from __future__ import annotations
import contextlib
import math
import warnings
import tlz as toolz
from fsspec.core import get_fs_token_paths
from fsspec.utils import stringify_path
from packaging.version import parse as parse_version
import dask
from dask.base import tokenize
from dask.blockwise import BlockIndex
from dask.dataframe.backends import dataframe_creation_dispatch
from dask.dataframe.core import DataFrame, Scalar
from dask.dataframe.io.io import from_map
from dask.dataframe.io.parquet.utils import (
Engine,
_sort_and_analyze_paths,
_split_user_options,
)
from dask.dataframe.io.utils import DataFrameIOFunction, _is_local_fs
from dask.dataframe.methods import concat
from dask.delayed import Delayed
from dask.highlevelgraph import HighLevelGraph
from dask.layers import DataFrameIOLayer
from dask.utils import apply, import_required, natural_sort_key, parse_bytes
__all__ = ("read_parquet", "to_parquet")
NONE_LABEL = "__null_dask_index__"
# ----------------------------------------------------------------------
# User API
class ParquetFunctionWrapper(DataFrameIOFunction):
"""
Parquet Function-Wrapper Class
Reads parquet data from disk to produce a partition
(given a `part` argument).
"""
def __init__(
self,
engine,
fs,
meta,
columns,
index,
use_nullable_dtypes,
kwargs,
common_kwargs,
):
self.engine = engine
self.fs = fs
self.meta = meta
self._columns = columns
self.index = index
self.use_nullable_dtypes = use_nullable_dtypes
# `kwargs` = user-defined kwargs to be passed
# identically for all partitions.
#
# `common_kwargs` = kwargs set by engine to be
# passed identically for all
# partitions.
self.common_kwargs = toolz.merge(common_kwargs, kwargs or {})
@property
def columns(self):
return self._columns
def project_columns(self, columns):
"""Return a new ParquetFunctionWrapper object
with a sub-column projection.
"""
if columns == self.columns:
return self
return ParquetFunctionWrapper(
self.engine,
self.fs,
self.meta,
columns,
self.index,
self.use_nullable_dtypes,
None, # Already merged into common_kwargs
self.common_kwargs,
)
def __call__(self, part):
if not isinstance(part, list):
part = [part]
return read_parquet_part(
self.fs,
self.engine,
self.meta,
[
# Temporary workaround for HLG serialization bug
# (see: https://github.com/dask/dask/issues/8581)
(p.data["piece"], p.data.get("kwargs", {}))
if hasattr(p, "data")
else (p["piece"], p.get("kwargs", {}))
for p in part
],
self.columns,
self.index,
self.use_nullable_dtypes,
self.common_kwargs,
)
class ToParquetFunctionWrapper:
"""
Parquet Function-Wrapper Class
Writes a DataFrame partition into a distinct parquet
file. When called, the function also requires the
current block index (via ``blockwise.BlockIndex``).
"""
def __init__(
self,
engine,
path,
fs,
partition_on,
write_metadata_file,
i_offset,
name_function,
kwargs_pass,
):
self.engine = engine
self.path = path
self.fs = fs
self.partition_on = partition_on
self.write_metadata_file = write_metadata_file
self.i_offset = i_offset
self.name_function = name_function
self.kwargs_pass = kwargs_pass
# NOTE: __name__ must be with "to-parquet"
# for the name of the resulting `Blockwise`
# layer to begin with "to-parquet"
self.__name__ = "to-parquet"
def __dask_tokenize__(self):
return (
self.engine,
self.path,
self.fs,
self.partition_on,
self.write_metadata_file,
self.i_offset,
self.name_function,
self.kwargs_pass,
)
def __call__(self, df, block_index: tuple[int]):
# Get partition index from block index tuple
part_i = block_index[0]
filename = (
f"part.{part_i + self.i_offset}.parquet"
if self.name_function is None
else self.name_function(part_i + self.i_offset)
)
# Write out data
return self.engine.write_partition(
df,
self.path,
self.fs,
filename,
self.partition_on,
self.write_metadata_file,
**(dict(self.kwargs_pass, head=True) if part_i == 0 else self.kwargs_pass),
)
[docs]@dataframe_creation_dispatch.register_inplace("pandas")
def read_parquet(
path,
columns=None,
filters=None,
categories=None,
index=None,
storage_options=None,
engine="auto",
use_nullable_dtypes: bool = False,
calculate_divisions=None,
ignore_metadata_file=False,
metadata_task_size=None,
split_row_groups=False,
chunksize=None,
aggregate_files=None,
parquet_file_extension=(".parq", ".parquet", ".pq"),
filesystem=None,
**kwargs,
):
"""
Read a Parquet file into a Dask DataFrame
This reads a directory of Parquet data into a Dask.dataframe, one file per
partition. It selects the index among the sorted columns if any exist.
Parameters
----------
path : str or list
Source directory for data, or path(s) to individual parquet files.
Prefix with a protocol like ``s3://`` to read from alternative
filesystems. To read from multiple files you can pass a globstring or a
list of paths, with the caveat that they must all have the same
protocol.
columns : str or list, default None
Field name(s) to read in as columns in the output. By default all
non-index fields will be read (as determined by the pandas parquet
metadata, if present). Provide a single field name instead of a list to
read in the data as a Series.
filters : Union[List[Tuple[str, str, Any]], List[List[Tuple[str, str, Any]]]], default None
List of filters to apply, like ``[[('col1', '==', 0), ...], ...]``.
Using this argument will NOT result in row-wise filtering of the final
partitions unless ``engine="pyarrow"`` is also specified. For
other engines, filtering is only performed at the partition level, that is,
to prevent the loading of some row-groups and/or files.
For the "pyarrow" engine, predicates can be expressed in disjunctive
normal form (DNF). This means that the inner-most tuple describes a single
column predicate. These inner predicates are combined with an AND
conjunction into a larger predicate. The outer-most list then combines all
of the combined filters with an OR disjunction.
Predicates can also be expressed as a ``List[Tuple]``. These are evaluated
as an AND conjunction. To express OR in predicates, one must use the
(preferred for "pyarrow") ``List[List[Tuple]]`` notation.
