coder_impl.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.
#

"""Coder implementations.

The actual encode/decode implementations are split off from coders to
allow conditional (compiled/pure) implementations, which can be used to
encode many elements with minimal overhead.

This module may be optionally compiled with Cython, using the corresponding
coder_impl.pxd file for type hints.  In particular, because CoderImpls are
never pickled and sent across the wire (unlike Coders themselves) the workers
can use compiled Impls even if the main program does not (or vice versa).

For internal use only; no backwards-compatibility guarantees.
"""
# pytype: skip-file

import dataclasses
import decimal
import enum
import itertools
import json
import logging
import pickle
from io import BytesIO
from typing import TYPE_CHECKING
from typing import Any
from typing import Callable
from typing import Dict
from typing import Iterable
from typing import Iterator
from typing import List
from typing import Optional
from typing import Sequence
from typing import Set
from typing import Tuple
from typing import Type

import numpy as np
from fastavro import parse_schema
from fastavro import schemaless_reader
from fastavro import schemaless_writer

from apache_beam.coders import observable
from apache_beam.coders.avro_record import AvroRecord
from apache_beam.internal import cloudpickle_pickler
from apache_beam.internal.cloudpickle import cloudpickle
from apache_beam.typehints.schemas import named_tuple_from_schema
from apache_beam.utils import proto_utils
from apache_beam.utils import windowed_value
from apache_beam.utils.sharded_key import ShardedKey
from apache_beam.utils.timestamp import MAX_TIMESTAMP
from apache_beam.utils.timestamp import MIN_TIMESTAMP
from apache_beam.utils.timestamp import Timestamp

try:
  import dill
except ImportError:
  dill = None

if TYPE_CHECKING:
  import proto

  from apache_beam.transforms import userstate
  from apache_beam.transforms.window import IntervalWindow

try:
  from . import stream  # pylint: disable=unused-import
except ImportError:
  SLOW_STREAM = True
else:
  SLOW_STREAM = False

is_compiled = False
fits_in_64_bits = lambda x: -(1 << 63) <= x <= (1 << 63) - 1

if TYPE_CHECKING or SLOW_STREAM:
  from .slow_stream import ByteCountingOutputStream
  from .slow_stream import InputStream as create_InputStream
  from .slow_stream import OutputStream as create_OutputStream
  from .slow_stream import get_varint_size

  try:
    import cython
    is_compiled = cython.compiled
  except ImportError:
    globals()['cython'] = type('fake_cython', (), {'cast': lambda typ, x: x})

else:
  # pylint: disable=wrong-import-order, wrong-import-position, ungrouped-imports
  from .stream import ByteCountingOutputStream
  from .stream import InputStream as create_InputStream
  from .stream import OutputStream as create_OutputStream
  from .stream import get_varint_size

  # Make it possible to import create_InputStream and other cdef-classes
  # from apache_beam.coders.coder_impl when Cython codepath is used.
  globals()['create_InputStream'] = create_InputStream
  globals()['create_OutputStream'] = create_OutputStream
  globals()['ByteCountingOutputStream'] = ByteCountingOutputStream
  # pylint: enable=wrong-import-order, wrong-import-position, ungrouped-imports
  is_compiled = True

_LOGGER = logging.getLogger(__name__)

_TIME_SHIFT = 1 << 63
MIN_TIMESTAMP_micros = MIN_TIMESTAMP.micros
MAX_TIMESTAMP_micros = MAX_TIMESTAMP.micros

IterableStateReader = Callable[[bytes, 'CoderImpl'], Iterable]
IterableStateWriter = Callable[[Iterable, 'CoderImpl'], bytes]
Observables = List[Tuple[observable.ObservableMixin, 'CoderImpl']]


class CoderImpl(object):
  """For internal use only; no backwards-compatibility guarantees."""
  def encode_to_stream(self, value, stream, nested):
    # type: (Any, create_OutputStream, bool) -> None

    """Reads object from potentially-nested encoding in stream."""
    raise NotImplementedError

  def decode_from_stream(self, stream, nested):
    # type: (create_InputStream, bool) -> Any

    """Reads object from potentially-nested encoding in stream."""
    raise NotImplementedError

  def encode(self, value):
    # type: (Any) -> bytes

    """Encodes an object to an unnested string."""
    raise NotImplementedError

  def decode(self, encoded):
    # type: (bytes) -> Any

    """Decodes an object to an unnested string."""
    raise NotImplementedError

  def encode_all(self, values):
    # type: (Iterable[Any]) -> bytes
    out = create_OutputStream()
    for value in values:
      self.encode_to_stream(value, out, True)
    return out.get()

  def decode_all(self, encoded):
    # type: (bytes) -> Iterator[Any]
    input_stream = create_InputStream(encoded)
    while input_stream.size() > 0:
      yield self.decode_from_stream(input_stream, True)

  def encode_nested(self, value):
    # type: (Any) -> bytes
    out = create_OutputStream()
    self.encode_to_stream(value, out, True)
    return out.get()

  def decode_nested(self, encoded):
    # type: (bytes) -> Any
    return self.decode_from_stream(create_InputStream(encoded), True)

  def estimate_size(self, value, nested=False):
    # type: (Any, bool) -> int

    """Estimates the encoded size of the given value, in bytes."""
    out = ByteCountingOutputStream()
    self.encode_to_stream(value, out, nested)
    return out.get_count()

  def _get_nested_size(self, inner_size, nested):
    if not nested:
      return inner_size
    varint_size = get_varint_size(inner_size)
    return varint_size + inner_size

  def get_estimated_size_and_observables(self, value, nested=False):
    # type: (Any, bool) -> Tuple[int, Observables]

    """Returns estimated size of value along with any nested observables.

    The list of nested observables is returned as a list of 2-tuples of
    (obj, coder_impl), where obj is an instance of observable.ObservableMixin,
    and coder_impl is the CoderImpl that can be used to encode elements sent by
    obj to its observers.

    Arguments:
      value: the value whose encoded size is to be estimated.
      nested: whether the value is nested.

    Returns:
      The estimated encoded size of the given value and a list of observables
      whose elements are 2-tuples of (obj, coder_impl) as described above.
    """
    return self.estimate_size(value, nested), []


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

  Subclass of CoderImpl implementing stream methods using encode/decode."""
  def encode_to_stream(self, value, stream, nested):
    # type: (Any, create_OutputStream, bool) -> None

    """Reads object from potentially-nested encoding in stream."""
    stream.write(self.encode(value), nested)

  def decode_from_stream(self, stream, nested):
    # type: (create_InputStream, bool) -> Any

    """Reads object from potentially-nested encoding in stream."""
    return self.decode(stream.read_all(nested))


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

  Subclass of CoderImpl implementing encode/decode using stream methods."""
  def encode(self, value):
    # type: (Any) -> bytes
    out = create_OutputStream()
    self.encode_to_stream(value, out, False)
    return out.get()

  def decode(self, encoded):
    # type: (bytes) -> Any
    return self.decode_from_stream(create_InputStream(encoded), False)

  def estimate_size(self, value, nested=False):
    # type: (Any, bool) -> int

    """Estimates the encoded size of the given value, in bytes."""
    out = ByteCountingOutputStream()
    self.encode_to_stream(value, out, nested)
    return out.get_count()


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

  A CoderImpl that calls back to the _impl methods on the Coder itself.

  This is the default implementation used if Coder._get_impl()
  is not overwritten.
  """
  def __init__(self, encoder, decoder, size_estimator=None):
    self._encoder = encoder
    self._decoder = decoder
    self._size_estimator = size_estimator or self._default_size_estimator

  def _default_size_estimator(self, value):
    return len(self.encode(value))

  def encode_to_stream(self, value, stream, nested):
    # type: (Any, create_OutputStream, bool) -> None
    return stream.write(self._encoder(value), nested)

  def decode_from_stream(self, stream, nested):
    # type: (create_InputStream, bool) -> Any
    return self._decoder(stream.read_all(nested))

  def encode(self, value):
    return self._encoder(value)

  def decode(self, encoded):
    return self._decoder(encoded)

  def estimate_size(self, value, nested=False):
    # type: (Any, bool) -> int
    return self._get_nested_size(self._size_estimator(value), nested)

  def get_estimated_size_and_observables(self, value, nested=False):
    # type: (Any, bool) -> Tuple[int, Observables]
    # TODO(robertwb): Remove this once all coders are correct.
    if isinstance(value, observable.ObservableMixin):
      # CallbackCoderImpl can presumably encode the elements too.
      return 1, [(value, self)]

    return self.estimate_size(value, nested), []

  def __repr__(self):
    return 'CallbackCoderImpl[encoder=%s, decoder=%s]' % (
        self._encoder, self._decoder)


class ProtoCoderImpl(SimpleCoderImpl):
  """For internal use only; no backwards-compatibility guarantees."""
  def __init__(self, proto_message_type):
    self.proto_message_type = proto_message_type

  def encode(self, value):
    return value.SerializePartialToString()

  def decode(self, encoded):
    proto_message = self.proto_message_type()
    proto_message.ParseFromString(encoded)  # This is in effect "ParsePartial".
    return proto_message


class DeterministicProtoCoderImpl(ProtoCoderImpl):
  """For internal use only; no backwards-compatibility guarantees."""
  def encode(self, value):
    return value.SerializePartialToString(deterministic=True)


class ProtoPlusCoderImpl(SimpleCoderImpl):
  """For internal use only; no backwards-compatibility guarantees."""
  def __init__(self, proto_plus_type):
    # type: (Type[proto.Message]) -> None
    self.proto_plus_type = proto_plus_type

  def encode(self, value):
    return value._pb.SerializePartialToString(deterministic=True)

  def decode(self, value):
    return self.proto_plus_type.deserialize(value)


