bigquery_file_loads.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.
#

"""
Functionality to perform file loads into BigQuery for Batch and Streaming
pipelines.

This source is able to work around BigQuery load quotas and limitations. When
destinations are dynamic, or when data for a single job is too large, the data
will be split into multiple jobs.

NOTHING IN THIS FILE HAS BACKWARDS COMPATIBILITY GUARANTEES.
"""

# pytype: skip-file

from __future__ import absolute_import

import hashlib
import logging
import random
import uuid

from future.utils import iteritems

import apache_beam as beam
from apache_beam import pvalue
from apache_beam.io import filesystems as fs
from apache_beam.io.gcp import bigquery_tools
from apache_beam.io.gcp.bigquery_io_metadata import create_bigquery_io_metadata
from apache_beam.options import value_provider as vp
from apache_beam.options.pipeline_options import GoogleCloudOptions
from apache_beam.transforms import trigger
from apache_beam.transforms.display import DisplayDataItem
from apache_beam.transforms.window import GlobalWindows

_LOGGER = logging.getLogger(__name__)

ONE_TERABYTE = (1 << 40)

# The maximum file size for imports is 5TB. We keep our files under that.
_DEFAULT_MAX_FILE_SIZE = 4 * ONE_TERABYTE

_DEFAULT_MAX_WRITERS_PER_BUNDLE = 20

# The maximum size for a single load job is one terabyte
_MAXIMUM_LOAD_SIZE = 15 * ONE_TERABYTE

# Big query only supports up to 10 thousand URIs for a single load job.
_MAXIMUM_SOURCE_URIS = 10 * 1000

# If triggering_frequency is supplied, we will trigger the file write after
# this many records are written.
_FILE_TRIGGERING_RECORD_COUNT = 500000


def _generate_job_name(job_name, job_type, step_name):
  return bigquery_tools.generate_bq_job_name(
      job_name=job_name,
      step_id=step_name,
      job_type=job_type,
      random=random.randint(0, 1000))


def file_prefix_generator(
    with_validation=True, pipeline_gcs_location=None, temp_location=None):
  def _generate_file_prefix(unused_elm):
    # If a gcs location is provided to the pipeline, then we shall use that.
    # Otherwise, we shall use the temp_location from pipeline options.
    gcs_base = pipeline_gcs_location.get()
    if not gcs_base:
      gcs_base = temp_location

    # This will fail at pipeline execution time, but will fail early, as this
    # step doesn't have any dependencies (and thus will be one of the first
    # stages to be run).
    if with_validation and (not gcs_base or not gcs_base.startswith('gs://')):
      raise ValueError(
          'Invalid GCS location: %r.\n'
          'Writing to BigQuery with FILE_LOADS method requires a'
          ' GCS location to be provided to write files to be loaded'
          ' into BigQuery. Please provide a GCS bucket through'
          ' custom_gcs_temp_location in the constructor of WriteToBigQuery'
          ' or the fallback option --temp_location, or pass'
          ' method="STREAMING_INSERTS" to WriteToBigQuery.' % gcs_base)

    prefix_uuid = _bq_uuid()
    return fs.FileSystems.join(gcs_base, 'bq_load', prefix_uuid)

  return _generate_file_prefix


def _make_new_file_writer(
    file_prefix,
    destination,
    file_format,
    schema=None,
    schema_side_inputs=tuple()):
  destination = bigquery_tools.get_hashable_destination(destination)

  # Windows does not allow : on filenames. Replacing with underscore.
  # Other disallowed characters are:
  # https://docs.microsoft.com/en-us/windows/desktop/fileio/naming-a-file
  destination = destination.replace(':', '.')

  directory = fs.FileSystems.join(file_prefix, destination)

  if not fs.FileSystems.exists(directory):
    fs.FileSystems.mkdirs(directory)

  file_name = str(uuid.uuid4())
  file_path = fs.FileSystems.join(file_prefix, destination, file_name)

  if file_format == bigquery_tools.FileFormat.AVRO:
    if callable(schema):
      schema = schema(destination, *schema_side_inputs)
    elif isinstance(schema, vp.ValueProvider):
      schema = schema.get()

    writer = bigquery_tools.AvroRowWriter(
        fs.FileSystems.create(file_path, "application/avro"), schema)
  elif file_format == bigquery_tools.FileFormat.JSON:
    writer = bigquery_tools.JsonRowWriter(
        fs.FileSystems.create(file_path, "application/text"))
  else:
    raise ValueError((
        'Only AVRO and JSON are supported as intermediate formats for '
        'BigQuery WriteRecordsToFile, got: {}.').format(file_format))

  return file_path, writer


def _bq_uuid(seed=None):
  if not seed:
    return str(uuid.uuid4()).replace("-", "")
  else:
    return str(hashlib.md5(seed.encode('utf8')).hexdigest())


class _ShardDestinations(beam.DoFn):
  """Adds a shard number to the key of the KV element.

