Permalink
1168 lines (921 sloc) 47.6 KB
#
# 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.
#
"""Sources and sinks.
A Source manages record-oriented data input from a particular kind of source
(e.g. a set of files, a database table, etc.). The reader() method of a source
returns a reader object supporting the iterator protocol; iteration yields
raw records of unprocessed, serialized data.
A Sink manages record-oriented data output to a particular kind of sink
(e.g. a set of files, a database table, etc.). The writer() method of a sink
returns a writer object supporting writing records of serialized data to
the sink.
"""
from __future__ import absolute_import
import logging
import math
import random
import uuid
from builtins import object
from builtins import range
from collections import namedtuple
from apache_beam import coders
from apache_beam import pvalue
from apache_beam.portability import common_urns
from apache_beam.portability import python_urns
from apache_beam.portability.api import beam_runner_api_pb2
from apache_beam.pvalue import AsIter
from apache_beam.pvalue import AsSingleton
from apache_beam.transforms import core
from apache_beam.transforms import ptransform
from apache_beam.transforms import window
from apache_beam.transforms.display import DisplayDataItem
from apache_beam.transforms.display import HasDisplayData
from apache_beam.utils import timestamp
from apache_beam.utils import urns
from apache_beam.utils.windowed_value import WindowedValue
__all__ = ['BoundedSource', 'RangeTracker', 'Read', 'RestrictionTracker',
'Sink', 'Write', 'Writer']
# Encapsulates information about a bundle of a source generated when method
# BoundedSource.split() is invoked.
# This is a named 4-tuple that has following fields.
# * weight - a number that represents the size of the bundle. This value will
# be used to compare the relative sizes of bundles generated by the
# current source.
# The weight returned here could be specified using a unit of your
# choice (for example, bundles of sizes 100MB, 200MB, and 700MB may
# specify weights 100, 200, 700 or 1, 2, 7) but all bundles of a
# source should specify the weight using the same unit.
# * source - a BoundedSource object for the bundle.
# * start_position - starting position of the bundle
# * stop_position - ending position of the bundle.
#
# Type for start and stop positions are specific to the bounded source and must
# be consistent throughout.
SourceBundle = namedtuple(
'SourceBundle',
'weight source start_position stop_position')
class SourceBase(HasDisplayData, urns.RunnerApiFn):
"""Base class for all sources that can be passed to beam.io.Read(...).
"""
urns.RunnerApiFn.register_pickle_urn(python_urns.PICKLED_SOURCE)
class BoundedSource(SourceBase):
"""A source that reads a finite amount of input records.
This class defines following operations which can be used to read the source
efficiently.
* Size estimation - method ``estimate_size()`` may return an accurate
estimation in bytes for the size of the source.
* Splitting into bundles of a given size - method ``split()`` can be used to
split the source into a set of sub-sources (bundles) based on a desired
bundle size.
* Getting a RangeTracker - method ``get_range_tracker()`` should return a
``RangeTracker`` object for a given position range for the position type
of the records returned by the source.
* Reading the data - method ``read()`` can be used to read data from the
source while respecting the boundaries defined by a given
``RangeTracker``.
A runner will perform reading the source in two steps.
(1) Method ``get_range_tracker()`` will be invoked with start and end
positions to obtain a ``RangeTracker`` for the range of positions the
runner intends to read. Source must define a default initial start and end
position range. These positions must be used if the start and/or end
positions passed to the method ``get_range_tracker()`` are ``None``
(2) Method read() will be invoked with the ``RangeTracker`` obtained in the
previous step.
**Mutability**
A ``BoundedSource`` object should not be mutated while
its methods (for example, ``read()``) are being invoked by a runner. Runner
implementations may invoke methods of ``BoundedSource`` objects through
multi-threaded and/or reentrant execution modes.
"""
def estimate_size(self):
"""Estimates the size of source in bytes.
An estimate of the total size (in bytes) of the data that would be read
from this source. This estimate is in terms of external storage size,
before performing decompression or other processing.
Returns:
estimated size of the source if the size can be determined, ``None``
otherwise.
"""
raise NotImplementedError
def split(self, desired_bundle_size, start_position=None, stop_position=None):
"""Splits the source into a set of bundles.
Bundles should be approximately of size ``desired_bundle_size`` bytes.
Args:
desired_bundle_size: the desired size (in bytes) of the bundles returned.
start_position: if specified the given position must be used as the
starting position of the first bundle.
stop_position: if specified the given position must be used as the ending
position of the last bundle.
Returns:
an iterator of objects of type 'SourceBundle' that gives information about
the generated bundles.
"""
raise NotImplementedError
def get_range_tracker(self, start_position, stop_position):
"""Returns a RangeTracker for a given position range.
Framework may invoke ``read()`` method with the RangeTracker object returned
here to read data from the source.
