sdks/python/apache_beam/transforms/external.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.
#

"""Defines Transform whose expansion is implemented elsewhere."""
# pytype: skip-file

import contextlib
import copy
import functools
import glob
import logging
import re
import subprocess
import threading
import uuid
from collections import OrderedDict
from collections import namedtuple

import grpc
import yaml

from apache_beam import pvalue
from apache_beam.coders import RowCoder
from apache_beam.options import pipeline_options
from apache_beam.portability import common_urns
from apache_beam.portability.api import beam_artifact_api_pb2_grpc
from apache_beam.portability.api import beam_expansion_api_pb2
from apache_beam.portability.api import beam_expansion_api_pb2_grpc
from apache_beam.portability.api import beam_runner_api_pb2
from apache_beam.portability.api import external_transforms_pb2
from apache_beam.portability.api import schema_pb2
from apache_beam.portability.common_urns import ManagedTransforms
from apache_beam.runners import pipeline_context
from apache_beam.runners.portability import artifact_service
from apache_beam.transforms import environments
from apache_beam.transforms import ptransform
from apache_beam.typehints import WithTypeHints
from apache_beam.typehints import native_type_compatibility
from apache_beam.typehints import row_type
from apache_beam.typehints.schemas import named_fields_to_schema
from apache_beam.typehints.schemas import named_tuple_from_schema
from apache_beam.typehints.schemas import named_tuple_to_schema
from apache_beam.typehints.schemas import typing_from_runner_api
from apache_beam.typehints.trivial_inference import instance_to_type
from apache_beam.typehints.typehints import Union
from apache_beam.typehints.typehints import UnionConstraint
from apache_beam.utils import subprocess_server
from apache_beam.utils import transform_service_launcher

DEFAULT_EXPANSION_SERVICE = 'localhost:8097'

MANAGED_SCHEMA_TRANSFORM_IDENTIFIER = "beam:transform:managed:v1"

_IO_EXPANSION_SERVICE_JAR_TARGET = "sdks:java:io:expansion-service:shadowJar"

_GCP_EXPANSION_SERVICE_JAR_TARGET = (
    "sdks:java:io:google-cloud-platform:expansion-service:shadowJar")

# A mapping from supported managed transforms URNs to expansion service jars
# that include the corresponding transforms.
MANAGED_TRANSFORM_URN_TO_JAR_TARGET_MAPPING = {
    ManagedTransforms.Urns.ICEBERG_READ.urn: _IO_EXPANSION_SERVICE_JAR_TARGET,
    ManagedTransforms.Urns.ICEBERG_WRITE.urn: _IO_EXPANSION_SERVICE_JAR_TARGET,
    ManagedTransforms.Urns.ICEBERG_CDC_READ.urn: _IO_EXPANSION_SERVICE_JAR_TARGET,  # pylint: disable=line-too-long
    ManagedTransforms.Urns.KAFKA_READ.urn: _IO_EXPANSION_SERVICE_JAR_TARGET,
    ManagedTransforms.Urns.KAFKA_WRITE.urn: _IO_EXPANSION_SERVICE_JAR_TARGET,
    ManagedTransforms.Urns.BIGQUERY_READ.urn: _GCP_EXPANSION_SERVICE_JAR_TARGET,
    ManagedTransforms.Urns.BIGQUERY_WRITE.urn: _GCP_EXPANSION_SERVICE_JAR_TARGET,  # pylint: disable=line-too-long
    ManagedTransforms.Urns.POSTGRES_READ.urn: _GCP_EXPANSION_SERVICE_JAR_TARGET,
    ManagedTransforms.Urns.POSTGRES_WRITE.urn: _GCP_EXPANSION_SERVICE_JAR_TARGET,  # pylint: disable=line-too-long
    ManagedTransforms.Urns.MYSQL_READ.urn: _GCP_EXPANSION_SERVICE_JAR_TARGET,
    ManagedTransforms.Urns.MYSQL_WRITE.urn: _GCP_EXPANSION_SERVICE_JAR_TARGET,
    ManagedTransforms.Urns.SQL_SERVER_READ.urn: _GCP_EXPANSION_SERVICE_JAR_TARGET,  # pylint: disable=line-too-long
    ManagedTransforms.Urns.SQL_SERVER_WRITE.urn: _GCP_EXPANSION_SERVICE_JAR_TARGET,  # pylint: disable=line-too-long
    ManagedTransforms.Urns.DELTA_LAKE_READ.urn: _IO_EXPANSION_SERVICE_JAR_TARGET,
    ManagedTransforms.Urns.DELTA_LAKE_CDC_READ.urn: _IO_EXPANSION_SERVICE_JAR_TARGET,
}


def convert_to_typing_type(type_):
  if isinstance(type_, row_type.RowTypeConstraint):
    return named_tuple_from_schema(named_fields_to_schema(type_._fields))
  else:
    return native_type_compatibility.convert_to_python_type(type_)


def _is_optional_or_none(typehint):
  return (
      type(None) in typehint.union_types if isinstance(
          typehint, UnionConstraint) else typehint is type(None))


def _strip_optional(typehint):
  if not _is_optional_or_none(typehint):
    return typehint
  new_types = typehint.union_types.difference({type(None)})
  if len(new_types) == 1:
    return list(new_types)[0]
  return Union[new_types]


def iter_urns(coder, context=None):
  yield coder.to_runner_api_parameter(context)[0]
  for child in coder._get_component_coders():
    for urn in iter_urns(child, context):
      yield urn


class PayloadBuilder(object):
  """
  Abstract base class for building payloads to pass to ExternalTransform.
  """
  def build(self):
    """
    :return: ExternalConfigurationPayload
    """
    raise NotImplementedError