Note that the "fastparquet" engine does not currently support DNF for
the filtering of partitioned columns (``List[Tuple]`` is required).
index : str, list or False, default None
Field name(s) to use as the output frame index. By default will be
inferred from the pandas parquet file metadata, if present. Use ``False``
to read all fields as columns.
categories : list or dict, default None
For any fields listed here, if the parquet encoding is Dictionary,
the column will be created with dtype category. Use only if it is
guaranteed that the column is encoded as dictionary in all row-groups.
If a list, assumes up to 2**16-1 labels; if a dict, specify the number
of labels expected; if None, will load categories automatically for
data written by dask/fastparquet, not otherwise.
storage_options : dict, default None
Key/value pairs to be passed on to the file-system backend, if any.
Note that the default file-system backend can be configured with the
``filesystem`` argument, described below.
open_file_options : dict, default None
Key/value arguments to be passed along to ``AbstractFileSystem.open``
when each parquet data file is open for reading. Experimental
(optimized) "precaching" for remote file systems (e.g. S3, GCS) can
be enabled by adding ``{"method": "parquet"}`` under the
``"precache_options"`` key. Also, a custom file-open function can be
used (instead of ``AbstractFileSystem.open``), by specifying the
desired function under the ``"open_file_func"`` key.
engine : {'auto', 'pyarrow', 'fastparquet'}, default 'auto'
Parquet library to use. Defaults to 'auto', which uses ``pyarrow`` if
it is installed, and falls back to ``fastparquet`` otherwise.
use_nullable_dtypes : {False, True}
Whether to use extension dtypes for the resulting ``DataFrame``.
``use_nullable_dtypes=True`` is only supported when ``engine="pyarrow"``.
.. note::
Use the ``dataframe.dtype_backend`` config option to select which
dtype implementation to use.
``dataframe.dtype_backend="pandas"`` (the default) will use
pandas' ``numpy``-backed nullable dtypes (e.g. ``Int64``,
``string[python]``, etc.) while ``dataframe.dtype_backend="pyarrow"``
will use ``pyarrow``-backed extension dtypes (e.g. ``int64[pyarrow]``,
``string[pyarrow]``, etc.). ``dataframe.dtype_backend="pyarrow"``
requires ``pandas`` 1.5+.
calculate_divisions : bool, default False
Whether to use min/max statistics from the footer metadata (or global
``_metadata`` file) to calculate divisions for the output DataFrame
collection. Divisions will not be calculated if statistics are missing.
This option will be ignored if ``index`` is not specified and there is
no physical index column specified in the custom "pandas" Parquet
metadata. Note that ``calculate_divisions=True`` may be extremely slow
when no global ``_metadata`` file is present, especially when reading
from remote storage. Set this to ``True`` only when known divisions
are needed for your workload (see :ref:`dataframe-design-partitions`).
ignore_metadata_file : bool, default False
Whether to ignore the global ``_metadata`` file (when one is present).
If ``True``, or if the global ``_metadata`` file is missing, the parquet
metadata may be gathered and processed in parallel. Parallel metadata
processing is currently supported for ``ArrowDatasetEngine`` only.
metadata_task_size : int, default configurable
If parquet metadata is processed in parallel (see ``ignore_metadata_file``
description above), this argument can be used to specify the number of
dataset files to be processed by each task in the Dask graph. If this
argument is set to ``0``, parallel metadata processing will be disabled.
The default values for local and remote filesystems can be specified
with the "metadata-task-size-local" and "metadata-task-size-remote"
config fields, respectively (see "dataframe.parquet").
split_row_groups : bool or int, default False
If True, then each output dataframe partition will correspond to a single
parquet-file row-group. If False, each partition will correspond to a
complete file. If a positive integer value is given, each dataframe
partition will correspond to that number of parquet row-groups (or fewer).
chunksize : int or str, default None
WARNING: The ``chunksize`` argument will be deprecated in the future.
Please use ``split_row_groups`` to specify how many row-groups should be
mapped to each output partition. If you strongly oppose the deprecation of
``chunksize``, please comment at https://github.com/dask/dask/issues/9043".
The desired size of each output ``DataFrame`` partition in terms of total
(uncompressed) parquet storage space. If specified, adjacent row-groups
and/or files will be aggregated into the same output partition until the
cumulative ``total_byte_size`` parquet-metadata statistic reaches this
value. Use `aggregate_files` to enable/disable inter-file aggregation.
aggregate_files : bool or str, default None
WARNING: The ``aggregate_files`` argument will be deprecated in the future.
Please consider using ``from_map`` to create a DataFrame collection with a
custom file-to-partition mapping. If you strongly oppose the deprecation of
``aggregate_files``, comment at https://github.com/dask/dask/issues/9051".
Whether distinct file paths may be aggregated into the same output
partition. This parameter is only used when `chunksize` is specified
or when `split_row_groups` is an integer >1. A setting of True means
that any two file paths may be aggregated into the same output partition,
while False means that inter-file aggregation is prohibited.
For "hive-partitioned" datasets, a "partition"-column name can also be
specified. In this case, we allow the aggregation of any two files
sharing a file path up to, and including, the corresponding directory name.
For example, if ``aggregate_files`` is set to ``"section"`` for the
directory structure below, ``03.parquet`` and ``04.parquet`` may be
aggregated together, but ``01.parquet`` and ``02.parquet`` cannot be.
If, however, ``aggregate_files`` is set to ``"region"``, ``01.parquet``
may be aggregated with ``02.parquet``, and ``03.parquet`` may be aggregated
with ``04.parquet``::
dataset-path/
├── region=1/
│ ├── section=a/
│ │ └── 01.parquet
│ ├── section=b/
│ └── └── 02.parquet
└── region=2/
├── section=a/
│ ├── 03.parquet
└── └── 04.parquet
Note that the default behavior of ``aggregate_files`` is ``False``.
parquet_file_extension: str, tuple[str], or None, default (".parq", ".parquet", ".pq")
A file extension or an iterable of extensions to use when discovering
parquet files in a directory. Files that don't match these extensions
will be ignored. This argument only applies when ``paths`` corresponds
to a directory and no ``_metadata`` file is present (or
``ignore_metadata_file=True``). Passing in ``parquet_file_extension=None``
will treat all files in the directory as parquet files.
The purpose of this argument is to ensure that the engine will ignore
unsupported metadata files (like Spark's '_SUCCESS' and 'crc' files).