Args:
start_position: starting position of the range. If 'None' default start
position of the source must be used.
stop_position: ending position of the range. If 'None' default stop
position of the source must be used.
Returns:
a ``RangeTracker`` for the given position range.
"""
raise NotImplementedError
def read(self, range_tracker):
"""Returns an iterator that reads data from the source.
The returned set of data must respect the boundaries defined by the given
``RangeTracker`` object. For example:
* Returned set of data must be for the range
``[range_tracker.start_position, range_tracker.stop_position)``. Note
that a source may decide to return records that start after
``range_tracker.stop_position``. See documentation in class
``RangeTracker`` for more details. Also, note that framework might
invoke ``range_tracker.try_split()`` to perform dynamic split
operations. range_tracker.stop_position may be updated
dynamically due to successful dynamic split operations.
* Method ``range_tracker.try_split()`` must be invoked for every record
that starts at a split point.
* Method ``range_tracker.record_current_position()`` may be invoked for
records that do not start at split points.
Args:
range_tracker: a ``RangeTracker`` whose boundaries must be respected
when reading data from the source. A runner that reads this
source muss pass a ``RangeTracker`` object that is not
``None``.
Returns:
an iterator of data read by the source.
"""
raise NotImplementedError
def default_output_coder(self):
"""Coder that should be used for the records returned by the source.
Should be overridden by sources that produce objects that can be encoded
more efficiently than pickling.
"""
return coders.registry.get_coder(object)
def is_bounded(self):
return True
class RangeTracker(object):
"""A thread safe object used by Dataflow source framework.
A Dataflow source is defined using a ''BoundedSource'' and a ''RangeTracker''
pair. A ''RangeTracker'' is used by Dataflow source framework to perform
dynamic work rebalancing of position-based sources.
**Position-based sources**
A position-based source is one where the source can be described by a range
of positions of an ordered type and the records returned by the reader can be
described by positions of the same type.
In case a record occupies a range of positions in the source, the most
important thing about the record is the position where it starts.
Defining the semantics of positions for a source is entirely up to the source
class, however the chosen definitions have to obey certain properties in order
to make it possible to correctly split the source into parts, including
dynamic splitting. Two main aspects need to be defined:
1. How to assign starting positions to records.
2. Which records should be read by a source with a range '[A, B)'.
Moreover, reading a range must be *efficient*, i.e., the performance of
reading a range should not significantly depend on the location of the range.
For example, reading the range [A, B) should not require reading all data
before 'A'.
The sections below explain exactly what properties these definitions must
satisfy, and how to use a ``RangeTracker`` with a properly defined source.
**Properties of position-based sources**
The main requirement for position-based sources is *associativity*: reading
records from '[A, B)' and records from '[B, C)' should give the same
records as reading from '[A, C)', where 'A <= B <= C'. This property
ensures that no matter how a range of positions is split into arbitrarily many
sub-ranges, the total set of records described by them stays the same.
The other important property is how the source's range relates to positions of
records in the source. In many sources each record can be identified by a
unique starting position. In this case:
* All records returned by a source '[A, B)' must have starting positions in
this range.
* All but the last record should end within this range. The last record may or
may not extend past the end of the range.
* Records should not overlap.
Such sources should define "read '[A, B)'" as "read from the first record
starting at or after 'A', up to but not including the first record starting
at or after 'B'".
Some examples of such sources include reading lines or CSV from a text file,
reading keys and values from a BigTable, etc.
The concept of *split points* allows to extend the definitions for dealing
with sources where some records cannot be identified by a unique starting
position.
In all cases, all records returned by a source '[A, B)' must *start* at or
after 'A'.
**Split points**
Some sources may have records that are not directly addressable. For example,
imagine a file format consisting of a sequence of compressed blocks. Each
block can be assigned an offset, but records within the block cannot be
directly addressed without decompressing the block. Let us refer to this
hypothetical format as <i>CBF (Compressed Blocks Format)</i>.
Many such formats can still satisfy the associativity property. For example,
in CBF, reading '[A, B)' can mean "read all the records in all blocks whose
starting offset is in '[A, B)'".
To support such complex formats, we introduce the notion of *split points*. We
say that a record is a split point if there exists a position 'A' such that
the record is the first one to be returned when reading the range
'[A, infinity)'. In CBF, the only split points would be the first records
in each block.
Split points allow us to define the meaning of a record's position and a
source's range in all cases:
* For a record that is at a split point, its position is defined to be the
largest 'A' such that reading a source with the range '[A, infinity)'
returns this record.
* Positions of other records are only required to be non-decreasing.
* Reading the source '[A, B)' must return records starting from the first
split point at or after 'A', up to but not including the first split point
at or after 'B'. In particular, this means that the first record returned
by a source MUST always be a split point.