  Experimental; no backwards compatibility guarantees."""
  DEFAULT_SHARDING_FACTOR = 10

  def __init__(self, sharding_factor=DEFAULT_SHARDING_FACTOR):
    self.sharding_factor = sharding_factor

  def start_bundle(self):
    self._shard_count = random.randrange(self.sharding_factor)

  def process(self, element):
    destination = element[0]
    row = element[1]

    sharded_destination = (
        destination, self._shard_count % self.sharding_factor)
    self._shard_count += 1
    yield (sharded_destination, row)


class WriteRecordsToFile(beam.DoFn):
  """Write input records to files before triggering a load job.

  This transform keeps up to ``max_files_per_bundle`` files open to write to. It
  receives (destination, record) tuples, and it writes the records to different
  files for each destination.

  If there are more than ``max_files_per_bundle`` destinations that we need to
  write to, then those records are grouped by their destination, and later
  written to files by ``WriteGroupedRecordsToFile``.

  It outputs two PCollections.
  """

  UNWRITTEN_RECORD_TAG = 'UnwrittenRecords'
  WRITTEN_FILE_TAG = 'WrittenFiles'

  def __init__(
      self,
      schema,
      max_files_per_bundle=_DEFAULT_MAX_WRITERS_PER_BUNDLE,
      max_file_size=_DEFAULT_MAX_FILE_SIZE,
      file_format=None):
    """Initialize a :class:`WriteRecordsToFile`.

    Args:
      max_files_per_bundle (int): The maximum number of files that can be kept
        open during execution of this step in a worker. This is to avoid over-
        whelming the worker memory.
      max_file_size (int): The maximum size in bytes for a file to be used in
        an export job.

    """
    self.schema = schema
    self.max_files_per_bundle = max_files_per_bundle
    self.max_file_size = max_file_size
    self.file_format = file_format or bigquery_tools.FileFormat.JSON

  def display_data(self):
    return {
        'max_files_per_bundle': self.max_files_per_bundle,
        'max_file_size': str(self.max_file_size),
        'file_format': self.file_format,
    }

  def start_bundle(self):
    self._destination_to_file_writer = {}

  def process(self, element, file_prefix, *schema_side_inputs):
    """Take a tuple with (destination, row) and write to file or spill out.

    Destination may be a ``TableReference`` or a string, and row is a
    Python dictionary for a row to be inserted to BigQuery."""
    destination = bigquery_tools.get_hashable_destination(element[0])
    row = element[1]

    if destination not in self._destination_to_file_writer:
      if len(self._destination_to_file_writer) < self.max_files_per_bundle:
        self._destination_to_file_writer[destination] = _make_new_file_writer(
            file_prefix,
            destination,
            self.file_format,
            self.schema,
            schema_side_inputs)
      else:
        yield pvalue.TaggedOutput(
            WriteRecordsToFile.UNWRITTEN_RECORD_TAG, element)
        return

    (file_path, writer) = self._destination_to_file_writer[destination]

    # TODO(pabloem): Is it possible for this to throw exception?
    writer.write(row)

    file_size = writer.tell()
    if file_size > self.max_file_size:
      writer.close()
      self._destination_to_file_writer.pop(destination)
      yield pvalue.TaggedOutput(
          WriteRecordsToFile.WRITTEN_FILE_TAG,
          (element[0], (file_path, file_size)))

  def finish_bundle(self):
    for destination, file_path_writer in \
      iteritems(self._destination_to_file_writer):
      (file_path, writer) = file_path_writer
      file_size = writer.tell()
      writer.close()
      yield pvalue.TaggedOutput(
          WriteRecordsToFile.WRITTEN_FILE_TAG,
          GlobalWindows.windowed_value((destination, (file_path, file_size))))
    self._destination_to_file_writer = {}


class WriteGroupedRecordsToFile(beam.DoFn):
  """Receives collection of dest-iterable(records), writes it to files.

  This is different from ``WriteRecordsToFile`` because it receives records
  grouped by destination. This means that it's not necessary to keep multiple
  file descriptors open, because we know for sure when records for a single
  destination have been written out.

