def _read_pf_simple(fs, path, base, index_names, all_columns, is_series,
categories, cats, scheme, storage_name_mapping):
"""Read dataset with fastparquet using ParquetFile machinery"""
from fastparquet import ParquetFile
pf = ParquetFile(path, open_with=fs.open)
relpath = path.replace(base, '').lstrip('/')
for rg in pf.row_groups:
for ch in rg.columns:
ch.file_path = relpath
pf.file_scheme = scheme
pf.cats = cats
pf.fn = base
df = pf.to_pandas(all_columns, categories, index=index_names)
if df.index.nlevels == 1:
if index_names:
df.index.name = storage_name_mapping.get(index_names[0],
index_names[0])
else:
if index_names:
df.index.names = [storage_name_mapping.get(name, name)
for name in index_names]
df.columns = [storage_name_mapping.get(col, col)
for col in all_columns
if col not in (index_names or [])]
if is_series:
return df[df.columns[0]]
else:
return df
def read_partition(fs,
piece,
columns,
index,
categories=(),
pf=None,
**kwargs):
if isinstance(index, list):
columns += index
if pf:
df = pf.read_row_group_file(piece,
columns,
categories,
index=index,
**kwargs.get("read", {}))
else:
base, fns = _analyze_paths([piece], fs)
scheme = get_file_scheme(fns)
pf = ParquetFile(piece, open_with=fs.open)
relpath = piece.replace(base, "").lstrip("/")
for rg in pf.row_groups:
for ch in rg.columns:
ch.file_path = relpath
pf.file_scheme = scheme
pf.cats = _paths_to_cats(fns, scheme)
pf.fn = base
df = pf.to_pandas(columns, categories, index=index)
return df
def _read_pf_simple(fs, path, base, index_names, all_columns, is_series,
categories, cats, scheme, storage_name_mapping):
"""Read dataset with fastparquet using ParquetFile machinery"""
from fastparquet import ParquetFile
pf = ParquetFile(path, open_with=fs.open)
relpath = path.replace(base, '').lstrip('/')
for rg in pf.row_groups:
for ch in rg.columns:
ch.file_path = relpath
pf.file_scheme = scheme
pf.cats = cats
pf.fn = base
df = pf.to_pandas(all_columns, categories, index=index_names)
if df.index.nlevels == 1:
if index_names:
df.index.name = storage_name_mapping.get(index_names[0],
index_names[0])
else:
if index_names:
df.index.names = [storage_name_mapping.get(name, name)
for name in index_names]
df.columns = [storage_name_mapping.get(col, col)
for col in all_columns
if col not in (index_names or [])]
if is_series:
return df[df.columns[0]]
else:
return df
def read_partition(fs,
piece,
columns,
index,
categories=(),
pf=None,
**kwargs):
if isinstance(index, list):
columns += index
if pf is None:
base, fns = _analyze_paths([piece], fs)
scheme = get_file_scheme(fns)
pf = ParquetFile(piece, open_with=fs.open)
relpath = piece.replace(base, "").lstrip("/")
for rg in pf.row_groups:
for ch in rg.columns:
ch.file_path = relpath
pf.file_scheme = scheme
pf.cats = _paths_to_cats(fns, scheme)
pf.fn = base
return pf.to_pandas(columns, categories, index=index)
else:
if isinstance(pf, tuple):
pf = _determine_pf_parts(fs, pf[0], pf[1], **kwargs)[1]
pf._dtypes = lambda *args: pf.dtypes # ugly patch, could be fixed
pf.fmd.row_groups = None
rg_piece = pf.row_groups[piece]
pf.fmd.key_value_metadata = None
return pf.read_row_group_file(rg_piece,
columns,
categories,
index=index,
**kwargs.get("read", {}))
def read_partition(
cls, fs, piece, columns, index, categories=(), pf=None, **kwargs
):
null_index_name = False
if isinstance(index, list):
if index == [None]:
