  def __missing__(self, tag):
    self[tag] = receiver = ConsumerSet.create(
        self._counter_factory, self._step_name, tag, [], None, None, None)
    return receiver

  def total_output_bytes(self):
    # type: () -> int
    total = 0
    for receiver in self.values():
      elements = receiver.opcounter.element_counter.value()
      if elements > 0:
        mean = (receiver.opcounter.mean_byte_counter.value())[0]
        total += elements * mean
    return total


OpInputInfo = NamedTuple(
    'OpInputInfo',
    [
        ('transform_id', str),
        ('main_input_tag', str),
        ('main_input_coder', coders.WindowedValueCoder),
        ('outputs', Iterable[str]),
    ])


class DoOperation(Operation):
  """A Do operation that will execute a custom DoFn for each input element."""
  def __init__(
      self,
      name,  # type: common.NameContext
      spec,  # operation_specs.WorkerDoFn  # need to fix this type
      counter_factory,
      sampler,
      side_input_maps=None,
      user_state_context=None,
  ):
    super(DoOperation, self).__init__(name, spec, counter_factory, sampler)
    self.side_input_maps = side_input_maps
    self.user_state_context = user_state_context
    self.tagged_receivers = None  # type: Optional[_TaggedReceivers]
    # A mapping of timer tags to the input "PCollections" they come in on.
    self.input_info = None  # type: Optional[OpInputInfo]
    self.scoped_timer_processing_state = self.state_sampler.scoped_state(
        self.name_context,
        'process-timers',
        metrics_container=self.metrics_container)
    # See fn_data in dataflow_runner.py
    # TODO: Store all the items from spec?
    self.fn, _, _, _, _ = (pickler.loads(self.spec.serialized_fn))

  def _read_side_inputs(self, tags_and_types):
    # type: (...) -> Iterator[apache_sideinputs.SideInputMap]

    """Generator reading side inputs in the order prescribed by tags_and_types.

    Args:
      tags_and_types: List of tuples (tag, type). Each side input has a string
        tag that is specified in the worker instruction. The type is actually
        a boolean which is True for singleton input (read just first value)
        and False for collection input (read all values).

    Yields:
      With each iteration it yields the result of reading an entire side source
      either in singleton or collection mode according to the tags_and_types
      argument.
    """
    # Only call this on the old path where side_input_maps was not
    # provided directly.
    assert self.side_input_maps is None

    # We will read the side inputs in the order prescribed by the
    # tags_and_types argument because this is exactly the order needed to
    # replace the ArgumentPlaceholder objects in the args/kwargs of the DoFn
    # getting the side inputs.
    #
    # Note that for each tag there could be several read operations in the
    # specification. This can happen for instance if the source has been
    # sharded into several files.
    for i, (side_tag, view_class, view_options) in enumerate(tags_and_types):
      sources = []
      # Using the side_tag in the lambda below will trigger a pylint warning.
      # However in this case it is fine because the lambda is used right away
      # while the variable has the value assigned by the current iteration of
      # the for loop.
      # pylint: disable=cell-var-from-loop
      for si in filter(lambda o: o.tag == side_tag, self.spec.side_inputs):
        if not isinstance(si, operation_specs.WorkerSideInputSource):
          raise NotImplementedError('Unknown side input type: %r' % si)
        sources.append(si.source)
      si_counter = opcounters.SideInputReadCounter(
          self.counter_factory,
          self.state_sampler,
          declaring_step=self.name_context.step_name,
          # Inputs are 1-indexed, so we add 1 to i in the side input id
          input_index=i + 1)
      element_counter = opcounters.OperationCounters(
          self.counter_factory,
          self.name_context.step_name,
          view_options['coder'],
          i,
          suffix='side-input')
      iterator_fn = sideinputs.get_iterator_fn_for_sources(
          sources, read_counter=si_counter, element_counter=element_counter)
      yield apache_sideinputs.SideInputMap(
          view_class, view_options, sideinputs.EmulatedIterable(iterator_fn))

  def setup(self, data_sampler=None):
    # type: (Optional[DataSampler]) -> None
    with self.scoped_start_state:
      super(DoOperation, self).setup(data_sampler)