It may be necessary to change this argument if the data files in your
parquet dataset do not end in ".parq", ".parquet", or ".pq".
filesystem: "fsspec", "arrow", fsspec.AbstractFileSystem, or pyarrow.fs.FileSystem
Filesystem backend to use. Note that the "fastparquet" engine only
supports "fsspec" or an explicit ``pyarrow.fs.FileSystem`` object.
Default is "fsspec".
dataset: dict, default None
Dictionary of options to use when creating a ``pyarrow.dataset.Dataset``
or ``fastparquet.ParquetFile`` object. These options may include a
"filesystem" key (or "fs" for the "fastparquet" engine) to configure
the desired file-system backend. However, the top-level ``filesystem``
argument will always take precedence.
read: dict, default None
Dictionary of options to pass through to ``engine.read_partitions``
using the ``read`` key-word argument.
arrow_to_pandas: dict, default None
Dictionary of options to use when converting from ``pyarrow.Table`` to
a pandas ``DataFrame`` object. Only used by the "arrow" engine.
**kwargs: dict (of dicts)
Options to pass through to ``engine.read_partitions`` as stand-alone
key-word arguments. Note that these options will be ignored by the
engines defined in ``dask.dataframe``, but may be used by other custom
implementations.
Examples
--------
>>> df = dd.read_parquet('s3://bucket/my-parquet-data') # doctest: +SKIP
See Also
--------
to_parquet
pyarrow.parquet.ParquetDataset
"""
if use_nullable_dtypes:
use_nullable_dtypes = dask.config.get("dataframe.dtype_backend")
# "Pre-deprecation" warning for `chunksize`
if chunksize:
warnings.warn(
"The `chunksize` argument will be deprecated in the future. "
"Please use `split_row_groups` to specify how many row-groups "
"should be mapped to each output partition.\n\n"
"If you strongly oppose the deprecation of `chunksize`, please "
"comment at https://github.com/dask/dask/issues/9043",
FutureWarning,
)
# "Pre-deprecation" warning for `aggregate_files`
if aggregate_files:
warnings.warn(
"The `aggregate_files` argument will be deprecated in the future. "
"Please consider using `from_map` to create a DataFrame collection "
"with a custom file-to-partition mapping.\n\n"
"If you strongly oppose the deprecation of `aggregate_files`, "
"please comment at https://github.com/dask/dask/issues/9051",
FutureWarning,
)
if "read_from_paths" in kwargs:
kwargs.pop("read_from_paths")
warnings.warn(
"`read_from_paths` is no longer supported and will be ignored.",
FutureWarning,
)
# Handle gather_statistics deprecation
if "gather_statistics" in kwargs:
if calculate_divisions is None:
calculate_divisions = kwargs.pop("gather_statistics")
warnings.warn(
"``gather_statistics`` is deprecated and will be removed in a "
"future release. Please use ``calculate_divisions`` instead.",
FutureWarning,
)
else:
warnings.warn(
f"``gather_statistics`` is deprecated. Ignoring this option "
f"in favor of ``calculate_divisions={calculate_divisions}``",
FutureWarning,
)
calculate_divisions = bool(calculate_divisions)
# We support a top-level `parquet_file_extension` kwarg, but
# must check if the deprecated `require_extension` option is
# being passed to the engine. If `parquet_file_extension` is
# set to the default value, and `require_extension` was also
# specified, we will use `require_extension` but warn the user.
if (
"dataset" in kwargs
and "require_extension" in kwargs["dataset"]
and parquet_file_extension == (".parq", ".parquet", ".pq")
):
parquet_file_extension = kwargs["dataset"].pop("require_extension")
warnings.warn(
"require_extension is deprecated, and will be removed from "
"read_parquet in a future release. Please use the top-level "
"parquet_file_extension argument instead.",
FutureWarning,
)
# Store initial function arguments
input_kwargs = {
"columns": columns,
"filters": filters,
"categories": categories,
"index": index,
"storage_options": storage_options,
"engine": engine,
"use_nullable_dtypes": use_nullable_dtypes,
"calculate_divisions": calculate_divisions,
"ignore_metadata_file": ignore_metadata_file,
"metadata_task_size": metadata_task_size,
"split_row_groups": split_row_groups,
"chunksize": chunksize,
"aggregate_files": aggregate_files,
"parquet_file_extension": parquet_file_extension,
"filesystem": filesystem,
**kwargs,
}
if isinstance(columns, str):
input_kwargs["columns"] = [columns]
df = read_parquet(path, **input_kwargs)
return df[columns]
if columns is not None:
columns = list(columns)
if isinstance(engine, str):
engine = get_engine(engine)
if hasattr(path, "name"):
path = stringify_path(path)
# Update input_kwargs
input_kwargs.update({"columns": columns, "engine": engine})
# Process and split user options
(
dataset_options,
read_options,
open_file_options,
other_options,
) = _split_user_options(**kwargs)
# Extract global filesystem and paths
fs, paths, dataset_options, open_file_options = engine.extract_filesystem(
path,
filesystem,
dataset_options,
open_file_options,
storage_options,
)
read_options["open_file_options"] = open_file_options
paths = sorted(paths, key=natural_sort_key) # numeric rather than glob ordering
auto_index_allowed = False
if index is None:
# User is allowing auto-detected index
auto_index_allowed = True
if index and isinstance(index, str):
index = [index]
read_metadata_result = engine.read_metadata(
fs,
paths,
categories=categories,
index=index,
use_nullable_dtypes=use_nullable_dtypes,
gather_statistics=calculate_divisions,
filters=filters,
split_row_groups=split_row_groups,
chunksize=chunksize,
aggregate_files=aggregate_files,
ignore_metadata_file=ignore_metadata_file,
metadata_task_size=metadata_task_size,
parquet_file_extension=parquet_file_extension,
dataset=dataset_options,
read=read_options,
**other_options,
)
# In the future, we may want to give the engine the
# option to return a dedicated element for `common_kwargs`.
# However, to avoid breaking the API, we just embed this
# data in the first element of `parts` for now.
# The logic below is inteded to handle backward and forward
# compatibility with a user-defined engine.
meta, statistics, parts, index = read_metadata_result[:4]
common_kwargs = {}
aggregation_depth = False
if len(parts):
# For now, `common_kwargs` and `aggregation_depth`
# may be stored in the first element of `parts`
common_kwargs = parts[0].pop("common_kwargs", {})
aggregation_depth = parts[0].pop("aggregation_depth", aggregation_depth)
# Parse dataset statistics from metadata (if available)
parts, divisions, index = process_statistics(
parts,
statistics,
filters,
index,
chunksize,
split_row_groups,
fs,
aggregation_depth,
)