# Handling a None-labeled index...
# The pandas metadata told us to read in an index
# labeled `None`. If this corresponds to a `RangeIndex`,
# fastparquet will need use the pandas metadata to
# construct the index. Otherwise, the index will correspond
# to a column named "__index_level_0__". We will need to
# check the `ParquetFile` object for this column below.
index = []
null_index_name = True
columns += index
if pf is None:
base, fns = _analyze_paths([piece], fs)
scheme = get_file_scheme(fns)
pf = ParquetFile(piece, open_with=fs.open)
relpath = piece.replace(base, "").lstrip("/")
for rg in pf.row_groups:
for ch in rg.columns:
ch.file_path = relpath
pf.file_scheme = scheme
pf.cats = paths_to_cats(fns, scheme)
pf.fn = base
if null_index_name and "__index_level_0__" in pf.columns:
# See "Handling a None-labeled index" comment above
index = ["__index_level_0__"]
columns += index
return pf.to_pandas(columns, categories, index=index)
else:
if isinstance(pf, tuple):
if isinstance(pf[0], list):
pf = _determine_pf_parts(fs, pf[0], pf[1], **kwargs)[1]
else:
pf = ParquetFile(
pf[0], open_with=fs.open, sep=fs.sep, **kwargs.get("file", {})
)
pf._dtypes = lambda *args: pf.dtypes # ugly patch, could be fixed
pf.fmd.row_groups = None
rg_piece = pf.row_groups[piece]
if null_index_name:
if "__index_level_0__" in pf.columns:
# See "Handling a None-labeled index" comment above
index = ["__index_level_0__"]
columns += index
pf.fmd.key_value_metadata = None
else:
pf.fmd.key_value_metadata = None
return pf.read_row_group_file(
rg_piece, columns, categories, index=index, **kwargs.get("read", {})
)
def _read_parquet_file(
fs,
base,
fn,
index,
columns,
series,
categories,
cs,
dt,
scheme,
storage_name_mapping,
*args
):
"""Read a single file with fastparquet, to be used in a task"""
from fastparquet.api import ParquetFile
from collections import OrderedDict
name_storage_mapping = {v: k for k, v in storage_name_mapping.items()}
if not isinstance(columns, (tuple, list)):
columns = [columns]
series = True
if index:
index, = index
if index not in columns:
columns = columns + [index]
columns = [name_storage_mapping.get(col, col) for col in columns]
index = name_storage_mapping.get(index, index)
cs = OrderedDict([(k, v) for k, v in cs.items() if k in columns])
pf = ParquetFile(fn, open_with=fs.open)
pf.file_scheme = scheme
for rg in pf.row_groups:
for ch in rg.columns:
ch.file_path = fn.replace(base, "").lstrip("/")
pf.fn = base
df = pf.to_pandas(columns=columns, index=index, categories=categories)
if df.index.nlevels == 1:
if index:
df.index.name = storage_name_mapping.get(index, index)
else:
if index:
df.index.names = [storage_name_mapping.get(name, name) for name in index]
df.columns = [storage_name_mapping.get(col, col) for col in columns if col != index]
if series:
return df[df.columns[0]]
else:
return df
def _read_fp_multifile(fs,
fs_token,
paths,
columns=None,
categories=None,
index=None):
"""Read dataset with fastparquet by assuming metadata from first file"""
from fastparquet import ParquetFile
from fastparquet.util import analyse_paths, get_file_scheme, join_path
base, fns = analyse_paths(paths)
parsed_paths = [join_path(p) for p in paths]
scheme = get_file_scheme(fns)
pf = ParquetFile(paths[0], open_with=fs.open)
pf.file_scheme = scheme
pf.cats = _paths_to_cats(fns, scheme)
(
meta,
_,
index_name,
out_type,
all_columns,
index_names,
storage_name_mapping,
) = _pf_validation(pf, columns, index, categories, [])
name = "read-parquet-" + tokenize(fs_token, paths, all_columns, categories)
dsk = {(name, i): (
_read_pf_simple,
fs,
path,
base,
index_names,
all_columns,
out_type == Series,
categories,
pf.cats,
pf.file_scheme,
storage_name_mapping,
)
for i, path in enumerate(parsed_paths)}
divisions = (None, ) * (len(paths) + 1)
return out_type(dsk, name, meta, divisions)
