#
# 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.
#
"""PValue, PCollection: one node of a dataflow graph.
A node of a dataflow processing graph is a PValue. Currently, there is only
one type: PCollection (a potentially very large set of arbitrary values).
Once created, a PValue belongs to a pipeline and has an associated
transform (of type PTransform), which describes how the value will be
produced when the pipeline gets executed.
"""
# pytype: skip-file
import collections
import itertools
from typing import TYPE_CHECKING
from typing import Any
from typing import Dict
from typing import Generic
from typing import Iterator
from typing import Optional
from typing import Sequence
from typing import TypeVar
from typing import Union
from apache_beam import coders
from apache_beam import typehints
from apache_beam.internal import pickler
from apache_beam.portability import common_urns
from apache_beam.portability import python_urns
from apache_beam.portability.api import beam_runner_api_pb2
if TYPE_CHECKING:
from apache_beam.transforms import sideinputs
from apache_beam.transforms.core import ParDo
from apache_beam.transforms.core import Windowing
from apache_beam.pipeline import AppliedPTransform
from apache_beam.pipeline import Pipeline
from apache_beam.runners.pipeline_context import PipelineContext
__all__ = [
'PCollection',
'TaggedOutput',
'AsSideInput',
'AsSingleton',
'AsIter',
'AsList',
'AsDict',
'EmptySideInput',
'Row',
]
T = TypeVar('T')
class PValue(object):
"""Base class for PCollection.
Dataflow users should not construct PValue objects directly in their
pipelines.
A PValue has the following main characteristics:
(1) Belongs to a pipeline. Added during object initialization.
(2) Has a transform that can compute the value if executed.
(3) Has a value which is meaningful if the transform was executed.
"""
def __init__(self,
pipeline, # type: Pipeline
tag=None, # type: Optional[str]
element_type=None, # type: Optional[Union[type,typehints.TypeConstraint]]
windowing=None, # type: Optional[Windowing]
is_bounded=True,
):
"""Initializes a PValue with all arguments hidden behind keyword arguments.
Args:
pipeline: Pipeline object for this PValue.
tag: Tag of this PValue.
element_type: The type of this PValue.
"""
self.pipeline = pipeline
self.tag = tag
self.element_type = element_type
# The AppliedPTransform instance for the application of the PTransform
# generating this PValue. The field gets initialized when a transform
# gets applied.
self.producer = None # type: Optional[AppliedPTransform]
self.is_bounded = is_bounded
if windowing:
self._windowing = windowing
self.requires_deterministic_key_coder = None
def __str__(self):
return self._str_internal()
def __repr__(self):
return '<%s at %s>' % (self._str_internal(), hex(id(self)))
def _str_internal(self):
return "%s[%s.%s]" % (
self.__class__.__name__,
self.producer.full_label if self.producer else None,
self.tag)
def apply(self, *args, **kwargs):
"""Applies a transform or callable to a PValue.
Args:
*args: positional arguments.
**kwargs: keyword arguments.
The method will insert the pvalue as the next argument following an
optional first label and a transform/callable object. It will call the
pipeline.apply() method with this modified argument list.
"""
arglist = list(args)
arglist.insert(1, self)
return self.pipeline.apply(*arglist, **kwargs)
def __or__(self, ptransform):
return self.pipeline.apply(ptransform, self)
[docs]class PCollection(PValue, Generic[T]):
"""A multiple values (potentially huge) container.
Dataflow users should not construct PCollection objects directly in their
pipelines.
"""
def __eq__(self, other):
if isinstance(other, PCollection):
return self.tag == other.tag and self.producer == other.producer
def __hash__(self):
return hash((self.tag, self.producer))
@property
def windowing(self):
# type: () -> Windowing
if not hasattr(self, '_windowing'):
assert self.producer is not None and self.producer.transform is not None
self._windowing = self.producer.transform.get_windowing(
self.producer.inputs)
return self._windowing
def __reduce_ex__(self, unused_version):
# Pickling a PCollection is almost always the wrong thing to do, but we
# can't prohibit it as it often gets implicitly picked up (e.g. as part
# of a closure).
return _InvalidUnpickledPCollection, ()
[docs] @staticmethod
def from_(pcoll, is_bounded=None):
# type: (PValue, Optional[bool]) -> PCollection
"""Create a PCollection, using another PCollection as a starting point.
Transfers relevant attributes.