# Account for index and columns arguments.
# Modify `meta` dataframe accordingly
meta, index, columns = set_index_columns(meta, index, columns, auto_index_allowed)
if meta.index.name == NONE_LABEL:
meta.index.name = None
if len(divisions) < 2:
# empty dataframe - just use meta
divisions = (None, None)
io_func = lambda x: x
parts = [meta]
else:
# Use IO function wrapper
io_func = ParquetFunctionWrapper(
engine,
fs,
meta,
columns,
index,
use_nullable_dtypes,
{}, # All kwargs should now be in `common_kwargs`
common_kwargs,
)
# If we are using a remote filesystem and retries is not set, bump it
# to be more fault tolerant, as transient transport errors can occur.
# The specific number 5 isn't hugely motivated: it's less than ten and more
# than two.
annotations = dask.config.get("annotations", {})
if "retries" not in annotations and not _is_local_fs(fs):
ctx = dask.annotate(retries=5)
else:
ctx = contextlib.nullcontext() # type: ignore
with ctx:
# Construct the output collection with from_map
return from_map(
io_func,
parts,
meta=meta,
divisions=divisions,
label="read-parquet",
token=tokenize(path, **input_kwargs),
enforce_metadata=False,
creation_info={
"func": read_parquet,
"args": (path,),
"kwargs": input_kwargs,
},
)
def check_multi_support(engine):
# Helper function to check that the engine
# supports a multi-partition read
return hasattr(engine, "multi_support") and engine.multi_support()
def read_parquet_part(
fs, engine, meta, part, columns, index, use_nullable_dtypes, kwargs
):
"""Read a part of a parquet dataset
This function is used by `read_parquet`."""
if isinstance(part, list):
if len(part) == 1 or part[0][1] or not check_multi_support(engine):
# Part kwargs expected
func = engine.read_partition
dfs = [
func(
fs,
rg,
columns.copy(),
index,
use_nullable_dtypes=use_nullable_dtypes,
**toolz.merge(kwargs, kw),
)
for (rg, kw) in part
]
df = concat(dfs, axis=0) if len(dfs) > 1 else dfs[0]
else:
# No part specific kwargs, let engine read
# list of parts at once
df = engine.read_partition(
fs,
[p[0] for p in part],
columns.copy(),
index,
use_nullable_dtypes=use_nullable_dtypes,
**kwargs,
)
else:
# NOTE: `kwargs` are the same for all parts, while `part_kwargs` may
# be different for each part.
rg, part_kwargs = part
df = engine.read_partition(
fs,
rg,
columns,
index,
use_nullable_dtypes=use_nullable_dtypes,
**toolz.merge(kwargs, part_kwargs),
)
if meta.columns.name:
df.columns.name = meta.columns.name
columns = columns or []
index = index or []
df = df[[c for c in columns if c not in index]]
if index == [NONE_LABEL]:
df.index.name = None
return df
[docs]def to_parquet(
df,
path,
engine="auto",
compression="snappy",
write_index=True,
append=False,
overwrite=False,
ignore_divisions=False,
partition_on=None,
storage_options=None,
custom_metadata=None,
write_metadata_file=None,
compute=True,
compute_kwargs=None,
schema="infer",
name_function=None,
**kwargs,
):
"""Store Dask.dataframe to Parquet files
Notes
-----
Each partition will be written to a separate file.
Parameters
----------
df : dask.dataframe.DataFrame
path : string or pathlib.Path
Destination directory for data. Prepend with protocol like ``s3://``
or ``hdfs://`` for remote data.
engine : {'auto', 'pyarrow', 'fastparquet'}, default 'auto'
Parquet library to use. Defaults to 'auto', which uses ``pyarrow`` if
it is installed, and falls back to ``fastparquet`` otherwise.
compression : string or dict, default 'snappy'
Either a string like ``"snappy"`` or a dictionary mapping column names
to compressors like ``{"name": "gzip", "values": "snappy"}``. Defaults
to ``"snappy"``.
write_index : boolean, default True
Whether or not to write the index. Defaults to True.
append : bool, default False
If False (default), construct data-set from scratch. If True, add new
row-group(s) to an existing data-set. In the latter case, the data-set
must exist, and the schema must match the input data.
overwrite : bool, default False
Whether or not to remove the contents of `path` before writing the dataset.
The default is False. If True, the specified path must correspond to
a directory (but not the current working directory). This option cannot
be set to True if `append=True`.
NOTE: `overwrite=True` will remove the original data even if the current
write operation fails. Use at your own risk.
ignore_divisions : bool, default False
If False (default) raises error when previous divisions overlap with
the new appended divisions. Ignored if append=False.
partition_on : list, default None
Construct directory-based partitioning by splitting on these fields'
values. Each dask partition will result in one or more datafiles,
there will be no global groupby.
storage_options : dict, default None
Key/value pairs to be passed on to the file-system backend, if any.
custom_metadata : dict, default None
Custom key/value metadata to include in all footer metadata (and
in the global "_metadata" file, if applicable). Note that the custom
metadata may not contain the reserved b"pandas" key.
write_metadata_file : bool or None, default None
Whether to write the special ``_metadata`` file. If ``None`` (the
default), a ``_metadata`` file will only be written if ``append=True``
and the dataset already has a ``_metadata`` file.
compute : bool, default True
If ``True`` (default) then the result is computed immediately. If
``False`` then a ``dask.dataframe.Scalar`` object is returned for
future computation.
compute_kwargs : dict, default True
Options to be passed in to the compute method
schema : pyarrow.Schema, dict, "infer", or None, default "infer"
Global schema to use for the output dataset. Defaults to "infer", which
will infer the schema from the dask dataframe metadata. This is usually
sufficient for common schemas, but notably will fail for ``object``
dtype columns that contain things other than strings. These columns
will require an explicit schema be specified. The schema for a subset
of columns can be overridden by passing in a dict of column names to
pyarrow types (for example ``schema={"field": pa.string()}``); columns
not present in this dict will still be automatically inferred.
Alternatively, a full ``pyarrow.Schema`` may be passed, in which case
no schema inference will be done. Passing in ``schema=None`` will
disable the use of a global file schema - each written file may use a
different schema dependent on the dtypes of the corresponding
partition. Note that this argument is ignored by the "fastparquet"
engine.
name_function : callable, default None
Function to generate the filename for each output partition.
