sdks/python/apache_beam/io/parquetio.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.
#
"""``PTransforms`` for reading from and writing to Parquet files.

Provides two read ``PTransform``\\s, ``ReadFromParquet`` and
``ReadAllFromParquet``, that produces a ``PCollection`` of records.
Each record of this ``PCollection`` will contain a single record read from
a Parquet file. Records that are of simple types will be mapped into
corresponding Python types. The actual parquet file operations are done by
pyarrow. Source splitting is supported at row group granularity.

Additionally, this module provides a write ``PTransform`` ``WriteToParquet``
that can be used to write a given ``PCollection`` of Python objects to a
Parquet file.
"""
from __future__ import absolute_import

from functools import partial

from apache_beam.io import filebasedsink
from apache_beam.io import filebasedsource
from apache_beam.io.filesystem import CompressionTypes
from apache_beam.io.iobase import RangeTracker
from apache_beam.io.iobase import Read
from apache_beam.io.iobase import Write
from apache_beam.transforms import DoFn
from apache_beam.transforms import ParDo
from apache_beam.transforms import PTransform

try:
  import pyarrow as pa
  import pyarrow.parquet as pq
except ImportError:
  pa = None
  pq = None

__all__ = ['ReadFromParquet', 'ReadAllFromParquet', 'ReadFromParquetBatched',
           'ReadAllFromParquetBatched', 'WriteToParquet']


class _ArrowTableToRowDictionaries(DoFn):
  """ A DoFn that consumes an Arrow table and yields a python dictionary for
  each row in the table."""
  def process(self, table):
    num_rows = table.num_rows
    data_items = table.to_pydict().items()
    for n in range(num_rows):
      row = {}
      for column, values in data_items:
        row[column] = values[n]
      yield row


class ReadFromParquetBatched(PTransform):
  """A :class:`~apache_beam.transforms.ptransform.PTransform` for reading
     Parquet files as a `PCollection` of `pyarrow.Table`. This `PTransform` is
     currently experimental. No backward-compatibility guarantees."""

  def __init__(self, file_pattern=None, min_bundle_size=0,
               validate=True, columns=None):
    """ Initializes :class:`~ReadFromParquetBatched`

    An alternative to :class:`~ReadFromParquet` that yields each row group from
    the Parquet file as a `pyarrow.Table`.  These Table instances can be
    processed directly, or converted to a pandas DataFrame for processing.  For
    more information on supported types and schema, please see the pyarrow
    documentation.

    .. testcode::

      with beam.Pipeline() as p:
        dataframes = p \\
            | 'Read' >> beam.io.ReadFromParquetBatched('/mypath/mypqfiles*') \\
            | 'Convert to pandas' >> beam.Map(lambda table: table.to_pandas())

    .. NOTE: We're not actually interested in this error; but if we get here,
       it means that the way of calling this transform hasn't changed.

    .. testoutput::
      :hide:

      Traceback (most recent call last):
       ...
      IOError: No files found based on the file pattern

    See also: :class:`~ReadFromParquet`.

    Args:
      file_pattern (str): the file glob to read
      min_bundle_size (int): the minimum size in bytes, to be considered when
        splitting the input into bundles.
      validate (bool): flag to verify that the files exist during the pipeline
        creation time.
      columns (List[str]): list of columns that will be read from files.
        A column name may be a prefix of a nested field, e.g. 'a' will select
        'a.b', 'a.c', and 'a.d.e'
    """

    super(ReadFromParquetBatched, self).__init__()
    self._source = _create_parquet_source(
        file_pattern,
        min_bundle_size,
        validate=validate,
        columns=columns,
    )

  def expand(self, pvalue):
    return pvalue.pipeline | Read(self._source)

  def display_data(self):
    return {'source_dd': self._source}


class ReadFromParquet(PTransform):
  """A :class:`~apache_beam.transforms.ptransform.PTransform` for reading
     Parquet files as a `PCollection` of dictionaries. This `PTransform` is
     currently experimental. No backward-compatibility guarantees."""

  def __init__(self, file_pattern=None, min_bundle_size=0,
               validate=True, columns=None):
    """Initializes :class:`ReadFromParquet`.

    Uses source ``_ParquetSource`` to read a set of Parquet files defined by
    a given file pattern.

    If ``/mypath/myparquetfiles*`` is a file-pattern that points to a set of
    Parquet files, a :class:`~apache_beam.pvalue.PCollection` for the records in
    these Parquet files can be created in the following manner.