      # See fn_data in dataflow_runner.py
      fn, args, kwargs, tags_and_types, window_fn = (
          pickler.loads(self.spec.serialized_fn))

      state = common.DoFnState(self.counter_factory)
      state.step_name = self.name_context.logging_name()

      # Tag to output index map used to dispatch the output values emitted
      # by the DoFn function to the appropriate receivers. The main output is
      # either the only output or the output tagged with 'None' and is
      # associated with its corresponding index.
      self.tagged_receivers = _TaggedReceivers(
          self.counter_factory, self.name_context.logging_name())

      if len(self.spec.output_tags) == 1:
        self.tagged_receivers[None] = self.receivers[0]
        self.tagged_receivers[self.spec.output_tags[0]] = self.receivers[0]
      else:
        for index, tag in enumerate(self.spec.output_tags):
          self.tagged_receivers[tag] = self.receivers[index]
          if tag == 'None':
            self.tagged_receivers[None] = self.receivers[index]

      if self.user_state_context:
        self.timer_specs = {
            spec.name: spec
            for spec in userstate.get_dofn_specs(fn)[1]
        }  # type: Dict[str, TimerSpec]

      if self.side_input_maps is None:
        if tags_and_types:
          self.side_input_maps = list(self._read_side_inputs(tags_and_types))
        else:
          self.side_input_maps = []

      self.dofn_runner = common.DoFnRunner(
          fn,
          args,
          kwargs,
          self.side_input_maps,
          window_fn,
          tagged_receivers=self.tagged_receivers,
          step_name=self.name_context.logging_name(),
          state=state,
          user_state_context=self.user_state_context,
          transform_id=self.name_context.transform_id,
          operation_name=self.name_context.metrics_name())
      self.dofn_runner.setup()

  def start(self):
    # type: () -> None
    with self.scoped_start_state:
      super(DoOperation, self).start()
      self.dofn_runner.execution_context = self.execution_context
      self.dofn_runner.start()

  def get_batching_preference(self):
    if self.fn._process_batch_defined:
      if self.fn._process_defined:
        return common.BatchingPreference.DO_NOT_CARE
      else:
        return common.BatchingPreference.BATCH_REQUIRED
    else:
      return common.BatchingPreference.BATCH_FORBIDDEN

  def get_input_batch_converter(self) -> Optional[BatchConverter]:
    return getattr(self.fn, 'input_batch_converter', None)

  def get_output_batch_converter(self) -> Optional[BatchConverter]:
    return getattr(self.fn, 'output_batch_converter', None)

  def process(self, o):
    # type: (WindowedValue) -> None
    with self.scoped_process_state:
      delayed_applications = self.dofn_runner.process(o)
      if delayed_applications:
        assert self.execution_context is not None
        for delayed_application in delayed_applications:
          self.execution_context.delayed_applications.append(
              (self, delayed_application))

  def process_batch(self, windowed_batch: WindowedBatch) -> None:
    self.dofn_runner.process_batch(windowed_batch)

  def finalize_bundle(self):
    # type: () -> None
    self.dofn_runner.finalize()

  def needs_finalization(self):
    # type: () -> bool
    return self.dofn_runner.bundle_finalizer_param.has_callbacks()

  def add_timer_info(self, timer_family_id, timer_info):
    self.user_state_context.add_timer_info(timer_family_id, timer_info)

  def process_timer(self, tag, timer_data):
    with self.scoped_timer_processing_state:
      timer_spec = self.timer_specs[tag]
      self.dofn_runner.process_user_timer(
          timer_spec,
          timer_data.user_key,
          timer_data.windows[0],
          timer_data.fire_timestamp,
          timer_data.paneinfo,
          timer_data.dynamic_timer_tag)

  def finish(self):
    # type: () -> None
    super(DoOperation, self).finish()
    with self.scoped_finish_state:
      self.dofn_runner.finish()
      if self.user_state_context:
        self.user_state_context.commit()

  def teardown(self):
    # type: () -> None
    with self.scoped_finish_state:
      self.dofn_runner.teardown()
    if self.user_state_context:
      self.user_state_context.reset()

  def reset(self):
    # type: () -> None
    super(DoOperation, self).reset()
    for side_input_map in self.side_input_maps:
      side_input_map.reset()
    if self.user_state_context:
      self.user_state_context.reset()
    self.dofn_runner.bundle_finalizer_param.reset()

  def pcollection_count_monitoring_infos(self, tag_to_pcollection_id):
    # type: (Dict[str, str]) -> Dict[FrozenSet, metrics_pb2.MonitoringInfo]