The function should accept an integer (partition index) as input and
return a string which will be used as the filename for the corresponding
partition. Should preserve the lexicographic order of partitions.
If not specified, files will created using the convention
``part.0.parquet``, ``part.1.parquet``, ``part.2.parquet``, ...
and so on for each partition in the DataFrame.
**kwargs :
Extra options to be passed on to the specific backend.
Examples
--------
>>> df = dd.read_csv(...) # doctest: +SKIP
>>> df.to_parquet('/path/to/output/', ...) # doctest: +SKIP
By default, files will be created in the specified output directory using the
convention ``part.0.parquet``, ``part.1.parquet``, ``part.2.parquet``, ... and so on for
each partition in the DataFrame. To customize the names of each file, you can use the
``name_function=`` keyword argument. The function passed to ``name_function`` will be
used to generate the filename for each partition and should expect a partition's index
integer as input and return a string which will be used as the filename for the corresponding
partition. Strings produced by ``name_function`` must preserve the order of their respective
partition indices.
For example:
>>> name_function = lambda x: f"data-{x}.parquet"
>>> df.to_parquet('/path/to/output/', name_function=name_function) # doctest: +SKIP
will result in the following files being created::
/path/to/output/
├── data-0.parquet
├── data-1.parquet
├── data-2.parquet
└── ...
See Also
--------
read_parquet: Read parquet data to dask.dataframe
"""
compute_kwargs = compute_kwargs or {}
partition_on = partition_on or []
if isinstance(partition_on, str):
partition_on = [partition_on]
if set(partition_on) - set(df.columns):
raise ValueError(
"Partitioning on non-existent column. "
"partition_on=%s ."
"columns=%s" % (str(partition_on), str(list(df.columns)))
)
if df.columns.inferred_type not in {"string", "empty"}:
raise ValueError("parquet doesn't support non-string column names")
if isinstance(engine, str):
engine = get_engine(engine)
if hasattr(path, "name"):
path = stringify_path(path)
fs, _, _ = get_fs_token_paths(path, mode="wb", storage_options=storage_options)
# Trim any protocol information from the path before forwarding
path = fs._strip_protocol(path)
if overwrite:
if append:
raise ValueError("Cannot use both `overwrite=True` and `append=True`!")
if fs.exists(path) and fs.isdir(path):
# Check for any previous parquet layers reading from a file in the
# output directory, since deleting those files now would result in
# errors or incorrect results.
for layer_name, layer in df.dask.layers.items():
if layer_name.startswith("read-parquet-") and isinstance(
layer, DataFrameIOLayer
):
path_with_slash = path.rstrip("/") + "/" # ensure trailing slash
for input in layer.inputs:
if input["piece"][0].startswith(path_with_slash):
raise ValueError(
"Reading and writing to the same parquet file within the "
"same task graph is not supported."
)
# Don't remove the directory if it's the current working directory
if _is_local_fs(fs):
working_dir = fs.expand_path(".")[0]
if path.rstrip("/") == working_dir.rstrip("/"):
raise ValueError(
"Cannot clear the contents of the current working directory!"
)
# It's safe to clear the output directory
fs.rm(path, recursive=True)
# Always skip divisions checks if divisions are unknown
if not df.known_divisions:
ignore_divisions = True
# Save divisions and corresponding index name. This is necessary,
# because we may be resetting the index to write the file
division_info = {"divisions": df.divisions, "name": df.index.name}
if division_info["name"] is None:
# As of 0.24.2, pandas will rename an index with name=None
# when df.reset_index() is called. The default name is "index",
# but dask will always change the name to the NONE_LABEL constant
if NONE_LABEL not in df.columns:
division_info["name"] = NONE_LABEL
elif write_index:
raise ValueError(
"Index must have a name if __null_dask_index__ is a column."
)
else:
warnings.warn(
"If read back by Dask, column named __null_dask_index__ "
"will be set to the index (and renamed to None)."
)
# There are some "resrved" names that may be used as the default column
# name after resetting the index. However, we don't want to treat it as
# a "special" name if the string is already used as a "real" column name.
reserved_names = []
for name in ["index", "level_0"]:
if name not in df.columns:
reserved_names.append(name)
# If write_index==True (default), reset the index and record the
# name of the original index in `index_cols` (we will set the name
# to the NONE_LABEL constant if it is originally `None`).
# `fastparquet` will use `index_cols` to specify the index column(s)
# in the metadata. `pyarrow` will revert the `reset_index` call
# below if `index_cols` is populated (because pyarrow will want to handle
# index preservation itself). For both engines, the column index
# will be written to "pandas metadata" if write_index=True
index_cols = []
if write_index:
real_cols = set(df.columns)
none_index = list(df._meta.index.names) == [None]
df = df.reset_index()
if none_index:
df.columns = [
c if c not in reserved_names else NONE_LABEL for c in df.columns
]
index_cols = [c for c in set(df.columns) - real_cols]
else:
# Not writing index - might as well drop it
df = df.reset_index(drop=True)
if custom_metadata and b"pandas" in custom_metadata.keys():
raise ValueError(
"User-defined key/value metadata (custom_metadata) can not "
"contain a b'pandas' key. This key is reserved by Pandas, "
"and overwriting the corresponding value can render the "
"entire dataset unreadable."
)
# Engine-specific initialization steps to write the dataset.
# Possibly create parquet metadata, and load existing stuff if appending
i_offset, fmd, metadata_file_exists, extra_write_kwargs = engine.initialize_write(
df,
fs,
path,
append=append,
ignore_divisions=ignore_divisions,
partition_on=partition_on,
division_info=division_info,
index_cols=index_cols,
schema=schema,
custom_metadata=custom_metadata,
**kwargs,
)
# By default we only write a metadata file when appending if one already
# exists
if append and write_metadata_file is None:
write_metadata_file = metadata_file_exists
# Check that custom name_function is valid,
# and that it will produce unique names
if name_function is not None:
if not callable(name_function):
raise ValueError("``name_function`` must be a callable with one argument.")
filenames = [name_function(i + i_offset) for i in range(df.npartitions)]
if len(set(filenames)) < len(filenames):
raise ValueError("``name_function`` must produce unique filenames.")