UNKNOWN_TYPE = 0xFF
NONE_TYPE = 0
INT_TYPE = 1
FLOAT_TYPE = 2
BYTES_TYPE = 3
UNICODE_TYPE = 4
BOOL_TYPE = 9
LIST_TYPE = 5
TUPLE_TYPE = 6
DICT_TYPE = 7
SET_TYPE = 8
ITERABLE_LIKE_TYPE = 10

PROTO_TYPE = 100
DATACLASS_TYPE = 101
NAMED_TUPLE_TYPE = 102
ENUM_TYPE = 103
NESTED_STATE_TYPE = 104
DATACLASS_KW_ONLY_TYPE = 105

# Types that can be encoded as iterables, but are not literally
# lists, etc. due to being lazy.  The actual type is not preserved
# through encoding, only the elements. This is particularly useful
# for the value list types created in GroupByKey.
_ITERABLE_LIKE_TYPES = set()  # type: Set[Type]


def _verify_dill_compat():
  base_error = (
      "This pipeline runs with the pipeline option "
      "--update_compatibility_version=2.67.0 or earlier. "
      "When running with this option on SDKs 2.68.0 or "
      "later, you must ensure dill==0.3.1.1 is installed.")
  if not dill:
    raise RuntimeError(base_error + ". Dill is not installed.")
  if dill.__version__ != "0.3.1.1":
    raise RuntimeError(base_error + f". Found dill version '{dill.__version__}")


class FastPrimitivesCoderImpl(StreamCoderImpl):
  """For internal use only; no backwards-compatibility guarantees."""
  def __init__(
      self,
      fallback_coder_impl,
      requires_deterministic_step_label=None,
      force_use_dill=False,
      use_relative_filepaths=True):
    self.fallback_coder_impl = fallback_coder_impl
    self.iterable_coder_impl = IterableCoderImpl(self)
    self.requires_deterministic_step_label = requires_deterministic_step_label
    self.warn_deterministic_fallback = True
    self.force_use_dill = force_use_dill
    self.use_relative_filepaths = use_relative_filepaths

  @staticmethod
  def register_iterable_like_type(t):
    _ITERABLE_LIKE_TYPES.add(t)

  def get_estimated_size_and_observables(self, value, nested=False):
    # type: (Any, bool) -> Tuple[int, Observables]
    if isinstance(value, observable.ObservableMixin):
      # FastPrimitivesCoderImpl can presumably encode the elements too.
      return 1, [(value, self)]

    out = ByteCountingOutputStream()
    self.encode_to_stream(value, out, nested)
    return out.get_count(), []

  def encode_to_stream(self, value, stream, nested):
    # type: (Any, create_OutputStream, bool) -> None
    t = type(value)
    if value is None:
      stream.write_byte(NONE_TYPE)
    elif t is int:
      # In Python 3, an int may be larger than 64 bits.
      # We need to check whether value fits into a 64 bit integer before
      # writing the marker byte.
      try:
        # In Cython-compiled code this will throw an overflow error
        # when value does not fit into int64.
        int_value = value
        # If Cython is not used, we must do a (slower) check ourselves.
        if not TYPE_CHECKING and not is_compiled:
          if not fits_in_64_bits(value):
            raise OverflowError()
        stream.write_byte(INT_TYPE)
        stream.write_var_int64(int_value)
      except OverflowError:
        stream.write_byte(UNKNOWN_TYPE)
        self.fallback_coder_impl.encode_to_stream(value, stream, nested)
    elif t is float:
      stream.write_byte(FLOAT_TYPE)
      stream.write_bigendian_double(value)
    elif t is bytes:
      stream.write_byte(BYTES_TYPE)
      stream.write(value, nested)
    elif t is str:
      unicode_value = value  # for typing
      stream.write_byte(UNICODE_TYPE)
      stream.write(unicode_value.encode('utf-8'), nested)
    elif t is list or t is tuple:
      stream.write_byte(LIST_TYPE if t is list else TUPLE_TYPE)
      stream.write_var_int64(len(value))
      for e in value:
        self.encode_to_stream(e, stream, True)
    elif t is bool:
      stream.write_byte(BOOL_TYPE)
      stream.write_byte(value)
    elif t in _ITERABLE_LIKE_TYPES:
      stream.write_byte(ITERABLE_LIKE_TYPE)
      self.iterable_coder_impl.encode_to_stream(value, stream, nested)
    elif t is dict:
      dict_value = value  # for typing
      stream.write_byte(DICT_TYPE)
      stream.write_var_int64(len(dict_value))
      if self.requires_deterministic_step_label is not None:
        try:
          ordered_kvs = sorted(dict_value.items())
        except Exception as exn:
          raise TypeError(
              "Unable to deterministically order keys of dict for '%s'" %
              self.requires_deterministic_step_label) from exn
        for k, v in ordered_kvs:
          self.encode_to_stream(k, stream, True)
          self.encode_to_stream(v, stream, True)
      else:
        # Loop over dict.items() is optimized by Cython.
        for k, v in dict_value.items():
          self.encode_to_stream(k, stream, True)
          self.encode_to_stream(v, stream, True)
    elif t is set:
      stream.write_byte(SET_TYPE)
      stream.write_var_int64(len(value))
      if self.requires_deterministic_step_label is not None:
        try:
          value = sorted(value)
        except Exception as exn:
          raise TypeError(
              "Unable to deterministically order element of set for '%s'" %
              self.requires_deterministic_step_label) from exn
      for e in value:
        self.encode_to_stream(e, stream, True)
    # All possibly deterministic encodings should be above this clause,
    # all non-deterministic ones below.
    elif self.requires_deterministic_step_label is not None:
      self.encode_special_deterministic(value, stream)
    else:
      stream.write_byte(UNKNOWN_TYPE)
      self.fallback_coder_impl.encode_to_stream(value, stream, nested)

  def encode_special_deterministic(self, value, stream):
    if self.warn_deterministic_fallback:
      _LOGGER.warning(
          "Using fallback deterministic coder for type '%s' in '%s'. ",
          type(value),
          self.requires_deterministic_step_label)
      self.warn_deterministic_fallback = False
    if isinstance(value, proto_utils.message_types):
      stream.write_byte(PROTO_TYPE)
      self.encode_type(type(value), stream)
      stream.write(value.SerializePartialToString(deterministic=True), True)
    elif dataclasses.is_dataclass(value):
      if not type(value).__dataclass_params__.frozen:
        raise TypeError(
            "Unable to deterministically encode non-frozen '%s' of type '%s' "
            "for the input of '%s'" %
            (value, type(value), self.requires_deterministic_step_label))
      init_fields = [field for field in dataclasses.fields(value) if field.init]
      try:
        if any(field.kw_only for field in init_fields):
          stream.write_byte(DATACLASS_KW_ONLY_TYPE)
          self.encode_type(type(value), stream)
          stream.write_var_int64(len(init_fields))
          for field in init_fields:
            stream.write(field.name.encode("utf-8"), True)
            self.encode_to_stream(getattr(value, field.name), stream, True)
        else:  # Not using kw_only, we can pass parameters by position.
          stream.write_byte(DATACLASS_TYPE)
          self.encode_type(type(value), stream)
          values = [getattr(value, field.name) for field in init_fields]
          self.iterable_coder_impl.encode_to_stream(values, stream, True)
      except Exception as e:
        raise TypeError(self._deterministic_encoding_error_msg(value)) from e
    elif isinstance(value, tuple) and hasattr(type(value), '_fields'):
      stream.write_byte(NAMED_TUPLE_TYPE)
      self.encode_type(type(value), stream)
      try:
        self.iterable_coder_impl.encode_to_stream(value, stream, True)
      except Exception as e:
        raise TypeError(self._deterministic_encoding_error_msg(value)) from e
    elif isinstance(value, enum.Enum):
      stream.write_byte(ENUM_TYPE)
      self.encode_type(type(value), stream)
      # Enum values can be of any type.
      try:
        self.encode_to_stream(value.value, stream, True)
      except Exception as e:
        raise TypeError(self._deterministic_encoding_error_msg(value)) from e
    elif (hasattr(value, "__getstate__") and
          # https://github.com/apache/beam/issues/33020
          type(value).__reduce__ == object.__reduce__):
      if not hasattr(value, "__setstate__"):
        raise TypeError(
            "Unable to deterministically encode '%s' of type '%s', "
            "for the input of '%s'. The object defines __getstate__ but not "
            "__setstate__." %
            (value, type(value), self.requires_deterministic_step_label))
      stream.write_byte(NESTED_STATE_TYPE)
      self.encode_type(type(value), stream)
      state_value = value.__getstate__()
      try:
        self.encode_to_stream(state_value, stream, True)
      except Exception as e:
        raise TypeError(self._deterministic_encoding_error_msg(value)) from e
    else:
      raise TypeError(self._deterministic_encoding_error_msg(value))