    """Returns the element count MonitoringInfo collected by this operation."""
    infos = super(
        DoOperation,
        self).pcollection_count_monitoring_infos(tag_to_pcollection_id)

    if self.tagged_receivers:
      for tag, receiver in self.tagged_receivers.items():
        if str(tag) not in tag_to_pcollection_id:
          continue
        pcollection_id = tag_to_pcollection_id[str(tag)]
        mi = monitoring_infos.int64_counter(
            monitoring_infos.ELEMENT_COUNT_URN,
            receiver.opcounter.element_counter.value(),
            pcollection=pcollection_id)
        infos[monitoring_infos.to_key(mi)] = mi
        (unused_mean, sum, count, min, max) = (
            receiver.opcounter.mean_byte_counter.value())
        sampled_byte_count = monitoring_infos.int64_distribution(
            monitoring_infos.SAMPLED_BYTE_SIZE_URN,
            DistributionData(sum, count, min, max),
            pcollection=pcollection_id)
        infos[monitoring_infos.to_key(sampled_byte_count)] = sampled_byte_count
    return infos

  def _get_runtime_performance_hints(self):
    fns = pickler.loads(self.spec.serialized_fn)
    if fns and hasattr(fns[0], '_runtime_output_constraints'):
      return fns[0]._runtime_output_constraints

    return {}


class SdfTruncateSizedRestrictions(DoOperation):
  def __init__(self, *args, **kwargs):
    super(SdfTruncateSizedRestrictions, self).__init__(*args, **kwargs)

  def current_element_progress(self):
    # type: () -> Optional[iobase.RestrictionProgress]
    return self.receivers[0].current_element_progress()

  def try_split(
      self, fraction_of_remainder
  ):  # type: (...) -> Optional[Tuple[Iterable[SdfSplitResultsPrimary], Iterable[SdfSplitResultsResidual]]]
    return self.receivers[0].try_split(fraction_of_remainder)


class SdfProcessSizedElements(DoOperation):
  def __init__(self, *args, **kwargs):
    super(SdfProcessSizedElements, self).__init__(*args, **kwargs)
    self.lock = threading.RLock()
    self.element_start_output_bytes = None  # type: Optional[int]

  def process(self, o):
    # type: (WindowedValue) -> None
    assert self.tagged_receivers is not None
    with self.scoped_process_state:
      try:
        with self.lock:
          self.element_start_output_bytes = \
            self.tagged_receivers.total_output_bytes()
          for receiver in self.tagged_receivers.values():
            receiver.opcounter.restart_sampling()
        # Actually processing the element can be expensive; do it without
        # the lock.
        delayed_applications = self.dofn_runner.process_with_sized_restriction(
            o)
        if delayed_applications:
          assert self.execution_context is not None
          for delayed_application in delayed_applications:
            self.execution_context.delayed_applications.append(
                (self, delayed_application))
      finally:
        with self.lock:
          self.element_start_output_bytes = None

  def try_split(self, fraction_of_remainder):
    # type: (...) -> Optional[Tuple[Iterable[SdfSplitResultsPrimary], Iterable[SdfSplitResultsResidual]]]
    split = self.dofn_runner.try_split(fraction_of_remainder)
    if split:
      primaries, residuals = split
      return [(self, primary) for primary in primaries
              ], [(self, residual) for residual in residuals]
    return None

  def current_element_progress(self):
    # type: () -> Optional[iobase.RestrictionProgress]
    with self.lock:
      if self.element_start_output_bytes is not None:
        progress = self.dofn_runner.current_element_progress()
        if progress is not None:
          assert self.tagged_receivers is not None
          return progress.with_completed(
              self.tagged_receivers.total_output_bytes() -
              self.element_start_output_bytes)
      return None