# If we are using a remote filesystem and retries is not set, bump it
# to be more fault tolerant, as transient transport errors can occur.
# The specific number 5 isn't hugely motivated: it's less than ten and more
# than two.
annotations = dask.config.get("annotations", {})
if "retries" not in annotations and not _is_local_fs(fs):
ctx = dask.annotate(retries=5)
else:
ctx = contextlib.nullcontext()
# Create Blockwise layer for parquet-data write
with ctx:
data_write = df.map_partitions(
ToParquetFunctionWrapper(
engine,
path,
fs,
partition_on,
write_metadata_file,
i_offset,
name_function,
toolz.merge(
kwargs,
{"compression": compression, "custom_metadata": custom_metadata},
extra_write_kwargs,
),
),
BlockIndex((df.npartitions,)),
# Pass in the original metadata to avoid
# metadata emulation in `map_partitions`.
# This is necessary, because we are not
# expecting a dataframe-like output.
meta=df._meta,
enforce_metadata=False,
transform_divisions=False,
align_dataframes=False,
)
# Collect metadata and write _metadata.
# TODO: Use tree-reduction layer (when available)
if write_metadata_file:
final_name = "metadata-" + data_write._name
dsk = {
(final_name, 0): (
apply,
engine.write_metadata,
[
data_write.__dask_keys__(),
fmd,
fs,
path,
],
{"append": append, "compression": compression},
)
}
else:
# NOTE: We still define a single task to tie everything together
# when we are not writing a _metadata file. We do not want to
# return `data_write` (or a `data_write.to_bag()`), because calling
# `compute()` on a multi-partition collection requires the overhead
# of trying to concatenate results on the client.
final_name = "store-" + data_write._name
dsk = {(final_name, 0): (lambda x: None, data_write.__dask_keys__())}
# Convert data_write + dsk to computable collection
graph = HighLevelGraph.from_collections(final_name, dsk, dependencies=(data_write,))
out = Scalar(graph, final_name, "")
if compute:
out = out.compute(**compute_kwargs)
# Invalidate the filesystem listing cache for the output path after write.
# We do this before returning, even if `compute=False`. This helps ensure
# that reading files that were just written succeeds.
fs.invalidate_cache(path)
return out
def create_metadata_file(
paths,
root_dir=None,
out_dir=None,
engine="pyarrow",
storage_options=None,
split_every=32,
compute=True,
compute_kwargs=None,
fs=None,
):
"""Construct a global _metadata file from a list of parquet files.
Dask's read_parquet function is designed to leverage a global
_metadata file whenever one is available. The to_parquet
function will generate this file automatically by default, but it
may not exist if the dataset was generated outside of Dask. This
utility provides a mechanism to generate a _metadata file from a
list of existing parquet files.
NOTE: This utility is not yet supported for the "fastparquet" engine.
Parameters
----------
paths : list(string)
List of files to collect footer metadata from.
root_dir : string, optional
Root directory of dataset. The `file_path` fields in the new
_metadata file will relative to this directory. If None, a common
root directory will be inferred.
out_dir : string or False, optional
Directory location to write the final _metadata file. By default,
this will be set to `root_dir`. If False is specified, the global
metadata will be returned as an in-memory object (and will not be
written to disk).
engine : str or Engine, default 'pyarrow'
Parquet Engine to use. Only 'pyarrow' is supported if a string
is passed.
storage_options : dict, optional
Key/value pairs to be passed on to the file-system backend, if any.
split_every : int, optional
The final metadata object that is written to _metadata can be much
smaller than the list of footer metadata. In order to avoid the
aggregation of all metadata within a single task, a tree reduction
is used. This argument specifies the maximum number of metadata
inputs to be handled by any one task in the tree. Defaults to 32.
compute : bool, optional
If True (default) then the result is computed immediately. If False
then a ``dask.delayed`` object is returned for future computation.
compute_kwargs : dict, optional
Options to be passed in to the compute method
fs : fsspec object, optional
File-system instance to use for file handling. If prefixes have
been removed from the elements of ``paths`` before calling this
function, an ``fs`` argument must be provided to ensure correct
behavior on remote file systems ("naked" paths cannot be used
to infer file-system information).
"""
# Get engine.
# Note that "fastparquet" is not yet supported
if isinstance(engine, str):
if engine not in ("pyarrow", "arrow"):
raise ValueError(
f"{engine} is not a supported engine for create_metadata_file "
"Try engine='pyarrow'."
)
engine = get_engine(engine)
# Process input path list
if fs is None:
# Only do this if an fsspec file-system object is not
# already defined. The prefixes may already be stripped.
fs, _, paths = get_fs_token_paths(
paths, mode="rb", storage_options=storage_options
)
ap_kwargs = {"root": root_dir} if root_dir else {}
paths, root_dir, fns = _sort_and_analyze_paths(paths, fs, **ap_kwargs)
out_dir = root_dir if out_dir is None else out_dir
# Start constructing a raw graph
dsk = {}
name = "gen-metadata-" + tokenize(paths, fs)
collect_name = "collect-" + name
agg_name = "agg-" + name
# Define a "collect" task for each file in the input list.
# Each tasks will:
# 1. Extract the footer metadata from a distinct file
# 2. Populate the `file_path` field in the metadata
# 3. Return the extracted/modified metadata
for p, (fn, path) in enumerate(zip(fns, paths)):
key = (collect_name, p, 0)
dsk[key] = (engine.collect_file_metadata, path, fs, fn)
# Build a reduction tree to aggregate all footer metadata
# into a single metadata object. Each task in the tree
# will take in a list of metadata objects as input, and will
# usually output a single (aggregated) metadata object.
# The final task in the tree will write the result to disk
# instead of returning it (this behavior is triggered by
# passing a file path to `engine.aggregate_metadata`).
parts = len(paths)
widths = [parts]
while parts > 1:
parts = math.ceil(parts / split_every)
widths.append(parts)
height = len(widths)
for depth in range(1, height):
for group in range(widths[depth]):
p_max = widths[depth - 1]
lstart = split_every * group
lstop = min(lstart + split_every, p_max)
dep_task_name = collect_name if depth == 1 else agg_name
node_list = [(dep_task_name, p, depth - 1) for p in range(lstart, lstop)]
if depth == height - 1:
assert group == 0
dsk[name] = (engine.aggregate_metadata, node_list, fs, out_dir)
else:
dsk[(agg_name, group, depth)] = (
engine.aggregate_metadata,
node_list,
None,
None,
)
# There will be no aggregation tasks if there is only one file
if len(paths) == 1:
dsk[name] = (engine.aggregate_metadata, [(collect_name, 0, 0)], fs, out_dir)
# Convert the raw graph to a `Delayed` object
graph = HighLevelGraph.from_collections(name, dsk, dependencies=[])
out = Delayed(name, graph)
# Optionally compute the result
if compute:
if compute_kwargs is None:
compute_kwargs = dict()
out = out.compute(**compute_kwargs)
return out
_ENGINES: dict[str, Engine] = {}
def get_engine(engine):
"""Get the parquet engine backend implementation.