  def _deterministic_encoding_error_msg(self, value):
    return (
        "Unable to deterministically encode '%s' of type '%s', "
        "please provide a type hint for the input of '%s'" %
        (value, type(value), self.requires_deterministic_step_label))

  def encode_type_2_67_0(self, t, stream):
    """
    Encode special type with <=2.67.0 compatibility.
    """
    if t not in _pickled_types:
      _verify_dill_compat()
      _pickled_types[t] = dill.dumps(t)
    stream.write(_pickled_types[t], True)

  def encode_type(self, t, stream):
    if self.force_use_dill:
      return self.encode_type_2_67_0(t, stream)

    if t not in _pickled_types:
      config = cloudpickle.CloudPickleConfig(
          id_generator=None,
          skip_reset_dynamic_type_state=True,
          filepath_interceptor=cloudpickle.get_relative_path)
      if not self.use_relative_filepaths:
        config.filepath_interceptor = None
      _pickled_types[t] = cloudpickle_pickler.dumps(t, config=config)
    stream.write(_pickled_types[t], True)

  def decode_type(self, stream):
    if self.force_use_dill:
      return _unpickle_type_2_67_0(stream.read_all(True))
    return _unpickle_type(stream.read_all(True))

  def decode_from_stream(self, stream, nested):
    # type: (create_InputStream, bool) -> Any
    t = stream.read_byte()
    if t == NONE_TYPE:
      return None
    elif t == INT_TYPE:
      return stream.read_var_int64()
    elif t == FLOAT_TYPE:
      return stream.read_bigendian_double()
    elif t == BYTES_TYPE:
      return stream.read_all(nested)
    elif t == UNICODE_TYPE:
      return stream.read_all(nested).decode('utf-8')
    elif t == LIST_TYPE or t == TUPLE_TYPE or t == SET_TYPE:
      vlen = stream.read_var_int64()
      vlist = [self.decode_from_stream(stream, True) for _ in range(vlen)]
      if t == LIST_TYPE:
        return vlist
      elif t == TUPLE_TYPE:
        return tuple(vlist)
      return set(vlist)
    elif t == DICT_TYPE:
      vlen = stream.read_var_int64()
      v = {}
      for _ in range(vlen):
        k = self.decode_from_stream(stream, True)
        v[k] = self.decode_from_stream(stream, True)
      return v
    elif t == BOOL_TYPE:
      return not not stream.read_byte()
    elif t == ITERABLE_LIKE_TYPE:
      return self.iterable_coder_impl.decode_from_stream(stream, nested)
    elif t == PROTO_TYPE:
      cls = self.decode_type(stream)
      msg = cls()
      msg.ParseFromString(stream.read_all(True))
      return msg
    elif t == DATACLASS_KW_ONLY_TYPE:
      cls = self.decode_type(stream)
      vlen = stream.read_var_int64()
      fields = {}
      for _ in range(vlen):
        field_name = stream.read_all(True).decode('utf-8')
        fields[field_name] = self.decode_from_stream(stream, True)
      return cls(**fields)
    elif t == DATACLASS_TYPE or t == NAMED_TUPLE_TYPE:
      cls = self.decode_type(stream)
      return cls(*self.iterable_coder_impl.decode_from_stream(stream, True))
    elif t == ENUM_TYPE:
      cls = self.decode_type(stream)
      return cls(self.decode_from_stream(stream, True))
    elif t == NESTED_STATE_TYPE:
      cls = self.decode_type(stream)
      state = self.decode_from_stream(stream, True)
      value = cls.__new__(cls)
      value.__setstate__(state)
      return value
    elif t == UNKNOWN_TYPE:
      return self.fallback_coder_impl.decode_from_stream(stream, nested)
    else:
      raise ValueError('Unknown type tag %x' % t)


_pickled_types = {}  # type: Dict[type, bytes]
_unpickled_types = {}  # type: Dict[bytes, type]


def _unpickle_type_2_67_0(bs):
  """
  Decode special type with <=2.67.0 compatibility.
  """
  t = _unpickled_types.get(bs, None)
  if t is None:
    _verify_dill_compat()
    t = _unpickled_types[bs] = dill.loads(bs)
    # Fix unpicklable anonymous named tuples for Python 3.6.
    if t.__base__ is tuple and hasattr(t, '_fields'):
      try:
        pickle.loads(pickle.dumps(t))
      except pickle.PicklingError:
        t.__reduce__ = lambda self: (
            _unpickle_named_tuple_2_67_0, (bs, tuple(self)))
  return t


def _unpickle_named_tuple_2_67_0(bs, items):
  return _unpickle_type_2_67_0(bs)(*items)


def _unpickle_type(bs):
  if not _unpickled_types.get(bs, None):
    _unpickled_types[bs] = cloudpickle_pickler.loads(bs)

  return _unpickled_types[bs]


def _unpickle_named_tuple(bs, items):
  return _unpickle_type(bs)(*items)


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

  A coder for bytes/str objects."""
  def encode_to_stream(self, value, out, nested):
    # type: (bytes, create_OutputStream, bool) -> None

    # value might be of type np.bytes if passed from encode_batch, and cython
    # does not recognize it as bytes.
    if is_compiled and isinstance(value, np.bytes_):
      value = bytes(value)

    out.write(value, nested)

  def decode_from_stream(self, in_stream, nested):
    # type: (create_InputStream, bool) -> bytes
    return in_stream.read_all(nested)

  def encode(self, value):
    assert isinstance(value, bytes), (value, type(value))
    return value

  def decode(self, encoded):
    return encoded


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

  A coder for bool objects."""
  def encode_to_stream(self, value, out, nested):
    out.write_byte(1 if value else 0)

  def decode_from_stream(self, in_stream, nested):
    value = in_stream.read_byte()
    if value == 0:
      return False
    elif value == 1:
      return True
    raise ValueError("Expected 0 or 1, got %s" % value)

  def encode(self, value):
    return b'\x01' if value else b'\x00'

  def decode(self, encoded):
    value = ord(encoded)
    if value == 0:
      return False
    elif value == 1:
      return True
    raise ValueError("Expected 0 or 1, got %s" % value)

  def estimate_size(self, unused_value, nested=False):
    # Note that booleans are encoded the same way regardless of nesting.
    return 1


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

  Note this implementation always uses nested context when encoding keys
  and values. This differs from Java's MapCoder, which uses
  nested=False if possible for the last value encoded.

  This difference is acceptable because MapCoder is not standard. It is only
  used in a standard context by RowCoder which always uses nested context for
  attribute values.

  A coder for typing.Mapping objects."""
  def __init__(
      self,
      key_coder,  # type: CoderImpl
      value_coder,  # type: CoderImpl
      is_deterministic=False):
    self._key_coder = key_coder
    self._value_coder = value_coder
    self._is_deterministic = is_deterministic

  def encode_to_stream(self, dict_value, out, nested):
    out.write_bigendian_int32(len(dict_value))
    # Note this implementation always uses nested context when encoding keys
    # and values which differs from Java. See note in docstring.
    if self._is_deterministic:
      for key, value in sorted(dict_value.items()):
        self._key_coder.encode_to_stream(key, out, True)
        self._value_coder.encode_to_stream(value, out, True)
    else:
      # This loop is separate from the above so the loop over dict.items()
      # will be optimized by Cython.
      for key, value in dict_value.items():
        self._key_coder.encode_to_stream(key, out, True)
        self._value_coder.encode_to_stream(value, out, True)

  def decode_from_stream(self, in_stream, nested):
    size = in_stream.read_bigendian_int32()
    result = {}
    for _ in range(size):
      # Note this implementation always uses nested context when encoding keys
      # and values which differs from Java. See note in docstring.
      key = self._key_coder.decode_from_stream(in_stream, True)
      value = self._value_coder.decode_from_stream(in_stream, True)
      result[key] = value

    return result

  def estimate_size(self, unused_value, nested=False):
    estimate = 4  # 4 bytes for int32 size prefix
    for key, value in unused_value.items():
      estimate += self._key_coder.estimate_size(key, True)
      estimate += self._value_coder.estimate_size(value, True)
    return estimate


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

  A coder for typing.Optional objects."""