  def payload(self):
    """
    The serialized ExternalConfigurationPayload

    :return: bytes
    """
    return self.build().SerializeToString()

  def _get_schema_proto_and_payload(self, **kwargs):
    named_fields = []
    fields_to_values = OrderedDict()

    for key, value in kwargs.items():
      if not key:
        raise ValueError('Parameter name cannot be empty')
      if value is None:
        raise ValueError(
            'Received value None for key %s. None values are currently not '
            'supported' % key)
      named_fields.append(
          (key, convert_to_typing_type(instance_to_type(value))))
      fields_to_values[key] = value

    schema_proto = named_fields_to_schema(named_fields)
    row = named_tuple_from_schema(schema_proto)(**fields_to_values)
    schema = named_tuple_to_schema(type(row))

    payload = RowCoder(schema).encode(row)
    return (schema_proto, payload)


class SchemaBasedPayloadBuilder(PayloadBuilder):
  """
  Base class for building payloads based on a schema that provides
  type information for each configuration value to encode.
  """
  def _get_named_tuple_instance(self):
    raise NotImplementedError()

  def build(self):
    row = self._get_named_tuple_instance()
    schema = named_tuple_to_schema(type(row))
    return external_transforms_pb2.ExternalConfigurationPayload(
        schema=schema, payload=RowCoder(schema).encode(row))


class ImplicitSchemaPayloadBuilder(SchemaBasedPayloadBuilder):
  """
  Build a payload that generates a schema from the provided values.
  """
  def __init__(self, values):
    self._values = values

  def _get_named_tuple_instance(self):
    # omit fields with value=None since we can't infer their type
    values = {
        key: value
        for key, value in self._values.items() if value is not None
    }

    schema = named_fields_to_schema([
        (key, convert_to_typing_type(instance_to_type(value)))
        for key, value in values.items()
    ])
    return named_tuple_from_schema(schema)(**values)


class NamedTupleBasedPayloadBuilder(SchemaBasedPayloadBuilder):
  """
  Build a payload based on a NamedTuple schema.
  """
  def __init__(self, tuple_instance):
    """
    :param tuple_instance: an instance of a typing.NamedTuple
    """
    super().__init__()
    self._tuple_instance = tuple_instance

  def _get_named_tuple_instance(self):
    return self._tuple_instance


class SchemaTransformPayloadBuilder(PayloadBuilder):
  def __init__(self, identifier, **kwargs):
    self._identifier = identifier
    self._kwargs = kwargs

  def identifier(self):
    """
    The URN referencing this SchemaTransform

    :return: str
    """
    return self._identifier

  def build(self):
    schema_proto, payload = self._get_schema_proto_and_payload(**self._kwargs)
    payload = external_transforms_pb2.SchemaTransformPayload(
        identifier=self._identifier,
        configuration_schema=schema_proto,
        configuration_row=payload)
    return payload


class ExplicitSchemaTransformPayloadBuilder(SchemaTransformPayloadBuilder):
  def __init__(self, identifier, schema_proto, **kwargs):
    self._identifier = identifier
    self._schema_proto = schema_proto
    self._kwargs = kwargs

  def build(self):
    def dict_to_row_recursive(field_type, py_value):
      if py_value is None:
        return None
      type_info = field_type.WhichOneof('type_info')
      if type_info == 'row_type':
        return dict_to_row(field_type.row_type.schema, py_value)
      elif type_info == 'array_type':
        return [
            dict_to_row_recursive(field_type.array_type.element_type, value)
            for value in py_value
        ]
      elif type_info == 'map_type':
        return {
            key: dict_to_row_recursive(field_type.map_type.value_type, value)
            for key, value in py_value.items()
        }
      else:
        return py_value

    def dict_to_row(schema_proto, py_value):
      row_type = named_tuple_from_schema(schema_proto)
      if isinstance(py_value, dict):
        extra = set(py_value.keys()) - set(row_type._fields)
        if extra:
          raise ValueError(
              f"Transform '{self.identifier()}' was configured with unknown "
              f"fields: {extra}. Valid fields: {set(row_type._fields)}")
        return row_type(
            *[
                dict_to_row_recursive(
                    field.type, py_value.get(field.name, None))
                for field in schema_proto.fields
            ])
      else:
        return row_type(py_value)

    return external_transforms_pb2.SchemaTransformPayload(
        identifier=self._identifier,
        configuration_schema=self._schema_proto,
        configuration_row=RowCoder(self._schema_proto).encode(
            dict_to_row(self._schema_proto, self._kwargs)))


class JavaClassLookupPayloadBuilder(PayloadBuilder):
  """
  Builds a payload for directly instantiating a Java transform using a
  constructor and builder methods.
  """

  IGNORED_ARG_FORMAT = 'ignore%d'

  def __init__(self, class_name):
    """
    :param class_name: fully qualified name of the transform class.
    """
    if not class_name:
      raise ValueError('Class name must not be empty')

    self._class_name = class_name
    self._constructor_method = None
    self._constructor_param_args = None
    self._constructor_param_kwargs = None
    self._builder_methods_and_params = OrderedDict()

  def _args_to_named_fields(self, args):
    next_field_id = 0
    named_fields = OrderedDict()
    for value in args:
      if value is None:
        raise ValueError(
            'Received value None. None values are currently not supported')
      named_fields[(
          JavaClassLookupPayloadBuilder.IGNORED_ARG_FORMAT %
          next_field_id)] = value
      next_field_id += 1
    return named_fields

  def build(self):
    all_constructor_param_kwargs = self._args_to_named_fields(
        self._constructor_param_args)
    if self._constructor_param_kwargs:
      all_constructor_param_kwargs.update(self._constructor_param_kwargs)
    constructor_schema, constructor_payload = (
      self._get_schema_proto_and_payload(**all_constructor_param_kwargs))
    payload = external_transforms_pb2.JavaClassLookupPayload(
        class_name=self._class_name,
        constructor_schema=constructor_schema,
        constructor_payload=constructor_payload)
    if self._constructor_method:
      payload.constructor_method = self._constructor_method