Parameters
----------
engine : {'auto', 'pyarrow', 'fastparquet'}, default 'auto'
Parquet library to use. Defaults to 'auto', which uses ``pyarrow`` if
it is installed, and falls back to ``fastparquet`` otherwise.
Returns
-------
A dict containing a ``'read'`` and ``'write'`` function.
"""
if engine in _ENGINES:
return _ENGINES[engine]
if engine == "auto":
try:
return get_engine("pyarrow")
except RuntimeError:
pass
try:
return get_engine("fastparquet")
except RuntimeError:
raise RuntimeError("Please install either pyarrow or fastparquet") from None
elif engine == "fastparquet":
import_required("fastparquet", "`fastparquet` not installed")
from dask.dataframe.io.parquet.fastparquet import FastParquetEngine
_ENGINES["fastparquet"] = eng = FastParquetEngine
return eng
elif engine in ("pyarrow", "arrow", "pyarrow-dataset"):
pa = import_required("pyarrow", "`pyarrow` not installed")
pa_version = parse_version(pa.__version__)
if engine in ("pyarrow-dataset", "arrow"):
engine = "pyarrow"
if engine == "pyarrow":
if pa_version.major < 1:
raise ImportError(
f"pyarrow-{pa_version.major} does not support the "
f"pyarrow.dataset API. Please install pyarrow>=1."
)
from dask.dataframe.io.parquet.arrow import ArrowDatasetEngine
_ENGINES[engine] = eng = ArrowDatasetEngine
return eng
else:
raise ValueError(
f'Unsupported engine: "{engine}".'
' Valid choices include "pyarrow" and "fastparquet".'
)
#####################
# Utility Functions #
#####################
def sorted_columns(statistics, columns=None):
"""Find sorted columns given row-group statistics
This finds all columns that are sorted, along with the
appropriate ``divisions`` for those columns. If the (optional)
``columns`` argument is used, the search will be restricted
to the specified column set.
Returns
-------
out: List of {'name': str, 'divisions': List[str]} dictionaries
"""
if not statistics:
return []
out = []
for i, c in enumerate(statistics[0]["columns"]):
if columns and c["name"] not in columns:
continue
if not all(
"min" in s["columns"][i] and "max" in s["columns"][i] for s in statistics
):
continue
divisions = [c["min"]]
max = c["max"]
success = c["min"] is not None
for stats in statistics[1:]:
c = stats["columns"][i]
if c["min"] is None:
success = False
break
if c["min"] >= max:
divisions.append(c["min"])
max = c["max"]
else:
success = False
break
if success:
divisions.append(max)
assert divisions == sorted(divisions)
out.append({"name": c["name"], "divisions": divisions})
return out
def apply_filters(parts, statistics, filters):
"""Apply filters onto parts/statistics pairs
Parameters
----------
parts: list
Tokens corresponding to row groups to read in the future
statistics: List[dict]
List of statistics for each part, including min and max values
filters: Union[List[Tuple[str, str, Any]], List[List[Tuple[str, str, Any]]]]
List of filters to apply, like ``[[('x', '=', 0), ...], ...]``. This
implements partition-level (hive) filtering only, i.e., to prevent the
loading of some row-groups and/or files.
Predicates can be expressed in disjunctive normal form (DNF). This means
that the innermost tuple describes a single column predicate. These
inner predicates are combined with an AND conjunction into a larger
predicate. The outer-most list then combines all of the combined
filters with an OR disjunction.
Predicates can also be expressed as a List[Tuple]. These are evaluated
as an AND conjunction. To express OR in predicates, one must use the
(preferred) List[List[Tuple]] notation.
Note that the "fastparquet" engine does not currently support DNF for
the filtering of partitioned columns (List[Tuple] is required).
Returns
-------
parts, statistics: the same as the input, but possibly a subset
"""
def apply_conjunction(parts, statistics, conjunction):
for column, operator, value in conjunction:
if operator == "in" and not isinstance(value, (list, set, tuple)):
raise TypeError("Value of 'in' filter must be a list, set, or tuple.")
out_parts = []
out_statistics = []
for part, stats in zip(parts, statistics):
if "filter" in stats and stats["filter"]:
continue # Filtered by engine
try:
c = toolz.groupby("name", stats["columns"])[column][0]
min = c["min"]
max = c["max"]
except KeyError:
out_parts.append(part)
out_statistics.append(stats)
else:
if (
operator in ("==", "=")
and min <= value <= max
or operator == "!="
and (min != value or max != value)
or operator == "<"
and min < value
or operator == "<="
and min <= value
or operator == ">"
and max > value
or operator == ">="
and max >= value
or operator == "in"
and any(min <= item <= max for item in value)
):
out_parts.append(part)
out_statistics.append(stats)
parts, statistics = out_parts, out_statistics
return parts, statistics
conjunction, *disjunction = filters if isinstance(filters[0], list) else [filters]
out_parts, out_statistics = apply_conjunction(parts, statistics, conjunction)
for conjunction in disjunction:
for part, stats in zip(*apply_conjunction(parts, statistics, conjunction)):
if part not in out_parts:
out_parts.append(part)
out_statistics.append(stats)
return out_parts, out_statistics
def process_statistics(
parts,
statistics,
filters,
index,
chunksize,
split_row_groups,
fs,
aggregation_depth,
):
"""Process row-group column statistics in metadata
Used in read_parquet.
"""
if statistics and len(parts) != len(statistics):
# It is up to the Engine to guarantee that these
# lists are the same length (if statistics are defined).
# This misalignment may be indicative of a bug or
# incorrect read_parquet usage, so throw a warning.
warnings.warn(
f"Length of partition statistics ({len(statistics)}) "
f"does not match the partition count ({len(parts)}). "
f"This may indicate a bug or incorrect read_parquet "
f"usage. We must ignore the statistics and disable: "
f"filtering, divisions, and/or file aggregation."