* Positions of split points must be unique.
As a result, for any decomposition of the full range of the source into
position ranges, the total set of records will be the full set of records in
the source, and each record will be read exactly once.
**Consumed positions**
As the source is being read, and records read from it are being passed to the
downstream transforms in the pipeline, we say that positions in the source are
being *consumed*. When a reader has read a record (or promised to a caller
that a record will be returned), positions up to and including the record's
start position are considered *consumed*.
Dynamic splitting can happen only at *unconsumed* positions. If the reader
just returned a record at offset 42 in a file, dynamic splitting can happen
only at offset 43 or beyond, as otherwise that record could be read twice (by
the current reader and by a reader of the task starting at 43).
"""
SPLIT_POINTS_UNKNOWN = object()
def start_position(self):
"""Returns the starting position of the current range, inclusive."""
raise NotImplementedError(type(self))
def stop_position(self):
"""Returns the ending position of the current range, exclusive."""
raise NotImplementedError(type(self))
def try_claim(self, position): # pylint: disable=unused-argument
"""Atomically determines if a record at a split point is within the range.
This method should be called **if and only if** the record is at a split
point. This method may modify the internal state of the ``RangeTracker`` by
updating the last-consumed position to ``position``.
** Thread safety **
Methods of the class ``RangeTracker`` including this method may get invoked
by different threads, hence must be made thread-safe, e.g. by using a single
lock object.
Args:
position: starting position of a record being read by a source.
Returns:
``True``, if the given position falls within the current range, returns
``False`` otherwise.
"""
raise NotImplementedError
def set_current_position(self, position):
"""Updates the last-consumed position to the given position.
A source may invoke this method for records that do not start at split
points. This may modify the internal state of the ``RangeTracker``. If the
record starts at a split point, method ``try_claim()`` **must** be invoked
instead of this method.
Args:
position: starting position of a record being read by a source.
"""
raise NotImplementedError
def position_at_fraction(self, fraction):
"""Returns the position at the given fraction.
Given a fraction within the range [0.0, 1.0) this method will return the
position at the given fraction compared to the position range
[self.start_position, self.stop_position).
** Thread safety **
Methods of the class ``RangeTracker`` including this method may get invoked
by different threads, hence must be made thread-safe, e.g. by using a single
lock object.
Args:
fraction: a float value within the range [0.0, 1.0).
Returns:
a position within the range [self.start_position, self.stop_position).
"""
raise NotImplementedError
def try_split(self, position):
"""Atomically splits the current range.
Determines a position to split the current range, split_position, based on
the given position. In most cases split_position and position will be the
same.
Splits the current range '[self.start_position, self.stop_position)'
into a "primary" part '[self.start_position, split_position)' and a
"residual" part '[split_position, self.stop_position)', assuming the
current last-consumed position is within
'[self.start_position, split_position)' (i.e., split_position has not been
consumed yet).
If successful, updates the current range to be the primary and returns a
tuple (split_position, split_fraction). split_fraction should be the
fraction of size of range '[self.start_position, split_position)' compared
to the original (before split) range
'[self.start_position, self.stop_position)'.
If the split_position has already been consumed, returns ``None``.
** Thread safety **
Methods of the class ``RangeTracker`` including this method may get invoked
by different threads, hence must be made thread-safe, e.g. by using a single
lock object.
Args:
position: suggested position where the current range should try to
be split at.
Returns:
a tuple containing the split position and split fraction if split is
successful. Returns ``None`` otherwise.
"""
raise NotImplementedError
def fraction_consumed(self):
"""Returns the approximate fraction of consumed positions in the source.
** Thread safety **
Methods of the class ``RangeTracker`` including this method may get invoked
by different threads, hence must be made thread-safe, e.g. by using a single
lock object.
Returns:
the approximate fraction of positions that have been consumed by
successful 'try_split()' and 'report_current_position()' calls, or
0.0 if no such calls have happened.
"""
raise NotImplementedError
def split_points(self):
"""Gives the number of split points consumed and remaining.
For a ``RangeTracker`` used by a ``BoundedSource`` (within a
``BoundedSource.read()`` invocation) this method produces a 2-tuple that
gives the number of split points consumed by the ``BoundedSource`` and the
number of split points remaining within the range of the ``RangeTracker``
that has not been consumed by the ``BoundedSource``.
More specifically, given that the position of the current record being read
by ``BoundedSource`` is current_position this method produces a tuple that
consists of
(1) number of split points in the range [self.start_position(),
current_position) without including the split point that is currently being
consumed. This represents the total amount of parallelism in the consumed
part of the source.
(2) number of split points within the range
[current_position, self.stop_position()) including the split point that is
currently being consumed. This represents the total amount of parallelism in
the unconsumed part of the source.