def _read_parquet_file(fs, base, fn, index, columns, series, categories,
cs, dt, scheme, storage_name_mapping, *args):
"""Read a single file with fastparquet, to be used in a task"""
from fastparquet.api import ParquetFile
from collections import OrderedDict
name_storage_mapping = {v: k for k, v in storage_name_mapping.items()}
if not isinstance(columns, (tuple, list)):
columns = [columns,]
series = True
if index:
index, = index
if index not in columns:
columns = columns + [index]
columns = [name_storage_mapping.get(col, col) for col in columns]
index = name_storage_mapping.get(index, index)
cs = OrderedDict([(k, v) for k, v in cs.items() if k in columns])
pf = ParquetFile(fn, open_with=fs.open)
pf.file_scheme = scheme
for rg in pf.row_groups:
for ch in rg.columns:
ch.file_path = fn.replace(base, "").lstrip('/')
pf.fn = base
df = pf.to_pandas(columns=columns, index=index, categories=categories)
if df.index.nlevels == 1:
if index:
df.index.name = storage_name_mapping.get(index, index)
else:
if index:
df.index.names = [storage_name_mapping.get(name, name)
for name in index]
df.columns = [storage_name_mapping.get(col, col)
for col in columns
if col != index]
if series:
return df[df.columns[0]]
else:
return df
def _read_fp_multifile(fs, fs_token, paths, columns=None,
categories=None, index=None):
"""Read dataset with fastparquet by assuming metadata from first file"""
from fastparquet import ParquetFile
from fastparquet.util import analyse_paths, get_file_scheme
base, fns = analyse_paths(paths)
scheme = get_file_scheme(fns)
pf = ParquetFile(paths[0], open_with=fs.open)
pf.file_scheme = scheme
pf.cats = _paths_to_cats(fns, scheme)
(meta, _, index_name, out_type, all_columns, index_names,
storage_name_mapping) = _pf_validation(
pf, columns, index, categories, [])
name = 'read-parquet-' + tokenize(fs_token, paths, all_columns,
categories)
dsk = {(name, i): (_read_pf_simple, fs, path, base,
index_names, all_columns, out_type == Series,
categories, pf.cats,
pf.file_scheme, storage_name_mapping)
for i, path in enumerate(paths)}
divisions = (None, ) * (len(paths) + 1)
return out_type(dsk, name, meta, divisions)
def _determine_pf_parts(fs, paths, gather_statistics, **kwargs):
""" Determine how to access metadata and break read into ``parts``
This logic is mostly to handle `gather_statistics=False` cases,
because this also means we should avoid scanning every file in the
dataset. If _metadata is available, set `gather_statistics=True`
(if `gather_statistics=None`).
"""
parts = []
if len(paths) > 1:
if gather_statistics is not False:
# This scans all the files, allowing index/divisions
# and filtering
pf = ParquetFile(paths,
open_with=fs.open,
sep=fs.sep,
**kwargs.get("file", {}))
else:
base, fns = _analyze_paths(paths, fs)
relpaths = [path.replace(base, "").lstrip("/") for path in paths]
if "_metadata" in relpaths:
# We have a _metadata file, lets use it
pf = ParquetFile(base + fs.sep + "_metadata",
open_with=fs.open,
sep=fs.sep,
**kwargs.get("file", {}))
else:
# Rely on metadata for 0th file.
# Will need to pass a list of paths to read_partition
scheme = get_file_scheme(fns)
pf = ParquetFile(paths[0],
open_with=fs.open,
**kwargs.get("file", {}))
pf.file_scheme = scheme
pf.cats = _paths_to_cats(fns, scheme)
parts = paths.copy()
else:
if fs.isdir(paths[0]):
# This is a directory, check for _metadata, then _common_metadata
paths = fs.glob(paths[0] + fs.sep + "*")
base, fns = _analyze_paths(paths, fs)
relpaths = [path.replace(base, "").lstrip("/") for path in paths]
if "_metadata" in relpaths:
# Using _metadata file (best-case scenario)
pf = ParquetFile(base + fs.sep + "_metadata",
open_with=fs.open,
sep=fs.sep,
**kwargs.get("file", {}))
if gather_statistics is None:
gather_statistics = True
elif gather_statistics is not False:
# Scan every file
pf = ParquetFile(paths,
open_with=fs.open,
**kwargs.get("file", {}))
else:
# Use _common_metadata file if it is available.
# Otherwise, just use 0th file
if "_common_metadata" in relpaths:
pf = ParquetFile(base + fs.sep + "_common_metadata",
open_with=fs.open,
**kwargs.get("file", {}))
else:
pf = ParquetFile(paths[0],
open_with=fs.open,
**kwargs.get("file", {}))
scheme = get_file_scheme(fns)
pf.file_scheme = scheme
pf.cats = _paths_to_cats(fns, scheme)
parts = paths.copy()
else:
# There is only one file to read
pf = ParquetFile(paths[0],
open_with=fs.open,
sep=fs.sep,
**kwargs.get("file", {}))
return parts, pf, gather_statistics
def _determine_pf_parts(fs, paths, gather_statistics, **kwargs):
""" Determine how to access metadata and break read into ``parts``
This logic is mostly to handle `gather_statistics=False` cases,
because this also means we should avoid scanning every file in the
dataset. If _metadata is available, set `gather_statistics=True`
(if `gather_statistics=None`).
The `fast_metadata` output specifies that ParquetFile metadata parsing
is fast enough for each worker to perform during `read_partition`. The
value will be set to True if: (1) The path is a directory containing
_metadta, (2) the path is a list of files containing _metadata, (3)
there is only one file to read, or (4) `gather_statistics` is False.
In other cases, the ParquetFile object will need to be stored in the
task graph, because metadata parsing is too expensive.
"""
parts = []
fast_metadata = True
if len(paths) > 1:
base, fns = _analyze_paths(paths, fs)
if gather_statistics is not False:
# This scans all the files, allowing index/divisions
# and filtering
if "_metadata" not in fns:
paths_use = paths
fast_metadata = False
else:
paths_use = base + fs.sep + "_metadata"
pf = ParquetFile(paths_use,
open_with=fs.open,
sep=fs.sep,
**kwargs.get("file", {}))
else:
if "_metadata" in fns:
# We have a _metadata file, lets use it
pf = ParquetFile(base + fs.sep + "_metadata",
open_with=fs.open,
sep=fs.sep,
**kwargs.get("file", {}))
else:
# Rely on metadata for 0th file.
# Will need to pass a list of paths to read_partition
scheme = get_file_scheme(fns)
pf = ParquetFile(paths[0],
open_with=fs.open,
**kwargs.get("file", {}))
pf.file_scheme = scheme
pf.cats = paths_to_cats(fns, scheme)
parts = paths.copy()
elif fs.isdir(paths[0]):
# This is a directory, check for _metadata, then _common_metadata
paths = fs.glob(paths[0] + fs.sep + "*")
base, fns = _analyze_paths(paths, fs)
if "_metadata" in fns:
# Using _metadata file (best-case scenario)
pf = ParquetFile(base + fs.sep + "_metadata",
open_with=fs.open,
sep=fs.sep,
**kwargs.get("file", {}))
if gather_statistics is None:
gather_statistics = True
elif gather_statistics is not False:
# Scan every file
pf = ParquetFile(paths,
open_with=fs.open,
**kwargs.get("file", {}))
fast_metadata = False
else:
# Use _common_metadata file if it is available.
# Otherwise, just use 0th file
if "_common_metadata" in fns:
pf = ParquetFile(base + fs.sep + "_common_metadata",
open_with=fs.open,
**kwargs.get("file", {}))
else:
pf = ParquetFile(paths[0],
open_with=fs.open,
**kwargs.get("file", {}))
scheme = get_file_scheme(fns)
pf.file_scheme = scheme
pf.cats = paths_to_cats(fns, scheme)
parts = paths.copy()
else:
# There is only one file to read
base = None
pf = ParquetFile(paths[0],
open_with=fs.open,
sep=fs.sep,
**kwargs.get("file", {}))
return parts, pf, gather_statistics, fast_metadata, base
def _determine_pf_parts(fs, paths, gather_statistics, **kwargs):
"""Determine how to access metadata and break read into ``parts``
This logic is mostly to handle `gather_statistics=False` cases,
because this also means we should avoid scanning every file in the
dataset. If _metadata is available, set `gather_statistics=True`
(if `gather_statistics=None`).