  Experimental; no backwards compatibility guarantees.
  """
  def __init__(
      self, schema, max_file_size=_DEFAULT_MAX_FILE_SIZE, file_format=None):
    self.schema = schema
    self.max_file_size = max_file_size
    self.file_format = file_format or bigquery_tools.FileFormat.JSON

  def process(self, element, file_prefix, *schema_side_inputs):
    destination = element[0]
    rows = element[1]

    file_path, writer = None, None

    for row in rows:
      if writer is None:
        (file_path, writer) = _make_new_file_writer(
            file_prefix,
            destination,
            self.file_format,
            self.schema,
            schema_side_inputs)

      writer.write(row)

      file_size = writer.tell()
      if file_size > self.max_file_size:
        writer.close()
        yield (destination, (file_path, file_size))
        file_path, writer = None, None
    if writer is not None:
      writer.close()
      yield (destination, (file_path, file_size))


class TriggerCopyJobs(beam.DoFn):
  """Launches jobs to copy from temporary tables into the main target table.

  When a job needs to write to multiple destination tables, or when a single
  destination table needs to have multiple load jobs to write to it, files are
  loaded into temporary tables, and those tables are later copied to the
  destination tables.

  This transform emits (destination, job_reference) pairs.

  TODO(BEAM-7822): In file loads method of writing to BigQuery,
    copying from temp_tables to destination_table is not atomic.
    See: https://issues.apache.org/jira/browse/BEAM-7822
  """
  def __init__(
      self,
      create_disposition=None,
      write_disposition=None,
      test_client=None,
      step_name=None):
    self.create_disposition = create_disposition
    self.write_disposition = write_disposition
    self.test_client = test_client
    self._observed_tables = set()
    self.bq_io_metadata = None
    self._step_name = step_name

  def display_data(self):
    return {
        'launchesBigQueryJobs': DisplayDataItem(
            True, label="This Dataflow job launches bigquery jobs.")
    }

  def start_bundle(self):
    self._observed_tables = set()
    self.bq_wrapper = bigquery_tools.BigQueryWrapper(client=self.test_client)
    if not self.bq_io_metadata:
      self.bq_io_metadata = create_bigquery_io_metadata(self._step_name)

  def process(self, element, job_name_prefix=None):
    destination = element[0]
    job_reference = element[1]

    copy_to_reference = bigquery_tools.parse_table_reference(destination)
    if copy_to_reference.projectId is None:
      copy_to_reference.projectId = vp.RuntimeValueProvider.get_value(
          'project', str, '')

    copy_from_reference = bigquery_tools.parse_table_reference(destination)
    copy_from_reference.tableId = job_reference.jobId
    if copy_from_reference.projectId is None:
      copy_from_reference.projectId = vp.RuntimeValueProvider.get_value(
          'project', str, '')

    copy_job_name = '%s_%s' % (
        job_name_prefix,
        _bq_uuid(
            '%s:%s.%s' % (
                copy_from_reference.projectId,
                copy_from_reference.datasetId,
                copy_from_reference.tableId)))

    _LOGGER.info(
        "Triggering copy job from %s to %s",
        copy_from_reference,
        copy_to_reference)
    if copy_to_reference.tableId not in self._observed_tables:
      # When the write_disposition for a job is WRITE_TRUNCATE,
      # multiple copy jobs to the same destination can stump on
      # each other, truncate data, and write to the BQ table over and
      # over.
      # Thus, the first copy job runs with the user's write_disposition,
      # but afterwards, all jobs must always WRITE_APPEND to the table.
      # If they do not, subsequent copy jobs will clear out data appended
      # by previous jobs.
      write_disposition = self.write_disposition
      wait_for_job = True
      self._observed_tables.add(copy_to_reference.tableId)
    else:
      wait_for_job = False
      write_disposition = 'WRITE_APPEND'

    if not self.bq_io_metadata:
      self.bq_io_metadata = create_bigquery_io_metadata(self._step_name)
    job_reference = self.bq_wrapper._insert_copy_job(
        copy_to_reference.projectId,
        copy_job_name,
        copy_from_reference,
        copy_to_reference,
        create_disposition=self.create_disposition,
        write_disposition=write_disposition,
        job_labels=self.bq_io_metadata.add_additional_bq_job_labels())

    if wait_for_job:
      self.bq_wrapper.wait_for_bq_job(job_reference, sleep_duration_sec=10)

    yield (destination, job_reference)


class TriggerLoadJobs(beam.DoFn):
  """Triggers the import jobs to BQ.