  def monitoring_infos(self, transform_id, tag_to_pcollection_id):
    # type: (str, Dict[str, str]) -> Dict[FrozenSet, metrics_pb2.MonitoringInfo]

    progress_coder = coders.IterableCoder(coders.FloatCoder())

    with self.lock:
      infos = super(SdfProcessSizedElements,
                    self).monitoring_infos(transform_id, tag_to_pcollection_id)
      current_element_progress = self.current_element_progress()
      if current_element_progress:
        if current_element_progress.completed_work:
          completed = current_element_progress.completed_work
          remaining = current_element_progress.remaining_work
        else:
          completed = current_element_progress.fraction_completed
          remaining = current_element_progress.fraction_remaining
        assert completed is not None
        assert remaining is not None
        completed_mi = metrics_pb2.MonitoringInfo(
            urn=monitoring_infos.WORK_COMPLETED_URN,
            type=monitoring_infos.PROGRESS_TYPE,
            labels=monitoring_infos.create_labels(ptransform=transform_id),
            payload=progress_coder.encode([completed]))
        remaining_mi = metrics_pb2.MonitoringInfo(
            urn=monitoring_infos.WORK_REMAINING_URN,
            type=monitoring_infos.PROGRESS_TYPE,
            labels=monitoring_infos.create_labels(ptransform=transform_id),
            payload=progress_coder.encode([remaining]))
        infos[monitoring_infos.to_key(completed_mi)] = completed_mi
        infos[monitoring_infos.to_key(remaining_mi)] = remaining_mi
    return infos


class CombineOperation(Operation):
  """A Combine operation executing a CombineFn for each input element."""
  def __init__(self, name_context, spec, counter_factory, state_sampler):
    super(CombineOperation,
          self).__init__(name_context, spec, counter_factory, state_sampler)
    # Combiners do not accept deferred side-inputs (the ignored fourth argument)
    # and therefore the code to handle the extra args/kwargs is simpler than for
    # the DoFn's of ParDo.
    fn, args, kwargs = pickler.loads(self.spec.serialized_fn)[:3]
    self.phased_combine_fn = (
        PhasedCombineFnExecutor(self.spec.phase, fn, args, kwargs))

  def setup(self, data_sampler=None):
    # type: (Optional[DataSampler]) -> None
    with self.scoped_start_state:
      _LOGGER.debug('Setup called for %s', self)
      super(CombineOperation, self).setup(data_sampler)
      self.phased_combine_fn.combine_fn.setup()

  def process(self, o):
    # type: (WindowedValue) -> None
    with self.scoped_process_state:
      if self.debug_logging_enabled:
        _LOGGER.debug('Processing [%s] in %s', o, self)
      key, values = o.value
      self.output(o.with_value((key, self.phased_combine_fn.apply(values))))

  def finish(self):
    # type: () -> None
    _LOGGER.debug('Finishing %s', self)
    super(CombineOperation, self).finish()

  def teardown(self):
    # type: () -> None
    with self.scoped_finish_state:
      _LOGGER.debug('Teardown called for %s', self)
      super(CombineOperation, self).teardown()
      self.phased_combine_fn.combine_fn.teardown()


def create_pgbk_op(step_name, spec, counter_factory, state_sampler):
  if spec.combine_fn:
    return PGBKCVOperation(step_name, spec, counter_factory, state_sampler)
  else:
    return PGBKOperation(step_name, spec, counter_factory, state_sampler)


class PGBKOperation(Operation):
  """Partial group-by-key operation.