"""
parts = []
if len(paths) > 1:
paths, base, fns = _sort_and_analyze_paths(paths, fs)
if gather_statistics is not False:
# This scans all the files, allowing index/divisions
# and filtering
if "_metadata" not in fns:
paths_use = paths
else:
paths_use = base + fs.sep + "_metadata"
pf = ParquetFile(paths_use,
open_with=fs.open,
sep=fs.sep,
**kwargs.get("file", {}))
else:
if "_metadata" in fns:
# We have a _metadata file, lets use it
pf = ParquetFile(
base + fs.sep + "_metadata",
open_with=fs.open,
sep=fs.sep,
**kwargs.get("file", {}),
)
else:
# Rely on metadata for 0th file.
# Will need to pass a list of paths to read_partition
scheme = get_file_scheme(fns)
pf = ParquetFile(paths[0],
open_with=fs.open,
**kwargs.get("file", {}))
pf.file_scheme = scheme
pf.cats = paths_to_cats(fns, scheme)
parts = paths.copy()
elif fs.isdir(paths[0]):
# This is a directory, check for _metadata, then _common_metadata
paths = fs.glob(paths[0] + fs.sep + "*")
paths, base, fns = _sort_and_analyze_paths(paths, fs)
if "_metadata" in fns:
# Using _metadata file (best-case scenario)
pf = ParquetFile(
base + fs.sep + "_metadata",
open_with=fs.open,
sep=fs.sep,
**kwargs.get("file", {}),
)
if gather_statistics is None:
gather_statistics = True
elif gather_statistics is not False:
# Scan every file
pf = ParquetFile(paths,
open_with=fs.open,
**kwargs.get("file", {}))
else:
# Use _common_metadata file if it is available.
# Otherwise, just use 0th file
if "_common_metadata" in fns:
pf = ParquetFile(
base + fs.sep + "_common_metadata",
open_with=fs.open,
**kwargs.get("file", {}),
)
fns.remove("_common_metadata")
else:
pf = ParquetFile(paths[0],
open_with=fs.open,
**kwargs.get("file", {}))
scheme = get_file_scheme(fns)
pf.file_scheme = scheme
pf.cats = paths_to_cats(fns, scheme)
parts = [fs.sep.join([base, fn]) for fn in fns]
else:
# There is only one file to read
base = None
pf = ParquetFile(paths[0],
open_with=fs.open,
sep=fs.sep,
**kwargs.get("file", {}))
# Ensure that there is no overlap between partition columns
# and explicit columns in `pf`
if pf.cats:
_partitions = [p for p in pf.cats if p not in pf.columns]
if not _partitions:
pf.cats = {}
elif len(_partitions) != len(pf.cats):
raise ValueError("No partition-columns should be written in the \n"
"file unless they are ALL written in the file.\n"
"columns: {} | partitions: {}".format(
pf.columns, pf.cats.keys()))
return parts, pf, gather_statistics, base
def read_metadata(fs,
paths,
categories=None,
index=None,
gather_statistics=None,
filters=None,
**kwargs):
if len(paths) > 1:
if gather_statistics is not False:
# this scans all the files, allowing index/divisions
# and filtering
pf = fastparquet.ParquetFile(paths,
open_with=fs.open,
sep=fs.sep,
**kwargs.get("file", {}))
else:
base, fns = analyse_paths(paths)
scheme = get_file_scheme(fns)
pf = ParquetFile(paths[0],
open_with=fs.open,
**kwargs.get("file", {}))
pf.file_scheme = scheme
pf.cats = _paths_to_cats(fns, scheme)
relpath = paths.replace(base, "").lstrip("/")
for rg in pf.row_groups:
rg.cats = pf.cats
rg.schema = pf.schema
for ch in rg.columns:
ch.file_path = relpath
else:
try:
pf = fastparquet.ParquetFile(paths[0] + fs.sep + "_metadata",
open_with=fs.open,
sep=fs.sep,
**kwargs.get("file", {}))
if gather_statistics is None:
gather_statistics = True
except Exception:
pf = fastparquet.ParquetFile(paths[0],
open_with=fs.open,
sep=fs.sep,
**kwargs.get("file", {}))
columns = None
if pf.fmd.key_value_metadata:
pandas_md = [
x.value for x in pf.fmd.key_value_metadata if x.key == "pandas"
]
else:
pandas_md = []
if len(pandas_md) == 0:
index_names = []
column_names = pf.columns + list(pf.cats)
storage_name_mapping = {k: k for k in column_names}
column_index_names = [None]
elif len(pandas_md) == 1:
(
index_names,
column_names,
storage_name_mapping,
column_index_names,
) = _parse_pandas_metadata(json.loads(pandas_md[0]))
# auto-ranges should not be created by fastparquet
index_names = [n for n in index_names if n is not None]
column_names.extend(pf.cats)
else:
raise ValueError("File has multiple entries for 'pandas' metadata")
if index is None and len(index_names) > 0:
if len(index_names) == 1:
index = index_names[0]
else:
index = index_names
# Normalize user inputs
column_names, index_names = _normalize_index_columns(
columns, column_names, index, index_names)
all_columns = index_names + column_names
categories_dict = None
if isinstance(categories, dict):
categories_dict = categories
if categories is None:
categories = pf.categories
elif isinstance(categories, string_types):
categories = [categories]
else:
categories = list(categories)
# Check that categories are included in columns
if categories and not set(categories).intersection(all_columns):
raise ValueError("categories not in available columns.\n"
"categories: {} | columns: {}".format(
categories, list(all_columns)))
dtypes = pf._dtypes(categories)
dtypes = {storage_name_mapping.get(k, k): v for k, v in dtypes.items()}
index_cols = index or ()
meta = _meta_from_dtypes(all_columns, dtypes, index_cols,
column_index_names)
# fastparquet doesn't handle multiindex
if len(index_names) > 1:
raise ValueError("Cannot read DataFrame with MultiIndex.")
