#
# 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 for the Beam DataFrame API.
Sources
#######
This module provides analogs for pandas ``read`` methods, like
:func:`pandas.read_csv`. However Beam sources like :func:`read_csv`
create a Beam :class:`~apache_beam.PTransform`, and return a
:class:`~apache_beam.dataframe.frames.DeferredDataFrame` or
:class:`~apache_beam.dataframe.frames.DeferredSeries` representing the contents
of the referenced file(s) or data source.
The result of these methods must be applied to a :class:`~apache_beam.Pipeline`
object, for example::
df = p | beam.dataframe.io.read_csv(...)
Sinks
#####
This module also defines analogs for pandas sink, or ``to``, methods that
generate a Beam :class:`~apache_beam.PTransform`. Users should prefer calling
these operations from :class:`~apache_beam.dataframe.frames.DeferredDataFrame`
instances (for example with
:meth:`DeferredDataFrame.to_csv
<apache_beam.dataframe.frames.DeferredDataFrame.to_csv>`).
"""
import itertools
import math
import re
from io import BytesIO
from io import StringIO
from io import TextIOWrapper
import pandas as pd
import apache_beam as beam
from apache_beam import io
from apache_beam.dataframe import frame_base
from apache_beam.io import fileio
_DEFAULT_LINES_CHUNKSIZE = 10_000
_DEFAULT_BYTES_CHUNKSIZE = 1 << 20
[docs]@frame_base.with_docs_from(pd)
def read_csv(path, *args, splittable=False, **kwargs):
"""If your files are large and records do not contain quoted newlines, you may
pass the extra argument ``splittable=True`` to enable dynamic splitting for
this read on newlines. Using this option for records that do contain quoted
newlines may result in partial records and data corruption."""
if 'nrows' in kwargs:
raise ValueError('nrows not yet supported')
return _ReadFromPandas(
pd.read_csv,
path,
args,
kwargs,
incremental=True,
splitter=_TextFileSplitter(args, kwargs) if splittable else None)
def _as_pc(df, label=None):
from apache_beam.dataframe import convert # avoid circular import
# TODO(roberwb): Amortize the computation for multiple writes?
return convert.to_pcollection(df, yield_elements='pandas', label=label)
[docs]@frame_base.with_docs_from(pd.DataFrame)
def to_csv(df, path, transform_label=None, *args, **kwargs):
label_pc = f"{transform_label} - ToPCollection" if transform_label \
else f"ToPCollection(df) - {path}"
label_pd = f"{transform_label} - ToPandasDataFrame" if transform_label \
else f"WriteToPandas(df) - {path}"
return _as_pc(df, label_pc) | label_pd >> _WriteToPandas(
'to_csv', path, args, kwargs, incremental=True, binary=False)
[docs]@frame_base.with_docs_from(pd)
def read_fwf(path, *args, **kwargs):
return _ReadFromPandas(
pd.read_fwf,
path,
args,
kwargs,
incremental=True,
binary=False,
splitter=_TextFileSplitter(args, kwargs))
[docs]@frame_base.with_docs_from(pd)
def read_json(path, *args, **kwargs):
if 'nrows' in kwargs:
raise NotImplementedError('nrows not yet supported')
elif kwargs.get('lines', False):
# Work around https://github.com/pandas-dev/pandas/issues/34548.
kwargs = dict(kwargs, nrows=1 << 63)
return _ReadFromPandas(
pd.read_json,
path,
args,
kwargs,
incremental=kwargs.get('lines', False),
splitter=_DelimSplitter(b'\n', _DEFAULT_BYTES_CHUNKSIZE) if kwargs.get(
'lines', False) else None,
binary=False)
[docs]@frame_base.with_docs_from(pd.DataFrame)
def to_json(df, path, orient=None, *args, **kwargs):
if orient is None:
if isinstance(df._expr.proxy(), pd.DataFrame):
orient = 'columns'
elif isinstance(df._expr.proxy(), pd.Series):
orient = 'index'
else:
raise frame_base.WontImplementError('not dataframes or series')
kwargs['orient'] = orient
return _as_pc(df) | _WriteToPandas(
'to_json',
path,
args,
kwargs,
incremental=orient in ('index', 'records', 'values'),
binary=False)
[docs]@frame_base.with_docs_from(pd)
def read_html(path, *args, **kwargs):
return _ReadFromPandas(
lambda *args,
**kwargs: pd.read_html(*args, **kwargs)[0],
path,
args,
kwargs)
[docs]@frame_base.with_docs_from(pd.DataFrame)
def to_html(df, path, *args, **kwargs):
return _as_pc(df) | _WriteToPandas(
'to_html',
path,
args,
kwargs,
incremental=(
df._expr.proxy().index.nlevels == 1 or
not kwargs.get('sparsify', True)),
binary=False)
def _binary_reader(format):
func = getattr(pd, 'read_%s' % format)
result = lambda path, *args, **kwargs: _ReadFromPandas(func, path, args,
kwargs)
result.__name__ = f'read_{format}'
return result
def _binary_writer(format):
result = (
lambda df,
path,
*args,
**kwargs: _as_pc(df) | _WriteToPandas(f'to_{format}', path, args, kwargs))
result.__name__ = f'to_{format}'
return result
for format in ('excel', 'feather', 'parquet', 'stata'):
globals()['read_%s' % format] = frame_base.with_docs_from(pd)(
_binary_reader(format))
globals()['to_%s' % format] = frame_base.with_docs_from(pd.DataFrame)(
_binary_writer(format))
for format in ('sas', 'spss'):
if hasattr(pd, 'read_%s' % format): # Depends on pandas version.
