pipeline.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.
#

"""Pipeline, the top-level Beam object.

A pipeline holds a DAG of data transforms. Conceptually the nodes of the DAG
are transforms (:class:`~apache_beam.transforms.ptransform.PTransform` objects)
and the edges are values (mostly :class:`~apache_beam.pvalue.PCollection`
objects). The transforms take as inputs one or more PValues and output one or
more :class:`~apache_beam.pvalue.PValue` s.

The pipeline offers functionality to traverse the graph.  The actual operation
to be executed for each node visited is specified through a runner object.

Typical usage::

  # Create a pipeline object using a local runner for execution.
  with beam.Pipeline('DirectRunner') as p:

    # Add to the pipeline a "Create" transform. When executed this
    # transform will produce a PCollection object with the specified values.
    pcoll = p | 'Create' >> beam.Create([1, 2, 3])

    # Another transform could be applied to pcoll, e.g., writing to a text file.
    # For other transforms, refer to transforms/ directory.
    pcoll | 'Write' >> beam.io.WriteToText('./output')

    # run() will execute the DAG stored in the pipeline.  The execution of the
    # nodes visited is done using the specified local runner.

"""

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

import abc
import contextlib
import copy
import logging
import os
import re
import shutil
import tempfile
import threading
import unicodedata
import uuid
from collections import defaultdict
from collections.abc import Iterable
from collections.abc import Mapping
from collections.abc import Sequence
from typing import TYPE_CHECKING
from typing import Any
from typing import Optional
from typing import Type
from typing import Union

from google.protobuf import message

from apache_beam import pvalue
from apache_beam.coders import typecoders
from apache_beam.internal import pickler
from apache_beam.io.filesystems import FileSystems
from apache_beam.options.pipeline_options import CrossLanguageOptions
from apache_beam.options.pipeline_options import DebugOptions
from apache_beam.options.pipeline_options import PipelineOptions
from apache_beam.options.pipeline_options import SetupOptions
from apache_beam.options.pipeline_options import StandardOptions
from apache_beam.options.pipeline_options import StreamingOptions
from apache_beam.options.pipeline_options import TypeOptions
from apache_beam.options.pipeline_options_context import scoped_pipeline_options
from apache_beam.options.pipeline_options_validator import PipelineOptionsValidator
from apache_beam.portability import common_urns
from apache_beam.portability.api import beam_runner_api_pb2
from apache_beam.runners import PipelineRunner
from apache_beam.runners import create_runner
from apache_beam.runners import pipeline_utils
from apache_beam.transforms import ParDo
from apache_beam.transforms import ptransform
from apache_beam.transforms.display import DisplayData
from apache_beam.transforms.display import HasDisplayData
from apache_beam.transforms.resources import merge_resource_hints
from apache_beam.transforms.resources import resource_hints_from_options
from apache_beam.transforms.sideinputs import get_sideinput_index
from apache_beam.typehints import TypeCheckError
from apache_beam.typehints import typehints
from apache_beam.utils import proto_utils
from apache_beam.utils import subprocess_server
from apache_beam.utils.annotations import deprecated
from apache_beam.utils.interactive_utils import alter_label_if_ipython
from apache_beam.utils.interactive_utils import is_in_ipython

if TYPE_CHECKING:
  from types import TracebackType

  from apache_beam.runners.pipeline_context import PipelineContext
  from apache_beam.runners.runner import PipelineResult
  from apache_beam.transforms import environments

__all__ = ['Pipeline', 'transform_annotations']


class Pipeline(HasDisplayData):
  """A pipeline object that manages a DAG of
  :class:`~apache_beam.transforms.ptransform.PTransform` s
  and their :class:`~apache_beam.pvalue.PValue` s.

  Conceptually the :class:`~apache_beam.transforms.ptransform.PTransform` s are
  the DAG's nodes and the :class:`~apache_beam.pvalue.PValue` s are the edges.

  All the transforms applied to the pipeline must have distinct full labels.
  If same transform instance needs to be applied then the right shift operator
  should be used to designate new names
  (e.g. ``input | "label" >> my_transform``).
  """
  @classmethod
  def runner_implemented_transforms(cls) -> frozenset[str]:

    # This set should only contain transforms which are required to be
    # implemented by a runner.
    return frozenset([
        common_urns.primitives.GROUP_BY_KEY.urn,
        common_urns.primitives.IMPULSE.urn,
    ])

  def __init__(
      self,
      runner: Optional[Union[str, PipelineRunner]] = None,
      options: Optional[PipelineOptions] = None,
      argv: Optional[list[str]] = None,
      display_data: Optional[dict[str, Any]] = None):
    """Initialize a pipeline object.

    Args:
      runner (~apache_beam.runners.runner.PipelineRunner): An object of
        type :class:`~apache_beam.runners.runner.PipelineRunner` that will be
        used to execute the pipeline. For registered runners, the runner name
        can be specified, otherwise a runner object must be supplied.
      options (~apache_beam.options.pipeline_options.PipelineOptions):
        A configured
        :class:`~apache_beam.options.pipeline_options.PipelineOptions` object
        containing arguments that should be used for running the Beam job.
      argv (list[str]): a list of arguments (such as :data:`sys.argv`)
        to be used for building a
        :class:`~apache_beam.options.pipeline_options.PipelineOptions` object.
        This will only be used if argument **options** is :data:`None`.
      display_data (dict[str, Any]): a dictionary of static data associated
        with this pipeline that can be displayed when it runs.

    Raises:
      ValueError: if either the runner or options argument is not
        of the expected type.
    """
    # Initializing logging configuration in case the user did not set it up.
    logging.basicConfig()

    if options is not None:
      if isinstance(options, PipelineOptions):
        # Make a deep copy of options since they could be overwritten in later
        # steps. However, the 'runner' object within 'options' is excluded from
        # the deep copy (it is shallow copied) due to potential issues with deep
        # copying specific runner instances, such as FlumeRunner.
        saved_runner = options.view_as(StandardOptions).runner
        options.view_as(StandardOptions).runner = None
        self._options = copy.deepcopy(options)
        self._options.view_as(StandardOptions).runner = saved_runner
        options.view_as(StandardOptions).runner = saved_runner
      else:
        raise ValueError(
            'Parameter options, if specified, must be of type PipelineOptions. '
            'Received : %r' % options)
    elif argv is not None:
      if isinstance(argv, list):
        self._options = PipelineOptions(argv)
      else:
        raise ValueError(
            'Parameter argv, if specified, must be a list. Received : %r' %
            argv)
    else:
      self._options = PipelineOptions([])

    FileSystems.set_options(self._options)

    if runner is None:
      runner = self._options.view_as(StandardOptions).runner
      if runner is None:
        runner = StandardOptions.DEFAULT_RUNNER
        logging.info((
            'Missing pipeline option (runner). Executing pipeline '
            'using the default runner: %s.'),
                     runner)

    if isinstance(runner, str):
      runner = create_runner(runner)
    elif not isinstance(runner, PipelineRunner):
      raise TypeError(
          'Runner %s is not a PipelineRunner object or the '
          'name of a registered runner.' % runner)

    # Runner can override the default pickler to be used.
    if (self._options.view_as(SetupOptions).pickle_library == 'default' and
        runner.default_pickle_library_override()):
      logging.info(
          "Runner defaulting to pickling library: %s.",
          runner.default_pickle_library_override())
      self._options.view_as(
          SetupOptions).pickle_library = runner.default_pickle_library_override(
          )
    pickler.set_library(self._options.view_as(SetupOptions).pickle_library)

    # Validate pipeline options
    errors = PipelineOptionsValidator(self._options, runner).validate()
    if errors:
      raise ValueError(
          'Pipeline has validations errors: \n' + '\n'.join(errors))

    typecoders.registry.update_compatibility_version = self._options.view_as(
        StreamingOptions).update_compatibility_version