  ENCODE_NULL = 0
  ENCODE_PRESENT = 1

  def __init__(
      self,
      value_coder  # type: CoderImpl
  ):
    self._value_coder = value_coder

  def encode_to_stream(self, value, out, nested):
    if value is None:
      out.write_byte(self.ENCODE_NULL)
    else:
      out.write_byte(self.ENCODE_PRESENT)
      self._value_coder.encode_to_stream(value, out, nested)

  def decode_from_stream(self, in_stream, nested):
    null_indicator = in_stream.read_byte()
    if null_indicator == self.ENCODE_NULL:
      return None
    elif null_indicator == self.ENCODE_PRESENT:
      return self._value_coder.decode_from_stream(in_stream, nested)
    else:
      raise ValueError(
          "Encountered unexpected value for null indicator: '%s'" %
          null_indicator)

  def estimate_size(self, unused_value, nested=False):
    return 1 + (
        self._value_coder.estimate_size(unused_value)
        if unused_value is not None else 0)


class BigEndianShortCoderImpl(StreamCoderImpl):
  """For internal use only; no backwards-compatibility guarantees."""
  def encode_to_stream(self, value, out, nested):
    # type: (int, create_OutputStream, bool) -> None
    out.write_bigendian_int16(value)

  def decode_from_stream(self, in_stream, nested):
    # type: (create_InputStream, bool) -> float
    return in_stream.read_bigendian_int16()

  def estimate_size(self, unused_value, nested=False):
    # type: (Any, bool) -> int
    # A short is encoded as 2 bytes, regardless of nesting.
    return 2


class SinglePrecisionFloatCoderImpl(StreamCoderImpl):
  """For internal use only; no backwards-compatibility guarantees."""
  def encode_to_stream(self, value, out, nested):
    # type: (float, create_OutputStream, bool) -> None
    out.write_bigendian_float(value)

  def decode_from_stream(self, in_stream, nested):
    # type: (create_InputStream, bool) -> float
    return in_stream.read_bigendian_float()

  def estimate_size(self, unused_value, nested=False):
    # type: (Any, bool) -> int
    # A float is encoded as 4 bytes, regardless of nesting.
    return 4


class FloatCoderImpl(StreamCoderImpl):
  """For internal use only; no backwards-compatibility guarantees."""
  def encode_to_stream(self, value, out, nested):
    # type: (float, create_OutputStream, bool) -> None
    out.write_bigendian_double(value)

  def decode_from_stream(self, in_stream, nested):
    # type: (create_InputStream, bool) -> float
    return in_stream.read_bigendian_double()

  def estimate_size(self, unused_value, nested=False):
    # type: (Any, bool) -> int
    # A double is encoded as 8 bytes, regardless of nesting.
    return 8


if not TYPE_CHECKING:
  IntervalWindow = None


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

  # TODO: Fn Harness only supports millis. Is this important enough to fix?
  def _to_normal_time(self, value):
    """Convert "lexicographically ordered unsigned" to signed."""
    return value - _TIME_SHIFT

  def _from_normal_time(self, value):
    """Convert signed to "lexicographically ordered unsigned"."""
    return value + _TIME_SHIFT

  def encode_to_stream(self, value, out, nested):
    # type: (IntervalWindow, create_OutputStream, bool) -> None
    typed_value = value
    span_millis = (
        typed_value._end_micros // 1000 - typed_value._start_micros // 1000)
    out.write_bigendian_uint64(
        self._from_normal_time(typed_value._end_micros // 1000))
    out.write_var_int64(span_millis)

  def decode_from_stream(self, in_, nested):
    # type: (create_InputStream, bool) -> IntervalWindow
    if not TYPE_CHECKING:
      global IntervalWindow  # pylint: disable=global-variable-not-assigned
      if IntervalWindow is None:
        from apache_beam.transforms.window import IntervalWindow
    # instantiating with None is not part of the public interface
    # pylint: disable=too-many-function-args
    typed_value = IntervalWindow(None, None)  # type: ignore[arg-type]
    typed_value._end_micros = (
        1000 * self._to_normal_time(in_.read_bigendian_uint64()))
    typed_value._start_micros = (
        typed_value._end_micros - 1000 * in_.read_var_int64())
    return typed_value

  def estimate_size(self, value, nested=False):
    # type: (Any, bool) -> int
    # An IntervalWindow is context-insensitive, with a timestamp (8 bytes)
    # and a varint timespam.
    typed_value = value
    span_millis = (
        typed_value._end_micros // 1000 - typed_value._start_micros // 1000)
    return 8 + get_varint_size(span_millis)


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

  TODO: SDK agnostic encoding
  For interoperability with Java SDK, encoding needs to match
  that of the Java SDK InstantCoder.
  https://github.com/apache/beam/blob/f5029b4f0dfff404310b2ef55e2632bbacc7b04f/sdks/java/core/src/main/java/org/apache/beam/sdk/coders/InstantCoder.java#L79
  """
  def encode_to_stream(self, value, out, nested):
    # type: (Timestamp, create_OutputStream, bool) -> None
    millis = value.micros // 1000
    if millis >= 0:
      millis = millis - _TIME_SHIFT
    else:
      millis = millis + _TIME_SHIFT
    out.write_bigendian_int64(millis)

  def decode_from_stream(self, in_stream, nested):
    # type: (create_InputStream, bool) -> Timestamp
    millis = in_stream.read_bigendian_int64()
    if millis < 0:
      millis = millis + _TIME_SHIFT
    else:
      millis = millis - _TIME_SHIFT
    return Timestamp(micros=millis * 1000)

  def estimate_size(self, unused_value, nested=False):
    # A Timestamp is encoded as a 64-bit integer in 8 bytes, regardless of
    # nesting.
    return 8


class TimerCoderImpl(StreamCoderImpl):
  """For internal use only; no backwards-compatibility guarantees."""
  def __init__(self, key_coder_impl, window_coder_impl):
    self._timestamp_coder_impl = TimestampCoderImpl()
    self._boolean_coder_impl = BooleanCoderImpl()
    self._pane_info_coder_impl = PaneInfoCoderImpl()
    self._key_coder_impl = key_coder_impl
    self._windows_coder_impl = TupleSequenceCoderImpl(window_coder_impl)
    from apache_beam.coders.coders import StrUtf8Coder
    self._tag_coder_impl = StrUtf8Coder().get_impl()

  def encode_to_stream(self, value, out, nested):
    # type: (userstate.Timer, create_OutputStream, bool) -> None
    self._key_coder_impl.encode_to_stream(value.user_key, out, True)
    self._tag_coder_impl.encode_to_stream(value.dynamic_timer_tag, out, True)
    self._windows_coder_impl.encode_to_stream(value.windows, out, True)
    self._boolean_coder_impl.encode_to_stream(value.clear_bit, out, True)
    if not value.clear_bit:
      self._timestamp_coder_impl.encode_to_stream(
          value.fire_timestamp, out, True)
      self._timestamp_coder_impl.encode_to_stream(
          value.hold_timestamp, out, True)
      self._pane_info_coder_impl.encode_to_stream(value.paneinfo, out, True)

  def decode_from_stream(self, in_stream, nested):
    # type: (create_InputStream, bool) -> userstate.Timer
    from apache_beam.transforms import userstate
    user_key = self._key_coder_impl.decode_from_stream(in_stream, True)
    dynamic_timer_tag = self._tag_coder_impl.decode_from_stream(in_stream, True)
    windows = self._windows_coder_impl.decode_from_stream(in_stream, True)
    clear_bit = self._boolean_coder_impl.decode_from_stream(in_stream, True)
    if clear_bit:
      return userstate.Timer(
          user_key=user_key,
          dynamic_timer_tag=dynamic_timer_tag,
          windows=windows,
          clear_bit=clear_bit,
          fire_timestamp=None,
          hold_timestamp=None,
          paneinfo=None)

    return userstate.Timer(
        user_key=user_key,
        dynamic_timer_tag=dynamic_timer_tag,
        windows=windows,
        clear_bit=clear_bit,
        fire_timestamp=self._timestamp_coder_impl.decode_from_stream(
            in_stream, True),
        hold_timestamp=self._timestamp_coder_impl.decode_from_stream(
            in_stream, True),
        paneinfo=self._pane_info_coder_impl.decode_from_stream(in_stream, True))


small_ints = [chr(_).encode('latin-1') for _ in range(128)]


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

  A coder for int objects."""
  def encode_to_stream(self, value, out, nested):
    # type: (int, create_OutputStream, bool) -> None
    try:
      out.write_var_int64(value)
    except OverflowError as e:
      raise OverflowError(
          f"Integer value '{value}' is out of the encodable range for "
          f"VarIntCoder. This coder is limited to values that fit "
          f"within a 64-bit signed integer (-(2**63) to 2**63 - 1). "
          f"Original error: {e}") from e

  def decode_from_stream(self, in_stream, nested):
    # type: (create_InputStream, bool) -> int
    return in_stream.read_var_int64()

  def encode(self, value):
    ivalue = value  # type cast
    if 0 <= ivalue < len(small_ints):
      return small_ints[ivalue]
    return StreamCoderImpl.encode(self, value)

  def decode(self, encoded):
    if len(encoded) == 1:
      i = ord(encoded)
      if 0 <= i < 128:
        return i
    return StreamCoderImpl.decode(self, encoded)

  def estimate_size(self, value, nested=False):
    # type: (Any, bool) -> int
    # Note that VarInts are encoded the same way regardless of nesting.
    try:
      return get_varint_size(value)
    except OverflowError as e:
      raise OverflowError(
          f"Cannot estimate size for integer value '{value}'. "
          f"Value is out of the range for VarIntCoder (64-bit signed integer). "
          f"Original error: {e}") from e


class VarInt32CoderImpl(StreamCoderImpl):
  """For internal use only; no backwards-compatibility guarantees.