  Experimental; no backwards compatibility guarantees.
  """

  TEMP_TABLES = 'TemporaryTables'

  def __init__(
      self,
      schema=None,
      create_disposition=None,
      write_disposition=None,
      test_client=None,
      temporary_tables=False,
      additional_bq_parameters=None,
      source_format=None,
      step_name=None):
    self.schema = schema
    self.test_client = test_client
    self.temporary_tables = temporary_tables
    self.additional_bq_parameters = additional_bq_parameters or {}
    self.source_format = source_format
    self.bq_io_metadata = None
    self._step_name = step_name
    if self.temporary_tables:
      # If we are loading into temporary tables, we rely on the default create
      # and write dispositions, which mean that a new table will be created.
      self.create_disposition = None
      self.write_disposition = None
    else:
      self.create_disposition = create_disposition
      self.write_disposition = write_disposition

  def display_data(self):
    result = {
        'create_disposition': str(self.create_disposition),
        'write_disposition': str(self.write_disposition),
    }
    result['schema'] = str(self.schema)
    result['launchesBigQueryJobs'] = DisplayDataItem(
        True, label="This Dataflow job launches bigquery jobs.")
    return result

  def start_bundle(self):
    self.bq_wrapper = bigquery_tools.BigQueryWrapper(client=self.test_client)
    if not self.bq_io_metadata:
      self.bq_io_metadata = create_bigquery_io_metadata(self._step_name)

  def process(self, element, load_job_name_prefix, *schema_side_inputs):
    # Each load job is assumed to have files respecting these constraints:
    # 1. Total size of all files < 15 TB (Max size for load jobs)
    # 2. Total no. of files in a single load job < 10,000
    # This assumption means that there will always be a single load job
    # triggered for each partition of files.
    destination = element[0]
    files = element[1]

    if callable(self.schema):
      schema = self.schema(destination, *schema_side_inputs)
    elif isinstance(self.schema, vp.ValueProvider):
      schema = self.schema.get()
    else:
      schema = self.schema

    if callable(self.additional_bq_parameters):
      additional_parameters = self.additional_bq_parameters(destination)
    elif isinstance(self.additional_bq_parameters, vp.ValueProvider):
      additional_parameters = self.additional_bq_parameters.get()
    else:
      additional_parameters = self.additional_bq_parameters

    table_reference = bigquery_tools.parse_table_reference(destination)
    if table_reference.projectId is None:
      table_reference.projectId = vp.RuntimeValueProvider.get_value(
          'project', str, '')
    # Load jobs for a single destination are always triggered from the same
    # worker. This means that we can generate a deterministic numbered job id,
    # and not need to worry.
    destination_hash = _bq_uuid(
        '%s:%s.%s' % (
            table_reference.projectId,
            table_reference.datasetId,
            table_reference.tableId))
    uid = _bq_uuid()
    job_name = '%s_%s_%s' % (load_job_name_prefix, destination_hash, uid)
    _LOGGER.debug(
        'Load job has %s files. Job name is %s.', len(files), job_name)

    create_disposition = self.create_disposition
    if self.temporary_tables:
      # If we are using temporary tables, then we must always create the
      # temporary tables, so we replace the create_disposition.
      create_disposition = 'CREATE_IF_NEEDED'
      # For temporary tables, we create a new table with the name with JobId.
      table_reference.tableId = job_name
      yield pvalue.TaggedOutput(TriggerLoadJobs.TEMP_TABLES, table_reference)

    _LOGGER.info(
        'Triggering job %s to load data to BigQuery table %s.'
        'Schema: %s. Additional parameters: %s',
        job_name,
        table_reference,
        schema,
        additional_parameters)
    if not self.bq_io_metadata:
      self.bq_io_metadata = create_bigquery_io_metadata(self._step_name)
    job_reference = self.bq_wrapper.perform_load_job(
        table_reference,
        files,
        job_name,
        schema=schema,
        write_disposition=self.write_disposition,
        create_disposition=create_disposition,
        additional_load_parameters=additional_parameters,
        source_format=self.source_format,
        job_labels=self.bq_io_metadata.add_additional_bq_job_labels())
    yield (destination, job_reference)


class PartitionFiles(beam.DoFn):