    # set default experiments for portable runners
    # (needs to occur prior to pipeline construction)
    if runner.is_fnapi_compatible():
      self._options.view_as(DebugOptions).add_experiment('beam_fn_api')

    self.local_tempdir = tempfile.mkdtemp(prefix='beam-pipeline-temp')

    # Default runner to be used.
    self.runner = runner

    # Stack of transforms generated by nested apply() calls. The stack will
    # contain a root node as an enclosing (parent) node for top transforms.
    self.transforms_stack = [
        AppliedPTransform(None, None, '', None, None, None)
    ]
    # Set of transform labels (full labels) applied to the pipeline.
    # If a transform is applied and the full label is already in the set
    # then the transform will have to be cloned with a new label.
    self.applied_labels: set[str] = set()
    # Hints supplied via pipeline options are considered the outermost hints.
    self._root_transform().resource_hints = resource_hints_from_options(options)
    # Create a ComponentIdMap for assigning IDs to components. Ensures that any
    # components that receive an ID during pipeline construction (for example in
    # ExternalTransform), will receive the same component ID when generating the
    # full pipeline proto.
    self.component_id_map = ComponentIdMap()

    # Records whether this pipeline contains any external transforms.
    self.contains_external_transforms = False

    self._display_data = display_data or {}
    self._error_handlers = []
    self._annotations_stack = [{}]

  def display_data(self) -> dict[str, Any]:
    return self._display_data

  @property  # type: ignore[misc]  # decorated property not supported
  def options(self) -> PipelineOptions:
    return self._options

  @property
  def allow_unsafe_triggers(self) -> bool:
    return self._options.view_as(TypeOptions).allow_unsafe_triggers

  def _register_error_handler(self, error_handler):
    self._error_handlers.append(error_handler)

  def _current_transform(self) -> 'AppliedPTransform':
    """Returns the transform currently on the top of the stack."""
    return self.transforms_stack[-1]

  @contextlib.contextmanager
  def transform_annotations(self, **annotations):
    """A context manager for attaching annotations to a set of transforms.

    All transforms applied while this context is active will have these
    annotations attached. This includes  sub-transforms applied within
    composite transforms.
    """
    self._annotations_stack.append({
        **self._annotations_stack[-1], **encode_annotations(annotations)
    })
    yield
    self._annotations_stack.pop()

  def _current_annotations(self):
    """Returns the set of annotations that should be used on apply."""
    return {**_global_annotations_stack()[-1], **self._annotations_stack[-1]}

  def _root_transform(self) -> 'AppliedPTransform':
    """Returns the root transform of the transform stack."""
    return self.transforms_stack[0]

  def _remove_labels_recursively(
      self, applied_transform: 'AppliedPTransform') -> None:
    for part in applied_transform.parts:
      if part.full_label in self.applied_labels:
        self.applied_labels.remove(part.full_label)
        self._remove_labels_recursively(part)

  def _replace(self, override: 'PTransformOverride') -> None:
    assert isinstance(override, PTransformOverride)

    # From original transform output --> replacement transform output
    output_map: dict[pvalue.PValue, pvalue.PValue] = {}
    output_replacements: dict[AppliedPTransform,
                              list[tuple[pvalue.PValue, Optional[str]]]] = {}
    input_replacements: dict[AppliedPTransform,
                             Mapping[str,
                                     Union[pvalue.PBegin,
                                           pvalue.PCollection]]] = {}
    side_input_replacements: dict[AppliedPTransform,
                                  list[pvalue.AsSideInput]] = {}

    class TransformUpdater(PipelineVisitor):  # pylint: disable=used-before-assignment
      """"A visitor that replaces the matching PTransforms."""
      def __init__(self, pipeline: Pipeline) -> None:
        self.pipeline = pipeline

      def _replace_if_needed(
          self, original_transform_node: AppliedPTransform) -> None:
        if override.matches(original_transform_node):
          assert isinstance(original_transform_node, AppliedPTransform)
          replacement_transform = (
              override.get_replacement_transform_for_applied_ptransform(
                  original_transform_node))
          if replacement_transform is original_transform_node.transform:
            return
          replacement_transform.side_inputs = tuple(
              getattr(original_transform_node.transform, 'side_inputs', ()))

          replacement_transform_node = AppliedPTransform(
              original_transform_node.parent,
              replacement_transform,
              original_transform_node.full_label,
              original_transform_node.main_inputs,
              None,
              annotations=original_transform_node.annotations)

          # TODO(https://github.com/apache/beam/issues/21178): Merge rather
          # than override.
          replacement_transform_node.resource_hints = (
              original_transform_node.resource_hints)

          # Transform execution could depend on order in which nodes are
          # considered. Hence we insert the replacement transform node to same
          # index as the original transform node. Note that this operation
          # removes the original transform node.
          if original_transform_node.parent:
            assert isinstance(original_transform_node.parent, AppliedPTransform)
            parent_parts = original_transform_node.parent.parts
            parent_parts[parent_parts.index(original_transform_node)] = (
                replacement_transform_node)
          else:
            # Original transform has to be a root.
            roots = self.pipeline.transforms_stack[0].parts
            assert original_transform_node in roots
            roots[roots.index(original_transform_node)] = (
                replacement_transform_node)

          inputs = override.get_replacement_inputs(original_transform_node)
          if len(inputs) > 1:
            transform_input = inputs
          elif len(inputs) == 1:
            transform_input = inputs[0]
          elif len(inputs) == 0:
            transform_input = pvalue.PBegin(self.pipeline)
          try:
            # We have to add the new AppliedTransform to the stack before
            # expand() and pop it out later to make sure that parts get added
            # correctly.
            self.pipeline.transforms_stack.append(replacement_transform_node)

            # Keeping the same label for the replaced node but recursively
            # removing labels of child transforms of original transform since
            # they will be replaced during the expand below. This is needed in
            # case the replacement contains children that have labels that
            # conflicts with labels of the children of the original.
            self.pipeline._remove_labels_recursively(original_transform_node)

            new_output = replacement_transform.expand(transform_input)
            assert isinstance(
                new_output, (dict, pvalue.PValue, pvalue.DoOutputsTuple))

            if isinstance(new_output, pvalue.PValue):
              new_output.element_type = None
              self.pipeline._infer_result_type(
                  replacement_transform, inputs, new_output)

            if isinstance(new_output, dict):
              for new_tag, new_pcoll in new_output.items():
                replacement_transform_node.add_output(new_pcoll, new_tag)
            elif isinstance(new_output, pvalue.DoOutputsTuple):
              replacement_transform_node.add_output(
                  new_output, new_output._main_tag)
            else:
              replacement_transform_node.add_output(new_output, new_output.tag)

            # Recording updated outputs. This cannot be done in the same
            # visitor since if we dynamically update output type here, we'll
            # run into errors when visiting child nodes.
            #
            # NOTE: When replacing multiple outputs, the replacement
            # PCollection tags must have a matching tag in the original
            # transform.
            if isinstance(new_output, pvalue.PValue):
              if not new_output.producer:
                new_output.producer = replacement_transform_node
              output_map[original_transform_node.outputs[new_output.tag]] = \
                  new_output
            elif isinstance(new_output, (pvalue.DoOutputsTuple, tuple)):
              for pcoll in new_output:
                if not pcoll.producer:
                  pcoll.producer = replacement_transform_node
                output_map[original_transform_node.outputs[pcoll.tag]] = pcoll
            elif isinstance(new_output, dict):
              for tag, pcoll in new_output.items():
                if not pcoll.producer:
                  pcoll.producer = replacement_transform_node
                output_map[original_transform_node.outputs[tag]] = pcoll
          finally:
            self.pipeline.transforms_stack.pop()

      def enter_composite_transform(
          self, transform_node: AppliedPTransform) -> None:
        self._replace_if_needed(transform_node)

      def visit_transform(self, transform_node: AppliedPTransform) -> None:
        self._replace_if_needed(transform_node)

    self.visit(TransformUpdater(self))

    # Ensure no type information is lost.
    for old, new in output_map.items():
      if new.element_type == typehints.Any:
        # TODO(robertwb): Perhaps take the intersection?
        new.element_type = old.element_type

    # Adjusting inputs and outputs
    class InputOutputUpdater(PipelineVisitor):  # pylint: disable=used-before-assignment
      """"A visitor that records input and output values to be replaced.