    for builder_method_name, params in self._builder_methods_and_params.items():
      builder_method_args, builder_method_kwargs = params
      all_builder_method_kwargs = self._args_to_named_fields(
          builder_method_args)
      if builder_method_kwargs:
        all_builder_method_kwargs.update(builder_method_kwargs)
      builder_method_schema, builder_method_payload = (
        self._get_schema_proto_and_payload(**all_builder_method_kwargs))
      builder_method = external_transforms_pb2.BuilderMethod(
          name=builder_method_name,
          schema=builder_method_schema,
          payload=builder_method_payload)
      builder_method.name = builder_method_name
      payload.builder_methods.append(builder_method)
    return payload

  def with_constructor(self, *args, **kwargs):
    """
    Specifies the Java constructor to use.
    Arguments provided using args and kwargs will be applied to the Java
    transform constructor in the specified order.

    :param args: parameter values of the constructor.
    :param kwargs: parameter names and values of the constructor.
    """
    if self._has_constructor():
      raise ValueError(
          'Constructor or constructor method can only be specified once')

    self._constructor_param_args = args
    self._constructor_param_kwargs = kwargs

  def with_constructor_method(self, method_name, *args, **kwargs):
    """
    Specifies the Java constructor method to use.
    Arguments provided using args and kwargs will be applied to the Java
    transform constructor method in the specified order.

    :param method_name: name of the constructor method.
    :param args: parameter values of the constructor method.
    :param kwargs: parameter names and values of the constructor method.
    """
    if self._has_constructor():
      raise ValueError(
          'Constructor or constructor method can only be specified once')

    self._constructor_method = method_name
    self._constructor_param_args = args
    self._constructor_param_kwargs = kwargs

  def add_builder_method(self, method_name, *args, **kwargs):
    """
    Specifies a Java builder method to be invoked after instantiating the Java
    transform class. Specified builder method will be applied in order.
    Arguments provided using args and kwargs will be applied to the Java
    transform builder method in the specified order.

    :param method_name: name of the builder method.
    :param args: parameter values of the builder method.
    :param kwargs:  parameter names and values of the builder method.
    """
    self._builder_methods_and_params[method_name] = (args, kwargs)

  def _has_constructor(self):
    return (
        self._constructor_method or self._constructor_param_args or
        self._constructor_param_kwargs)


# Information regarding a SchemaTransform available in an external SDK.
SchemaTransformsConfig = namedtuple(
    'SchemaTransformsConfig',
    ['identifier', 'configuration_schema', 'inputs', 'outputs', 'description'])

ManagedReplacement = namedtuple(
    'ManagedReplacement',
    ['underlying_transform_identifier', 'update_compatibility_version'])


class SchemaAwareExternalTransform(ptransform.PTransform):
  """A proxy transform for SchemaTransforms implemented in external SDKs.

  This allows Python pipelines to directly use existing SchemaTransforms
  available to the expansion service without adding additional code in external
  SDKs.

  :param identifier: unique identifier of the SchemaTransform.
  :param expansion_service: an expansion service to use. This should already be
      available and the Schema-aware transforms to be used must already be
      deployed.
  :param rearrange_based_on_discovery: if this flag is set, the input kwargs
      will be rearranged to match the order of fields in the external
      SchemaTransform configuration. A discovery call will be made to fetch
      the configuration.
  :param classpath: (Optional) A list paths to additional jars to place on the
      expansion service classpath.
  :param managed_replacement: (Optional) a 'ManagedReplacement' namedtuple that
      defines information needed to replace the transform with an equivalent
      managed transform during the expansion. If an
      'updateCompatibilityBeamVersion' pipeline option is provided, we will
      only replace if the managed transform is update compatible with the
      provided version.
  :kwargs: field name to value mapping for configuring the schema transform.
      keys map to the field names of the schema of the SchemaTransform
      (in-order).
  """
  def __init__(
      self,
      identifier,
      expansion_service,
      rearrange_based_on_discovery=False,
      classpath=None,
      managed_replacement=None,
      **kwargs):
    self._expansion_service = expansion_service
    self._kwargs = kwargs
    self._classpath = classpath
    if managed_replacement:
      assert isinstance(managed_replacement, ManagedReplacement)
    self._managed_replacement = managed_replacement

    _kwargs = kwargs
    if rearrange_based_on_discovery:
      config = SchemaAwareExternalTransform.discover_config(
          self._expansion_service, identifier)
      self._payload_builder = ExplicitSchemaTransformPayloadBuilder(
          identifier,
          named_tuple_to_schema(config.configuration_schema),
          **_kwargs)

      if self._managed_replacement:
        # We have to do the replacement at the expansion instead of at
        # construction
        # since we don't have access to the PipelineOptions object at the
        # construction.
        underlying_transform_id = (
            self._managed_replacement.underlying_transform_identifier)
        if not (underlying_transform_id
                in MANAGED_TRANSFORM_URN_TO_JAR_TARGET_MAPPING):
          raise ValueError(
              'Could not find an expansion service jar for the managed ' +
              'transform ' + underlying_transform_id)
        managed_expansion_service_jar = (
            MANAGED_TRANSFORM_URN_TO_JAR_TARGET_MAPPING
        )[underlying_transform_id]
        self._managed_expansion_service = BeamJarExpansionService(
            managed_expansion_service_jar)
        managed_config = SchemaAwareExternalTransform.discover_config(
            self._managed_expansion_service,
            MANAGED_SCHEMA_TRANSFORM_IDENTIFIER)