    .. testcode::

      with beam.Pipeline() as p:
        records = p | 'Read' >> beam.io.ReadFromParquet('/mypath/mypqfiles*')

    .. NOTE: We're not actually interested in this error; but if we get here,
       it means that the way of calling this transform hasn't changed.

    .. testoutput::
      :hide:

      Traceback (most recent call last):
       ...
      IOError: No files found based on the file pattern

    Each element of this :class:`~apache_beam.pvalue.PCollection` will contain
    a Python dictionary representing a single record. The keys will be of type
    :class:`str` and named after their corresponding column names. The values
    will be of the type defined in the corresponding Parquet schema. Records
    that are of simple types will be mapped into corresponding Python types.
    Records that are of complex types like list and struct will be mapped to
    Python list and dictionary respectively. For more information on supported
    types and schema, please see the pyarrow documentation.

    See also: :class:`~ReadFromParquetBatched`.

    Args:
      file_pattern (str): the file glob to read
      min_bundle_size (int): the minimum size in bytes, to be considered when
        splitting the input into bundles.
      validate (bool): flag to verify that the files exist during the pipeline
        creation time.
      columns (List[str]): list of columns that will be read from files.
        A column name may be a prefix of a nested field, e.g. 'a' will select
        'a.b', 'a.c', and 'a.d.e'
    """
    super(ReadFromParquet, self).__init__()
    self._source = _create_parquet_source(
        file_pattern,
        min_bundle_size,
        validate=validate,
        columns=columns,
    )

  def expand(self, pvalue):
    return pvalue | Read(self._source) | ParDo(_ArrowTableToRowDictionaries())

  def display_data(self):
    return {'source_dd': self._source}


class ReadAllFromParquetBatched(PTransform):
  """A ``PTransform`` for reading ``PCollection`` of Parquet files.

   Uses source ``_ParquetSource`` to read a ``PCollection`` of Parquet files or
   file patterns and produce a ``PCollection`` of ``pyarrow.Table``, one for
   each Parquet file row group. This ``PTransform`` is currently experimental.
   No backward-compatibility guarantees.
  """

  DEFAULT_DESIRED_BUNDLE_SIZE = 64 * 1024 * 1024  # 64MB

  def __init__(self, min_bundle_size=0,
               desired_bundle_size=DEFAULT_DESIRED_BUNDLE_SIZE,
               columns=None,
               label='ReadAllFiles'):
    """Initializes ``ReadAllFromParquet``.

    Args:
      min_bundle_size: the minimum size in bytes, to be considered when
                       splitting the input into bundles.
      desired_bundle_size: the desired size in bytes, to be considered when
                       splitting the input into bundles.
      columns: list of columns that will be read from files. A column name
                       may be a prefix of a nested field, e.g. 'a' will select
                       'a.b', 'a.c', and 'a.d.e'
    """
    super(ReadAllFromParquetBatched, self).__init__()
    source_from_file = partial(
        _create_parquet_source,
        min_bundle_size=min_bundle_size,
        columns=columns
    )
    self._read_all_files = filebasedsource.ReadAllFiles(
        True, CompressionTypes.UNCOMPRESSED, desired_bundle_size,
        min_bundle_size, source_from_file)

    self.label = label

  def expand(self, pvalue):
    return pvalue | self.label >> self._read_all_files


class ReadAllFromParquet(PTransform):
  def __init__(self, **kwargs):
    self._read_batches = ReadAllFromParquetBatched(**kwargs)

  def expand(self, pvalue):
    return pvalue | self._read_batches | ParDo(_ArrowTableToRowDictionaries())


def _create_parquet_source(file_pattern=None,
                           min_bundle_size=0,
                           validate=False,
                           columns=None):
  return \
    _ParquetSource(
        file_pattern=file_pattern,
        min_bundle_size=min_bundle_size,
        validate=validate,
        columns=columns,
    )


class _ParquetUtils(object):
  @staticmethod
  def find_first_row_group_index(pf, start_offset):
    for i in range(_ParquetUtils.get_number_of_row_groups(pf)):
      row_group_start_offset = _ParquetUtils.get_offset(pf, i)
      if row_group_start_offset >= start_offset:
        return i
    return -1

  @staticmethod
  def get_offset(pf, row_group_index):
    first_column_metadata =\
      pf.metadata.row_group(row_group_index).column(0)
    if first_column_metadata.has_dictionary_page:
      return first_column_metadata.dictionary_page_offset
    else:
      return first_column_metadata.data_page_offset