      Input and output values that should be updated are recorded in maps
      input_replacements and output_replacements respectively.

      We cannot update input and output values while visiting since that results
      in validation errors.
      """
      def __init__(self, pipeline: Pipeline) -> None:
        self.pipeline = pipeline

      def enter_composite_transform(
          self, transform_node: AppliedPTransform) -> None:
        self.visit_transform(transform_node)

      def visit_transform(self, transform_node: AppliedPTransform) -> None:
        replace_output = False
        for tag in transform_node.outputs:
          if transform_node.outputs[tag] in output_map:
            replace_output = True
            break

        replace_input = False
        for input in transform_node.inputs:
          if input in output_map:
            replace_input = True
            break

        replace_side_inputs = False
        for side_input in transform_node.side_inputs:
          if side_input.pvalue in output_map:
            replace_side_inputs = True
            break

        if replace_output:
          output_replacements[transform_node] = []
          for original, replacement in output_map.items():
            for tag, output in transform_node.outputs.items():
              if output == original:
                output_replacements[transform_node].append((tag, replacement))

        if replace_input:
          new_inputs = {
              tag: input if not input in output_map else output_map[input]
              for (tag, input) in transform_node.main_inputs.items()
          }
          input_replacements[transform_node] = new_inputs

        if replace_side_inputs:
          new_side_inputs = []
          for side_input in transform_node.side_inputs:
            if side_input.pvalue in output_map:
              side_input.pvalue = output_map[side_input.pvalue]
              new_side_inputs.append(side_input)
            else:
              new_side_inputs.append(side_input)
          side_input_replacements[transform_node] = new_side_inputs

    self.visit(InputOutputUpdater(self))

    for transform, output_replacement in output_replacements.items():
      for tag, output in output_replacement:
        transform.replace_output(output, tag=tag)

    for transform, input_replacement in input_replacements.items():
      transform.replace_inputs(input_replacement)

    for transform, side_input_replacement in side_input_replacements.items():
      transform.replace_side_inputs(side_input_replacement)

  def _check_replacement(self, override: 'PTransformOverride') -> None:
    class ReplacementValidator(PipelineVisitor):
      def visit_transform(self, transform_node: AppliedPTransform) -> None:
        if override.matches(transform_node):
          raise RuntimeError(
              'Transform node %r was not replaced as expected.' %
              transform_node)

    self.visit(ReplacementValidator())

  def replace_all(self, replacements: Iterable['PTransformOverride']) -> None:
    """ Dynamically replaces PTransforms in the currently populated hierarchy.

    Currently this only works for replacements where input and output types
    are exactly the same.

    TODO: Update this to also work for transform overrides where input and
    output types are different.

    Args:
      replacements (list[~apache_beam.pipeline.PTransformOverride]): a list of
        :class:`~apache_beam.pipeline.PTransformOverride` objects.
    """
    for override in replacements:
      assert isinstance(override, PTransformOverride)
      self._replace(override)

    # Checking if the PTransforms have been successfully replaced. This will
    # result in a failure if a PTransform that was replaced in a given override
    # gets re-added in a subsequent override. This is not allowed and ordering
    # of PTransformOverride objects in 'replacements' is important.
    for override in replacements:
      self._check_replacement(override)

  def run(self, test_runner_api: Union[bool, str] = 'AUTO') -> 'PipelineResult':
    """Runs the pipeline. Returns whatever our runner returns after running."""
    with scoped_pipeline_options(self._options):
      return self._run_internal(test_runner_api)

  def _run_internal(
      self, test_runner_api: Union[bool, str] = 'AUTO') -> 'PipelineResult':
    """Internal implementation of run(), called within scoped options."""
    # All pipeline options are finalized at this point.
    # Call get_all_options to print warnings on invalid options.
    self.options.get_all_options(
        retain_unknown_options=True, display_warnings=True)

    for error_handler in self._error_handlers:
      error_handler.verify_closed()

    # Records whether this pipeline contains any cross-language transforms.
    self.contains_external_transforms = (
        ExternalTransformFinder.contains_external_transforms(self))

    try:
      if test_runner_api == 'AUTO':
        # Don't pay the cost of a round-trip if we're going to be going through
        # the FnApi anyway...
        is_fnapi_compatible = self.runner.is_fnapi_compatible() or (
            # DirectRunner uses the Fn API for batch only
            self.runner.__class__.__name__ == 'SwitchingDirectRunner' and
            not self._options.view_as(StandardOptions).streaming)

        # Multi-language pipelines that contain external pipeline segments may
        # not be able to create a Python pipeline object graph. Hence following
        # runner API check should be skipped for such pipelines.

        # The InteractiveRunner relies on a constant pipeline reference, skip
        # it.
        test_runner_api = (
            not is_fnapi_compatible and
            not self.contains_external_transforms and
            self.runner.__class__.__name__ != 'InteractiveRunner')

      # When possible, invoke a round trip through the runner API.
      if test_runner_api and self._verify_runner_api_compatible():
        return Pipeline.from_runner_api(
            self.to_runner_api(use_fake_coders=True),
            self.runner,
            self._options).run(False)

      if (self._options.view_as(TypeOptions).runtime_type_check and
          self._options.view_as(TypeOptions).performance_runtime_type_check):
        raise RuntimeError(
            'You cannot turn on runtime_type_check '
            'and performance_runtime_type_check simultaneously. '
            'Pick one or the other.')

      if self._options.view_as(TypeOptions).runtime_type_check:
        from apache_beam.typehints import typecheck
        self.visit(typecheck.TypeCheckVisitor())

      if self._options.view_as(TypeOptions).performance_runtime_type_check:
        from apache_beam.typehints import typecheck
        self.visit(typecheck.PerformanceTypeCheckVisitor())

      if self._options.view_as(SetupOptions).save_main_session:
        # If this option is chosen, verify we can pickle the main session early.
        tmpdir = tempfile.mkdtemp()
        try:
          pickler.dump_session(os.path.join(tmpdir, 'main_session.pickle'))
        finally:
          shutil.rmtree(tmpdir)
      return self.runner.run_pipeline(self, self._options)
    finally:
      if not is_in_ipython():
        shutil.rmtree(self.local_tempdir, ignore_errors=True)
      # else interactive beam handles the cleanup.

  def __enter__(self) -> 'Pipeline':
    self._extra_context = contextlib.ExitStack()
    self._extra_context.enter_context(
        subprocess_server.JavaJarServer.beam_services(
            self._options.view_as(CrossLanguageOptions).beam_services))
    self._extra_context.enter_context(
        subprocess_server.SubprocessServer.cache_subprocesses())
    return self

  def __exit__(
      self,
      exc_type: Optional[Type[BaseException]],
      exc_val: Optional[BaseException],
      exc_tb: Optional['TracebackType']) -> None:

    try:
      if not exc_type:
        self.result = self.run()
        if not self._options.view_as(StandardOptions).no_wait_until_finish:
          self.result.wait_until_finish()
        else:
          logging.info(
              'Job execution continues without waiting for completion.'
              ' Use "wait_until_finish" in PipelineResult to block'
              ' until finished.')
    finally:
      self._extra_context.__exit__(exc_type, exc_val, exc_tb)

  def visit(self, visitor: 'PipelineVisitor') -> None:
    """Visits depth-first every node of a pipeline's DAG.

    Runner-internal implementation detail; no backwards-compatibility guarantees

    Args:
      visitor (~apache_beam.pipeline.PipelineVisitor):
        :class:`~apache_beam.pipeline.PipelineVisitor` object whose callbacks
        will be called for each node visited. See
        :class:`~apache_beam.pipeline.PipelineVisitor` comments.