"""
if is_bounded is None:
is_bounded = pcoll.is_bounded
return PCollection(pcoll.pipeline, is_bounded=is_bounded)
[docs] def to_runner_api(self, context):
# type: (PipelineContext) -> beam_runner_api_pb2.PCollection
return beam_runner_api_pb2.PCollection(
unique_name=self._unique_name(),
coder_id=context.coder_id_from_element_type(
self.element_type, self.requires_deterministic_key_coder),
is_bounded=beam_runner_api_pb2.IsBounded.BOUNDED
if self.is_bounded else beam_runner_api_pb2.IsBounded.UNBOUNDED,
windowing_strategy_id=context.windowing_strategies.get_id(
self.windowing))
def _unique_name(self):
# type: () -> str
if self.producer:
return '%d%s.%s' % (
len(self.producer.full_label), self.producer.full_label, self.tag)
else:
return 'PCollection%s' % id(self)
[docs] @staticmethod
def from_runner_api(proto, context):
# type: (beam_runner_api_pb2.PCollection, PipelineContext) -> PCollection
# Producer and tag will be filled in later, the key point is that the same
# object is returned for the same pcollection id.
# We pass None for the PCollection's Pipeline to avoid a cycle during
# deserialization. It will be populated soon after this call, in
# Pipeline.from_runner_api(). This brief period is the only time that
# PCollection.pipeline is allowed to be None.
return PCollection(
None, # type: ignore[arg-type]
element_type=context.element_type_from_coder_id(proto.coder_id),
windowing=context.windowing_strategies.get_by_id(
proto.windowing_strategy_id),
is_bounded=proto.is_bounded == beam_runner_api_pb2.IsBounded.BOUNDED)
class _InvalidUnpickledPCollection(object):
pass
class PBegin(PValue):
"""A pipeline begin marker used as input to create/read transforms.
The class is used internally to represent inputs to Create and Read
transforms. This allows us to have transforms that uniformly take PValue(s)
as inputs.
"""
pass
class PDone(PValue):
"""PDone is the output of a transform that has a trivial result such as Write.
"""
pass
class DoOutputsTuple(object):
"""An object grouping the multiple outputs of a ParDo or FlatMap transform."""
def __init__(self,
pipeline, # type: Pipeline
transform, # type: ParDo
tags, # type: Sequence[str]
main_tag, # type: Optional[str]
allow_unknown_tags=None, # type: Optional[bool]
):
self._pipeline = pipeline
self._tags = tags
self._main_tag = main_tag
self._transform = transform
self._allow_unknown_tags = (
not tags if allow_unknown_tags is None else allow_unknown_tags)
# The ApplyPTransform instance for the application of the multi FlatMap
# generating this value. The field gets initialized when a transform
# gets applied.
self.producer = None # type: Optional[AppliedPTransform]
# Dictionary of PCollections already associated with tags.
self._pcolls = {} # type: Dict[Optional[str], PCollection]
def __str__(self):
return '<%s>' % self._str_internal()
def __repr__(self):
return '<%s at %s>' % (self._str_internal(), hex(id(self)))
def _str_internal(self):
return '%s main_tag=%s tags=%s transform=%s' % (
self.__class__.__name__, self._main_tag, self._tags, self._transform)
def __iter__(self):
# type: () -> Iterator[PCollection]
"""Iterates over tags returning for each call a (tag, pcollection) pair."""
if self._main_tag is not None:
yield self[self._main_tag]
for tag in self._tags:
yield self[tag]
def __getattr__(self, tag):
# type: (str) -> PCollection
# Special methods which may be accessed before the object is
# fully constructed (e.g. in unpickling).
if tag[:2] == tag[-2:] == '__':
return object.__getattr__(self, tag) # type: ignore
return self[tag]
def __getitem__(self, tag):
# type: (Union[int, str, None]) -> PCollection
# Accept int tags so that we can look at Partition tags with the
# same ints that we used in the partition function.
# TODO(gildea): Consider requiring string-based tags everywhere.
# This will require a partition function that does not return ints.
if isinstance(tag, int):
tag = str(tag)
if tag == self._main_tag:
tag = None
elif self._tags and tag not in self._tags and not self._allow_unknown_tags:
raise ValueError(
"Tag '%s' is neither the main tag '%s' "
"nor any of the tags %s" % (tag, self._main_tag, self._tags))
# Check if we accessed this tag before.
if tag in self._pcolls:
return self._pcolls[tag]
assert self.producer is not None
if tag is not None:
self._transform.output_tags.add(tag)
pcoll = PCollection(self._pipeline, tag=tag, element_type=typehints.Any)
# Transfer the producer from the DoOutputsTuple to the resulting
# PCollection.
pcoll.producer = self.producer.parts[0]
# Add this as an output to both the inner ParDo and the outer _MultiParDo
# PTransforms.
if tag not in self.producer.parts[0].outputs:
self.producer.parts[0].add_output(pcoll, tag)
self.producer.add_output(pcoll, tag)
else:
# Main output is output of inner ParDo.
pval = self.producer.parts[0].outputs[None]
assert isinstance(pval,
PCollection), ("DoOutputsTuple should follow a ParDo.")
pcoll = pval
self._pcolls[tag] = pcoll
return pcoll
[docs]class TaggedOutput(object):
"""An object representing a tagged value.