        yaml_config = yaml.dump(kwargs)
        self._managed_payload_builder = (
            ExplicitSchemaTransformPayloadBuilder(
                MANAGED_SCHEMA_TRANSFORM_IDENTIFIER,
                named_tuple_to_schema(managed_config.configuration_schema),
                transform_identifier=underlying_transform_id,
                config=yaml_config))
    else:
      self._payload_builder = SchemaTransformPayloadBuilder(
          identifier, **_kwargs)

  def expand(self, pcolls):
    # Expand the transform using the expansion service.
    payload_builder = self._payload_builder
    expansion_service = self._expansion_service

    if self._managed_replacement:
      compat_version_prior_to_current = (
          pcolls.pipeline._options.is_compat_version_prior_to(
              self._managed_replacement.update_compatibility_version))
      if not compat_version_prior_to_current:
        payload_builder = self._managed_payload_builder
        expansion_service = self._managed_expansion_service

    return pcolls | self._payload_builder.identifier() >> ExternalTransform(
        common_urns.schematransform_based_expand.urn,
        payload_builder,
        expansion_service)

  @classmethod
  @functools.lru_cache
  def discover(cls, expansion_service, ignore_errors=False):
    """Discover all SchemaTransforms available to the given expansion service.

    :return: a list of SchemaTransformsConfigs that represent the discovered
        SchemaTransforms.
    """
    return list(cls.discover_iter(expansion_service, ignore_errors))

  @staticmethod
  def discover_iter(expansion_service, ignore_errors=True):
    with ExternalTransform.service(expansion_service) as service:
      discover_response = service.DiscoverSchemaTransform(
          beam_expansion_api_pb2.DiscoverSchemaTransformRequest())

    for identifier in discover_response.schema_transform_configs:
      proto_config = discover_response.schema_transform_configs[identifier]
      try:
        schema = named_tuple_from_schema(proto_config.config_schema)
      except Exception as exn:
        if ignore_errors:
          truncated_schema = schema_pb2.Schema()
          truncated_schema.CopyFrom(proto_config.config_schema)
          for field in truncated_schema.fields:
            try:
              typing_from_runner_api(field.type)
            except Exception:
              if field.type.nullable:
                # Set it to an empty placeholder type.
                field.type.CopyFrom(
                    schema_pb2.FieldType(
                        nullable=True,
                        row_type=schema_pb2.RowType(
                            schema=schema_pb2.Schema())))
          try:
            schema = named_tuple_from_schema(truncated_schema)
          except Exception as exn:
            logging.info("Bad schema for %s: %s", identifier, str(exn)[:250])
            continue
        else:
          raise

      yield SchemaTransformsConfig(
          identifier=identifier,
          configuration_schema=schema,
          inputs=proto_config.input_pcollection_names,
          outputs=proto_config.output_pcollection_names,
          description=proto_config.description)

  @staticmethod
  def discover_config(expansion_service, name):
    """Discover one SchemaTransform by name in the given expansion service.

    :return: one SchemaTransformsConfig that represents the discovered
        SchemaTransform

    :raises:
      ValueError: if more than one SchemaTransform is discovered, or if none
      are discovered
    """

    schematransforms = SchemaAwareExternalTransform.discover(
        expansion_service, ignore_errors=True)
    matched = []

    for st in schematransforms:
      if name in st.identifier:
        matched.append(st)

    if not matched:
      raise ValueError(
          "Did not discover any SchemaTransforms resembling the name '%s'" %
          name)
    elif len(matched) > 1:
      raise ValueError(
          "Found multiple SchemaTransforms with the name '%s':\n%s\n" %
          (name, [st.identifier for st in matched]))

    return matched[0]


class JavaExternalTransform(ptransform.PTransform):
  """A proxy for Java-implemented external transforms.

  One builds these transforms just as one would in Java, e.g.::

      transform = JavaExternalTransform('fully.qualified.ClassName'
          )(contructorArg, ... ).builderMethod(...)

  or::

      JavaExternalTransform('fully.qualified.ClassName').staticConstructor(
          ...).builderMethod1(...).builderMethod2(...)

  :param class_name: fully qualified name of the java class
  :param expansion_service: (Optional) an expansion service to use.  If none is
      provided, a default expansion service will be started.
  :param classpath: (Optional) A list paths to additional jars to place on the
      expansion service classpath.
  """
  def __init__(self, class_name, expansion_service=None, classpath=None):
    if expansion_service and classpath:
      raise ValueError(
          f'Only one of expansion_service ({expansion_service}) '
          f'or classpath ({classpath}) may be provided.')
    self._payload_builder = JavaClassLookupPayloadBuilder(class_name)
    self._classpath = classpath
    self._expansion_service = expansion_service
    # Beam explicitly looks for following attributes. Hence adding
    # 'None' values here to prevent '__getattr__' from being called.
    self.inputs = None
    self._fn_api_payload = None

  def __call__(self, *args, **kwargs):
    self._payload_builder.with_constructor(*args, **kwargs)
    return self

  def __getattr__(self, name):
    # Don't try to emulate special methods.
    if name.startswith('__') and name.endswith('__'):
      return super().__getattr__(name)
    else:
      return self[name]

  def __getitem__(self, name):
    # Use directly for keywords or attribute conflicts.
    def construct(*args, **kwargs):
      if self._payload_builder._has_constructor():
        builder_method = self._payload_builder.add_builder_method
      else:
        builder_method = self._payload_builder.with_constructor_method
      builder_method(name, *args, **kwargs)
      return self