  A coder for int32 objects."""
  def encode_to_stream(self, value, out, nested):
    # type: (int, create_OutputStream, bool) -> None
    out.write_var_int32(value)

  def decode_from_stream(self, in_stream, nested):
    # type: (create_InputStream, bool) -> int
    return in_stream.read_var_int32()

  def encode(self, value):
    ivalue = value  # type cast
    if 0 <= ivalue < len(small_ints):
      return small_ints[ivalue]
    return StreamCoderImpl.encode(self, value)

  def decode(self, encoded):
    if len(encoded) == 1:
      i = ord(encoded)
      if 0 <= i < 128:
        return i
    return StreamCoderImpl.decode(self, encoded)

  def estimate_size(self, value, nested=False):
    # type: (Any, bool) -> int
    # Note that VarInts are encoded the same way regardless of nesting.
    return get_varint_size(int(value) & 0xFFFFFFFF)


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

  A coder that always encodes exactly one value."""
  def __init__(self, value):
    self._value = value

  def encode_to_stream(self, value, stream, nested):
    # type: (Any, create_OutputStream, bool) -> None
    pass

  def decode_from_stream(self, stream, nested):
    # type: (create_InputStream, bool) -> Any
    return self._value

  def encode(self, value):
    b = b''  # avoid byte vs str vs unicode error
    return b

  def decode(self, encoded):
    return self._value

  def estimate_size(self, value, nested=False):
    # type: (Any, bool) -> int
    return 0


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

  CoderImpl for coders that are comprised of several component coders."""
  def __init__(self, coder_impls):
    for c in coder_impls:
      assert isinstance(c, CoderImpl), c
    self._coder_impls = tuple(coder_impls)

  def _extract_components(self, value):
    raise NotImplementedError

  def _construct_from_components(self, components):
    raise NotImplementedError

  def encode_to_stream(self, value, out, nested):
    # type: (Any, create_OutputStream, bool) -> None
    values = self._extract_components(value)
    if len(self._coder_impls) != len(values):
      raise ValueError('Number of components does not match number of coders.')
    for i in range(0, len(self._coder_impls)):
      c = self._coder_impls[i]  # type cast
      c.encode_to_stream(
          values[i], out, nested or i + 1 < len(self._coder_impls))

  def decode_from_stream(self, in_stream, nested):
    # type: (create_InputStream, bool) -> Any
    return self._construct_from_components([
        c.decode_from_stream(
            in_stream, nested or i + 1 < len(self._coder_impls))
        for i, c in enumerate(self._coder_impls)
    ])

  def estimate_size(self, value, nested=False):
    # type: (Any, bool) -> int

    """Estimates the encoded size of the given value, in bytes."""
    # TODO(ccy): This ignores sizes of observable components.
    estimated_size, _ = (self.get_estimated_size_and_observables(value))
    return estimated_size

  def get_estimated_size_and_observables(self, value, nested=False):
    # type: (Any, bool) -> Tuple[int, Observables]

    """Returns estimated size of value along with any nested observables."""
    values = self._extract_components(value)
    estimated_size = 0
    observables = []  # type: Observables
    for i in range(0, len(self._coder_impls)):
      c = self._coder_impls[i]  # type cast
      child_size, child_observables = (
          c.get_estimated_size_and_observables(
              values[i], nested=nested or i + 1 < len(self._coder_impls)))
      estimated_size += child_size
      observables += child_observables
    return estimated_size, observables


class AvroCoderImpl(SimpleCoderImpl):
  """For internal use only; no backwards-compatibility guarantees."""
  def __init__(self, schema):
    self.parsed_schema = parse_schema(json.loads(schema))

  def encode(self, value):
    assert issubclass(type(value), AvroRecord)
    with BytesIO() as buf:
      schemaless_writer(buf, self.parsed_schema, value.record)
      return buf.getvalue()

  def decode(self, encoded):
    with BytesIO(encoded) as buf:
      return AvroRecord(schemaless_reader(buf, self.parsed_schema))


class TupleCoderImpl(AbstractComponentCoderImpl):
  """A coder for tuple objects."""
  def _extract_components(self, value):
    return tuple(value)

  def _construct_from_components(self, components):
    return tuple(components)


class _ConcatSequence(object):
  def __init__(self, head, tail):
    # type: (Iterable[Any], Iterable[Any]) -> None
    self._head = head
    self._tail = tail

  def __iter__(self):
    # type: () -> Iterator[Any]
    for elem in self._head:
      yield elem
    for elem in self._tail:
      yield elem

  def __eq__(self, other):
    return list(self) == list(other)

  def __hash__(self):
    raise NotImplementedError

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


FastPrimitivesCoderImpl.register_iterable_like_type(_ConcatSequence)


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

  A coder for sequences.

  If the length of the sequence in known we encode the length as a 32 bit
  ``int`` followed by the encoded bytes.

  If the length of the sequence is unknown, we encode the length as ``-1``
  followed by the encoding of elements buffered up to 64K bytes before prefixing
  the count of number of elements. A ``0`` is encoded at the end to indicate the
  end of stream.

  The resulting encoding would look like this::

    -1
    countA element(0) element(1) ... element(countA - 1)
    countB element(0) element(1) ... element(countB - 1)
    ...
    countX element(0) element(1) ... element(countX - 1)
    0

  If writing to state is enabled, the final terminating 0 will instead be
  repaced with::

    varInt64(-1)
    len(state_token)
    state_token

  where state_token is a bytes object used to retrieve the remainder of the
  iterable via the state API.
  """

  # Default buffer size of 64kB of handling iterables of unknown length.
  _DEFAULT_BUFFER_SIZE = 64 * 1024

  def __init__(
      self,
      elem_coder,  # type: CoderImpl
      read_state=None,  # type: Optional[IterableStateReader]
      write_state=None,  # type: Optional[IterableStateWriter]
      write_state_threshold=0  # type: int
  ):
    self._elem_coder = elem_coder
    self._read_state = read_state
    self._write_state = write_state
    self._write_state_threshold = write_state_threshold

  def _construct_from_sequence(self, values):
    raise NotImplementedError

  def encode_to_stream(self, value, out, nested):
    # type: (Sequence, create_OutputStream, bool) -> None
    # Compatible with Java's IterableLikeCoder.
    if hasattr(value, '__len__') and self._write_state is None:
      out.write_bigendian_int32(len(value))
      for elem in value:
        self._elem_coder.encode_to_stream(elem, out, True)
    else:
      # We don't know the size without traversing it so use a fixed size buffer
      # and encode as many elements as possible into it before outputting
      # the size followed by the elements.

      # -1 to indicate that the length is not known.
      out.write_bigendian_int32(-1)
      buffer = create_OutputStream()
      if self._write_state is None:
        target_buffer_size = self._DEFAULT_BUFFER_SIZE
      else:
        target_buffer_size = min(
            self._DEFAULT_BUFFER_SIZE, self._write_state_threshold)
      prev_index = index = -1
      # Don't want to miss out on fast list iteration optimization.
      value_iter = value if isinstance(value, (list, tuple)) else iter(value)
      start_size = out.size()
      for elem in value_iter:
        index += 1
        self._elem_coder.encode_to_stream(elem, buffer, True)
        if buffer.size() > target_buffer_size:
          out.write_var_int64(index - prev_index)
          out.write(buffer.get())
          prev_index = index
          buffer = create_OutputStream()
          if (self._write_state is not None and
              out.size() - start_size > self._write_state_threshold):
            tail = (
                value_iter[index +
                           1:] if isinstance(value_iter,
                                             (list, tuple)) else value_iter)
            state_token = self._write_state(tail, self._elem_coder)
            out.write_var_int64(-1)
            out.write(state_token, True)
            break
      else:
        if index > prev_index:
          out.write_var_int64(index - prev_index)
          out.write(buffer.get())
        out.write_var_int64(0)

  def decode_from_stream(self, in_stream, nested):
    # type: (create_InputStream, bool) -> Sequence
    size = in_stream.read_bigendian_int32()

    if size >= 0:
      elements = [
          self._elem_coder.decode_from_stream(in_stream, True)
          for _ in range(size)
      ]  # type: Iterable[Any]
    else:
      elements = []
      count = in_stream.read_var_int64()
      while count > 0:
        for _ in range(count):
          elements.append(self._elem_coder.decode_from_stream(in_stream, True))
        count = in_stream.read_var_int64()

      if count == -1:
        if self._read_state is None:
          raise ValueError(
              'Cannot read state-written iterable without state reader.')

        state_token = in_stream.read_all(True)
        elements = _ConcatSequence(
            elements, self._read_state(state_token, self._elem_coder))

    return self._construct_from_sequence(elements)

  def estimate_size(self, value, nested=False):
    # type: (Any, bool) -> int

    """Estimates the encoded size of the given value, in bytes."""
    # TODO(ccy): This ignores element sizes.
    estimated_size, _ = (self.get_estimated_size_and_observables(value))
    return estimated_size

  def get_estimated_size_and_observables(self, value, nested=False):
    # type: (Any, bool) -> Tuple[int, Observables]