ParDo, Map, and FlatMap transforms can emit values on multiple outputs which
are distinguished by string tags. The DoFn will return plain values
if it wants to emit on the main output and TaggedOutput objects
if it wants to emit a value on a specific tagged output.
"""
def __init__(self, tag, value):
# type: (str, Any) -> None
if not isinstance(tag, str):
raise TypeError(
'Attempting to create a TaggedOutput with non-string tag %s' %
(tag, ))
self.tag = tag
self.value = value
class _UnpickledSideInput(AsSideInput):
def __init__(self, side_input_data):
# type: (SideInputData) -> None
self._data = side_input_data
self._window_mapping_fn = side_input_data.window_mapping_fn
@staticmethod
def _from_runtime_iterable(it, options):
access_pattern = options['data'].access_pattern
if access_pattern == common_urns.side_inputs.ITERABLE.urn:
raw_view = it
elif access_pattern == common_urns.side_inputs.MULTIMAP.urn:
raw_view = collections.defaultdict(list)
for k, v in it:
raw_view[k].append(v)
else:
raise ValueError('Unknown access_pattern: %s' % access_pattern)
return options['data'].view_fn(raw_view)
def _view_options(self):
return {
'data': self._data,
# For non-fn-api runners.
'window_mapping_fn': self._data.window_mapping_fn,
'coder': self._windowed_coder(),
}
def _side_input_data(self):
return self._data
class SideInputData(object):
"""All of the data about a side input except for the bound PCollection."""
def __init__(self,
access_pattern, # type: str
window_mapping_fn, # type: sideinputs.WindowMappingFn
view_fn
):
self.access_pattern = access_pattern
self.window_mapping_fn = window_mapping_fn
self.view_fn = view_fn
def to_runner_api(self, context):
# type: (PipelineContext) -> beam_runner_api_pb2.SideInput
return beam_runner_api_pb2.SideInput(
access_pattern=beam_runner_api_pb2.FunctionSpec(
urn=self.access_pattern),
view_fn=beam_runner_api_pb2.FunctionSpec(
urn=python_urns.PICKLED_VIEWFN,
payload=pickler.dumps(self.view_fn)),
window_mapping_fn=beam_runner_api_pb2.FunctionSpec(
urn=python_urns.PICKLED_WINDOW_MAPPING_FN,
payload=pickler.dumps(self.window_mapping_fn)))
@staticmethod
def from_runner_api(proto, unused_context):
# type: (beam_runner_api_pb2.SideInput, PipelineContext) -> SideInputData
assert proto.view_fn.urn == python_urns.PICKLED_VIEWFN
assert (
proto.window_mapping_fn.urn == python_urns.PICKLED_WINDOW_MAPPING_FN)
return SideInputData(
proto.access_pattern.urn,
pickler.loads(proto.window_mapping_fn.payload),
pickler.loads(proto.view_fn.payload))
[docs]class AsSingleton(AsSideInput):
"""Marker specifying that an entire PCollection is to be used as a side input.
When a PCollection is supplied as a side input to a PTransform, it is
necessary to indicate whether the entire PCollection should be made available
as a PTransform side argument (in the form of an iterable), or whether just
one value should be pulled from the PCollection and supplied as the side
argument (as an ordinary value).
Wrapping a PCollection side input argument to a PTransform in this container
(e.g., data.apply('label', MyPTransform(), AsSingleton(my_side_input) )
selects the latter behavior.
The input PCollection must contain exactly one value per window, unless a
default is given, in which case it may be empty.
"""
_NO_DEFAULT = object()
def __init__(self, pcoll, default_value=_NO_DEFAULT):
# type: (PCollection, Any) -> None
super().__init__(pcoll)
self.default_value = default_value
def __repr__(self):
return 'AsSingleton(%s)' % self.pvalue
def _view_options(self):
base = super()._view_options()
if self.default_value != AsSingleton._NO_DEFAULT:
return dict(base, default=self.default_value)
return base
@staticmethod
def _from_runtime_iterable(it, options):
head = list(itertools.islice(it, 2))
if not head:
return options.get('default', EmptySideInput())
elif len(head) == 1:
return head[0]
raise ValueError(
'PCollection of size %d with more than one element accessed as a '
'singleton view. First two elements encountered are "%s", "%s".' %
(len(head), str(head[0]), str(head[1])))
@property
def element_type(self):
return self.pvalue.element_type
[docs]class AsIter(AsSideInput):
"""Marker specifying that an entire PCollection is to be used as a side input.