    return construct

  def expand(self, pcolls):
    if self._expansion_service is None:
      self._expansion_service = BeamJarExpansionService(
          ':sdks:java:expansion-service:app:shadowJar',
          extra_args=['{{PORT}}', '--javaClassLookupAllowlistFile=*'],
          classpath=self._classpath)
    return pcolls | ExternalTransform(
        common_urns.java_class_lookup.urn,
        self._payload_builder,
        self._expansion_service)


class AnnotationBasedPayloadBuilder(SchemaBasedPayloadBuilder):
  """
  Build a payload based on an external transform's type annotations.
  """
  def __init__(self, transform, **values):
    """
    :param transform: a PTransform instance or class. type annotations will
                      be gathered from its __init__ method
    :param values: values to encode
    """
    self._transform = transform
    self._values = values

  def _get_named_tuple_instance(self):
    schema = named_fields_to_schema([
        (k, convert_to_typing_type(v))
        for k, v in self._transform.__init__.__annotations__.items()
        if k in self._values
    ])
    return named_tuple_from_schema(schema)(**self._values)


class DataclassBasedPayloadBuilder(SchemaBasedPayloadBuilder):
  """
  Build a payload based on an external transform that uses dataclasses.
  """
  def __init__(self, transform):
    """
    :param transform: a dataclass-decorated PTransform instance from which to
                      gather type annotations and values
    """
    self._transform = transform

  def _get_named_tuple_instance(self):
    import dataclasses
    schema = named_fields_to_schema([
        (field.name, convert_to_typing_type(field.type))
        for field in dataclasses.fields(self._transform)
    ])
    return named_tuple_from_schema(schema)(
        **dataclasses.asdict(self._transform))


class ExternalTransform(ptransform.PTransform):
  """
    External provides a cross-language transform via expansion services in
    foreign SDKs.
  """
  _namespace_counter = 0

  # Variable name _namespace conflicts with DisplayData._namespace so we use
  # name _external_namespace here.
  _external_namespace = threading.local()

  _IMPULSE_PREFIX = 'impulse'

  def __init__(self, urn, payload, expansion_service=None):
    """Wrapper for an external transform with the given urn and payload.

    :param urn: the unique beam identifier for this transform
    :param payload: the payload, either as a byte string or a PayloadBuilder
    :param expansion_service: an expansion service implementing the beam
        ExpansionService protocol, either as an object with an Expand method
        or an address (as a str) to a grpc server that provides this method.
    """
    expansion_service = expansion_service or DEFAULT_EXPANSION_SERVICE
    if not urn and isinstance(payload, JavaClassLookupPayloadBuilder):
      urn = common_urns.java_class_lookup.urn
    self._urn = urn
    self._payload = (
        payload.payload() if isinstance(payload, PayloadBuilder) else payload)
    self._expansion_service = expansion_service
    self._external_namespace = self._fresh_namespace()
    self._inputs: dict[str, pvalue.PCollection] = {}
    self._outputs: dict[str, pvalue.PCollection] = {}

  def with_output_types(self, *args, **kwargs):
    return WithTypeHints.with_output_types(self, *args, **kwargs)

  def replace_named_inputs(self, named_inputs):
    self._inputs = named_inputs

  def replace_named_outputs(self, named_outputs):
    self._outputs = named_outputs

  def __post_init__(self, expansion_service):
    """
    This will only be invoked if ExternalTransform is used as a base class
    for a class decorated with dataclasses.dataclass
    """
    ExternalTransform.__init__(
        self, self.URN, DataclassBasedPayloadBuilder(self), expansion_service)

  def default_label(self):
    return '%s(%s)' % (self.__class__.__name__, self._urn)

  @classmethod
  def get_local_namespace(cls):
    return getattr(cls._external_namespace, 'value', 'external')

  @classmethod
  @contextlib.contextmanager
  def outer_namespace(cls, namespace):
    prev = cls.get_local_namespace()
    try:
      cls._external_namespace.value = namespace
      yield
    finally:
      cls._external_namespace.value = prev

  @classmethod
  def _fresh_namespace(cls) -> str:
    ExternalTransform._namespace_counter += 1
    return '%s_%d' % (cls.get_local_namespace(), cls._namespace_counter)

  def expand(self, pvalueish: pvalue.PCollection) -> pvalue.PCollection:
    if isinstance(pvalueish, pvalue.PBegin):
      self._inputs = {}
    elif isinstance(pvalueish, (list, tuple)):
      self._inputs = {str(ix): pvalue for ix, pvalue in enumerate(pvalueish)}
    elif isinstance(pvalueish, dict):
      self._inputs = pvalueish
    else:
      self._inputs = {'input': pvalueish}
    pipeline = (
        next(iter(self._inputs.values())).pipeline
        if self._inputs else pvalueish.pipeline)
    context = pipeline_context.PipelineContext(
        component_id_map=pipeline.component_id_map)
    transform_proto = beam_runner_api_pb2.PTransform(
        unique_name=pipeline._current_transform().full_label,
        spec=beam_runner_api_pb2.FunctionSpec(
            urn=self._urn, payload=self._payload))
    for tag, pcoll in self._inputs.items():
      transform_proto.inputs[tag] = context.pcollections.get_id(pcoll)
      # Conversion to/from proto assumes producers.
      # TODO: Possibly loosen this.
      context.transforms.put_proto(
          '%s_%s' % (self._IMPULSE_PREFIX, tag),
          beam_runner_api_pb2.PTransform(
              unique_name='%s_%s' % (self._IMPULSE_PREFIX, tag),
              spec=beam_runner_api_pb2.FunctionSpec(
                  urn=common_urns.primitives.IMPULSE.urn),
              outputs={'out': transform_proto.inputs[tag]}))