Methods of the class ``RangeTracker`` including this method may get invoked
by different threads, hence must be made thread-safe, e.g. by using a single
lock object.
** General information about consumed and remaining number of split
points returned by this method. **
* Before a source read (``BoundedSource.read()`` invocation) claims the
first split point, number of consumed split points is 0. This condition
holds independent of whether the input is "splittable". A splittable
source is a source that has more than one split point.
* Any source read that has only claimed one split point has 0 consumed
split points since the first split point is the current split point and
is still being processed. This condition holds independent of whether
the input is splittable.
* For an empty source read which never invokes
``RangeTracker.try_claim()``, the consumed number of split points is 0.
This condition holds independent of whether the input is splittable.
* For a source read which has invoked ``RangeTracker.try_claim()`` n
times, the consumed number of split points is n -1.
* If a ``BoundedSource`` sets a callback through function
``set_split_points_unclaimed_callback()``, ``RangeTracker`` can use that
callback when determining remaining number of split points.
* Remaining split points should include the split point that is currently
being consumed by the source read. Hence if the above callback returns
an integer value n, remaining number of split points should be (n + 1).
* After last split point is claimed remaining split points becomes 1,
because this unfinished read itself represents an unfinished split
point.
* After all records of the source has been consumed, remaining number of
split points becomes 0 and consumed number of split points becomes equal
to the total number of split points within the range being read by the
source. This method does not address this condition and will continue to
report number of consumed split points as
("total number of split points" - 1) and number of remaining split
points as 1. A runner that performs the reading of the source can
detect when all records have been consumed and adjust remaining and
consumed number of split points accordingly.
** Examples **
(1) A "perfectly splittable" input which can be read in parallel down to the
individual records.
Consider a perfectly splittable input that consists of 50 split points.
* Before a source read (``BoundedSource.read()`` invocation) claims the
first split point, number of consumed split points is 0 number of
remaining split points is 50.
* After claiming first split point, consumed number of split points is 0
and remaining number of split is 50.
* After claiming split point #30, consumed number of split points is 29
and remaining number of split points is 21.
* After claiming all 50 split points, consumed number of split points is
49 and remaining number of split points is 1.
(2) a "block-compressed" file format such as ``avroio``, in which a block of
records has to be read as a whole, but different blocks can be read in
parallel.
Consider a block compressed input that consists of 5 blocks.
* Before a source read (``BoundedSource.read()`` invocation) claims the
first split point (first block), number of consumed split points is 0
number of remaining split points is 5.
* After claiming first split point, consumed number of split points is 0
and remaining number of split is 5.
* After claiming split point #3, consumed number of split points is 2
and remaining number of split points is 3.
* After claiming all 5 split points, consumed number of split points is
4 and remaining number of split points is 1.
(3) an "unsplittable" input such as a cursor in a database or a gzip
compressed file.
Such an input is considered to have only a single split point. Number of
consumed split points is always 0 and number of remaining split points
is always 1.
By default ``RangeTracker` returns ``RangeTracker.SPLIT_POINTS_UNKNOWN`` for
both consumed and remaining number of split points, which indicates that the
number of split points consumed and remaining is unknown.
Returns:
A pair that gives consumed and remaining number of split points. Consumed
number of split points should be an integer larger than or equal to zero
or ``RangeTracker.SPLIT_POINTS_UNKNOWN``. Remaining number of split points
should be an integer larger than zero or
``RangeTracker.SPLIT_POINTS_UNKNOWN``.
"""
return (RangeTracker.SPLIT_POINTS_UNKNOWN,
RangeTracker.SPLIT_POINTS_UNKNOWN)
def set_split_points_unclaimed_callback(self, callback):
"""Sets a callback for determining the unclaimed number of split points.
By invoking this function, a ``BoundedSource`` can set a callback function
that may get invoked by the ``RangeTracker`` to determine the number of
unclaimed split points. A split point is unclaimed if
``RangeTracker.try_claim()`` method has not been successfully invoked for
that particular split point. The callback function accepts a single
parameter, a stop position for the BoundedSource (stop_position). If the
record currently being consumed by the ``BoundedSource`` is at position
current_position, callback should return the number of split points within
the range (current_position, stop_position). Note that, this should not
include the split point that is currently being consumed by the source.
This function must be implemented by subclasses before being used.
Args:
callback: a function that takes a single parameter, a stop position,
and returns unclaimed number of split points for the source read
operation that is calling this function. Value returned from
callback should be either an integer larger than or equal to
zero or ``RangeTracker.SPLIT_POINTS_UNKNOWN``.
"""
raise NotImplementedError
class Sink(HasDisplayData):
"""This class is deprecated, no backwards-compatibility guarantees.
A resource that can be written to using the ``beam.io.Write`` transform.
Here ``beam`` stands for Apache Beam Python code imported in following manner.