  MULTIPLE_PARTITIONS_TAG = 'MULTIPLE_PARTITIONS'
  SINGLE_PARTITION_TAG = 'SINGLE_PARTITION'

  class Partition(object):
    def __init__(self, max_size, max_files, files=None, size=0):
      self.max_size = max_size
      self.max_files = max_files
      self.files = files if files is not None else []
      self.size = size

    def can_accept(self, file_size, no_of_files=1):
      if (((self.size + file_size) <= self.max_size) and
          ((len(self.files) + no_of_files) <= self.max_files)):
        return True
      else:
        return False

    def add(self, file_path, file_size):
      self.files.append(file_path)
      self.size += file_size

  def __init__(self, max_partition_size, max_files_per_partition):
    self.max_partition_size = max_partition_size
    self.max_files_per_partition = max_files_per_partition

  def process(self, element):
    destination = element[0]
    files = element[1]
    partitions = []

    latest_partition = PartitionFiles.Partition(
        self.max_partition_size, self.max_files_per_partition)

    for file_path, file_size in files:
      if latest_partition.can_accept(file_size):
        latest_partition.add(file_path, file_size)
      else:
        partitions.append(latest_partition.files)
        latest_partition = PartitionFiles.Partition(
            self.max_partition_size, self.max_files_per_partition)
        latest_partition.add(file_path, file_size)
    partitions.append(latest_partition.files)

    if len(partitions) > 1:
      output_tag = PartitionFiles.MULTIPLE_PARTITIONS_TAG
    else:
      output_tag = PartitionFiles.SINGLE_PARTITION_TAG

    for partition in partitions:
      yield pvalue.TaggedOutput(output_tag, (destination, partition))


class WaitForBQJobs(beam.DoFn):
  """Takes in a series of BQ job names as side input, and waits for all of them.

  If any job fails, it will fail. If all jobs succeed, it will succeed.

  Experimental; no backwards compatibility guarantees.
  """
  def __init__(self, test_client=None):
    self.test_client = test_client

  def start_bundle(self):
    self.bq_wrapper = bigquery_tools.BigQueryWrapper(client=self.test_client)

  def process(self, element, dest_ids_list):
    job_references = [elm[1] for elm in dest_ids_list]
    for ref in job_references:
      # We must poll repeatedly until the job finishes or fails, thus setting
      # max_retries to 0.
      self.bq_wrapper.wait_for_bq_job(ref, sleep_duration_sec=10, max_retries=0)

    return dest_ids_list  # Pass the list of destination-jobs downstream


class DeleteTablesFn(beam.DoFn):
  def __init__(self, test_client=None):
    self.test_client = test_client

  def start_bundle(self):
    self.bq_wrapper = bigquery_tools.BigQueryWrapper(client=self.test_client)

  def process(self, table_reference):
    _LOGGER.info("Deleting table %s", table_reference)
    table_reference = bigquery_tools.parse_table_reference(table_reference)
    self.bq_wrapper._delete_table(
        table_reference.projectId,
        table_reference.datasetId,
        table_reference.tableId)


class BigQueryBatchFileLoads(beam.PTransform):
  """Takes in a set of elements, and inserts them to BigQuery via batch loads.

  """

  DESTINATION_JOBID_PAIRS = 'destination_load_jobid_pairs'
  DESTINATION_FILE_PAIRS = 'destination_file_pairs'
  DESTINATION_COPY_JOBID_PAIRS = 'destination_copy_jobid_pairs'
  COUNT = 0

  def __init__(
      self,
      destination,
      schema=None,
      custom_gcs_temp_location=None,
      create_disposition=None,
      write_disposition=None,
      triggering_frequency=None,
      temp_file_format=None,
      max_file_size=None,
      max_files_per_bundle=None,
      max_partition_size=None,
      max_files_per_partition=None,
      additional_bq_parameters=None,
      table_side_inputs=None,
      schema_side_inputs=None,
      test_client=None,
      validate=True,
      is_streaming_pipeline=False):
    self.destination = destination
    self.create_disposition = create_disposition
    self.write_disposition = write_disposition
    self.triggering_frequency = triggering_frequency
    self.max_file_size = max_file_size or _DEFAULT_MAX_FILE_SIZE
    self.max_files_per_bundle = (
        max_files_per_bundle or _DEFAULT_MAX_WRITERS_PER_BUNDLE)
    self.max_partition_size = max_partition_size or _MAXIMUM_LOAD_SIZE
    self.max_files_per_partition = (
        max_files_per_partition or _MAXIMUM_SOURCE_URIS)
    if (isinstance(custom_gcs_temp_location, str) or
        custom_gcs_temp_location is None):
      self._custom_gcs_temp_location = vp.StaticValueProvider(
          str, custom_gcs_temp_location or '')
    elif isinstance(custom_gcs_temp_location, vp.ValueProvider):
      self._custom_gcs_temp_location = custom_gcs_temp_location
    else:
      raise ValueError('custom_gcs_temp_location must be str or ValueProvider')

    self.test_client = test_client
    self.schema = schema
    self._temp_file_format = temp_file_format or bigquery_tools.FileFormat.JSON