    """Returns estimated size of value along with any nested observables."""
    estimated_size = 0
    # Size of 32-bit integer storing number of elements.
    estimated_size += 4
    if isinstance(value, observable.ObservableMixin):
      return estimated_size, [(value, self._elem_coder)]

    observables = []  # type: Observables
    for elem in value:
      child_size, child_observables = (
          self._elem_coder.get_estimated_size_and_observables(
              elem, nested=True))
      estimated_size += child_size
      observables += child_observables
    # TODO: (https://github.com/apache/beam/issues/18169) Update to use an
    # accurate count depending on size and count, currently we are
    # underestimating the size by up to 10 bytes per block of data since we are
    # not including the count prefix which occurs at most once per 64k of data
    # and is upto 10 bytes long. The upper bound of the underestimate is
    # 10 / 65536 ~= 0.0153% of the actual size.
    # TODO: More efficient size estimation in the case of state-backed
    # iterables.
    return estimated_size, observables


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

  A coder for homogeneous tuple objects."""
  def _construct_from_sequence(self, components):
    return tuple(components)


class _AbstractIterable(object):
  """Wraps an iterable hiding methods that might not always be available."""
  def __init__(self, contents):
    self._contents = contents

  def __iter__(self):
    return iter(self._contents)

  def __repr__(self):
    head = [repr(e) for e in itertools.islice(self, 4)]
    if len(head) == 4:
      head[-1] = '...'
    return '_AbstractIterable([%s])' % ', '.join(head)

  # Mostly useful for tests.
  def __eq__(left, right):
    end = object()
    for a, b in itertools.zip_longest(left, right, fillvalue=end):
      if a != b:
        return False
    return True


FastPrimitivesCoderImpl.register_iterable_like_type(_AbstractIterable)

# TODO(https://github.com/apache/beam/issues/21167): Enable using abstract
# iterables permanently
_iterable_coder_uses_abstract_iterable_by_default = False


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

  A coder for homogeneous iterable objects."""
  def __init__(self, *args, use_abstract_iterable=None, **kwargs):
    super().__init__(*args, **kwargs)
    if use_abstract_iterable is None:
      use_abstract_iterable = _iterable_coder_uses_abstract_iterable_by_default
    self._use_abstract_iterable = use_abstract_iterable

  def _construct_from_sequence(self, components):
    if self._use_abstract_iterable:
      return _AbstractIterable(components)
    else:
      return components


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

  A coder for homogeneous list objects."""
  def _construct_from_sequence(self, components):
    return components if isinstance(components, list) else list(components)


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

  Encoding used to describe a PaneInfo descriptor.  A PaneInfo descriptor
  can be encoded in three different ways: with a single byte (FIRST), with a
  single byte followed by a varint describing a single index (ONE_INDEX) or
  with a single byte followed by two varints describing two separate indices:
  the index and nonspeculative index.
  """

  FIRST = 0
  ONE_INDEX = 1
  TWO_INDICES = 2


# These are cdef'd to ints to optimized the common case.
PaneInfoTiming_UNKNOWN = windowed_value.PaneInfoTiming.UNKNOWN
PaneInfoEncoding_FIRST = PaneInfoEncoding.FIRST


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

  Coder for a PaneInfo descriptor."""
  def _choose_encoding(self, value):
    if ((value._index == 0 and value._nonspeculative_index == 0) or
        value._timing == PaneInfoTiming_UNKNOWN):
      return PaneInfoEncoding_FIRST
    elif (value._index == value._nonspeculative_index or
          value._timing == windowed_value.PaneInfoTiming.EARLY):
      return PaneInfoEncoding.ONE_INDEX
    else:
      return PaneInfoEncoding.TWO_INDICES

  def encode_to_stream(self, value, out, nested):
    # type: (windowed_value.PaneInfo, create_OutputStream, bool) -> None
    pane_info = value  # cast
    encoding_type = self._choose_encoding(pane_info)
    out.write_byte(pane_info._encoded_byte | (encoding_type << 4))
    if encoding_type == PaneInfoEncoding_FIRST:
      return
    elif encoding_type == PaneInfoEncoding.ONE_INDEX:
      out.write_var_int64(value.index)
    elif encoding_type == PaneInfoEncoding.TWO_INDICES:
      out.write_var_int64(value.index)
      out.write_var_int64(value.nonspeculative_index)
    else:
      raise NotImplementedError('Invalid PaneInfoEncoding: %s' % encoding_type)

  def decode_from_stream(self, in_stream, nested):
    # type: (create_InputStream, bool) -> windowed_value.PaneInfo
    encoded_first_byte = in_stream.read_byte()
    base = windowed_value._BYTE_TO_PANE_INFO[encoded_first_byte & 0xF]
    assert base is not None
    encoding_type = encoded_first_byte >> 4
    if encoding_type == PaneInfoEncoding_FIRST:
      return base
    elif encoding_type == PaneInfoEncoding.ONE_INDEX:
      index = in_stream.read_var_int64()
      if base.timing == windowed_value.PaneInfoTiming.EARLY:
        nonspeculative_index = -1
      else:
        nonspeculative_index = index
    elif encoding_type == PaneInfoEncoding.TWO_INDICES:
      index = in_stream.read_var_int64()
      nonspeculative_index = in_stream.read_var_int64()
    else:
      raise NotImplementedError('Invalid PaneInfoEncoding: %s' % encoding_type)
    return windowed_value.PaneInfo(
        base.is_first, base.is_last, base.timing, index, nonspeculative_index)

  def estimate_size(self, value, nested=False):
    # type: (Any, bool) -> int

    """Estimates the encoded size of the given value, in bytes."""
    size = 1
    encoding_type = self._choose_encoding(value)
    if encoding_type == PaneInfoEncoding.ONE_INDEX:
      size += get_varint_size(value.index)
    elif encoding_type == PaneInfoEncoding.TWO_INDICES:
      size += get_varint_size(value.index)
      size += get_varint_size(value.nonspeculative_index)
    return size


class _OrderedUnionCoderImpl(StreamCoderImpl):
  def __init__(self, coder_impl_types, fallback_coder_impl):
    assert len(coder_impl_types) < 128
    self._types, self._coder_impls = zip(*coder_impl_types)
    self._fallback_coder_impl = fallback_coder_impl

  def encode_to_stream(self, value, out, nested):
    value_t = type(value)
    for (ix, t) in enumerate(self._types):
      if value_t is t:
        out.write_byte(ix)
        c = self._coder_impls[ix]  # for typing
        c.encode_to_stream(value, out, nested)
        break
    else:
      if self._fallback_coder_impl is None:
        raise ValueError("No fallback.")
      out.write_byte(0xFF)
      self._fallback_coder_impl.encode_to_stream(value, out, nested)

  def decode_from_stream(self, in_stream, nested):
    ix = in_stream.read_byte()
    if ix == 0xFF:
      if self._fallback_coder_impl is None:
        raise ValueError("No fallback.")
      return self._fallback_coder_impl.decode_from_stream(in_stream, nested)
    else:
      c = self._coder_impls[ix]  # for typing
      return c.decode_from_stream(in_stream, nested)


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

  A coder for windowed values."""

  # Ensure that lexicographic ordering of the bytes corresponds to
  # chronological order of timestamps.
  # TODO(https://github.com/apache/beam/issues/18190): Clean this up once we
  # have a BEAM wide consensus on byte representation of timestamps.
  def _to_normal_time(self, value):
    """Convert "lexicographically ordered unsigned" to signed."""
    return value - _TIME_SHIFT

  def _from_normal_time(self, value):
    """Convert signed to "lexicographically ordered unsigned"."""
    return value + _TIME_SHIFT

  def __init__(self, value_coder, timestamp_coder, window_coder):
    # TODO(lcwik): Remove the timestamp coder field
    self._value_coder = value_coder
    self._timestamp_coder = timestamp_coder
    self._windows_coder = TupleSequenceCoderImpl(window_coder)
    self._pane_info_coder = PaneInfoCoderImpl()

  def encode_to_stream(self, value, out, nested):
    # type: (windowed_value.WindowedValue, create_OutputStream, bool) -> None
    wv = value  # type cast
    # Avoid creation of Timestamp object.
    restore_sign = -1 if wv.timestamp_micros < 0 else 1
    out.write_bigendian_uint64(
        # Convert to postive number and divide, since python rounds off to the
        # lower negative number. For ex: -3 / 2 = -2, but we expect it to be -1,
        # to be consistent across SDKs.
        # TODO(https://github.com/apache/beam/issues/18190): Clean this up once
        # we have a BEAM wide consensus on precision of timestamps.
        self._from_normal_time(
            restore_sign * (
                abs(
                    MIN_TIMESTAMP_micros if wv.timestamp_micros <
                    MIN_TIMESTAMP_micros else wv.timestamp_micros) // 1000)))
    self._windows_coder.encode_to_stream(wv.windows, out, True)
    # Default PaneInfo encoded byte representing NO_FIRING.
    self._pane_info_coder.encode_to_stream(wv.pane_info, out, True)
    self._value_coder.encode_to_stream(wv.value, out, nested)

  def decode_from_stream(self, in_stream, nested):
    # type: (create_InputStream, bool) -> windowed_value.WindowedValue
    timestamp = self._to_normal_time(in_stream.read_bigendian_uint64())
    # Restore MIN/MAX timestamps to their actual values as encoding incurs loss
    # of precision while converting to millis.
    # Note: This is only a best effort here as there is no way to know if these
    # were indeed MIN/MAX timestamps.
    # TODO(https://github.com/apache/beam/issues/18190): Clean this up once we
    # have a BEAM wide consensus on precision of timestamps.
    if timestamp <= -(abs(MIN_TIMESTAMP_micros) // 1000):
      timestamp = MIN_TIMESTAMP_micros
    elif timestamp >= MAX_TIMESTAMP_micros // 1000:
      timestamp = MAX_TIMESTAMP_micros
    else:
      timestamp *= 1000

    windows = self._windows_coder.decode_from_stream(in_stream, True)
    # Read PaneInfo encoded byte.
    pane_info = self._pane_info_coder.decode_from_stream(in_stream, True)
    value = self._value_coder.decode_from_stream(in_stream, nested)
    return windowed_value.create(
        value,
        timestamp,  # Avoid creation of Timestamp object.
        windows,
        pane_info)

  def get_estimated_size_and_observables(self, value, nested=False):
    # type: (Any, bool) -> Tuple[int, Observables]