``import apache_beam as beam``.
A parallel write to an ``iobase.Sink`` consists of three phases:
1. A sequential *initialization* phase (e.g., creating a temporary output
directory, etc.)
2. A parallel write phase where workers write *bundles* of records
3. A sequential *finalization* phase (e.g., committing the writes, merging
output files, etc.)
Implementing a new sink requires extending two classes.
1. iobase.Sink
``iobase.Sink`` is an immutable logical description of the location/resource
to write to. Depending on the type of sink, it may contain fields such as the
path to an output directory on a filesystem, a database table name,
etc. ``iobase.Sink`` provides methods for performing a write operation to the
sink described by it. To this end, implementors of an extension of
``iobase.Sink`` must implement three methods:
``initialize_write()``, ``open_writer()``, and ``finalize_write()``.
2. iobase.Writer
``iobase.Writer`` is used to write a single bundle of records. An
``iobase.Writer`` defines two methods: ``write()`` which writes a
single record from the bundle and ``close()`` which is called once
at the end of writing a bundle.
See also ``apache_beam.io.filebasedsink.FileBasedSink`` which provides a
simpler API for writing sinks that produce files.
**Execution of the Write transform**
``initialize_write()``, ``pre_finalize()``, and ``finalize_write()`` are
conceptually called once. However, implementors must
ensure that these methods are *idempotent*, as they may be called multiple
times on different machines in the case of failure/retry. A method may be
called more than once concurrently, in which case it's okay to have a
transient failure (such as due to a race condition). This failure should not
prevent subsequent retries from succeeding.
``initialize_write()`` should perform any initialization that needs to be done
prior to writing to the sink. ``initialize_write()`` may return a result
(let's call this ``init_result``) that contains any parameters it wants to
pass on to its writers about the sink. For example, a sink that writes to a
file system may return an ``init_result`` that contains a dynamically
generated unique directory to which data should be written.
To perform writing of a bundle of elements, Dataflow execution engine will
create an ``iobase.Writer`` using the implementation of
``iobase.Sink.open_writer()``. When invoking ``open_writer()`` execution
engine will provide the ``init_result`` returned by ``initialize_write()``
invocation as well as a *bundle id* (let's call this ``bundle_id``) that is
unique for each invocation of ``open_writer()``.
Execution engine will then invoke ``iobase.Writer.write()`` implementation for
each element that has to be written. Once all elements of a bundle are
written, execution engine will invoke ``iobase.Writer.close()`` implementation
which should return a result (let's call this ``write_result``) that contains
information that encodes the result of the write and, in most cases, some
encoding of the unique bundle id. For example, if each bundle is written to a
unique temporary file, ``close()`` method may return an object that contains
the temporary file name. After writing of all bundles is complete, execution
engine will invoke ``pre_finalize()`` and then ``finalize_write()``
implementation.
The execution of a write transform can be illustrated using following pseudo
code (assume that the outer for loop happens in parallel across many
machines)::
init_result = sink.initialize_write()
write_results = []
for bundle in partition(pcoll):
writer = sink.open_writer(init_result, generate_bundle_id())
for elem in bundle:
writer.write(elem)
write_results.append(writer.close())
pre_finalize_result = sink.pre_finalize(init_result, write_results)
sink.finalize_write(init_result, write_results, pre_finalize_result)
**init_result**
Methods of 'iobase.Sink' should agree on the 'init_result' type that will be
returned when initializing the sink. This type can be a client-defined object
or an existing type. The returned type must be picklable using Dataflow coder
``coders.PickleCoder``. Returning an init_result is optional.
**bundle_id**
In order to ensure fault-tolerance, a bundle may be executed multiple times
(e.g., in the event of failure/retry or for redundancy). However, exactly one
of these executions will have its result passed to the
``iobase.Sink.finalize_write()`` method. Each call to
``iobase.Sink.open_writer()`` is passed a unique bundle id when it is called
by the ``WriteImpl`` transform, so even redundant or retried bundles will have
a unique way of identifying their output.
The bundle id should be used to guarantee that a bundle's output is unique.
This uniqueness guarantee is important; if a bundle is to be output to a file,
for example, the name of the file must be unique to avoid conflicts with other
writers. The bundle id should be encoded in the writer result returned by the
writer and subsequently used by the ``finalize_write()`` method to identify
the results of successful writes.
For example, consider the scenario where a Writer writes files containing
serialized records and the ``finalize_write()`` is to merge or rename these
output files. In this case, a writer may use its unique id to name its output
file (to avoid conflicts) and return the name of the file it wrote as its
writer result. The ``finalize_write()`` will then receive an ``Iterable`` of
output file names that it can then merge or rename using some bundle naming
scheme.