    Raises:
      TypeError: if node is specified and is not a
        :class:`~apache_beam.pvalue.PValue`.
      ~apache_beam.error.PipelineError: if node is specified and does not
        belong to this pipeline instance.
    """

    visited: set[pvalue.PValue] = set()
    self._root_transform().visit(visitor, self, visited)

  def apply(
      self,
      transform: ptransform.PTransform,
      pvalueish: Optional[pvalue.PValue] = None,
      label: Optional[str] = None) -> pvalue.PValue:
    """Applies a custom transform using the pvalueish specified.

    Args:
      transform (~apache_beam.transforms.ptransform.PTransform): the
        :class:`~apache_beam.transforms.ptransform.PTransform` to apply.
      pvalueish (~apache_beam.pvalue.PCollection): the input for the
        :class:`~apache_beam.transforms.ptransform.PTransform` (typically a
        :class:`~apache_beam.pvalue.PCollection`).
      label (str): label of the
        :class:`~apache_beam.transforms.ptransform.PTransform`.

    Raises:
      TypeError: if the transform object extracted from the
        argument list is not a
        :class:`~apache_beam.transforms.ptransform.PTransform`.
      RuntimeError: if the transform object was already applied to
        this pipeline and needs to be cloned in order to apply again.
    """
    with scoped_pipeline_options(self._options):
      return self._apply_internal(transform, pvalueish, label)

  def _apply_internal(
      self,
      transform: ptransform.PTransform,
      pvalueish: Optional[pvalue.PValue] = None,
      label: Optional[str] = None) -> pvalue.PValue:
    """Internal implementation of apply(), called within scoped options."""
    if isinstance(transform, ptransform._NamedPTransform):
      return self.apply(
          transform.transform, pvalueish, label or transform.label)

    if not label and isinstance(transform, ptransform._PTransformFnPTransform):
      # This must be set before label is inspected.
      transform.set_options(self._options)

    if not isinstance(transform, ptransform.PTransform):
      raise TypeError("Expected a PTransform object, got %s" % transform)

    if label:
      # Fix self.label as it is inspected by some PTransform operations
      # (e.g. to produce error messages for type hint violations).
      old_label, transform.label = transform.label, label
      try:
        return self.apply(transform, pvalueish)
      finally:
        transform.label = old_label

    # Attempts to alter the label of the transform to be applied only when it's
    # a top-level transform so that the cell number will not be prepended to
    # every child transform in a composite.
    if self._current_transform() is self._root_transform():
      alter_label_if_ipython(transform, pvalueish)

    full_label = '/'.join(
        [self._current_transform().full_label, transform.label]).lstrip('/')
    if full_label in self.applied_labels:
      auto_unique_labels = self._options.view_as(
          StandardOptions).auto_unique_labels
      if auto_unique_labels:
        # If auto_unique_labels is set, we will append a unique suffix to the
        # label to make it unique.
        logging.warning(
            'Using --auto_unique_labels could cause data loss when '
            'updating a pipeline or reloading the job state. '
            'This is not recommended for streaming jobs.')
        unique_label = self._generate_unique_label(transform)
        return self.apply(transform, pvalueish, unique_label)
      else:
        raise RuntimeError(
            'A transform with label "%s" already exists in the pipeline. '
            'To apply a transform with a specified label, write '
            'pvalue | "label" >> transform or use the option '
            '"auto_unique_labels" to automatically generate unique '
            'transform labels. Note "auto_unique_labels" '
            'could cause data loss when updating a pipeline or '
            'reloading the job state. This is not recommended for '
            'streaming jobs.' % full_label)
    self.applied_labels.add(full_label)

    if pvalueish is None:
      full_label = self._current_transform().full_label
      raise TypeCheckError(
          f'Transform "{full_label}" was applied to the output of '
          f'an object of type None.')

    pvalueish, inputs = transform._extract_input_pvalues(pvalueish)
    try:
      if not isinstance(inputs, dict):
        inputs = {str(ix): input for (ix, input) in enumerate(inputs)}
    except TypeError:
      raise NotImplementedError(
          'Unable to extract PValue inputs from %s; either %s does not accept '
          'inputs of this format, or it does not properly override '
          '_extract_input_pvalues' % (pvalueish, transform))
    for t, leaf_input in inputs.items():
      if not isinstance(leaf_input, pvalue.PValue) or not isinstance(t, str):
        raise NotImplementedError(
            '%s does not properly override _extract_input_pvalues, '
            'returned %s from %s' % (transform, inputs, pvalueish))

    current = AppliedPTransform(
        self._current_transform(),
        transform,
        full_label,
        inputs,
        None,
        annotations=self._current_annotations())
    self._current_transform().add_part(current)

    try:
      self.transforms_stack.append(current)

      type_options = self._options.view_as(TypeOptions)
      if type_options.pipeline_type_check:
        transform.type_check_inputs(pvalueish)
      if isinstance(pvalueish, pvalue.PBegin) and isinstance(transform, ParDo):
        full_label = self._current_transform().full_label
        raise TypeCheckError(
            f"Transform '{full_label}' expects a PCollection as input. "
            "Got a PBegin/Pipeline instead.")

      self._assert_not_applying_PDone(pvalueish, transform)

      pvalueish_result = self.runner.apply(transform, pvalueish, self._options)

      if type_options is not None and type_options.pipeline_type_check:
        transform.type_check_outputs(pvalueish_result)

      for tag, result in ptransform.get_named_nested_pvalues(pvalueish_result):
        assert isinstance(result, (pvalue.PValue, pvalue.DoOutputsTuple))

        # Make sure we set the producer only for a leaf node in the transform
        # DAG. This way we preserve the last transform of a composite transform
        # as being the real producer of the result.
        if result.producer is None:
          result.producer = current

        # TODO(BEAM-1833): Pass full tuples dict.
        self._infer_result_type(transform, tuple(inputs.values()), result)

        assert isinstance(result.producer.inputs, tuple)
        if isinstance(result, pvalue.DoOutputsTuple):
          for tag, pc in list(result._pcolls.items()):
            if tag not in current.outputs:
              current.add_output(pc, tag)
          continue

        # If there is already a tag with the same name, increase a counter for
        # the name. This can happen, for example, when a composite outputs a
        # list of PCollections where all the tags are None.
        base = tag
        counter = 0
        while tag in current.outputs:
          counter += 1
          tag = '%s_%d' % (base, counter)

        current.add_output(result, tag)

      if (type_options is not None and
          type_options.type_check_strictness == 'ALL_REQUIRED' and
          transform.get_type_hints().output_types is None):
        ptransform_name = '%s(%s)' % (transform.__class__.__name__, full_label)
        raise TypeCheckError(
            'Pipeline type checking is enabled, however no '
            'output type-hint was found for the '
            'PTransform %s' % ptransform_name)
    finally:
      self.transforms_stack.pop()
    return pvalueish_result

  def _assert_not_applying_PDone(
      self,
      pvalueish: Optional[pvalue.PValue],
      transform: ptransform.PTransform):
    if isinstance(pvalueish, pvalue.PDone) and isinstance(transform, ParDo):
      # If the input is a PDone, we cannot apply a ParDo transform.
      full_label = self._current_transform().full_label
      producer_label = pvalueish.producer.full_label
      raise TypeCheckError(
          f'Transform "{full_label}" was applied to the output of '
          f'"{producer_label}" but "{producer_label.split("/")[-1]}" '
          'produces no PCollections.')

  def _generate_unique_label(self, transform: str) -> str:
    """
    Given a transform, generate a unique label for it based on current label.
    """
    unique_suffix = uuid.uuid4().hex[:6]
    return '%s_%s' % (transform.label, unique_suffix)

  def _infer_result_type(
      self,
      transform: ptransform.PTransform,
      inputs: Sequence[Union[pvalue.PBegin, pvalue.PCollection]],
      result_pcollection: Union[pvalue.PValue, pvalue.DoOutputsTuple]) -> None:
    """Infer and set the output element type for a PCollection.
    