    # Retrieve type hints and store them in variables
    # to avoid duplicate calls and AttributeError
    hints = self.get_type_hints()
    output_coders = None
    if hints.output_types:
      if hints.output_types[0]:
        output_coders = dict((str(k), context.coder_id_from_element_type(v))
                             for (k, v) in enumerate(hints.output_types[0]))
      elif hints.output_types[1]:
        output_coders = {
            k: context.coder_id_from_element_type(v)
            for (k, v) in hints.output_types[1].items()
        }
    components = context.to_runner_api()
    request = beam_expansion_api_pb2.ExpansionRequest(
        components=components,
        namespace=self._external_namespace,  # type: ignore[arg-type]
        transform=transform_proto,
        output_coder_requests=output_coders,
        pipeline_options=pipeline._options.to_runner_api())

    expansion_service = _maybe_use_transform_service(
        self._expansion_service, pipeline.options)

    with ExternalTransform.service(expansion_service) as service:
      response = service.Expand(request)
      if response.error:
        raise RuntimeError(_sanitize_java_traceback(response.error))
      self._expanded_components = response.components
      if any(e.dependencies
             for env in self._expanded_components.environments.values()
             for e in environments.expand_anyof_environments(env)):
        self._expanded_components = self._resolve_artifacts(
            self._expanded_components,
            service.artifact_service(),
            pipeline.local_tempdir)

    self._expanded_transform = response.transform
    self._expanded_requirements = response.requirements
    result_context = pipeline_context.PipelineContext(response.components)

    def fix_output(pcoll, tag):
      pcoll.pipeline = pipeline
      pcoll.tag = tag
      return pcoll

    self._outputs = {
        tag: fix_output(result_context.pcollections.get_by_id(pcoll_id), tag)
        for tag, pcoll_id in self._expanded_transform.outputs.items()
    }

    return self._output_to_pvalueish(self._outputs)

  @staticmethod
  @contextlib.contextmanager
  def service(expansion_service):
    if isinstance(expansion_service, str):
      channel_options = [("grpc.max_receive_message_length", -1),
                         ("grpc.max_send_message_length", -1)]
      if hasattr(grpc, 'local_channel_credentials'):
        # Some environments may not support insecure channels. Hence use a
        # secure channel with local credentials here.
        # TODO: update this to support secure non-local channels.
        channel_factory_fn = functools.partial(
            grpc.secure_channel,
            expansion_service,
            grpc.local_channel_credentials(),
            options=channel_options)
      else:
        # local_channel_credentials is an experimental API which is unsupported
        # by older versions of grpc which may be pulled in due to other project
        # dependencies.
        channel_factory_fn = functools.partial(
            grpc.insecure_channel, expansion_service, options=channel_options)
      with channel_factory_fn() as channel:
        yield ExpansionAndArtifactRetrievalStub(channel)
    elif hasattr(expansion_service, 'Expand'):
      yield expansion_service
    else:
      with expansion_service as stub:
        yield stub

  def _resolve_artifacts(self, components, service, dest):
    def _resolve_artifacts_for(env):
      if env.urn == common_urns.environments.ANYOF.urn:
        env.CopyFrom(
            environments.AnyOfEnvironment.create_proto([
                _resolve_artifacts_for(e)
                for e in environments.expand_anyof_environments(env)
            ]))
      elif env.dependencies:
        resolved = list(
            artifact_service.resolve_artifacts(env.dependencies, service, dest))
        del env.dependencies[:]
        env.dependencies.extend(resolved)
      return env

    for env in components.environments.values():
      _resolve_artifacts_for(env)
    return components

  def _output_to_pvalueish(self, output_dict):
    if len(output_dict) == 1:
      return next(iter(output_dict.values()))
    else:
      return output_dict

  def to_runner_api_transform(self, context, full_label):
    pcoll_renames = {}
    renamed_tag_seen = False
    for tag, pcoll in self._inputs.items():
      if tag not in self._expanded_transform.inputs:
        if renamed_tag_seen:
          raise RuntimeError(
              'Ambiguity due to non-preserved tags: %s vs %s' % (
                  sorted(self._expanded_transform.inputs.keys()),
                  sorted(self._inputs.keys())))
        else:
          renamed_tag_seen = True
          tag, = self._expanded_transform.inputs.keys()
      pcoll_renames[self._expanded_transform.inputs[tag]] = (
          context.pcollections.get_id(pcoll))
    for tag, pcoll in self._outputs.items():
      pcoll_renames[self._expanded_transform.outputs[tag]] = (
          context.pcollections.get_id(pcoll))

    def _equivalent(coder1, coder2):
      return coder1 == coder2 or _normalize(coder1) == _normalize(coder2)

    def _normalize(coder_proto):
      normalized = copy.copy(coder_proto)
      normalized.spec.environment_id = ''
      # TODO(robertwb): Normalize components as well.
      return normalized

    for id, proto in self._expanded_components.coders.items():
      if id.startswith(self._external_namespace):
        context.coders.put_proto(id, proto)
      elif id in context.coders:
        if not _equivalent(context.coders._id_to_proto[id], proto):
          raise RuntimeError(
              'Re-used coder id: %s\n%s\n%s' %
              (id, context.coders._id_to_proto[id], proto))
      else:
        context.coders.put_proto(id, proto)
    for id, proto in self._expanded_components.windowing_strategies.items():
      if id.startswith(self._external_namespace):
        context.windowing_strategies.put_proto(id, proto)
    for id, proto in self._expanded_components.environments.items():
      if id.startswith(self._external_namespace):
        context.environments.put_proto(id, proto)
    for id, proto in self._expanded_components.pcollections.items():
      id = pcoll_renames.get(id, id)
      if id not in context.pcollections._id_to_obj.keys():
        context.pcollections.put_proto(id, proto)