    # If we have multiple destinations, then we will have multiple load jobs,
    # thus we will need temporary tables for atomicity.
    self.dynamic_destinations = bool(callable(destination))

    self.additional_bq_parameters = additional_bq_parameters or {}
    self.table_side_inputs = table_side_inputs or ()
    self.schema_side_inputs = schema_side_inputs or ()

    self.is_streaming_pipeline = is_streaming_pipeline
    self._validate = validate
    if self._validate:
      self.verify()

  def verify(self):
    if (isinstance(self._custom_gcs_temp_location.get(), vp.StaticValueProvider)
        and not self._custom_gcs_temp_location.get().startswith('gs://')):
      # Only fail if the custom location is provided, and it is not a GCS
      # location.
      raise ValueError(
          'Invalid GCS location: %r.\n'
          'Writing to BigQuery with FILE_LOADS method requires a '
          'GCS location to be provided to write files to be '
          'loaded into BigQuery. Please provide a GCS bucket, or '
          'pass method="STREAMING_INSERTS" to WriteToBigQuery.' %
          self._custom_gcs_temp_location.get())
    if self.is_streaming_pipeline and not self.triggering_frequency:
      raise ValueError(
          'triggering_frequency must be specified to use file'
          'loads in streaming')
    elif not self.is_streaming_pipeline and self.triggering_frequency:
      raise ValueError(
          'triggering_frequency can only be used with file'
          'loads in streaming')

  def _window_fn(self):
    """Set the correct WindowInto PTransform"""

    # The user-supplied triggering_frequency is often chosen to control how
    # many BigQuery load jobs are triggered, to prevent going over BigQuery's
    # daily quota for load jobs. If this is set to a large value, currently we
    # have to buffer all the data until the trigger fires. Instead we ensure
    # that the files are written if a threshold number of records are ready.
    # We use only the user-supplied trigger on the actual BigQuery load.
    # This allows us to offload the data to the filesystem.
    if self.is_streaming_pipeline:
      return beam.WindowInto(beam.window.GlobalWindows(),
                             trigger=trigger.Repeatedly(
                                 trigger.AfterAny(
                                     trigger.AfterProcessingTime(
                                         self.triggering_frequency),
                                     trigger.AfterCount(
                                         _FILE_TRIGGERING_RECORD_COUNT))),
                             accumulation_mode=trigger.AccumulationMode\
                                 .DISCARDING)
    else:
      return beam.WindowInto(beam.window.GlobalWindows())

  def _write_files(self, destination_data_kv_pc, file_prefix_pcv):
    outputs = (
        destination_data_kv_pc
        | beam.ParDo(
            WriteRecordsToFile(
                schema=self.schema,
                max_files_per_bundle=self.max_files_per_bundle,
                max_file_size=self.max_file_size,
                file_format=self._temp_file_format),
            file_prefix_pcv,
            *self.schema_side_inputs).with_outputs(
                WriteRecordsToFile.UNWRITTEN_RECORD_TAG,
                WriteRecordsToFile.WRITTEN_FILE_TAG))

    # A PCollection of (destination, file) tuples. It lists files with records,
    # and the destination each file is meant to be imported into.
    destination_files_kv_pc = outputs[WriteRecordsToFile.WRITTEN_FILE_TAG]

    # A PCollection of (destination, record) tuples. These are later sharded,
    # grouped, and all records for each destination-shard is written to files.
    # This PCollection is necessary because not all records can be written into
    # files in ``WriteRecordsToFile``.
    unwritten_records_pc = outputs[WriteRecordsToFile.UNWRITTEN_RECORD_TAG]

    more_destination_files_kv_pc = (
        unwritten_records_pc
        | beam.ParDo(_ShardDestinations())
        | "GroupShardedRows" >> beam.GroupByKey()
        | "DropShardNumber" >> beam.Map(lambda x: (x[0][0], x[1]))
        | "WriteGroupedRecordsToFile" >> beam.ParDo(
            WriteGroupedRecordsToFile(
                schema=self.schema, file_format=self._temp_file_format),
            file_prefix_pcv,
            *self.schema_side_inputs))