)
statistics = []
divisions = None
if statistics:
result = list(
zip(
*[
(part, stats)
for part, stats in zip(parts, statistics)
if stats["num-rows"] > 0
]
)
)
parts, statistics = result or [[], []]
if filters:
parts, statistics = apply_filters(parts, statistics, filters)
# Aggregate parts/statistics if we are splitting by row-group
if chunksize or (split_row_groups and int(split_row_groups) > 1):
parts, statistics = aggregate_row_groups(
parts, statistics, chunksize, split_row_groups, fs, aggregation_depth
)
# Convert str index to list
index = [index] if isinstance(index, str) else index
# TODO: Remove `filters` criteria below after deprecation cycle.
# We can then remove the `sorted_col_names` logic and warning.
# See: https://github.com/dask/dask/pull/9661
process_columns = index if index and len(index) == 1 else None
if filters:
process_columns = None
# Use statistics to define divisions
if process_columns or filters:
sorted_col_names = []
for sorted_column_info in sorted_columns(
statistics, columns=process_columns
):
if index and sorted_column_info["name"] in index:
divisions = sorted_column_info["divisions"]
break
else:
# Filtered columns may also be sorted
sorted_col_names.append(sorted_column_info["name"])
if index is None and sorted_col_names:
assert bool(filters) # Should only get here when filtering
warnings.warn(
f"Sorted columns detected: {sorted_col_names}\n"
f"Use the `index` argument to set a sorted column as your "
f"index to create a DataFrame collection with known `divisions`.",
UserWarning,
)
divisions = divisions or (None,) * (len(parts) + 1)
return parts, divisions, index
def set_index_columns(meta, index, columns, auto_index_allowed):
"""Handle index/column arguments, and modify `meta`
Used in read_parquet.
"""
ignore_index_column_intersection = False
if columns is None:
# User didn't specify columns, so ignore any intersection
# of auto-detected values with the index (if necessary)
ignore_index_column_intersection = True
# Do not allow "un-named" fields to be read in as columns.
# These were intended to be un-named indices at write time.
_index = index or []
columns = [
c for c in meta.columns if c not in (None, NONE_LABEL) or c in _index
]
if not set(columns).issubset(set(meta.columns)):
raise ValueError(
"The following columns were not found in the dataset %s\n"
"The following columns were found %s"
% (set(columns) - set(meta.columns), meta.columns)
)
if index:
if isinstance(index, str):
index = [index]
if isinstance(columns, str):
columns = [columns]
if ignore_index_column_intersection:
columns = [col for col in columns if col not in index]
if set(index).intersection(columns):
if auto_index_allowed:
raise ValueError(
"Specified index and column arguments must not intersect"
" (set index=False or remove the detected index from columns).\n"
"index: {} | column: {}".format(index, columns)
)
else:
raise ValueError(
"Specified index and column arguments must not intersect.\n"
"index: {} | column: {}".format(index, columns)
)
return meta[list(columns)], index, columns
def aggregate_row_groups(
parts, stats, chunksize, split_row_groups, fs, aggregation_depth
):
if not stats or not stats[0].get("file_path_0", None):
return parts, stats
parts_agg = []
stats_agg = []
use_row_group_criteria = split_row_groups and int(split_row_groups) > 1
use_chunksize_criteria = bool(chunksize)
if use_chunksize_criteria:
chunksize = parse_bytes(chunksize)
next_part, next_stat = [parts[0].copy()], stats[0].copy()
for i in range(1, len(parts)):
stat, part = stats[i], parts[i]
# Criteria #1 for aggregating parts: parts are within the same file
same_path = stat["file_path_0"] == next_stat["file_path_0"]
multi_path_allowed = False
if aggregation_depth:
# Criteria #2 for aggregating parts: The part does not include
# row-group information, or both parts include the same kind
# of row_group aggregation (all None, or all indices)
multi_path_allowed = len(part["piece"]) == 1
if not (same_path or multi_path_allowed):
rgs = set(list(part["piece"][1]) + list(next_part[-1]["piece"][1]))
multi_path_allowed = (rgs == {None}) or (None not in rgs)
# Criteria #3 for aggregating parts: The parts share a
# directory at the "depth" allowed by `aggregation_depth`
if not same_path and multi_path_allowed:
if aggregation_depth is True:
multi_path_allowed = True
elif isinstance(aggregation_depth, int):
# Make sure files share the same directory
root = stat["file_path_0"].split(fs.sep)[:-aggregation_depth]
next_root = next_stat["file_path_0"].split(fs.sep)[
:-aggregation_depth
]
multi_path_allowed = root == next_root
else:
raise ValueError(
f"{aggregation_depth} not supported for `aggregation_depth`"
)
def _check_row_group_criteria(stat, next_stat):
if use_row_group_criteria:
return (next_stat["num-row-groups"] + stat["num-row-groups"]) <= int(
split_row_groups
)
else:
return False
def _check_chunksize_criteria(stat, next_stat):
if use_chunksize_criteria:
return (
next_stat["total_byte_size"] + stat["total_byte_size"]
) <= chunksize
else:
return False
stat["num-row-groups"] = stat.get("num-row-groups", 1)
next_stat["num-row-groups"] = next_stat.get("num-row-groups", 1)
if (same_path or multi_path_allowed) and (
_check_row_group_criteria(stat, next_stat)
or _check_chunksize_criteria(stat, next_stat)
):
# Update part list
next_part.append(part)
# Update Statistics
next_stat["total_byte_size"] += stat["total_byte_size"]
next_stat["num-rows"] += stat["num-rows"]
next_stat["num-row-groups"] += stat["num-row-groups"]
for col, col_add in zip(next_stat["columns"], stat["columns"]):
if col["name"] != col_add["name"]:
raise ValueError("Columns are different!!")
if "min" in col:
col["min"] = min(col["min"], col_add["min"])
if "max" in col:
col["max"] = max(col["max"], col_add["max"])
else:
parts_agg.append(next_part)
stats_agg.append(next_stat)
next_part, next_stat = [part.copy()], stat.copy()
parts_agg.append(next_part)
stats_agg.append(next_stat)
return parts_agg, stats_agg
DataFrame.to_parquet.__doc__ = to_parquet.__doc__