    This function determines the output types of transforms by combining:
    1. Concrete input types from previous transforms
    2. Type hints declared on the current transform
    3. Type variable binding and substitution

    TYPE VARIABLE BINDING
    ---------------------
    Type variables (K, V, T, etc.) act as placeholders that get bound to
    concrete types through pattern matching. This requires both an input
    pattern and an output template:
    
    Input Pattern (from .with_input_types()):
        Defines where in the input to find each type variable
        Example: Tuple[K, V] means "K is the first element, V is the second"
    
    Output Template (from .with_output_types()):
        Defines how to use the bound variables in the output
        Example: Tuple[V, K] means "swap the positions"
    
    CONCRETE TYPES VS TYPE VARIABLES
    ---------------------------------
    The system handles these differently:
    
    Concrete Types (e.g., str, int, Tuple[str, int]):
        - Used as-is without any binding
        - Do not fall back to Any
        - Example: .with_output_types(Tuple[str, int]) → Tuple[str, int]
    
    Type Variables (e.g., K, V, T):
        - Must be bound through pattern matching
        - Require .with_input_types() to provide the pattern
        - Fall back to Any if not bound
        - Example without pattern: Tuple[K, V] → Tuple[Any, Any]
        - Example with pattern: Tuple[K, V] → Tuple[str, int]
    
    BINDING ALGORITHM
    -----------------
    1. Match: Compare input pattern to concrete input
       Pattern:  Tuple[K, V]
       Concrete: Tuple[str, int]
       Result:   {K: str, V: int}  ← Bindings created
    
    2. Substitute: Apply bindings to output template
       Template: Tuple[V, K]  ← Note: swapped!
       Bindings: {K: str, V: int}
       Result:   Tuple[int, str]  ← Swapped concrete types
    
    Each transform operates in its own type inference scope. Type variables
    declared in a parent composite transform do NOT automatically propagate
    to child transforms.

    Parent scope (composite):
        @with_input_types(Tuple[K, V])  ← K, V defined here
        class MyComposite(PTransform):
            def expand(self, pcoll):
                # Child scope - parent's K, V are NOT available
                return pcoll | ChildTransform()
    
    Type variables that remain unbound after inference fall back to Any:
    
    EXAMPLES
    --------
    Example 1: Concrete types (no variables)
        Input:  Tuple[str, int]
        Transform: .with_output_types(Tuple[str, int])
        Output: Tuple[str, int]  ← Used as-is
    
    Example 2: Type variables with pattern (correct)
        Input:  Tuple[str, int]
        Transform: .with_input_types(Tuple[K, V])
                   .with_output_types(Tuple[V, K])
        Binding: {K: str, V: int}
        Output: Tuple[int, str]  ← Swapped!
    
    Example 3: Type variables without pattern (falls back to Any)
        Input:  Tuple[str, int]
        Transform: .with_output_types(Tuple[K, V])  ← No input pattern!
        Binding: None (can't match)
        Output: Tuple[Any, Any]  ← Fallback
    
    Example 4: Mixed concrete and variables
        Input:  Tuple[str, int]
        Transform: .with_input_types(Tuple[str, V])
                   .with_output_types(Tuple[str, V])
        Binding: {V: int}  ← Only V needs binding
        Output: Tuple[str, int]  ← str passed through, V bound to int
    
    Args:
        transform: The PTransform being applied
        inputs: Input PCollections (provides concrete types)
        result_pcollection: Output PCollection to set type on
    """
    # TODO(robertwb): Multi-input inference.
    type_options = self._options.view_as(TypeOptions)
    if type_options is None or not type_options.pipeline_type_check:
      return
    if (isinstance(result_pcollection, pvalue.PCollection) and
        (not result_pcollection.element_type
         # TODO(robertwb): Ideally we'd do intersection here.
         or result_pcollection.element_type == typehints.Any)):
      # {Single, multi}-input, single-output inference.
      input_element_types_tuple = tuple(i.element_type for i in inputs)
      input_element_type = (
          input_element_types_tuple[0] if len(input_element_types_tuple) == 1
          else typehints.Union[input_element_types_tuple])
      type_hints = transform.get_type_hints()
      if not result_pcollection.tag:
        declared_output_type = type_hints.simple_output_type(transform.label)
      else:
        declared_output_type = type_hints.tagged_output_types().get(
            result_pcollection.tag, typehints.Any)

      if declared_output_type:
        input_types = type_hints.input_types
        if input_types and input_types[0]:
          declared_input_type = input_types[0][0]
          result_element_type = typehints.bind_type_variables(
              declared_output_type,
              typehints.match_type_variables(
                  declared_input_type, input_element_type))
        else:
          result_element_type = declared_output_type
      else:
        result_element_type = transform.infer_output_type(input_element_type)

      # Any remaining type variables have no bindings higher than this scope.
      result_pcollection.element_type = typehints.bind_type_variables(
          result_element_type, {'*': typehints.Any})
    elif isinstance(result_pcollection, pvalue.DoOutputsTuple):
      # {Single, multi}-input, multi-output inference.
      # TODO(https://github.com/apache/beam/issues/18957): Add support for
      #   tagged type hints.
      #   https://github.com/apache/beam/pull/9810#discussion_r338765251
      for pcoll in result_pcollection:
        if pcoll.element_type is None:
          pcoll.element_type = typehints.Any

  def __reduce__(self) -> tuple[Type, tuple[str, ...]]:
    # Some transforms contain a reference to their enclosing pipeline,
    # which in turn reference all other transforms (resulting in quadratic
    # time/space to pickle each transform individually).  As we don't
    # require pickled pipelines to be executable, break the chain here.
    return str, ('Pickled pipeline stub.', )

  def _verify_runner_api_compatible(self) -> bool:
    if self._options.view_as(TypeOptions).runtime_type_check:
      # This option is incompatible with the runner API as it requires
      # the runner to inspect non-serialized hints on the transform
      # itself.
      return False

    class Visitor(PipelineVisitor):  # pylint: disable=used-before-assignment
      ok = True  # Really a nonlocal.

      def enter_composite_transform(
          self, transform_node: AppliedPTransform) -> None:
        pass

      def visit_transform(self, transform_node: AppliedPTransform) -> None:
        try:
          # Transforms must be picklable.
          pickler.loads(
              pickler.dumps(transform_node.transform, enable_trace=False),
              enable_trace=False)
        except Exception:
          Visitor.ok = False

      def visit_value(self, value: pvalue.PValue, _: AppliedPTransform) -> None:
        if isinstance(value, pvalue.PDone):
          Visitor.ok = False

    self.visit(Visitor())
    return Visitor.ok

  def to_runner_api(
      self,
      return_context: bool = False,
      context: Optional['PipelineContext'] = None,
      use_fake_coders: bool = False,
      default_environment: Optional['environments.Environment'] = None
  ) -> beam_runner_api_pb2.Pipeline:
    """For internal use only; no backwards-compatibility guarantees."""
    from apache_beam.runners import pipeline_context
    if context is None:
      context = pipeline_context.PipelineContext(
          use_fake_coders=use_fake_coders,
          component_id_map=self.component_id_map,
          default_environment=default_environment)
    elif default_environment is not None:
      raise ValueError(
          'Only one of context or default_environment may be specified.')