**write_result**
``iobase.Writer.close()`` and ``finalize_write()`` implementations must agree
on type of the ``write_result`` object returned when invoking
``iobase.Writer.close()``. This type can be a client-defined object or
an existing type. The returned type must be picklable using Dataflow coder
``coders.PickleCoder``. Returning a ``write_result`` when
``iobase.Writer.close()`` is invoked is optional but if unique
``write_result`` objects are not returned, sink should, guarantee idempotency
when same bundle is written multiple times due to failure/retry or redundancy.
**More information**
For more information on creating new sinks please refer to the official
documentation at
``https://beam.apache.org/documentation/sdks/python-custom-io#creating-sinks``
"""
def initialize_write(self):
"""Initializes the sink before writing begins.
Invoked before any data is written to the sink.
Please see documentation in ``iobase.Sink`` for an example.
Returns:
An object that contains any sink specific state generated by
initialization. This object will be passed to open_writer() and
finalize_write() methods.
"""
raise NotImplementedError
def open_writer(self, init_result, uid):
"""Opens a writer for writing a bundle of elements to the sink.
Args:
init_result: the result of initialize_write() invocation.
uid: a unique identifier generated by the system.
Returns:
an ``iobase.Writer`` that can be used to write a bundle of records to the
current sink.
"""
raise NotImplementedError
def pre_finalize(self, init_result, writer_results):
"""Pre-finalization stage for sink.
Called after all bundle writes are complete and before finalize_write.
Used to setup and verify filesystem and sink states.
Args:
init_result: the result of ``initialize_write()`` invocation.
writer_results: an iterable containing results of ``Writer.close()``
invocations. This will only contain results of successful writes, and
will only contain the result of a single successful write for a given
bundle.
Returns:
An object that contains any sink specific state generated.
This object will be passed to finalize_write().
"""
raise NotImplementedError
def finalize_write(self, init_result, writer_results,
pre_finalize_result):
"""Finalizes the sink after all data is written to it.
Given the result of initialization and an iterable of results from bundle
writes, performs finalization after writing and closes the sink. Called
after all bundle writes are complete.
The bundle write results that are passed to finalize are those returned by
bundles that completed successfully. Although bundles may have been run
multiple times (for fault-tolerance), only one writer result will be passed
to finalize for each bundle. An implementation of finalize should perform
clean up of any failed and successfully retried bundles. Note that these
failed bundles will not have their writer result passed to finalize, so
finalize should be capable of locating any temporary/partial output written
by failed bundles.
If all retries of a bundle fails, the whole pipeline will fail *without*
finalize_write() being invoked.
A best practice is to make finalize atomic. If this is impossible given the
semantics of the sink, finalize should be idempotent, as it may be called
multiple times in the case of failure/retry or for redundancy.
Note that the iteration order of the writer results is not guaranteed to be
consistent if finalize is called multiple times.
Args:
init_result: the result of ``initialize_write()`` invocation.
writer_results: an iterable containing results of ``Writer.close()``
invocations. This will only contain results of successful writes, and
will only contain the result of a single successful write for a given
bundle.
pre_finalize_result: the result of ``pre_finalize()`` invocation.
"""
raise NotImplementedError
class Writer(object):
"""This class is deprecated, no backwards-compatibility guarantees.
Writes a bundle of elements from a ``PCollection`` to a sink.
A Writer ``iobase.Writer.write()`` writes and elements to the sink while
``iobase.Writer.close()`` is called after all elements in the bundle have been
written.
See ``iobase.Sink`` for more detailed documentation about the process of
writing to a sink.
"""
def write(self, value):
"""Writes a value to the sink using the current writer."""
raise NotImplementedError
def close(self):
"""Closes the current writer.
Please see documentation in ``iobase.Sink`` for an example.
Returns:
An object representing the writes that were performed by the current
writer.
"""
raise NotImplementedError
class Read(ptransform.PTransform):
"""A transform that reads a PCollection."""
def __init__(self, source):
"""Initializes a Read transform.
Args:
source: Data source to read from.
"""
super(Read, self).__init__()
self.source = source
def expand(self, pbegin):
from apache_beam.options.pipeline_options import DebugOptions
from apache_beam.transforms import util
assert isinstance(pbegin, pvalue.PBegin)
self.pipeline = pbegin.pipeline
debug_options = self.pipeline._options.view_as(DebugOptions)
if debug_options.experiments and 'beam_fn_api' in debug_options.experiments:
source = self.source
def split_source(unused_impulse):
total_size = source.estimate_size()
if total_size:
# 1MB = 1 shard, 1GB = 32 shards, 1TB = 1000 shards, 1PB = 32k shards
chunk_size = max(1 << 20, 1000 * int(math.sqrt(total_size)))
else:
chunk_size = 64 << 20 # 64mb
return source.split(chunk_size)
return (
pbegin
| core.Impulse()
| 'Split' >> core.FlatMap(split_source)
| util.Reshuffle()
| 'ReadSplits' >> core.FlatMap(lambda split: split.source.read(
split.source.get_range_tracker(
split.start_position, split.stop_position))))
else:
# Treat Read itself as a primitive.
return pvalue.PCollection(self.pipeline)
def get_windowing(self, unused_inputs):
return core.Windowing(window.GlobalWindows())
def _infer_output_coder(self, input_type=None, input_coder=None):
if isinstance(self.source, BoundedSource):
return self.source.default_output_coder()
else:
return self.source.coder
def display_data(self):
return {'source': DisplayDataItem(self.source.__class__,
label='Read Source'),
'source_dd': self.source}
def to_runner_api_parameter(self, context):
return (common_urns.deprecated_primitives.READ.urn,
beam_runner_api_pb2.ReadPayload(
source=self.source.to_runner_api(context),
is_bounded=beam_runner_api_pb2.IsBounded.BOUNDED
if self.source.is_bounded()
else beam_runner_api_pb2.IsBounded.UNBOUNDED))
@staticmethod
def from_runner_api_parameter(parameter, context):
return Read(SourceBase.from_runner_api(parameter.source, context))
ptransform.PTransform.register_urn(
common_urns.deprecated_primitives.READ.urn,
beam_runner_api_pb2.ReadPayload,
Read.from_runner_api_parameter)
class Write(ptransform.PTransform):
"""A ``PTransform`` that writes to a sink.
A sink should inherit ``iobase.Sink``. Such implementations are
handled using a composite transform that consists of three ``ParDo``s -
(1) a ``ParDo`` performing a global initialization (2) a ``ParDo`` performing
a parallel write and (3) a ``ParDo`` performing a global finalization. In the
case of an empty ``PCollection``, only the global initialization and
finalization will be performed. Currently only batch workflows support custom
sinks.
Example usage::
pcollection | beam.io.Write(MySink())
This returns a ``pvalue.PValue`` object that represents the end of the
Pipeline.
The sink argument may also be a full PTransform, in which case it will be
applied directly. This allows composite sink-like transforms (e.g. a sink
with some pre-processing DoFns) to be used the same as all other sinks.
This transform also supports sinks that inherit ``iobase.NativeSink``. These
are sinks that are implemented natively by the Dataflow service and hence
should not be updated by users. These sinks are processed using a Dataflow
native write transform.
"""
def __init__(self, sink):
"""Initializes a Write transform.
Args:
sink: Data sink to write to.
"""
super(Write, self).__init__()
self.sink = sink
def display_data(self):
return {'sink': self.sink.__class__,
'sink_dd': self.sink}
def expand(self, pcoll):
from apache_beam.runners.dataflow.native_io import iobase as dataflow_io
if isinstance(self.sink, dataflow_io.NativeSink):
# A native sink
return pcoll | 'NativeWrite' >> dataflow_io._NativeWrite(self.sink)
elif isinstance(self.sink, Sink):
# A custom sink
return pcoll | WriteImpl(self.sink)
elif isinstance(self.sink, ptransform.PTransform):
# This allows "composite" sinks to be used like non-composite ones.
return pcoll | self.sink
else:
raise ValueError('A sink must inherit iobase.Sink, iobase.NativeSink, '
'or be a PTransform. Received : %r' % self.sink)
class WriteImpl(ptransform.PTransform):
"""Implements the writing of custom sinks."""
def __init__(self, sink):
super(WriteImpl, self).__init__()
self.sink = sink
def expand(self, pcoll):
do_once = pcoll.pipeline | 'DoOnce' >> core.Create([None])
init_result_coll = do_once | 'InitializeWrite' >> core.Map(
lambda _, sink: sink.initialize_write(), self.sink)
if getattr(self.sink, 'num_shards', 0):
min_shards = self.sink.num_shards
if min_shards == 1:
keyed_pcoll = pcoll | core.Map(lambda x: (None, x))
else:
keyed_pcoll = pcoll | core.ParDo(_RoundRobinKeyFn(min_shards))
write_result_coll = (keyed_pcoll
| core.WindowInto(window.GlobalWindows())
| core.GroupByKey()
| 'WriteBundles' >> core.ParDo(
_WriteKeyedBundleDoFn(self.sink),
AsSingleton(init_result_coll)))
else:
min_shards = 1
write_result_coll = (pcoll
| 'WriteBundles' >>
core.ParDo(_WriteBundleDoFn(self.sink),
AsSingleton(init_result_coll))
| 'Pair' >> core.Map(lambda x: (None, x))
| core.WindowInto(window.GlobalWindows())
| core.GroupByKey()
| 'Extract' >> core.FlatMap(lambda x: x[1]))
# PreFinalize should run before FinalizeWrite, and the two should not be
# fused.
pre_finalize_coll = do_once | 'PreFinalize' >> core.FlatMap(
_pre_finalize,
self.sink,
AsSingleton(init_result_coll),
AsIter(write_result_coll))
return do_once | 'FinalizeWrite' >> core.FlatMap(
_finalize_write,
self.sink,
AsSingleton(init_result_coll),
AsIter(write_result_coll),
min_shards,
AsSingleton(pre_finalize_coll))
class _WriteBundleDoFn(core.DoFn):
"""A DoFn for writing elements to an iobase.Writer.