  This takes (windowed) input (key, value) tuples and outputs
  (key, [value]) tuples, performing a best effort group-by-key for
  values in this bundle, memory permitting.
  """
  def __init__(self, name_context, spec, counter_factory, state_sampler):
    super(PGBKOperation,
          self).__init__(name_context, spec, counter_factory, state_sampler)
    assert not self.spec.combine_fn
    self.table = collections.defaultdict(list)
    self.size = 0
    # TODO(robertwb) Make this configurable.
    self.max_size = 10 * 1000

  def process(self, o):
    # type: (WindowedValue) -> None
    with self.scoped_process_state:
      # TODO(robertwb): Structural (hashable) values.
      key = o.value[0], tuple(o.windows)
      self.table[key].append(o)
      self.size += 1
      if self.size > self.max_size:
        self.flush(9 * self.max_size // 10)

  def finish(self):
    # type: () -> None
    self.flush(0)
    super().finish()

  def flush(self, target):
    # type: (int) -> None
    limit = self.size - target
    for ix, (kw, vs) in enumerate(list(self.table.items())):
      if ix >= limit:
        break
      del self.table[kw]
      key, windows = kw
      output_value = [v.value[1] for v in vs]
      windowed_value = WindowedValue((key, output_value),
                                     vs[0].timestamp,
                                     windows)
      self.output(windowed_value)


class PGBKCVOperation(Operation):
  """Partial group-by-key operation.

  This operation handles grouped values with
a combine function applied.
  """
  def __init__(
      self, name_context, spec, counter_factory, state_sampler, windowing=None):
    super(PGBKCVOperation,
          self).__init__(name_context, spec, counter_factory, state_sampler)
    # Combiners do not accept deferred side-inputs (the ignored fourth
    # argument) and therefore the code to handle the extra args/kwargs is
    # simpler than for the DoFn's of ParDo.
    fn, args, kwargs = pickler.loads(self.spec.combine_fn)[:3]
    self.combine_fn = curry_combine_fn(fn, args, kwargs)
    self.combine_fn_add_input = self.combine_fn.add_input
    if self.combine_fn.compact.__func__ is core.CombineFn.compact:
      self.combine_fn_compact = None
    else:
      self.combine_fn_compact = self.combine_fn.compact
    if windowing:
      self.is_default_windowing = windowing.is_default()
      tsc_type = windowing.timestamp_combiner
      self.timestamp_combiner = (
          None if tsc_type == window.TimestampCombiner.OUTPUT_AT_EOW else
          window.TimestampCombiner.get_impl(tsc_type, windowing.windowfn))
    else:
      self.is_default_windowing = False  # unknown
      self.timestamp_combiner = None
    # Optimization for the (known tiny accumulator, often wide keyspace)
    # combine functions.
    # TODO(b/36567833): Bound by in-memory size rather than key count.
    self.max_keys = (
        1000 * 1000 if
        isinstance(fn, (combiners.CountCombineFn, combiners.MeanCombineFn)) or
        # TODO(b/36597732): Replace this 'or' part by adding the 'cy' optimized
        # combiners to the short list above.
        (
            isinstance(fn, core.CallableWrapperCombineFn) and
            fn._fn in (min, max, sum)) else 100 * 1000)  # pylint: disable=protected-access
    self.key_count = 0
    self.table = {}

  def setup(self, data_sampler=None):
    # type: (Optional[DataSampler]) -> None
    with self.scoped_start_state:
      _LOGGER.debug('Setup called for %s', self)
      super(PGBKCVOperation, self).setup(data_sampler)
      self.combine_fn.setup()

  def process(self, wkv):
    # type: (WindowedValue) -> None
    with self.scoped_process_state:
      key, value = wkv.value
      # pylint: disable=unidiomatic-typecheck
      # Optimization for the global window case.
      if self.is_default_windowing:
        self.add_key_value(key, value, None)
      else:
        for window in wkv.windows:
          self.add_key_value((window, key),
                             value,
                             wkv.timestamp if self.timestamp_combiner else None)

  def add_key_value(self, wkey, value, timestamp):
    entry = self.table.get(wkey, None)
    if entry is None:
      if self.key_count >= self.max_keys:
        target = self.key_count * 9 // 10
        old_wkeys = []
        # TODO(robertwb): Use an LRU cache?
        for old_wkey, old_wvalue in self.table.items():
          old_wkeys.append(old_wkey)  # Can't mutate while iterating.
          self.output_key(old_wkey, old_wvalue[0], old_wvalue[1])
          self.key_count -= 1
          if self.key_count <= target:
            break
        for old_wkey in reversed(old_wkeys):
          del self.table[old_wkey]
      self.key_count += 1
      # We save the accumulator as a one element list so we can efficiently
      # mutate when new values are added without searching the cache again.
      entry = self.table[wkey] = [
          self.combine_fn.create_accumulator(), timestamp
      ]
    entry[0] = self.combine_fn_add_input(entry[0], value)
    if not self.is_default_windowing and self.timestamp_combiner:
      entry[1] = self.timestamp_combiner.combine(entry[1], timestamp)