for cat in categories:
if cat in meta:
meta[cat] = pd.Series(
pd.Categorical([], categories=[UNKNOWN_CATEGORIES]),
index=meta.index,
)
for catcol in pf.cats:
if catcol in meta.columns:
meta[catcol] = meta[catcol].cat.set_categories(pf.cats[catcol])
elif meta.index.name == catcol:
meta.index = meta.index.set_categories(pf.cats[catcol])
if gather_statistics and pf.row_groups:
stats = []
if filters is None:
filters = []
# make statistics conform in layout
for (i, row_group) in enumerate(pf.row_groups):
s = {"num-rows": row_group.num_rows, "columns": []}
for col in pf.columns:
d = {"name": col}
if pf.statistics["min"][col][0] is not None:
cs_min = pf.statistics["min"][col][i]
cs_max = pf.statistics["max"][col][i]
if isinstance(cs_min, np.datetime64):
cs_min = pd.Timestamp(cs_min)
cs_max = pd.Timestamp(cs_max)
d.update({
"min":
cs_min,
"max":
cs_max,
"null_count":
pf.statistics["null_count"][col][i],
})
s["columns"].append(d)
# Need this to filter out partitioned-on categorical columns
s["filter"] = fastparquet.api.filter_out_cats(
row_group, filters)
stats.append(s)
else:
stats = None
pf._dtypes = lambda *args: pf.dtypes # ugly patch, could be fixed
pf.fmd.row_groups = None
# Create `parts` (list of row-group-descriptor dicts)
parts = [{
"piece": rg,
"kwargs": {
"pf": pf,
"categories": categories_dict or categories
},
} for rg in pf.row_groups]
return (meta, stats, parts)
def read_partition(
cls,
fs,
pieces,
columns,
index,
categories=(),
root_cats=None,
root_file_scheme=None,
base_path=None,
**kwargs,
):
null_index_name = False
base_path = False if not root_cats else base_path
if isinstance(index, list):
if index == [None]:
# Handling a None-labeled index...
# The pandas metadata told us to read in an index
# labeled `None`. If this corresponds to a `RangeIndex`,
# fastparquet will need use the pandas metadata to
# construct the index. Otherwise, the index will correspond
# to a column named "__index_level_0__". We will need to
# check the `ParquetFile` object for this column below.
index = []
null_index_name = True
columns += index
# Use global `parquet_file` object. Need to reattach
# the desired row_group
parquet_file = kwargs.pop("parquet_file", None)
# Always convert pieces to list
if not isinstance(pieces, list):
pieces = [pieces]
sample = pieces[0]
if isinstance(sample, tuple):
if isinstance(sample[0], str):
# We have paths to read from
assert parquet_file is None
row_groups = []
rg_offset = 0
parquet_file = ParquetFile(
[p[0] for p in pieces],
open_with=fs.open,
root=base_path or False,
**kwargs.get("file", {}),
)
for piece in pieces:
_pf = (parquet_file if len(pieces) == 1 else ParquetFile(
piece[0],
open_with=fs.open,
root=base_path or False,
**kwargs.get("file", {}),
))
n_local_row_groups = len(_pf.row_groups)
local_rg_indices = piece[1] or list(
range(n_local_row_groups))
row_groups += [
parquet_file.row_groups[rg + rg_offset]
for rg in local_rg_indices
]
rg_offset += n_local_row_groups
update_parquet_file = len(row_groups) < len(
parquet_file.row_groups)
elif parquet_file:
row_groups = []
for piece in pieces:
# `piece[1]` will contain actual row-group objects,
# but they may be pickled
rgs = piece[0]
if isinstance(rgs, bytes):
rgs = pickle.loads(rgs)
row_groups += rgs
update_parquet_file = True
else:
raise ValueError("Neither path nor ParquetFile detected!")