    # The FlumeRunner is the only runner setting this option. Use getattr
    # because other runners do not have this option.
    context.enable_best_effort_deterministic_pickling = getattr(
        self.runner, 'enable_best_effort_deterministic_pickling', False)

    # The RunnerAPI spec requires certain transforms and side-inputs to have KV
    # inputs (and corresponding outputs).
    # Currently we only upgrade to KV pairs.  If there is a need for more
    # general shapes, potential conflicts will have to be resolved.
    # We also only handle single-input, and (for fixing the output) single
    # output, which is sufficient.
    # Also marks such values as requiring deterministic key coders.
    deterministic_key_coders = not self._options.view_as(
        TypeOptions).allow_non_deterministic_key_coders

    class ForceKvInputTypes(PipelineVisitor):
      def enter_composite_transform(
          self, transform_node: AppliedPTransform) -> None:
        self.visit_transform(transform_node)

      def visit_transform(self, transform_node: AppliedPTransform) -> None:
        if not transform_node.transform:
          return
        if hasattr(
            transform_node.transform, 'runner_api_requires_keyed_input'
        ) and transform_node.transform.runner_api_requires_keyed_input():
          pcoll = transform_node.inputs[0]
          pcoll.element_type = typehints.coerce_to_kv_type(
              pcoll.element_type, transform_node.full_label)
          pcoll.requires_deterministic_key_coder = (
              deterministic_key_coders and transform_node.full_label)
          if len(transform_node.outputs) == 1:
            # The runner often has expectations about the output types as well.
            output, = transform_node.outputs.values()
            if not output.element_type:
              output.element_type = transform_node.transform.infer_output_type(
                  pcoll.element_type)
            if (isinstance(output.element_type,
                           typehints.TupleHint.TupleConstraint) and
                len(output.element_type.tuple_types) == 2 and
                pcoll.element_type.tuple_types[0]
                == output.element_type.tuple_types[0]):
              output.requires_deterministic_key_coder = (
                  deterministic_key_coders and transform_node.full_label)
        for side_input in getattr(transform_node.transform, 'side_inputs', []):
          if side_input.requires_keyed_input():
            side_input.pvalue.element_type = typehints.coerce_to_kv_type(
                side_input.pvalue.element_type,
                transform_node.full_label,
                side_input_producer=side_input.pvalue.producer.full_label)
            side_input.pvalue.requires_deterministic_key_coder = (
                deterministic_key_coders and transform_node.full_label)

    self.visit(ForceKvInputTypes())

    # Mutates context; placing inline would force dependence on
    # argument evaluation order.
    root_transform_id = context.transforms.get_id(self._root_transform())
    proto = beam_runner_api_pb2.Pipeline(
        root_transform_ids=[root_transform_id],
        components=context.to_runner_api(),
        requirements=context.requirements(),
        display_data=DisplayData('', self._display_data).to_proto())
    proto.components.transforms[root_transform_id].unique_name = (
        root_transform_id)
    self.merge_compatible_environments(proto)
    if return_context:
      return proto, context  # type: ignore  # too complicated for now
    else:
      return proto

  @staticmethod
  def merge_compatible_environments(proto):
    """Tries to minimize the number of distinct environments by merging
    those that are compatible (currently defined as identical).

    Mutates proto as contexts may have references to proto.components.
    """
    pipeline_utils.merge_common_environments(proto, inplace=True)

  @staticmethod
  def from_runner_api(
      proto: beam_runner_api_pb2.Pipeline,
      runner: PipelineRunner,
      options: PipelineOptions,
      return_context: bool = False,
  ) -> 'Pipeline':
    """For internal use only; no backwards-compatibility guarantees."""
    p = Pipeline(
        runner=runner,
        options=options,
        display_data={
            str(ix): d
            for ix, d in enumerate(proto.display_data)
        })
    from apache_beam.runners import pipeline_context
    context = pipeline_context.PipelineContext(
        proto.components, requirements=proto.requirements)
    if proto.root_transform_ids:
      root_transform_id, = proto.root_transform_ids
      p.transforms_stack = [context.transforms.get_by_id(root_transform_id)]
    else:
      p.transforms_stack = [AppliedPTransform(None, None, '', None, None, None)]
    # TODO(robertwb): These are only needed to continue construction. Omit?
    p.applied_labels = {
        t.unique_name
        for t in proto.components.transforms.values()
    }
    for id in proto.components.pcollections:
      pcollection = context.pcollections.get_by_id(id)
      pcollection.pipeline = p
      if not pcollection.producer:
        raise ValueError('No producer for %s' % id)

    # Inject PBegin input where necessary.
    from apache_beam.io.iobase import Read
    from apache_beam.transforms.core import Create
    has_pbegin = [Read, Create]
    for id in proto.components.transforms:
      transform = context.transforms.get_by_id(id)
      if not transform.inputs and transform.transform.__class__ in has_pbegin:
        transform.main_inputs = {'None': pvalue.PBegin(p)}

    if return_context:
      return p, context  # type: ignore  # too complicated for now
    else:
      return p


class PipelineVisitor(object):
  """For internal use only; no backwards-compatibility guarantees.

  Visitor pattern class used to traverse a DAG of transforms
  (used internally by Pipeline for bookkeeping purposes).
  """
  def visit_value(
      self, value: pvalue.PValue, producer_node: 'AppliedPTransform') -> None:
    """Callback for visiting a PValue in the pipeline DAG.

    Args:
      value: PValue visited (typically a PCollection instance).
      producer_node: AppliedPTransform object whose transform produced the
        pvalue.
    """
    pass

  def visit_transform(self, transform_node: 'AppliedPTransform') -> None:
    """Callback for visiting a transform leaf node in the pipeline DAG."""
    pass

  def enter_composite_transform(
      self, transform_node: 'AppliedPTransform') -> None:
    """Callback for entering traversal of a composite transform node."""
    pass

  def leave_composite_transform(
      self, transform_node: 'AppliedPTransform') -> None:
    """Callback for leaving traversal of a composite transform node."""
    pass


class ExternalTransformFinder(PipelineVisitor):
  """Looks for any external transforms in the pipeline and if found records
  it.
  """
  def __init__(self):
    self._contains_external_transforms = False

  @staticmethod
  def contains_external_transforms(pipeline):
    visitor = ExternalTransformFinder()
    pipeline.visit(visitor)
    return visitor._contains_external_transforms

  def _perform_exernal_transform_test(self, transform):
    if not transform:
      return
    from apache_beam.transforms import ExternalTransform
    if isinstance(transform, ExternalTransform):
      self._contains_external_transforms = True

  def visit_transform(self, transform_node: 'AppliedPTransform') -> None:
    self._perform_exernal_transform_test(transform_node.transform)

  def enter_composite_transform(
      self, transform_node: 'AppliedPTransform') -> None:
    # Python SDK object graph may represent an external transform that is a leaf
    # of the pipeline graph as a composite without sub-transforms.
    # Note that this visitor is just used to identify pipelines with external
    # transforms. A Runner API pipeline proto generated from the Pipeline object
    # will include external sub-transform.
    self._perform_exernal_transform_test(transform_node.transform)


class AppliedPTransform(object):
  """For internal use only; no backwards-compatibility guarantees.

  A transform node representing an instance of applying a PTransform
  (used internally by Pipeline for bookeeping purposes).
  """
  def __init__(
      self,
      parent: Optional['AppliedPTransform'],
      transform: Optional[ptransform.PTransform],
      full_label: str,
      main_inputs: Optional[Mapping[str,
                                    Union[pvalue.PBegin, pvalue.PCollection]]],
      environment_id: Optional[str],
      annotations: Optional[dict[str, bytes]],
  ) -> None:
    self.parent = parent
    self.transform = transform
    # Note that we want the PipelineVisitor classes to use the full_label,
    # inputs, side_inputs, and outputs fields from this instance instead of the
    # ones of the PTransform instance associated with it. Doing this permits
    # reusing PTransform instances in different contexts (apply() calls) without
    # any interference. This is particularly useful for composite transforms.
    self.full_label = full_label
    self.main_inputs = dict(main_inputs or {})

    self.side_inputs = (
        tuple() if transform is None else getattr(
            transform, 'side_inputs', tuple()))
    self.outputs: dict[Union[str, int, None], pvalue.PValue] = {}
    self.parts: list[AppliedPTransform] = []
    self.environment_id: Optional[
        str] = environment_id if environment_id else None
    # We may need to merge the hints with environment-provided hints here
    # once environment is a first-class citizen in Beam graph and we have
    # access to actual environment, not just an id.
    self.resource_hints: dict[str, bytes] = dict(
        transform.get_resource_hints()) if transform and hasattr(
            transform, 'get_resource_hints') else {}

    if transform and hasattr(transform, 'annotations'):
      annotations = {
          **(annotations or {}), **encode_annotations(transform.annotations())
      }

    self.annotations = annotations

    self.display_data = {}

  @property
  def inputs(self):
    return tuple(self.main_inputs.values())

  def __repr__(self) -> str:
    return "%s(%s, %s)" % (
        self.__class__.__name__, self.full_label, type(self.transform).__name__)

  def replace_output(
      self,
      output: Union[pvalue.PValue, pvalue.DoOutputsTuple],
      tag: Union[str, int, None] = None) -> None:
    """Replaces the output defined by the given tag with the given output.