  def finish(self):
    # type: () -> None
    for wkey, value in self.table.items():
      self.output_key(wkey, value[0], value[1])
    self.table = {}
    self.key_count = 0

  def teardown(self):
    # type: () -> None
    with self.scoped_finish_state:
      _LOGGER.debug('Teardown called for %s', self)
      super(PGBKCVOperation, self).teardown()
      self.combine_fn.teardown()

  def output_key(self, wkey, accumulator, timestamp):
    if self.combine_fn_compact is None:
      value = accumulator
    else:
      value = self.combine_fn_compact(accumulator)

    if self.is_default_windowing:
      self.output(_globally_windowed_value.with_value((wkey, value)))
    else:
      window, key = wkey
      if self.timestamp_combiner is None:
        timestamp = window.max_timestamp()
      self.output(WindowedValue((key, value), timestamp, (window, )))


class FlattenOperation(Operation):
  """Flatten operation.

  Receives one or more producer operations, outputs just one list
  with all the items.
  """
  def process(self, o):
    # type: (WindowedValue) -> None
    with self.scoped_process_state:
      if self.debug_logging_enabled:
        _LOGGER.debug('Processing [%s] in %s', o, self)
      self.output(o)


def create_operation(
    name_context,
    spec,
    counter_factory,
    step_name=None,
    state_sampler=None,
    test_shuffle_source=None,
    test_shuffle_sink=None,
    is_streaming=False):
  # type: (...) -> Operation

  """Create Operation object for given operation specification."""

  # TODO(pabloem): Document arguments to this function call.
  if not isinstance(name_context, common.NameContext):
    name_context = common.NameContext(step_name=name_context)

  if isinstance(spec, operation_specs.WorkerRead):
    if isinstance(spec.source, iobase.SourceBundle):
      op = ReadOperation(
          name_context, spec, counter_factory, state_sampler)  # type: Operation
    else:
      from dataflow_worker.native_operations import NativeReadOperation
      op = NativeReadOperation(
          name_context, spec, counter_factory, state_sampler)
  elif isinstance(spec, operation_specs.WorkerWrite):
    from dataflow_worker.native_operations import NativeWriteOperation
    op = NativeWriteOperation(
        name_context, spec, counter_factory, state_sampler)
  elif isinstance(spec, operation_specs.WorkerCombineFn):
    op = CombineOperation(name_context, spec, counter_factory, state_sampler)
  elif isinstance(spec, operation_specs.WorkerPartialGroupByKey):
    op = create_pgbk_op(name_context, spec, counter_factory, state_sampler)
  elif isinstance(spec, operation_specs.WorkerDoFn):
    op = DoOperation(name_context, spec, counter_factory, state_sampler)
  elif isinstance(spec, operation_specs.WorkerGroupingShuffleRead):
    from dataflow_worker.shuffle_operations import GroupedShuffleReadOperation
    op = GroupedShuffleReadOperation(
        name_context,
        spec,
        counter_factory,
        state_sampler,
        shuffle_source=test_shuffle_source)
  elif isinstance(spec, operation_specs.WorkerUngroupedShuffleRead):
    from dataflow_worker.shuffle_operations import UngroupedShuffleReadOperation
    op = UngroupedShuffleReadOperation(
        name_context,
        spec,
        counter_factory,
        state_sampler,
        shuffle_source=test_shuffle_source)
  elif isinstance(spec, operation_specs.WorkerInMemoryWrite):
    op = InMemoryWriteOperation(
        name_context, spec, counter_factory, state_sampler)
  elif isinstance(spec, operation_specs.WorkerShuffleWrite):
    from dataflow_worker.shuffle_operations import ShuffleWriteOperation
    op = ShuffleWriteOperation(
        name_context,
        spec,
        counter_factory,
        state_sampler,
        shuffle_sink=test_shuffle_sink)
  elif isinstance(spec, operation_specs.WorkerFlatten):
    op = FlattenOperation(name_context, spec, counter_factory, state_sampler)
  elif isinstance(spec, operation_specs.WorkerMergeWindows):
    from dataflow_worker.shuffle_operations import BatchGroupAlsoByWindowsOperation
    from dataflow_worker.shuffle_operations import StreamingGroupAlsoByWindowsOperation
    if is_streaming:
      op = StreamingGroupAlsoByWindowsOperation(
          name_context, spec, counter_factory, state_sampler)
    else:
      op = BatchGroupAlsoByWindowsOperation(
          name_context, spec, counter_factory, state_sampler)
  elif isinstance(spec, operation_specs.WorkerReifyTimestampAndWindows):
    from dataflow_worker.shuffle_operations import ReifyTimestampAndWindowsOperation
    op = ReifyTimestampAndWindowsOperation(
        name_context, spec, counter_factory, state_sampler)
  else:
    raise TypeError(
        'Expected an instance of operation_specs.Worker* class '
        'instead of %s' % (spec, ))
  return op