    """Returns estimated size of value along with any nested observables."""
    if isinstance(value, observable.ObservableMixin):
      # Should never be here.
      # TODO(robertwb): Remove when coders are set correctly.
      return 0, [(value, self._value_coder)]
    estimated_size = 0
    observables = []  # type: Observables
    value_estimated_size, value_observables = (
        self._value_coder.get_estimated_size_and_observables(
            value.value, nested=nested))
    estimated_size += value_estimated_size
    observables += value_observables
    estimated_size += (
        self._timestamp_coder.estimate_size(value.timestamp, nested=True))
    estimated_size += (
        self._windows_coder.estimate_size(value.windows, nested=True))
    estimated_size += (
        self._pane_info_coder.estimate_size(value.pane_info, nested=True))
    return estimated_size, observables


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

  A coder for windowed values with constant timestamp, windows and
  pane info. The coder drops timestamp, windows and pane info during
  encoding, and uses the supplied parameterized timestamp, windows
  and pane info values during decoding when reconstructing the windowed
  value."""
  def __init__(self, value_coder, window_coder, payload):
    super().__init__(value_coder, TimestampCoderImpl(), window_coder)
    self._timestamp, self._windows, self._pane_info = self._from_proto(
        payload, window_coder)

  def _from_proto(self, payload, window_coder):
    windowed_value_coder = WindowedValueCoderImpl(
        BytesCoderImpl(), TimestampCoderImpl(), window_coder)
    wv = windowed_value_coder.decode(payload)
    return wv.timestamp_micros, wv.windows, wv.pane_info

  def encode_to_stream(self, value, out, nested):
    wv = value  # type cast
    self._value_coder.encode_to_stream(wv.value, out, nested)

  def decode_from_stream(self, in_stream, nested):
    value = self._value_coder.decode_from_stream(in_stream, nested)
    return windowed_value.create(
        value, self._timestamp, self._windows, self._pane_info)

  def get_estimated_size_and_observables(self, value, nested=False):
    """Returns estimated size of value along with any nested observables."""
    if isinstance(value, observable.ObservableMixin):
      # Should never be here.
      # TODO(robertwb): Remove when coders are set correctly.
      return 0, [(value, self._value_coder)]
    estimated_size = 0
    observables = []
    value_estimated_size, value_observables = (
        self._value_coder.get_estimated_size_and_observables(
            value.value, nested=nested))
    estimated_size += value_estimated_size
    observables += value_observables
    return estimated_size, observables


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

  Coder which prefixes the length of the encoded object in the stream."""
  def __init__(self, value_coder):
    # type: (CoderImpl) -> None
    self._value_coder = value_coder

  def encode_to_stream(self, value, out, nested):
    # type: (Any, create_OutputStream, bool) -> None
    encoded_value = self._value_coder.encode(value)
    out.write_var_int64(len(encoded_value))
    out.write(encoded_value)

  def decode_from_stream(self, in_stream, nested):
    # type: (create_InputStream, bool) -> Any
    value_length = in_stream.read_var_int64()
    return self._value_coder.decode(in_stream.read(value_length))

  def estimate_size(self, value, nested=False):
    # type: (Any, bool) -> int
    value_size = self._value_coder.estimate_size(value)
    return get_varint_size(value_size) + value_size


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

  A coder for sharded user keys.

  The encoding and decoding should follow the order:
      shard id byte string
      encoded user key
  """
  def __init__(self, key_coder_impl):
    self._shard_id_coder_impl = BytesCoderImpl()
    self._key_coder_impl = key_coder_impl

  def encode_to_stream(self, value, out, nested):
    # type: (ShardedKey, create_OutputStream, bool) -> None
    self._shard_id_coder_impl.encode_to_stream(value._shard_id, out, True)
    self._key_coder_impl.encode_to_stream(value.key, out, True)

  def decode_from_stream(self, in_stream, nested):
    # type: (create_InputStream, bool) -> ShardedKey
    shard_id = self._shard_id_coder_impl.decode_from_stream(in_stream, True)
    key = self._key_coder_impl.decode_from_stream(in_stream, True)
    return ShardedKey(key=key, shard_id=shard_id)

  def estimate_size(self, value, nested=False):
    # type: (Any, bool) -> int
    estimated_size = 0
    estimated_size += (
        self._shard_id_coder_impl.estimate_size(value._shard_id, nested=True))
    estimated_size += (
        self._key_coder_impl.estimate_size(value.key, nested=True))
    return estimated_size


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

  A coder for custom window types, which prefix required max_timestamp to
  encoded original window.

  The coder encodes and decodes custom window types with following format:
    window's max_timestamp()
    encoded window using it's own coder.
  """
  def __init__(self, window_coder_impl: CoderImpl) -> None:
    self._window_coder_impl = window_coder_impl

  def encode_to_stream(self, value, stream, nested):
    TimestampCoderImpl().encode_to_stream(value.max_timestamp(), stream, nested)
    self._window_coder_impl.encode_to_stream(value, stream, nested)

  def decode_from_stream(self, stream, nested):
    TimestampCoderImpl().decode_from_stream(stream, nested)
    return self._window_coder_impl.decode_from_stream(stream, nested)

  def estimate_size(self, value: Any, nested: bool = False) -> int:
    return (
        TimestampCoderImpl().estimate_size(value.max_timestamp()) +
        self._window_coder_impl.estimate_size(value, nested))


_OpaqueWindow = None


def _create_opaque_window(end, encoded_window):
  # This is lazy to avoid circular import issues.
  global _OpaqueWindow
  if _OpaqueWindow is None:
    from apache_beam.transforms.window import BoundedWindow

    class _OpaqueWindow(BoundedWindow):
      def __init__(self, end, encoded_window):
        super().__init__(end)
        self.encoded_window = encoded_window

      def __repr__(self):
        return 'OpaqueWindow(%s, %s)' % (self.end, self.encoded_window)

      def __hash__(self):
        return hash(self.encoded_window)

      def __eq__(self, other):
        return (
            type(self) == type(other) and self.end == other.end and
            self.encoded_window == other.encoded_window)

  return _OpaqueWindow(end, encoded_window)


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

  A coder for unknown window types, which prefix required max_timestamp to
  encoded original window.

  The coder encodes and decodes custom window types with following format:
    window's max_timestamp()
    length prefixed encoded window
  """
  def __init__(self) -> None:
    pass

  def encode_to_stream(self, value, stream, nested):
    TimestampCoderImpl().encode_to_stream(value.max_timestamp(), stream, True)
    stream.write(value.encoded_window, True)

  def decode_from_stream(self, stream, nested):
    max_timestamp = TimestampCoderImpl().decode_from_stream(stream, True)
    return _create_opaque_window(
        max_timestamp.successor(), stream.read_all(True))

  def estimate_size(self, value: Any, nested: bool = False) -> int:
    return (
        TimestampCoderImpl().estimate_size(value.max_timestamp()) +
        len(value.encoded_window))


row_coders_registered = False


class RowColumnEncoder:
  ROW_ENCODERS = {1: 12345}

  @classmethod
  def register(cls, field_type, coder_impl):
    cls.ROW_ENCODERS[field_type, coder_impl] = cls

  @classmethod
  def create(cls, field_type, coder_impl, column):
    global row_coders_registered
    if not row_coders_registered:
      try:
        # pylint: disable=unused-import
        from apache_beam.coders import coder_impl_row_encoders
      except ImportError:
        pass
      row_coders_registered = True
    return cls.ROW_ENCODERS.get((field_type, column.dtype),
                                GenericRowColumnEncoder)(coder_impl, column)

  def null_flags(self):
    raise NotImplementedError(type(self))

  def encode_to_stream(self, index, out):
    raise NotImplementedError(type(self))

  def decode_from_stream(self, index, in_stream):
    raise NotImplementedError(type(self))

  def finalize_write(self):
    raise NotImplementedError(type(self))


class GenericRowColumnEncoder(RowColumnEncoder):
  def __init__(self, coder_impl, column):
    self.coder_impl = coder_impl
    self.column = column

  def null_flags(self):
    # pylint: disable=singleton-comparison
    return self.column == None  # This is an array.