    Args:
      output: replacement output
      tag: tag of the output to be replaced.
    """
    if isinstance(output, pvalue.DoOutputsTuple):
      self.replace_output(output[output._main_tag])
    elif isinstance(output, pvalue.PValue):
      self.outputs[tag] = output
    elif isinstance(output, dict):
      for output_tag, out in output.items():
        self.outputs[output_tag] = out
    else:
      raise TypeError("Unexpected output type: %s" % output)

    # Importing locally to prevent circular dependency issues.
    from apache_beam.transforms import external
    if isinstance(self.transform, external.ExternalTransform):
      self.transform.replace_named_outputs(self.named_outputs())

  def replace_inputs(self, main_inputs):
    self.main_inputs = main_inputs

    # Importing locally to prevent circular dependency issues.
    from apache_beam.transforms import external
    if isinstance(self.transform, external.ExternalTransform):
      self.transform.replace_named_inputs(self.named_inputs())

  def replace_side_inputs(self, side_inputs):
    self.side_inputs = side_inputs

    # Importing locally to prevent circular dependency issues.
    from apache_beam.transforms import external
    if isinstance(self.transform, external.ExternalTransform):
      self.transform.replace_named_inputs(self.named_inputs())

  def add_output(
      self,
      output: Union[pvalue.DoOutputsTuple, pvalue.PValue],
      tag: Union[str, int, None]) -> None:
    if isinstance(output, pvalue.DoOutputsTuple):
      self.add_output(output[tag], tag)
    elif isinstance(output, pvalue.PValue):
      assert tag not in self.outputs
      self.outputs[tag] = output
    else:
      raise TypeError("Unexpected output type: %s" % output)

  def add_part(self, part: 'AppliedPTransform') -> None:
    assert isinstance(part, AppliedPTransform)
    part._merge_outer_resource_hints()
    self.parts.append(part)

  def is_composite(self) -> bool:
    """Returns whether this is a composite transform.

    A composite transform has parts (inner transforms) or isn't the
    producer for any of its outputs. (An example of a transform that
    is not a producer is one that returns its inputs instead.)
    """
    return bool(self.parts) or all(
        pval.producer is not self for pval in self.outputs.values())

  def visit(
      self,
      visitor: PipelineVisitor,
      pipeline: Pipeline,
      visited: set[pvalue.PValue]) -> None:
    """Visits all nodes reachable from the current node."""

    for in_pval in self.inputs:
      if in_pval not in visited and not isinstance(in_pval, pvalue.PBegin):
        if in_pval.producer is not None:
          in_pval.producer.visit(visitor, pipeline, visited)
          # The value should be visited now since we visit outputs too.
          assert in_pval in visited, in_pval

    # Visit side inputs.
    for side_input in self.side_inputs:
      if isinstance(side_input, pvalue.AsSideInput) \
          and side_input.pvalue not in visited:
        pval = side_input.pvalue  # Unpack marker-object-wrapped pvalue.
        if pval.producer is not None:
          pval.producer.visit(visitor, pipeline, visited)
          # The value should be visited now since we visit outputs too.
          assert pval in visited
          # TODO(silviuc): Is there a way to signal that we are visiting a side
          # value? The issue is that the same PValue can be reachable through
          # multiple paths and therefore it is not guaranteed that the value
          # will be visited as a side value.

    # Visit a composite or primitive transform.
    if self.is_composite():
      visitor.enter_composite_transform(self)
      for part in self.parts:
        part.visit(visitor, pipeline, visited)
      visitor.leave_composite_transform(self)
    else:
      visitor.visit_transform(self)

    # Visit the outputs (one or more). It is essential to mark as visited the
    # tagged PCollections of the DoOutputsTuple object. A tagged PCollection is
    # connected directly with its producer (a multi-output ParDo), but the
    # output of such a transform is the containing DoOutputsTuple, not the
    # PCollection inside it. Without the code below a tagged PCollection will
    # not be marked as visited while visiting its producer.
    for out_pval in self.outputs.values():
      if isinstance(out_pval, pvalue.DoOutputsTuple):
        pvals = (v for v in out_pval)
      else:
        pvals = (out_pval, )
      for v in pvals:
        if v not in visited:
          visited.add(v)
          visitor.visit_value(v, self)

  def named_inputs(self) -> dict[str, pvalue.PValue]:
    if self.transform is None:
      assert not self.main_inputs and not self.side_inputs
      return {}
    else:
      if hasattr(self.transform, '_named_inputs'):
        named_inputs = self.transform._named_inputs(
            self.main_inputs, self.side_inputs)
      else:
        named_inputs = {}
      if not self.parts:
        for name, pc_out in self.outputs.items():
          if pc_out.producer is not self and pc_out not in named_inputs.values(
          ):
            named_inputs[f'__implicit_input_{name}'] = pc_out
      return named_inputs

  def named_outputs(self) -> dict[str, pvalue.PCollection]:
    if self.transform is None:
      assert not self.outputs
      return {}
    else:
      if hasattr(self.transform, '_named_outputs'):
        return self.transform._named_outputs(self.outputs)
      else:
        return {}

  def to_runner_api(
      self, context: 'PipelineContext') -> beam_runner_api_pb2.PTransform:
    # External transforms require more splicing than just setting the spec.
    from apache_beam.transforms import external
    if isinstance(self.transform, external.ExternalTransform):
      # TODO(https://github.com/apache/beam/issues/18371): Support resource
      # hints in XLang transforms. In particular, make sure hints on composites
      # are properly propagated.
      return self.transform.to_runner_api_transform(context, self.full_label)

    def transform_to_runner_api(
        transform: Optional[ptransform.PTransform], context: 'PipelineContext'
    ) -> Optional[beam_runner_api_pb2.FunctionSpec]:
      if transform is None:
        return None
      else:
        # We only populate inputs information to ParDo in order to expose
        # key_coder and window_coder to stateful DoFn.
        if isinstance(transform, ParDo):
          return transform.to_runner_api(
              context,
              has_parts=bool(self.parts),
              named_inputs=self.named_inputs())
        elif hasattr(transform, 'to_runner_api'):
          return transform.to_runner_api(context, has_parts=bool(self.parts))
        return None

    # Iterate over inputs and outputs by sorted key order, so that ids are
    # consistently generated for multiple runs of the same pipeline.
    try:
      transform_spec = transform_to_runner_api(self.transform, context)
    except Exception as exn:
      raise RuntimeError(f'Unable to translate {self.full_label}') from exn
    environment_id = self.environment_id
    transform_urn = transform_spec.urn if transform_spec else None
    if (not environment_id and
        (transform_urn not in Pipeline.runner_implemented_transforms())):
      environment_id = context.get_environment_id_for_resource_hints(
          self.resource_hints)
    if self.transform is not None:
      display_data = DisplayData.create_from(
          self.transform, extra_items=self.display_data)
    else:
      display_data = None

    return beam_runner_api_pb2.PTransform(
        unique_name=self.full_label,
        spec=transform_spec,
        subtransforms=[
            context.transforms.get_id(part, label=part.full_label)
            for part in self.parts
        ],
        inputs={
            tag: context.pcollections.get_id(pc)
            for tag, pc in sorted(self.named_inputs().items())
        },
        outputs={
            tag: context.pcollections.get_id(out)
            for tag, out in sorted(self.named_outputs().items())
        },
        environment_id=environment_id,
        annotations=self.annotations,
        # TODO(https://github.com/apache/beam/issues/18012): Add display_data.
        display_data=display_data.to_proto() if display_data else None)