globals()['read_%s' % format] = frame_base.with_docs_from(pd)(
_binary_reader(format))
read_clipboard = frame_base.not_implemented_method(
'read_clipboard', base_type=pd)
to_clipboard = frame_base.not_implemented_method(
'to_clipboard', base_type=pd.DataFrame)
read_msgpack = frame_base.wont_implement_method(
pd, 'read_msgpack', reason="deprecated")
to_msgpack = frame_base.wont_implement_method(
pd.DataFrame, 'to_msgpack', reason="deprecated")
read_hdf = frame_base.wont_implement_method(
pd, 'read_hdf', explanation="because HDF5 is a random access file format")
to_hdf = frame_base.wont_implement_method(
pd.DataFrame,
'to_hdf',
explanation="because HDF5 is a random access file format")
for name in dir(pd):
if name.startswith('read_') and name not in globals():
globals()[name] = frame_base.not_implemented_method(name, base_type=pd)
def _shift_range_index(offset, df):
if isinstance(df.index, pd.RangeIndex):
return df.set_index(df.index + offset)
else:
return df
class _ReadFromPandas(beam.PTransform):
def __init__(
self,
reader,
path,
args,
kwargs,
binary=True,
incremental=False,
splitter=False):
if 'compression' in kwargs:
raise NotImplementedError('compression')
if not isinstance(path, str):
raise frame_base.WontImplementError('non-deferred')
self.reader = reader
self.path = path
self.args = args
self.kwargs = kwargs
self.binary = binary
self.incremental = incremental
self.splitter = splitter
def expand(self, root):
paths_pcoll = root | beam.Create([self.path])
match = io.filesystems.FileSystems.match([self.path], limits=[1])[0]
if not match.metadata_list:
# TODO(https://github.com/apache/beam/issues/20858): This should be
# allowed for streaming pipelines if user provides an explicit schema.
raise FileNotFoundError(f"Found no files that match {self.path!r}")
first_path = match.metadata_list[0].path
with io.filesystems.FileSystems.open(first_path) as handle:
if not self.binary:
handle = TextIOWrapper(handle)
if self.incremental:
sample = next(
self.reader(handle, *self.args, **dict(self.kwargs, chunksize=100)))
else:
sample = self.reader(handle, *self.args, **self.kwargs)
matches_pcoll = paths_pcoll | fileio.MatchAll()
indices_pcoll = (
matches_pcoll.pipeline
| 'DoOnce' >> beam.Create([None])
| beam.Map(
lambda _,
paths: {path: ix
for ix, path in enumerate(sorted(paths))},
paths=beam.pvalue.AsList(
matches_pcoll | beam.Map(lambda match: match.path))))
pcoll = (
matches_pcoll
| beam.Reshuffle()
| fileio.ReadMatches()
| beam.ParDo(
_ReadFromPandasDoFn(
self.reader,
self.args,
self.kwargs,
self.binary,
self.incremental,
self.splitter),
path_indices=beam.pvalue.AsSingleton(indices_pcoll)))
from apache_beam.dataframe import convert
return convert.to_dataframe(pcoll, proxy=sample[:0])
class _Splitter:
def empty_buffer(self):
"""Returns an empty buffer of the right type (string or bytes).
"""
raise NotImplementedError(self)
def read_header(self, handle):
"""Reads the header from handle, which points to the start of the file.
Returns the pair (header, buffer) where buffer contains any part of the
file that was "overread" from handle while seeking the end of header.