Opens a writer at the first element and closes the writer at finish_bundle().
"""
def __init__(self, sink):
self.writer = None
self.sink = sink
def display_data(self):
return {'sink_dd': self.sink}
def process(self, element, init_result):
if self.writer is None:
# We ignore UUID collisions here since they are extremely rare.
self.writer = self.sink.open_writer(init_result, str(uuid.uuid4()))
self.writer.write(element)
def finish_bundle(self):
if self.writer is not None:
yield WindowedValue(self.writer.close(),
window.GlobalWindow().max_timestamp(),
[window.GlobalWindow()])
class _WriteKeyedBundleDoFn(core.DoFn):
def __init__(self, sink):
self.sink = sink
def display_data(self):
return {'sink_dd': self.sink}
def process(self, element, init_result):
bundle = element
writer = self.sink.open_writer(init_result, str(uuid.uuid4()))
for e in bundle[1]: # values
writer.write(e)
return [window.TimestampedValue(writer.close(), timestamp.MAX_TIMESTAMP)]
def _pre_finalize(unused_element, sink, init_result, write_results):
return sink.pre_finalize(init_result, write_results)
def _finalize_write(unused_element, sink, init_result, write_results,
min_shards, pre_finalize_results):
write_results = list(write_results)
extra_shards = []
if len(write_results) < min_shards:
logging.debug(
'Creating %s empty shard(s).', min_shards - len(write_results))
for _ in range(min_shards - len(write_results)):
writer = sink.open_writer(init_result, str(uuid.uuid4()))
extra_shards.append(writer.close())
outputs = sink.finalize_write(init_result, write_results + extra_shards,
pre_finalize_results)
if outputs:
return (
window.TimestampedValue(v, timestamp.MAX_TIMESTAMP) for v in outputs)
class _RoundRobinKeyFn(core.DoFn):
def __init__(self, count):
self.count = count
def start_bundle(self):
self.counter = random.randint(0, self.count - 1)
def process(self, element):
self.counter += 1
if self.counter >= self.count:
self.counter -= self.count
yield self.counter, element
class RestrictionTracker(object):
"""Manages concurrent access to a restriction.
Experimental; no backwards-compatibility guarantees.
Keeps track of the restrictions claimed part for a Splittable DoFn.
See following documents for more details.
* https://s.apache.org/splittable-do-fn
* https://s.apache.org/splittable-do-fn-python-sdk
"""
def current_restriction(self):
"""Returns the current restriction.
Returns a restriction accurately describing the full range of work the
current ``DoFn.process()`` call will do, including already completed work.
The current restriction returned by method may be updated dynamically due
to due to concurrent invocation of other methods of the
``RestrictionTracker``, For example, ``checkpoint()``.
** Thread safety **
Methods of the class ``RestrictionTracker`` including this method may get
invoked by different threads, hence must be made thread-safe, e.g. by using
a single lock object.
"""
raise NotImplementedError
def checkpoint(self):
"""Performs a checkpoint of the current restriction.
Signals that the current ``DoFn.process()`` call should terminate as soon as
possible. After this method returns, the tracker MUST refuse all future
claim calls, and ``RestrictionTracker.check_done()`` MUST succeed.
This invocation modifies the value returned by ``current_restriction()``
invocation and returns a restriction representing the rest of the work. The
old value of ``current_restriction()`` is equivalent to the new value of
``current_restriction()`` and the return value of this method invocation
combined.
** Thread safety **
Methods of the class ``RestrictionTracker`` including this method may get
invoked by different threads, hence must be made thread-safe, e.g. by using
a single lock object.
"""
raise NotImplementedError
def check_done(self):
"""Checks whether the restriction has been fully processed.
Called by the runner after iterator returned by ``DoFn.process()`` has been
fully read.
This method must raise a `ValueError` if there is still any unclaimed work
remaining in the restriction when this method is invoked. Exception raised
must have an informative error message.
** Thread safety **
Methods of the class ``RestrictionTracker`` including this method may get
invoked by different threads, hence must be made thread-safe, e.g. by using
a single lock object.
Returns: ``True`` if current restriction has been fully processed.
Raises:
~exceptions.ValueError: if there is still any unclaimed work remaining.
"""
raise NotImplementedError