    for id, proto in self._expanded_components.transforms.items():
      if id.startswith(self._IMPULSE_PREFIX):
        # Our fake inputs.
        continue
      assert id.startswith(
          self._external_namespace), (id, self._external_namespace)
      new_proto = beam_runner_api_pb2.PTransform(
          unique_name=proto.unique_name,
          # If URN is not set this is an empty spec.
          spec=proto.spec if proto.spec.urn else None,
          subtransforms=proto.subtransforms,
          inputs={
              tag: pcoll_renames.get(pcoll, pcoll)
              for tag, pcoll in proto.inputs.items()
          },
          outputs={
              tag: pcoll_renames.get(pcoll, pcoll)
              for tag, pcoll in proto.outputs.items()
          },
          display_data=proto.display_data,
          environment_id=proto.environment_id)
      context.transforms.put_proto(id, new_proto)

    for requirement in self._expanded_requirements:
      context.add_requirement(requirement)

    return beam_runner_api_pb2.PTransform(
        unique_name=full_label,
        spec=self._expanded_transform.spec,
        subtransforms=self._expanded_transform.subtransforms,
        inputs={
            tag: pcoll_renames.get(pcoll, pcoll)
            for tag, pcoll in self._expanded_transform.inputs.items()
        },
        outputs={
            tag: pcoll_renames.get(pcoll, pcoll)
            for tag, pcoll in self._expanded_transform.outputs.items()
        },
        annotations=self._expanded_transform.annotations,
        environment_id=self._expanded_transform.environment_id)


class ExpansionAndArtifactRetrievalStub(
    beam_expansion_api_pb2_grpc.ExpansionServiceStub):
  def __init__(self, channel, **kwargs):
    self._channel = channel
    self._kwargs = kwargs
    super().__init__(channel, **kwargs)

  def artifact_service(self):
    return beam_artifact_api_pb2_grpc.ArtifactRetrievalServiceStub(
        self._channel, **self._kwargs)

  def ready(self, timeout_sec):
    grpc.channel_ready_future(self._channel).result(timeout=timeout_sec)


class JavaJarExpansionService(object):
  """An expansion service based on an Java Jar file.

  This can be passed into an ExternalTransform as the expansion_service
  argument which will spawn a subprocess using this jar to expand the
  transform.

  Args:
    path_to_jar: the path to a locally available executable jar file to be used
      to start up the expansion service.
    extra_args: arguments to be provided when starting up the
      expansion service using the jar file. These arguments will replace the
      default arguments.
    classpath: Additional dependencies to be added to the classpath.
    append_args: arguments to be provided when starting up the
      expansion service using the jar file. These arguments will be appended to
      the default arguments.
    user_agent: HTTP user agent string used when downloading jars via
      `JavaJarServer.local_jar`, including the main jar and any classpath
      dependencies.
    maven_repository_url: Maven repository base URL to resolve artifacts when
      classpath entries or jars are specified as Maven coordinates
      (`group:artifact:version`). Defaults to Maven Central if not provided.
  """
  def __init__(
      self,
      path_to_jar,
      extra_args=None,
      classpath=None,
      append_args=None,
      user_agent=None,
      maven_repository_url=None):
    if extra_args and append_args:
      raise ValueError('Only one of extra_args or append_args may be provided')
    self.path_to_jar = path_to_jar
    self._extra_args = extra_args
    self._classpath = classpath or []
    self._service_count = 0
    self._append_args = append_args or []
    self._user_agent = user_agent
    self._maven_repository_url = maven_repository_url

  def is_existing_service(self):
    return subprocess_server.is_service_endpoint(self.path_to_jar)

  @staticmethod
  def _expand_jars(jar, user_agent=None, maven_repository_url=None):
    if glob.glob(jar):
      return glob.glob(jar)
    elif isinstance(jar, str) and (jar.startswith('http://') or
                                   jar.startswith('https://')):
      return [subprocess_server.JavaJarServer.local_jar(jar)]
    else:
      # If the input JAR is not a local glob, nor an http/https URL, then
      # we assume that it's a gradle-style Java artifact in Maven Central,
      # in the form group:artifact:version, so we attempt to parse that way.
      try:
        group_id, artifact_id, version = jar.split(':')
      except ValueError:
        # If we are not able to find a JAR, nor a JAR artifact, nor a URL for
        # a JAR path, we still choose to include it in the path.
        logging.warning('Unable to parse %s into group:artifact:version.', jar)
        return [jar]
      path = subprocess_server.JavaJarServer.local_jar(
          subprocess_server.JavaJarServer.path_to_maven_jar(
              artifact_id,
              group_id,
              version,
              repository=(
                  maven_repository_url or
                  subprocess_server.JavaJarServer.MAVEN_CENTRAL_REPOSITORY)),
          user_agent=user_agent)
      return [path]

  def _default_args(self):
    """Default arguments to be used by `JavaJarExpansionService`."""

    to_stage = ','.join([self.path_to_jar] + sum((
        JavaJarExpansionService._expand_jars(
            jar, self._user_agent, self._maven_repository_url)
        for jar in self._classpath or []), []))
    args = ['{{PORT}}', f'--filesToStage={to_stage}']
    # TODO(robertwb): See if it's possible to scope this per pipeline.
    # Checks to see if the cache is being used for this server.
    if subprocess_server.SubprocessServer._cache._live_owners:
      args.append('--alsoStartLoopbackWorker')
    return args

  def with_user_agent(self, user_agent: str):
    self._user_agent = user_agent
    return self