    # TODO(BEAM-9494): Remove the identity transform. We flatten both
    # PCollection paths and use an identity function to work around a
    # flatten optimization issue where the wrong coder is being used.
    all_destination_file_pairs_pc = (
        (destination_files_kv_pc, more_destination_files_kv_pc)
        | "DestinationFilesUnion" >> beam.Flatten()
        | "IdentityWorkaround" >> beam.Map(lambda x: x))

    if self.is_streaming_pipeline:
      # Apply the user's trigger back before we start triggering load jobs
      all_destination_file_pairs_pc = (
          all_destination_file_pairs_pc
          | "ApplyUserTrigger" >> beam.WindowInto(
              beam.window.GlobalWindows(),
              trigger=trigger.Repeatedly(
                  trigger.AfterAll(
                      trigger.AfterProcessingTime(self.triggering_frequency),
                      trigger.AfterCount(1))),
              accumulation_mode=trigger.AccumulationMode.DISCARDING))
    return all_destination_file_pairs_pc

  def _load_data(
      self,
      partitions_using_temp_tables,
      partitions_direct_to_destination,
      load_job_name_pcv,
      copy_job_name_pcv,
      singleton_pc,
      step_name):
    """Load data to BigQuery

    Data is loaded into BigQuery in the following two ways:
      1. Single partition:
         When there is a single partition of files destined to a single
         destination, a single load job is triggered.
      2. Multiple partitions and/or Dynamic Destinations:
         When there are multiple partitions of files destined for a single
         destination or when Dynamic Destinations are used, multiple load jobs
         need to be triggered for each partition/destination. Load Jobs are
         triggered to temporary tables, and those are later copied to the actual
         appropriate destination table. This ensures atomicity when only some
         of the load jobs would fail but not other. If any of them fails, then
         copy jobs are not triggered.
    """
    # Load data using temp tables
    trigger_loads_outputs = (
        partitions_using_temp_tables
        | "TriggerLoadJobsWithTempTables" >> beam.ParDo(
            TriggerLoadJobs(
                schema=self.schema,
                write_disposition=self.write_disposition,
                create_disposition=self.create_disposition,
                test_client=self.test_client,
                temporary_tables=True,
                additional_bq_parameters=self.additional_bq_parameters,
                source_format=self._temp_file_format,
                step_name=step_name),
            load_job_name_pcv,
            *self.schema_side_inputs).with_outputs(
                TriggerLoadJobs.TEMP_TABLES, main='main'))

    temp_tables_load_job_ids_pc = trigger_loads_outputs['main']
    temp_tables_pc = trigger_loads_outputs[TriggerLoadJobs.TEMP_TABLES]

    destination_copy_job_ids_pc = (
        singleton_pc
        | "WaitForTempTableLoadJobs" >> beam.ParDo(
            WaitForBQJobs(self.test_client),
            beam.pvalue.AsList(temp_tables_load_job_ids_pc))
        | beam.ParDo(
            TriggerCopyJobs(
                create_disposition=self.create_disposition,
                write_disposition=self.write_disposition,
                test_client=self.test_client,
                step_name=step_name),
            copy_job_name_pcv))

    finished_copy_jobs_pc = (
        singleton_pc
        | "WaitForCopyJobs" >> beam.ParDo(
            WaitForBQJobs(self.test_client),
            beam.pvalue.AsList(destination_copy_job_ids_pc)))

    _ = (
        finished_copy_jobs_pc
        | "RemoveTempTables/PassTables" >> beam.FlatMap(
            lambda x,
            deleting_tables: deleting_tables,
            pvalue.AsIter(temp_tables_pc))
        | "RemoveTempTables/AddUselessValue" >> beam.Map(lambda x: (x, None))
        | "RemoveTempTables/DeduplicateTables" >> beam.GroupByKey()
        | "RemoveTempTables/GetTableNames" >> beam.Map(lambda elm: elm[0])
        | "RemoveTempTables/Delete" >> beam.ParDo(
            DeleteTablesFn(self.test_client)))