class SimpleMapTaskExecutor(object):
  """An executor for map tasks.

   Stores progress of the read operation that is the first operation of a map
   task.
  """
  def __init__(
      self,
      map_task,
      counter_factory,
      state_sampler,
      test_shuffle_source=None,
      test_shuffle_sink=None):
    """Initializes SimpleMapTaskExecutor.

    Args:
      map_task: The map task we are to run. The maptask contains a list of
        operations, and aligned lists for step_names, original_names,
        system_names of pipeline steps.
      counter_factory: The CounterFactory instance for the work item.
      state_sampler: The StateSampler tracking the execution step.
      test_shuffle_source: Used during tests for dependency injection into
        shuffle read operation objects.
      test_shuffle_sink: Used during tests for dependency injection into
        shuffle write operation objects.
    """

    self._map_task = map_task
    self._counter_factory = counter_factory
    self._ops = []  # type: List[Operation]
    self._state_sampler = state_sampler
    self._test_shuffle_source = test_shuffle_source
    self._test_shuffle_sink = test_shuffle_sink

  def operations(self):
    # type: () -> List[Operation]
    return self._ops[:]

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

    """Executes all the operation_specs.Worker* instructions in a map task.

    We update the map_task with the execution status, expressed as counters.

    Raises:
      RuntimeError: if we find more than on read instruction in task spec.
      TypeError: if the spec parameter is not an instance of the recognized
        operation_specs.Worker* classes.
    """

    # operations is a list of operation_specs.Worker* instances.
    # The order of the elements is important because the inputs use
    # list indexes as references.
    for name_context, spec in zip(self._map_task.name_contexts,
                                  self._map_task.operations):
      # This is used for logging and assigning names to counters.
      op = create_operation(
          name_context,
          spec,
          self._counter_factory,
          None,
          self._state_sampler,
          test_shuffle_source=self._test_shuffle_source,
          test_shuffle_sink=self._test_shuffle_sink)
      self._ops.append(op)

      # Add receiver operations to the appropriate producers.
      if hasattr(op.spec, 'input'):
        producer, output_index = op.spec.input
        self._ops[producer].add_receiver(op, output_index)
      # Flatten has 'inputs', not 'input'
      if hasattr(op.spec, 'inputs'):
        for producer, output_index in op.spec.inputs:
          self._ops[producer].add_receiver(op, output_index)

    for ix, op in reversed(list(enumerate(self._ops))):
      _LOGGER.debug('Starting op %d %s', ix, op)
      op.start()
    for op in self._ops:
      op.finish()


class InefficientExecutionWarning(RuntimeWarning):
  """warning to indicate an inefficiency in a Beam pipeline."""


# Don't ignore InefficientExecutionWarning, but only log them once
warnings.simplefilter('once', InefficientExecutionWarning)