  def encode_to_stream(self, index, out):
    self.coder_impl.encode_to_stream(self.column[index], out, True)

  def decode_from_stream(self, index, in_stream):
    self.column[index] = self.coder_impl.decode_from_stream(in_stream, True)

  def finalize_write(self):
    pass


class RowCoderImpl(StreamCoderImpl):
  """For internal use only; no backwards-compatibility guarantees."""
  def __init__(self, schema, components):
    self.schema = schema
    self.num_fields = len(self.schema.fields)
    self.field_names = [f.name for f in self.schema.fields]
    self.field_nullable = [field.type.nullable for field in self.schema.fields]
    self.constructor = named_tuple_from_schema(schema)
    self.encoding_positions = list(range(len(self.schema.fields)))
    if self.schema.encoding_positions_set:
      # should never be duplicate encoding positions.
      enc_posx = list(
          set(field.encoding_position for field in self.schema.fields))
      if len(enc_posx) != len(self.schema.fields):
        names_no_pos = ", ".join(
            [f.name for f in self.schema.fields if f.encoding_position is None])
        raise ValueError(
            f'''Schema with id {schema.id} has encoding_positions_set=True,
            but found fields without encoding_position set: {names_no_pos}''')
      self.encoding_positions = list(
          field.encoding_position for field in self.schema.fields)
    self.encoding_positions_argsort = list(np.argsort(self.encoding_positions))
    self.encoding_positions_are_trivial = self.encoding_positions == list(
        range(len(self.encoding_positions)))
    self.components = list(
        components[self.encoding_positions.index(i)].get_impl()
        for i in self.encoding_positions)
    self.has_nullable_fields = any(
        field.type.nullable for field in self.schema.fields)

  def encode_to_stream(self, value, out, nested):
    out.write_var_int64(self.num_fields)
    attrs = [getattr(value, name) for name in self.field_names]

    if self.has_nullable_fields:
      any_nulls = False
      for attr in attrs:
        if attr is None:
          any_nulls = True
          break
      if any_nulls:
        out.write_var_int64((self.num_fields + 7) // 8)
        # Pack the bits, little-endian, in consecutive bytes.
        running = 0
        for i, attr in enumerate(attrs):
          if i and i % 8 == 0:
            out.write_byte(running)
            running = 0
          running |= (attr is None) << (i % 8)
        out.write_byte(running)
      else:
        out.write_byte(0)
    else:
      out.write_byte(0)

    for i in range(self.num_fields):
      if not self.encoding_positions_are_trivial:
        i = self.encoding_positions_argsort[i]
      attr = attrs[i]
      if attr is None:
        if not self.field_nullable[i]:
          raise ValueError(
              "Attempted to encode null for non-nullable field \"{}\".".format(
                  self.schema.fields[i].name))
        continue
      component_coder = self.components[i]  # for typing
      component_coder.encode_to_stream(attr, out, True)

  def _row_column_encoders(self, columns):
    return [
        RowColumnEncoder.create(
            self.schema.fields[i].type.atomic_type,
            self.components[i],
            columns[name]) for i, name in enumerate(self.field_names)
    ]

  def encode_batch_to_stream(self, columns: Dict[str, np.ndarray], out):
    attrs = self._row_column_encoders(columns)
    n = len(next(iter(columns.values())))
    if self.has_nullable_fields:
      null_flags_py = np.zeros((n, self.num_fields), dtype=np.uint8)
      null_bits_len = (self.num_fields + 7) // 8
      null_bits_py = np.zeros((n, null_bits_len), dtype=np.uint8)
      for i, attr in enumerate(attrs):
        attr_null_flags = attr.null_flags()
        if attr_null_flags is not None and attr_null_flags.any():
          null_flags_py[:, i] = attr_null_flags
          null_bits_py[:, i // 8] |= attr_null_flags << np.uint8(i % 8)
      has_null_bits = (null_bits_py.sum(axis=1) != 0).astype(np.uint8)
      null_bits = null_bits_py
      null_flags = null_flags_py
    else:
      has_null_bits = np.zeros((n, ), dtype=np.uint8)

    for k in range(n):
      out.write_var_int64(self.num_fields)
      if has_null_bits[k]:
        out.write_byte(null_bits_len)
        for i in range(null_bits_len):
          out.write_byte(null_bits[k, i])
      else:
        out.write_byte(0)
      for i in range(self.num_fields):
        if not self.encoding_positions_are_trivial:
          i = self.encoding_positions_argsort[i]
        if has_null_bits[k] and null_flags[k, i]:
          if not self.field_nullable[i]:
            raise ValueError(
                "Attempted to encode null for non-nullable field \"{}\".".
                format(self.schema.fields[i].name))
        else:
          cython.cast(RowColumnEncoder, attrs[i]).encode_to_stream(k, out)

  def decode_from_stream(self, in_stream, nested):
    nvals = in_stream.read_var_int64()
    null_mask_len = in_stream.read_var_int64()
    if null_mask_len:
      # pylint: disable=unused-variable
      null_mask_c = null_mask_py = in_stream.read(null_mask_len)

    # Note that if this coder's schema has *fewer* attributes than the encoded
    # value, we just need to ignore the additional values, which will occur
    # here because we only decode as many values as we have coders for.

    sorted_components = []
    for i in range(min(self.num_fields, nvals)):
      if not self.encoding_positions_are_trivial:
        i = self.encoding_positions_argsort[i]
      if (null_mask_len and i >> 3 < null_mask_len and
          null_mask_c[i >> 3] & (0x01 << (i & 0x07))):
        item = None
      else:
        component_coder = self.components[i]  # for typing
        item = component_coder.decode_from_stream(in_stream, True)
      sorted_components.append(item)

    # If this coder's schema has more attributes than the encoded value, then
    # the schema must have changed. Populate the unencoded fields with nulls.
    while len(sorted_components) < self.num_fields:
      sorted_components.append(None)

    return self.constructor(
        *(
            sorted_components if self.encoding_positions_are_trivial else
            [sorted_components[i] for i in self.encoding_positions]))

  def decode_batch_from_stream(self, dest: Dict[str, np.ndarray], in_stream):
    attrs = self._row_column_encoders(dest)
    n = len(next(iter(dest.values())))
    for k in range(n):
      if in_stream.size() == 0:
        break
      nvals = in_stream.read_var_int64()
      null_mask_len = in_stream.read_var_int64()
      if null_mask_len:
        # pylint: disable=unused-variable
        null_mask_c = null_mask = in_stream.read(null_mask_len)

      for i in range(min(self.num_fields, nvals)):
        if not self.encoding_positions_are_trivial:
          i = self.encoding_positions_argsort[i]
        if (null_mask_len and i >> 3 < null_mask_len and
            null_mask_c[i >> 3] & (0x01 << (i & 0x07))):
          continue
        else:
          cython.cast(RowColumnEncoder,
                      attrs[i]).decode_from_stream(k, in_stream)
    else:
      # Loop variable will be n-1 on normal exit.
      k = n

    for attr in attrs:
      attr.finalize_write()
    return k


class LogicalTypeCoderImpl(StreamCoderImpl):
  def __init__(self, logical_type, representation_coder):
    self.logical_type = logical_type
    self.representation_coder = representation_coder.get_impl()

  def encode_to_stream(self, value, out, nested):
    return self.representation_coder.encode_to_stream(
        self.logical_type.to_representation_type(value), out, nested)

  def decode_from_stream(self, in_stream, nested):
    return self.logical_type.to_language_type(
        self.representation_coder.decode_from_stream(in_stream, nested))


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

  For interoperability with Java SDK, encoding needs to match that of the Java
  SDK BigIntegerCoder."""
  def encode_to_stream(self, value, out, nested):
    # type: (int, create_OutputStream, bool) -> None
    if value < 0:
      byte_length = ((value + 1).bit_length() + 8) // 8
    else:
      byte_length = (value.bit_length() + 8) // 8
    encoded_value = value.to_bytes(
        length=byte_length, byteorder='big', signed=True)
    out.write(encoded_value, nested)

  def decode_from_stream(self, in_stream, nested):
    # type: (create_InputStream, bool) -> int
    encoded_value = in_stream.read_all(nested)
    return int.from_bytes(encoded_value, byteorder='big', signed=True)


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

  For interoperability with Java SDK, encoding needs to match that of the Java
  SDK BigDecimalCoder."""

  BIG_INT_CODER_IMPL = BigIntegerCoderImpl()

  def encode_to_stream(self, value, out, nested):
    # type: (decimal.Decimal, create_OutputStream, bool) -> None
    scale = -value.as_tuple().exponent  # type: ignore[operator]
    int_value = int(value.scaleb(scale))
    out.write_var_int64(scale)
    self.BIG_INT_CODER_IMPL.encode_to_stream(int_value, out, nested)

  def decode_from_stream(self, in_stream, nested):
    # type: (create_InputStream, bool) -> decimal.Decimal
    scale = in_stream.read_var_int64()
    int_value = self.BIG_INT_CODER_IMPL.decode_from_stream(in_stream, nested)
    value = decimal.Decimal(int_value).scaleb(-scale)
    return value