"""
raise NotImplementedError(self)
def read_to_record_boundary(self, buffered, handle):
"""Reads the given handle up to the end of the current record.
The buffer argument represents bytes that were read previously; logically
it's as if these were pushed back into handle for reading. If the
record end is within buffered, it's possible that no more bytes will be read
from handle at all.
Returns the pair (remaining_record_bytes, buffer) where buffer contains
any part of the file that was "overread" from handle while seeking the end
of the record.
"""
raise NotImplementedError(self)
class _DelimSplitter(_Splitter):
"""A _Splitter that splits on delimiters between records.
This delimiter is assumed ot never occur within a record.
"""
def __init__(self, delim, read_chunk_size=_DEFAULT_BYTES_CHUNKSIZE):
# Multi-char delimiters would require more care across chunk boundaries.
assert len(delim) == 1
self._delim = delim
self._empty = delim[:0]
self._read_chunk_size = read_chunk_size
def empty_buffer(self):
return self._empty
def read_header(self, handle):
return self._empty, self._empty
def read_to_record_boundary(self, buffered, handle):
if self._delim in buffered:
ix = buffered.index(self._delim) + len(self._delim)
return buffered[:ix], buffered[ix:]
else:
while True:
chunk = handle.read(self._read_chunk_size)
if self._delim in chunk:
ix = chunk.index(self._delim) + len(self._delim)
return buffered + chunk[:ix], chunk[ix:]
elif not chunk:
return buffered, self._empty
else:
buffered += chunk
def _maybe_encode(str_or_bytes):
if isinstance(str_or_bytes, str):
return str_or_bytes.encode('utf-8')
else:
return str_or_bytes
class _TextFileSplitter(_DelimSplitter):
"""Splitter for dynamically sharding CSV files and newline record boundaries.
Currently does not handle quoted newlines, so is off by default, but such
support could be added in the future.
"""
def __init__(self, args, kwargs, read_chunk_size=_DEFAULT_BYTES_CHUNKSIZE):
if args:
# TODO(robertwb): Automatically populate kwargs as we do for df methods.
raise ValueError(
'Non-path arguments must be passed by keyword '
'for splittable csv reads.')
if kwargs.get('skipfooter', 0):
raise ValueError('Splittablility incompatible with skipping footers.')
super().__init__(
_maybe_encode(kwargs.get('lineterminator', b'\n')),
_DEFAULT_BYTES_CHUNKSIZE)
self._kwargs = kwargs
def read_header(self, handle):
if self._kwargs.get('header', 'infer') == 'infer':
if 'names' in self._kwargs:
header = None
else:
header = 0
else:
header = self._kwargs['header']
if header is None:
return self._empty, self._empty
if isinstance(header, int):
max_header = header
else:
max_header = max(header)
skiprows = self._kwargs.get('skiprows', 0)
if isinstance(skiprows, int):
is_skiprow = lambda ix: ix < skiprows
elif callable(skiprows):
is_skiprow = skiprows
elif skiprows is None:
is_skiprow = lambda ix: False
else:
is_skiprow = lambda ix: ix in skiprows
comment = _maybe_encode(self._kwargs.get('comment', None))
if comment:
is_comment = lambda line: line.startswith(comment)
else:
is_comment = lambda line: False
skip_blank_lines = self._kwargs.get('skip_blank_lines', True)
if skip_blank_lines:
is_blank = lambda line: re.match(rb'^\s*$', line)
else:
is_blank = lambda line: False
text_header = b''
rest = b''
skipped = 0
for ix in itertools.count():
line, rest = self.read_to_record_boundary(rest, handle)
text_header += line
if is_skiprow(ix) or is_blank(line) or is_comment(line):
skipped += 1
continue
if ix - skipped == max_header:
return text_header, rest
class _TruncatingFileHandle(object):
"""A wrapper of a file-like object representing the restriction of the
underling handle according to the given SDF restriction tracker, breaking
the file only after the given delimiter.
For example, if the underling restriction is [103, 607) and each line were
exactly 10 characters long (i.e. every 10th charcter was a newline), then this
would give a view of a 500-byte file consisting of bytes bytes 110 to 609
(inclusive) of the underlying file.
As with all SDF trackers, the endpoint may change dynamically during reading.