When a PCollection is supplied as a side input to a PTransform, it is
necessary to indicate whether the entire PCollection should be made available
as a PTransform side argument (in the form of an iterable), or whether just
one value should be pulled from the PCollection and supplied as the side
argument (as an ordinary value).
Wrapping a PCollection side input argument to a PTransform in this container
(e.g., data.apply('label', MyPTransform(), AsIter(my_side_input) ) selects the
former behavor.
"""
def __repr__(self):
return 'AsIter(%s)' % self.pvalue
@staticmethod
def _from_runtime_iterable(it, options):
return it
def _side_input_data(self):
# type: () -> SideInputData
return SideInputData(
common_urns.side_inputs.ITERABLE.urn,
self._window_mapping_fn,
lambda iterable: iterable)
@property
def element_type(self):
return typehints.Iterable[self.pvalue.element_type]
[docs]class AsList(AsSideInput):
"""Marker specifying that an entire PCollection is to be used as a side input.
Intended for use in side-argument specification---the same places where
AsSingleton and AsIter are used, but forces materialization of this
PCollection as a list.
Args:
pcoll: Input pcollection.
Returns:
An AsList-wrapper around a PCollection whose one element is a list
containing all elements in pcoll.
"""
@staticmethod
def _from_runtime_iterable(it, options):
return list(it)
def _side_input_data(self):
# type: () -> SideInputData
return SideInputData(
common_urns.side_inputs.ITERABLE.urn, self._window_mapping_fn, list)
[docs]class AsDict(AsSideInput):
"""Marker specifying a PCollection to be used as an indexable side input.
Intended for use in side-argument specification---the same places where
AsSingleton and AsIter are used, but returns an interface that allows
key lookup.
Args:
pcoll: Input pcollection. All elements should be key-value pairs (i.e.
2-tuples) with unique keys.
Returns:
An AsDict-wrapper around a PCollection whose one element is a dict with
entries for uniquely-keyed pairs in pcoll.
"""
@staticmethod
def _from_runtime_iterable(it, options):
return dict(it)
def _side_input_data(self):
# type: () -> SideInputData
return SideInputData(
common_urns.side_inputs.ITERABLE.urn, self._window_mapping_fn, dict)
class AsMultiMap(AsSideInput):
"""Marker specifying a PCollection to be used as an indexable side input.
Similar to AsDict, but multiple values may be associated per key, and
the keys are fetched lazily rather than all having to fit in memory.
Intended for use in side-argument specification---the same places where
AsSingleton and AsIter are used, but returns an interface that allows
key lookup.
"""
@staticmethod
def _from_runtime_iterable(it, options):
# Legacy implementation.
result = collections.defaultdict(list)
for k, v in it:
result[k].append(v)
return result
def _side_input_data(self):
# type: () -> SideInputData
return SideInputData(
common_urns.side_inputs.MULTIMAP.urn,
self._window_mapping_fn,
lambda x: x)
def requires_keyed_input(self):
return True
[docs]class Row(object):
"""A dynamic schema'd row object.
This objects attributes are initialized from the keywords passed into its
constructor, e.g. Row(x=3, y=4) will create a Row with two attributes x and y.
More importantly, when a Row object is returned from a `Map`, `FlatMap`, or
`DoFn` type inference is able to deduce the schema of the resulting
PCollection, e.g.
pc | beam.Map(lambda x: Row(x=x, y=0.5 * x))
when applied to a PCollection of ints will produce a PCollection with schema
`(x=int, y=float)`.
Note that in Beam 2.30.0 and later, Row objects are sensitive to field order.
So `Row(x=3, y=4)` is not considered equal to `Row(y=4, x=3)`.
"""
def __init__(self, **kwargs):
self.__dict__.update(kwargs)
[docs] def as_dict(self):
return dict(self.__dict__)
def __iter__(self):
for _, value in self.__dict__.items():
yield value
def __repr__(self):
return 'Row(%s)' % ', '.join('%s=%r' % kv for kv in self.__dict__.items())
def __hash__(self):
return hash(self.__dict__.items())
def __eq__(self, other):
return (
type(self) == type(other) and
len(self.__dict__) == len(other.__dict__) and all(
s == o for s,
o in zip(self.__dict__.items(), other.__dict__.items())))
def __reduce__(self):
return _make_Row, tuple(self.__dict__.items())
def _make_Row(*items):
return Row(**dict(items))