  @staticmethod
  def from_runner_api(
      proto: beam_runner_api_pb2.PTransform,
      context: 'PipelineContext') -> 'AppliedPTransform':

    if common_urns.primitives.PAR_DO.urn == proto.spec.urn:
      # Preserving side input tags.
      pardo_payload = (
          proto_utils.parse_Bytes(
              proto.spec.payload, beam_runner_api_pb2.ParDoPayload))
      side_input_tags = list(pardo_payload.side_inputs.keys())
    else:
      pardo_payload = None
      side_input_tags = []

    main_inputs = {
        tag: context.pcollections.get_by_id(id)
        for (tag, id) in proto.inputs.items() if tag not in side_input_tags
    }

    transform = ptransform.PTransform.from_runner_api(proto, context)
    if transform and proto.environment_id:
      resource_hints = context.environments.get_by_id(
          proto.environment_id).resource_hints()
      if resource_hints:
        transform._resource_hints = dict(resource_hints)

    # Ordering is important here.
    # TODO(https://github.com/apache/beam/issues/20136): use key, value pairs
    # instead of depending on tags with index as a suffix.
    indexed_side_inputs = [
        (get_sideinput_index(tag), context.pcollections.get_by_id(id))
        for tag, id in proto.inputs.items() if tag in side_input_tags
    ]
    side_inputs = [si for _, si in sorted(indexed_side_inputs)]

    result = AppliedPTransform(
        parent=None,
        transform=transform,
        full_label=proto.unique_name,
        main_inputs=main_inputs,
        environment_id=None,
        annotations=proto.annotations)

    if result.transform and hasattr(
        result.transform, 'side_inputs') and result.transform.side_inputs:
      for si, pcoll in zip(result.transform.side_inputs, side_inputs):
        si.pvalue = pcoll
      result.side_inputs = tuple(result.transform.side_inputs)
    result.parts = []
    for transform_id in proto.subtransforms:
      part = context.transforms.get_by_id(transform_id)
      part.parent = result
      result.add_part(part)
    result.outputs = {
        None if tag == 'None' else tag: context.pcollections.get_by_id(id)
        for tag, id in proto.outputs.items()
    }
    # This annotation is expected by some runners.
    if proto.spec.urn == common_urns.primitives.PAR_DO.urn:
      result.transform.output_tags = set(proto.outputs.keys()).difference(
          {'None'})
    if not result.parts:
      for tag, pcoll_id in proto.outputs.items():
        if pcoll_id not in proto.inputs.values():
          pc = context.pcollections.get_by_id(pcoll_id)
          pc.producer = result
          pc.tag = None if tag == 'None' else tag
    return result

  def _merge_outer_resource_hints(self):
    if (self.parent is not None and self.parent.resource_hints):
      self.resource_hints = merge_resource_hints(
          outer_hints=self.parent.resource_hints,
          inner_hints=self.resource_hints)
    if self.resource_hints:
      for part in self.parts:
        part._merge_outer_resource_hints()


def encode_annotations(annotations: Optional[dict[str, Any]]):
  """Encodes non-byte annotation values as bytes."""
  if not annotations:
    return {}

  def annotation_to_bytes(key, a: Any) -> bytes:
    if isinstance(a, bytes):
      return a
    elif isinstance(a, str):
      return a.encode('ascii')
    elif isinstance(a, message.Message):
      return a.SerializeToString()
    else:
      raise TypeError(
          'Unknown annotation type %r (type %s) for %s' % (a, type(a), key))

  return {key: annotation_to_bytes(key, a) for (key, a) in annotations.items()}


_global_annotations_stack_data = threading.local()


def _global_annotations_stack():
  try:
    return _global_annotations_stack_data.stack
  except AttributeError:
    _global_annotations_stack_data.stack = [{}]
    return _global_annotations_stack_data.stack


@contextlib.contextmanager
def transform_annotations(**annotations):
  """A context manager for attaching annotations to a set of transforms.

  All transforms applied while this context is active will have these
  annotations attached. This includes  sub-transforms applied within
  composite transforms.
  """
  cur_stack = _global_annotations_stack()
  cur_stack.append({**cur_stack[-1], **encode_annotations(annotations)})
  yield
  cur_stack.pop()


class PTransformOverride(metaclass=abc.ABCMeta):
  """For internal use only; no backwards-compatibility guarantees.

  Gives a matcher and replacements for matching PTransforms.

  TODO: Update this to support cases where input and/our output types are
  different.
  """
  @abc.abstractmethod
  def matches(self, applied_ptransform: AppliedPTransform) -> bool:
    """Determines whether the given AppliedPTransform matches.

    Note that the matching will happen *after* Runner API proto translation.
    If matching is done via type checks, to/from_runner_api[_parameter] methods
    must be implemented to preserve the type (and other data) through proto
    serialization.

    Consider URN-based translation instead.

    Args:
      applied_ptransform: AppliedPTransform to be matched.

    Returns:
      a bool indicating whether the given AppliedPTransform is a match.
    """
    raise NotImplementedError

  def get_replacement_transform_for_applied_ptransform(
      self, applied_ptransform: AppliedPTransform) -> ptransform.PTransform:
    """Provides a runner specific override for a given `AppliedPTransform`.

    Args:
      applied_ptransform: `AppliedPTransform` containing the `PTransform` to be
        replaced.

    Returns:
      A `PTransform` that will be the replacement for the `PTransform` inside
      the `AppliedPTransform` given as an argument.
    """
    # Returns a PTransformReplacement
    return self.get_replacement_transform(applied_ptransform.transform)

  @deprecated(
      since='2.24', current='get_replacement_transform_for_applied_ptransform')
  def get_replacement_transform(
      self,
      ptransform: Optional[ptransform.PTransform]) -> ptransform.PTransform:
    """Provides a runner specific override for a given PTransform.

    Args:
      ptransform: PTransform to be replaced.

    Returns:
      A PTransform that will be the replacement for the PTransform given as an
      argument.
    """
    # Returns a PTransformReplacement
    raise NotImplementedError

  def get_replacement_inputs(
      self, applied_ptransform: AppliedPTransform) -> Iterable[pvalue.PValue]:
    """Provides inputs that will be passed to the replacement PTransform.

    Args:
      applied_ptransform: Original AppliedPTransform containing the PTransform
        to be replaced.

    Returns:
      An iterable of PValues that will be passed to the expand() method of the
      replacement PTransform.
    """
    return tuple(applied_ptransform.inputs) + tuple(
        side_input.pvalue for side_input in applied_ptransform.side_inputs)


class ComponentIdMap(object):
  """A utility for assigning unique component ids to Beam components.

  Component ID assignments are only guaranteed to be unique and consistent
  within the scope of a ComponentIdMap instance.
  """
  def __init__(self, namespace="ref"):
    self.namespace = namespace
    self._counters: dict[type, int] = defaultdict(lambda: 0)
    self._obj_to_id: dict[Any, str] = {}

  def get_or_assign(self, obj=None, obj_type=None, label=None):
    if obj not in self._obj_to_id:
      self._obj_to_id[obj] = self._unique_ref(obj, obj_type, label)

    return self._obj_to_id[obj]

  def _normalize(self, str_value):
    str_value = unicodedata.normalize('NFC', str_value)
    return re.sub(r'[^a-zA-Z0-9-_]+', '-', str_value)

  def _unique_ref(self, obj=None, obj_type=None, label=None):
    # Normalize, trim, and uniqify.
    prefix = self._normalize(
        '%s_%s_%s' %
        (self.namespace, obj_type.__name__, label or type(obj).__name__))[0:100]
    if isinstance(obj, typecoders.coders.Coder) and obj.version_tag():
      prefix = "%s_%s" % (prefix, obj.version_tag())
    self._counters[obj_type] += 1
    return '%s_%d' % (prefix, self._counters[obj_type])