    # Load data directly to destination table
    destination_load_job_ids_pc = (
        partitions_direct_to_destination
        | "TriggerLoadJobsWithoutTempTables" >> beam.ParDo(
            TriggerLoadJobs(
                schema=self.schema,
                write_disposition=self.write_disposition,
                create_disposition=self.create_disposition,
                test_client=self.test_client,
                temporary_tables=False,
                additional_bq_parameters=self.additional_bq_parameters,
                source_format=self._temp_file_format,
                step_name=step_name),
            load_job_name_pcv,
            *self.schema_side_inputs))

    _ = (
        singleton_pc
        | "WaitForDestinationLoadJobs" >> beam.ParDo(
            WaitForBQJobs(self.test_client),
            beam.pvalue.AsList(destination_load_job_ids_pc)))

    destination_load_job_ids_pc = (
        (temp_tables_load_job_ids_pc, destination_load_job_ids_pc)
        | beam.Flatten())

    return destination_load_job_ids_pc, destination_copy_job_ids_pc

  def expand(self, pcoll):
    p = pcoll.pipeline
    try:
      step_name = self.label
    except AttributeError:
      step_name = 'BigQueryBatchFileLoads_%d' % BigQueryBatchFileLoads.COUNT
      BigQueryBatchFileLoads.COUNT += 1

    temp_location = p.options.view_as(GoogleCloudOptions).temp_location
    job_name = (
        p.options.view_as(GoogleCloudOptions).job_name or 'AUTOMATIC_JOB_NAME')

    empty_pc = p | "ImpulseEmptyPC" >> beam.Create([])
    singleton_pc = p | "ImpulseSingleElementPC" >> beam.Create([None])

    load_job_name_pcv = pvalue.AsSingleton(
        singleton_pc
        | "LoadJobNamePrefix" >> beam.Map(
            lambda _: _generate_job_name(
                job_name, bigquery_tools.BigQueryJobTypes.LOAD, 'LOAD_STEP')))

    copy_job_name_pcv = pvalue.AsSingleton(
        singleton_pc
        | "CopyJobNamePrefix" >> beam.Map(
            lambda _: _generate_job_name(
                job_name, bigquery_tools.BigQueryJobTypes.COPY, 'COPY_STEP')))

    file_prefix_pcv = pvalue.AsSingleton(
        singleton_pc
        | "GenerateFilePrefix" >> beam.Map(
            file_prefix_generator(
                self._validate, self._custom_gcs_temp_location, temp_location)))

    destination_data_kv_pc = (
        pcoll
        | "RewindowIntoGlobal" >> self._window_fn()
        | "AppendDestination" >> beam.ParDo(
            bigquery_tools.AppendDestinationsFn(self.destination),
            *self.table_side_inputs))

    all_destination_file_pairs_pc = self._write_files(
        destination_data_kv_pc, file_prefix_pcv)

    grouped_files_pc = (
        all_destination_file_pairs_pc
        | "GroupFilesByTableDestinations" >> beam.GroupByKey())

    partitions = (
        grouped_files_pc
        | beam.ParDo(
            PartitionFiles(
                self.max_partition_size,
                self.max_files_per_partition)).with_outputs(
                    PartitionFiles.MULTIPLE_PARTITIONS_TAG,
                    PartitionFiles.SINGLE_PARTITION_TAG))

    multiple_partitions_per_destination_pc = partitions[
        PartitionFiles.MULTIPLE_PARTITIONS_TAG]
    single_partition_per_destination_pc = partitions[
        PartitionFiles.SINGLE_PARTITION_TAG]

    # When using dynamic destinations, elements with both single as well as
    # multiple partitions are loaded into BigQuery using temporary tables to
    # ensure atomicity.
    if self.dynamic_destinations:
      all_partitions = ((
          multiple_partitions_per_destination_pc,
          single_partition_per_destination_pc)
                        | "FlattenPartitions" >> beam.Flatten())
      destination_load_job_ids_pc, destination_copy_job_ids_pc = (
          self._load_data(all_partitions,
                          empty_pc,
                          load_job_name_pcv,
                          copy_job_name_pcv,
                          singleton_pc,
                          step_name))
    else:
      destination_load_job_ids_pc, destination_copy_job_ids_pc = (
          self._load_data(multiple_partitions_per_destination_pc,
                         single_partition_per_destination_pc,
                         load_job_name_pcv, copy_job_name_pcv, singleton_pc,
                         step_name))

    return {
        self.DESTINATION_JOBID_PAIRS: destination_load_job_ids_pc,
        self.DESTINATION_FILE_PAIRS: all_destination_file_pairs_pc,
        self.DESTINATION_COPY_JOBID_PAIRS: destination_copy_job_ids_pc,
    }