"""
def __init__(self, underlying, tracker, splitter):
self._underlying = underlying
self._tracker = tracker
self._splitter = splitter
self._empty = self._splitter.empty_buffer()
self._done = False
self._header, self._buffer = self._splitter.read_header(self._underlying)
self._buffer_start_pos = len(self._header)
self._iterator = None
start = self._tracker.current_restriction().start
# Seek to first delimiter after the start position.
if start > len(self._header):
if start > len(self._header) + len(self._buffer):
self._buffer_start_pos = start
self._buffer = self._empty
self._underlying.seek(start)
else:
self._buffer_start_pos = start
self._buffer = self._buffer[start - len(self._header):]
skip, self._buffer = self._splitter.read_to_record_boundary(
self._buffer, self._underlying)
self._buffer_start_pos += len(skip)
def readable(self):
return True
def writable(self):
return False
def seekable(self):
return False
@property
def closed(self):
return False
def __iter__(self):
# For pandas is_file_like.
return self
def __next__(self):
if self._iterator is None:
self._iterator = self._line_iterator()
return next(self._iterator)
def readline(self):
# This attribute is checked, but unused, by pandas.
return next(self)
def _line_iterator(self):
line_start = 0
chunk = self._read()
while True:
line_end = chunk.find(self._splitter._delim, line_start)
while line_end == -1:
more = self._read()
if not more:
if line_start < len(chunk):
yield chunk[line_start:]
return
chunk = chunk[line_start:] + more
line_start = 0
line_end = chunk.find(self._splitter._delim, line_start)
yield chunk[line_start:line_end + 1]
line_start = line_end + 1
def read(self, size=-1):
if self._iterator:
raise NotImplementedError('Cannot call read after iterating.')
return self._read(size)
def _read(self, size=-1):
if self._header:
res = self._header
self._header = None
return res
elif self._done:
return self._empty
elif size == -1:
self._buffer += self._underlying.read()
elif not self._buffer:
self._buffer = self._underlying.read(size)
if not self._buffer:
self._done = True
return self._empty
if self._tracker.try_claim(self._buffer_start_pos + len(self._buffer)):
res = self._buffer
self._buffer = self._empty
self._buffer_start_pos += len(res)
else:
offset = self._tracker.current_restriction().stop - self._buffer_start_pos
if offset <= 0:
res = self._empty
else:
rest, _ = self._splitter.read_to_record_boundary(
self._buffer[offset:], self._underlying)
res = self._buffer[:offset] + rest
self._done = True
return res
class _ReadFromPandasDoFn(beam.DoFn, beam.RestrictionProvider):
def __init__(self, reader, args, kwargs, binary, incremental, splitter):
# avoid pickling issues
if reader.__module__.startswith('pandas.'):
reader = reader.__name__
self.reader = reader
self.args = args
self.kwargs = kwargs
self.binary = binary
self.incremental = incremental
self.splitter = splitter
def initial_restriction(self, readable_file):
return beam.io.restriction_trackers.OffsetRange(
0, readable_file.metadata.size_in_bytes)
def restriction_size(self, readable_file, restriction):
return restriction.size()
def create_tracker(self, restriction):
tracker = beam.io.restriction_trackers.OffsetRestrictionTracker(restriction)
if self.splitter:
return tracker
else:
return beam.io.restriction_trackers.UnsplittableRestrictionTracker(
tracker)
def process(
self, readable_file, path_indices, tracker=beam.DoFn.RestrictionParam()):
reader = self.reader
if isinstance(reader, str):
reader = getattr(pd, self.reader)
indices_per_file = 10**int(math.log(2**63 // len(path_indices), 10))
if readable_file.metadata.size_in_bytes > indices_per_file:
raise RuntimeError(
f'Cannot safely index records from {len(path_indices)} files '
f'of size {readable_file.metadata.size_in_bytes} '
f'as their product is greater than 2^63.')
start_index = (
tracker.current_restriction().start +
path_indices[readable_file.metadata.path] * indices_per_file)
with readable_file.open() as handle:
if self.incremental:
# TODO(robertwb): We could consider trying to get progress for
# non-incremental sources that are read linearly, as long as they
# don't try to seek. This could be deceptive as progress would
# advance to 100% the instant the (large) read was done, discounting
# any downstream processing.
handle = _TruncatingFileHandle(
handle,
tracker,
splitter=self.splitter or
_DelimSplitter(b'\n', _DEFAULT_BYTES_CHUNKSIZE))
if not self.binary:
handle = TextIOWrapper(handle)
if self.incremental:
if 'chunksize' not in self.kwargs:
self.kwargs['chunksize'] = _DEFAULT_LINES_CHUNKSIZE
frames = reader(handle, *self.args, **self.kwargs)
else:
frames = [reader(handle, *self.args, **self.kwargs)]
for df in frames:
yield _shift_range_index(start_index, df)
if not self.incremental:
# Satisfy the SDF contract by claiming the whole range.
# Do this after emitting the frames to avoid advancing progress to 100%
# prior to that.
tracker.try_claim(tracker.current_restriction().stop)
class _WriteToPandas(beam.PTransform):
def __init__(
self, writer, path, args, kwargs, incremental=False, binary=True):
self.writer = writer
self.path = path
self.args = args
self.kwargs = kwargs
self.incremental = incremental
self.binary = binary
def expand(self, pcoll):
dir, name = io.filesystems.FileSystems.split(self.path)
return pcoll | fileio.WriteToFiles(
path=dir,
file_naming=fileio.default_file_naming(name),
sink=lambda _: _WriteToPandasFileSink(
self.writer, self.args, self.kwargs, self.incremental, self.binary))
class _WriteToPandasFileSink(fileio.FileSink):
def __init__(self, writer, args, kwargs, incremental, binary):
if 'compression' in kwargs:
raise NotImplementedError('compression')
self.writer = writer
self.args = args
self.kwargs = kwargs
self.incremental = incremental
self.binary = binary
self.StringOrBytesIO = BytesIO if binary else StringIO
if incremental:
self.write = self.write_record_incremental
self.flush = self.close_incremental
else:
self.write = self.buffer_record
self.flush = self.flush_buffer
def open(self, file_handle):
self.buffer = []
self.empty = self.header = self.footer = None
if not self.binary:
file_handle = TextIOWrapper(file_handle)
self.file_handle = file_handle
def write_to(self, df, file_handle=None):
non_none_handle = file_handle or self.StringOrBytesIO()
getattr(df, self.writer)(non_none_handle, *self.args, **self.kwargs)
if file_handle is None:
return non_none_handle.getvalue()
def write_record_incremental(self, value):
if self.empty is None:
self.empty = self.write_to(value[:0])
if self.header is None and len(value):
def new_value(ix):
if isinstance(ix, tuple):
return (new_value(ix[0]), ) + ix[1:]
else:
return str('x') + '_again'
def change_index(df):
df.index = df.index.map(new_value)
return df
one_row = self.write_to(value[:1])
another_row = self.write_to(change_index(value[:1]))
two_rows = self.write_to(pd.concat([value[:1], change_index(value[:1])]))
for ix, c in enumerate(self.empty):
if one_row[ix] != c:
break
else:
ix = len(self.empty)
self.header = self.empty[:ix]
self.footer = self.empty[ix:]
self.delimiter = two_rows[len(one_row) - len(self.footer):-(
len(another_row) - len(self.header)) or None]
self.file_handle.write(self.header)
self.first = True
if len(value):
if self.first:
self.first = False
else:
self.file_handle.write(self.delimiter)
# IDEA(robertwb): Construct a "truncating" stream wrapper to avoid the
# in-memory copy.
rows = self.write_to(value)
self.file_handle.write(rows[len(self.header):-len(self.footer) or None])
def close_incremental(self):
if self.footer is not None:
self.file_handle.write(self.footer)
elif self.empty is not None:
self.file_handle.write(self.empty)
self.file_handle.flush()
def buffer_record(self, value):
self.buffer.append(value)
def flush_buffer(self):
if self.buffer:
self.write_to(pd.concat(self.buffer), self.file_handle)
self.file_handle.flush()
[docs]class ReadViaPandas(beam.PTransform):
def __init__(
self,
format,
*args,
include_indexes=False,
objects_as_strings=True,
**kwargs):
self._reader = globals()['read_%s' % format](*args, **kwargs)
self._include_indexes = include_indexes
self._objects_as_strings = objects_as_strings
[docs] def expand(self, p):
from apache_beam.dataframe import convert # avoid circular import
df = p | self._reader
if self._objects_as_strings:
for col, t in zip(df.columns, df.dtypes):
if t == object:
df[col] = df[col].astype(pd.StringDtype())
return convert.to_pcollection(df, include_indexes=self._include_indexes)
[docs]class WriteViaPandas(beam.PTransform):
def __init__(self, format, *args, **kwargs):
self._writer_func = globals()['to_%s' % format]
self._args = args
self._kwargs = kwargs
[docs] def expand(self, pcoll):
from apache_beam.dataframe import convert # avoid circular import
return {
'files_written': self._writer_func(
convert.to_dataframe(pcoll), *self._args, **self._kwargs)
| beam.Map(lambda file_result: file_result.file_name).with_output_types(
str)
}