  @staticmethod
  def get_number_of_row_groups(pf):
    return pf.metadata.num_row_groups


class _ParquetSource(filebasedsource.FileBasedSource):
  """A source for reading Parquet files.
  """
  def __init__(self, file_pattern, min_bundle_size, validate, columns):
    super(_ParquetSource, self).__init__(
        file_pattern=file_pattern,
        min_bundle_size=min_bundle_size,
        validate=validate
    )
    self._columns = columns

  def read_records(self, file_name, range_tracker):
    next_block_start = -1

    def split_points_unclaimed(stop_position):
      if next_block_start >= stop_position:
        # Next block starts at or after the suggested stop position. Hence
        # there will not be split points to be claimed for the range ending at
        # suggested stop position.
        return 0
      return RangeTracker.SPLIT_POINTS_UNKNOWN

    range_tracker.set_split_points_unclaimed_callback(split_points_unclaimed)

    start_offset = range_tracker.start_position()
    if start_offset is None:
      start_offset = 0

    with self.open_file(file_name) as f:
      pf = pq.ParquetFile(f)

      # find the first dictionary page (or data page if there's no dictionary
      # page available) offset after the given start_offset. This offset is also
      # the starting offset of any row group since the Parquet specification
      # describes that the data pages always come first before the meta data in
      # each row group.
      index = _ParquetUtils.find_first_row_group_index(pf, start_offset)
      if index != -1:
        next_block_start = _ParquetUtils.get_offset(pf, index)
      else:
        next_block_start = range_tracker.stop_position()
      number_of_row_groups = _ParquetUtils.get_number_of_row_groups(pf)

      while range_tracker.try_claim(next_block_start):
        table = pf.read_row_group(index, self._columns)

        if index + 1 < number_of_row_groups:
          index = index + 1
          next_block_start = _ParquetUtils.get_offset(pf, index)
        else:
          next_block_start = range_tracker.stop_position()

        yield table


class WriteToParquet(PTransform):
  """A ``PTransform`` for writing parquet files.

    This ``PTransform`` is currently experimental. No backward-compatibility
    guarantees.
  """

  def __init__(self,
               file_path_prefix,
               schema,
               row_group_buffer_size=64*1024*1024,
               record_batch_size=1000,
               codec='none',
               use_deprecated_int96_timestamps=False,
               file_name_suffix='',
               num_shards=0,
               shard_name_template=None,
               mime_type='application/x-parquet'):
    """Initialize a WriteToParquet transform.

    Writes parquet files from a :class:`~apache_beam.pvalue.PCollection` of
    records. Each record is a dictionary with keys of a string type that
    represent column names. Schema must be specified like the example below.

    .. testsetup::

      from tempfile import NamedTemporaryFile
      import glob
      import os
      import pyarrow

      filename = NamedTemporaryFile(delete=False).name

    .. testcode::

      with beam.Pipeline() as p:
        records = p | 'Read' >> beam.Create(
            [{'name': 'foo', 'age': 10}, {'name': 'bar', 'age': 20}]
        )
        _ = records | 'Write' >> beam.io.WriteToParquet(filename,
            pyarrow.schema(
                [('name', pyarrow.binary()), ('age', pyarrow.int64())]
            )
        )

    .. testcleanup::

      for output in glob.glob('{}*'.format(filename)):
        os.remove(output)

    For more information on supported types and schema, please see the pyarrow
    document.

    Args:
      file_path_prefix: The file path to write to. The files written will begin
        with this prefix, followed by a shard identifier (see num_shards), and
        end in a common extension, if given by file_name_suffix. In most cases,
        only this argument is specified and num_shards, shard_name_template, and
        file_name_suffix use default values.
      schema: The schema to use, as type of ``pyarrow.Schema``.
      row_group_buffer_size: The byte size of the row group buffer. Note that
        this size is for uncompressed data on the memory and normally much
        bigger than the actual row group size written to a file.
      record_batch_size: The number of records in each record batch. Record
        batch is a basic unit used for storing data in the row group buffer.
        A higher record batch size implies low granularity on a row group buffer
        size. For configuring a row group size based on the number of records,
        set ``row_group_buffer_size`` to 1 and use ``record_batch_size`` to
        adjust the value.
      codec: The codec to use for block-level compression. Any string supported
        by the pyarrow specification is accepted.
      use_deprecated_int96_timestamps: Write nanosecond resolution timestamps to
        INT96 Parquet format. Defaults to False.
      file_name_suffix: Suffix for the files written.
      num_shards: The number of files (shards) used for output. If not set, the
        service will decide on the optimal number of shards.
        Constraining the number of shards is likely to reduce
        the performance of a pipeline.  Setting this value is not recommended
        unless you require a specific number of output files.
      shard_name_template: A template string containing placeholders for
        the shard number and shard count. When constructing a filename for a
        particular shard number, the upper-case letters 'S' and 'N' are
        replaced with the 0-padded shard number and shard count respectively.
        This argument can be '' in which case it behaves as if num_shards was
        set to 1 and only one file will be generated. The default pattern used
        is '-SSSSS-of-NNNNN' if None is passed as the shard_name_template.
      mime_type: The MIME type to use for the produced files, if the filesystem
        supports specifying MIME types.