  def __enter__(self):
    if self._service_count == 0:
      self.path_to_jar = subprocess_server.JavaJarServer.local_jar(
          self.path_to_jar, user_agent=self._user_agent)
      if self._extra_args is None:
        self._extra_args = self._default_args() + self._append_args
      # Consider memoizing these servers (with some timeout).
      logging.info(
          'Starting a JAR-based expansion service from JAR %s ' + (
              'and with classpath: %s' %
              self._classpath if self._classpath else ''),
          self.path_to_jar)
      classpath_urls = [
          subprocess_server.JavaJarServer.local_jar(path)
          for jar in self._classpath
          for path in JavaJarExpansionService._expand_jars(
              jar, user_agent=self._user_agent,
              maven_repository_url=self._maven_repository_url)
      ]
      self._service_provider = subprocess_server.JavaJarServer(
          ExpansionAndArtifactRetrievalStub,
          self.path_to_jar,
          self._extra_args,
          classpath=classpath_urls,
          logger="ExpansionService")
      self._service = self._service_provider.__enter__()
    self._service_count += 1
    return self._service

  def __exit__(self, *args):
    self._service_count -= 1
    if self._service_count == 0:
      self._service_provider.__exit__(*args)


class BeamJarExpansionService(JavaJarExpansionService):
  """An expansion service based on an Beam Java Jar file.

  Attempts to use a locally-built copy of the jar based on the gradle target,
  if it exists, otherwise attempts to download and cache the released artifact
  corresponding to this version of Beam from the apache maven repository.

  Args:
    gradle_target: Beam Gradle target for building an executable jar which will
      be used to start the expansion service.
    extra_args: arguments to be provided when starting up the
      expansion service using the jar file. These arguments will replace the
      default arguments.
    gradle_appendix: Gradle appendix of the artifact.
    classpath: Additional dependencies to be added to the classpath.
    append_args: arguments to be provided when starting up the
      expansion service using the jar file. These arguments will be appended to
      the default arguments.
    user_agent: HTTP user agent string used when downloading the Beam jar and
      any classpath dependencies.
    maven_repository_url: Maven repository base URL to resolve the Beam jar
      for the provided Gradle target. Defaults to Maven Central if not
      provided.
  """
  def __init__(
      self,
      gradle_target,
      extra_args=None,
      gradle_appendix=None,
      classpath=None,
      append_args=None,
      user_agent=None,
      maven_repository_url=None):
    path_to_jar = subprocess_server.JavaJarServer.path_to_beam_jar(
        gradle_target,
        gradle_appendix,
        maven_repository_url=maven_repository_url)
    self.gradle_target = gradle_target
    super().__init__(
        path_to_jar,
        extra_args,
        classpath=classpath,
        append_args=append_args,
        user_agent=user_agent,
        maven_repository_url=maven_repository_url)


def _maybe_use_transform_service(provided_service=None, options=None):
  # For anything other than 'JavaJarExpansionService' we just use the
  # provided service. For example, string address of an already available
  # service.
  if not isinstance(provided_service, JavaJarExpansionService):
    return provided_service

  if provided_service.is_existing_service():
    # This is an existing service supported through the 'beam_services'
    # PipelineOption.
    return provided_service

  def is_java_available():
    java = subprocess_server.JavaHelper.get_java()
    cmd = [java, '--version']

    try:
      subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
    except:  # pylint: disable=bare-except
      return False

    return True

  def is_docker_available():
    cmd = ['docker', '--version']

    try:
      subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
    except:  # pylint: disable=bare-except
      return False

    return True

  # We try java and docker based expansion services in that order.

  java_available = is_java_available()
  docker_available = is_docker_available()

  use_transform_service = options.view_as(
      pipeline_options.CrossLanguageOptions).use_transform_service
  user_agent = options.view_as(pipeline_options.SetupOptions).user_agent

  if (java_available and provided_service and not use_transform_service):
    return provided_service.with_user_agent(user_agent)
  elif docker_available:
    if use_transform_service:
      error_append = 'it was explicitly requested'
    elif not java_available:
      error_append = 'the Java executable is not available in the system'
    else:
      error_append = 'a Java expansion service was not provided.'

    project_name = str(uuid.uuid4())
    port = subprocess_server.pick_port(None)[0]

    logging.info(
        'Trying to expand the external transform using the Docker Compose '
        'based transform service since %s. Transform service will be under '
        'Docker Compose project name %s and will be made available at port %r.'
        % (error_append, project_name, str(port)))

    from apache_beam import version as beam_version
    beam_version = beam_version.__version__

    return transform_service_launcher.TransformServiceLauncher(
        project_name, port, beam_version, user_agent)
  else:
    raise ValueError(
        'Cannot start an expansion service since neither Java nor '
        'Docker executables are available in the system.')


def _sanitize_java_traceback(s):
  """Attempts to highlight the root cause in the error string.

  Java tracebacks read bottom to top, while Python tracebacks read top to
  bottom, resulting in the actual error message getting sandwiched between two
  walls of text.  This may result in the error being duplicated (as we don't
  want to remove relevant information) but should be clearer in most cases.

  Best-effort but non-destructive.
  """
  # We delete non-java-traceback lines.
  traceback_lines = [
      r'\tat \S+\(\S+\.java:\d+\)',
      r'\t\.\.\. \d+ more',
      # A bit more restrictive to avoid accidentally capturing non-java lines.
      r'Caused by: [a-z]+(\.\S+)?\.[A-Z][A-Za-z0-9_$]+(Error|Exception):[^\n]*'
  ]
  without_java_traceback = s + '\n'
  for p in traceback_lines:
    without_java_traceback = re.sub(
        fr'\n{p}$', '', without_java_traceback, flags=re.M)
  # If what's left is substantially smaller, duplicate it at the end for better
  # visibility.
  if len(without_java_traceback) < len(s) / 2:
    return s + '\n\n' + without_java_traceback.strip()
  else:
    return s


def memoize(func):
  cache = {}

  def wrapper(*args):
    if args not in cache:
      cache[args] = func(*args)
    return cache[args]

  return wrapper