    Returns:
      A WriteToParquet transform usable for writing.
    """
    super(WriteToParquet, self).__init__()
    self._sink = \
      _create_parquet_sink(
          file_path_prefix,
          schema,
          codec,
          row_group_buffer_size,
          record_batch_size,
          use_deprecated_int96_timestamps,
          file_name_suffix,
          num_shards,
          shard_name_template,
          mime_type
      )

  def expand(self, pcoll):
    return pcoll | Write(self._sink)

  def display_data(self):
    return {'sink_dd': self._sink}


def _create_parquet_sink(file_path_prefix,
                         schema,
                         codec,
                         row_group_buffer_size,
                         record_batch_size,
                         use_deprecated_int96_timestamps,
                         file_name_suffix,
                         num_shards,
                         shard_name_template,
                         mime_type):
  return \
    _ParquetSink(
        file_path_prefix,
        schema,
        codec,
        row_group_buffer_size,
        record_batch_size,
        use_deprecated_int96_timestamps,
        file_name_suffix,
        num_shards,
        shard_name_template,
        mime_type
    )


class _ParquetSink(filebasedsink.FileBasedSink):
  """A sink for parquet files."""

  def __init__(self,
               file_path_prefix,
               schema,
               codec,
               row_group_buffer_size,
               record_batch_size,
               use_deprecated_int96_timestamps,
               file_name_suffix,
               num_shards,
               shard_name_template,
               mime_type):
    super(_ParquetSink, self).__init__(
        file_path_prefix,
        file_name_suffix=file_name_suffix,
        num_shards=num_shards,
        shard_name_template=shard_name_template,
        coder=None,
        mime_type=mime_type,
        # Compression happens at the block level using the supplied codec, and
        # not at the file level.
        compression_type=CompressionTypes.UNCOMPRESSED)
    self._schema = schema
    self._codec = codec
    self._row_group_buffer_size = row_group_buffer_size
    self._use_deprecated_int96_timestamps = use_deprecated_int96_timestamps
    self._buffer = [[] for _ in range(len(schema.names))]
    self._buffer_size = record_batch_size
    self._record_batches = []
    self._record_batches_byte_size = 0
    self._file_handle = None

  def open(self, temp_path):
    self._file_handle = super(_ParquetSink, self).open(temp_path)
    return pq.ParquetWriter(
        self._file_handle, self._schema, compression=self._codec,
        use_deprecated_int96_timestamps=self._use_deprecated_int96_timestamps
    )

  def write_record(self, writer, value):
    if len(self._buffer[0]) >= self._buffer_size:
      self._flush_buffer()

    if self._record_batches_byte_size >= self._row_group_buffer_size:
      self._write_batches(writer)

    # reorder the data in columnar format.
    for i, n in enumerate(self._schema.names):
      self._buffer[i].append(value[n])

  def close(self, writer):
    if len(self._buffer[0]) > 0:
      self._flush_buffer()
    if self._record_batches_byte_size > 0:
      self._write_batches(writer)

    writer.close()
    if self._file_handle:
      self._file_handle.close()
      self._file_handle = None

  def display_data(self):
    res = super(_ParquetSink, self).display_data()
    res['codec'] = str(self._codec)
    res['schema'] = str(self._schema)
    res['row_group_buffer_size'] = str(self._row_group_buffer_size)
    return res

  def _write_batches(self, writer):
    table = pa.Table.from_batches(self._record_batches)
    self._record_batches = []
    self._record_batches_byte_size = 0
    writer.write_table(table)

  def _flush_buffer(self):
    arrays = [[] for _ in range(len(self._schema.names))]
    for x, y in enumerate(self._buffer):
      arrays[x] = pa.array(y, type=self._schema.types[x])
      self._buffer[x] = []
    rb = pa.RecordBatch.from_arrays(arrays, self._schema.names)
    self._record_batches.append(rb)
    size = 0
    for x in arrays:
      for b in x.buffers():
        size = size + b.size
    self._record_batches_byte_size = self._record_batches_byte_size + size