if update_parquet_file:
with _FP_FILE_LOCK:
parquet_file.fmd.row_groups = row_groups
# NOTE: May lose cats after `_set_attrs` call
save_cats = parquet_file.cats
parquet_file._set_attrs()
parquet_file.cats = save_cats
if null_index_name:
if "__index_level_0__" in parquet_file.columns:
# See "Handling a None-labeled index" comment above
index = ["__index_level_0__"]
columns += index
# Update hive-partitioning information if necessary
parquet_file.cats = root_cats or {}
if root_cats:
parquet_file.file_scheme = root_file_scheme
parquet_file._dtypes = (lambda *args: parquet_file.dtypes
) # ugly patch, could be fixed
if set(columns).issubset(parquet_file.columns +
list(parquet_file.cats.keys())):
# Convert ParquetFile to pandas
return parquet_file.to_pandas(
columns=columns,
categories=categories,
index=index,
)
else:
# Read necessary row-groups and concatenate
dfs = []
for row_group in row_groups:
dfs.append(
parquet_file.read_row_group_file(
row_group,
columns,
categories,
index=index,
**kwargs.get("read", {}),
))
return concat(dfs, axis=0) if len(dfs) > 1 else dfs[0]
else:
# `sample` is NOT a tuple
raise ValueError(f"Expected tuple, got {type(sample)}")
def _collect_file_parts(
cls,
pf_or_files,
dataset_info_kwargs,
):
# Collect necessary information from dataset_info
fs = dataset_info_kwargs["fs"]
split_row_groups = dataset_info_kwargs["split_row_groups"]
gather_statistics = dataset_info_kwargs["gather_statistics"]
stat_col_indices = dataset_info_kwargs["stat_col_indices"]
filters = dataset_info_kwargs["filters"]
dtypes = dataset_info_kwargs["dtypes"]
chunksize = dataset_info_kwargs["chunksize"]
aggregation_depth = dataset_info_kwargs["aggregation_depth"]
base_path = dataset_info_kwargs.get("base_path", None)
root_cats = dataset_info_kwargs.get("root_cats", None)
root_file_scheme = dataset_info_kwargs.get("root_file_scheme", None)
has_metadata_file = dataset_info_kwargs["has_metadata_file"]
# Get ParquetFile
if not isinstance(pf_or_files, fastparquet.api.ParquetFile):
# Construct local `ParquetFile` object
pf = ParquetFile(
pf_or_files,
open_with=fs.open,
root=base_path,
)
# Update hive-partitioning to match global cats/scheme
pf.cats = root_cats or {}
if root_cats:
pf.file_scheme = root_file_scheme
else:
# We already have a ParquetFile object to work with
pf = pf_or_files
# Organize row-groups by file
(
file_row_groups,
file_row_group_stats,
file_row_group_column_stats,
gather_statistics,
base_path,
) = cls._organize_row_groups(
pf,
split_row_groups,
gather_statistics,
stat_col_indices,
filters,
dtypes,
base_path,
has_metadata_file,
chunksize,
aggregation_depth,
)
# Convert organized row-groups to parts
parts, stats = _row_groups_to_parts(
gather_statistics,
split_row_groups,
aggregation_depth,
file_row_groups,
file_row_group_stats,
file_row_group_column_stats,
stat_col_indices,
cls._make_part,
make_part_kwargs={
"fs": fs,
"pf": pf,
"base_path": base_path,
"partitions": pf.info.get("partitions", None),
},
)
return parts, stats
def _collect_dataset_info(
cls,
paths,
fs,
categories,
index,
gather_statistics,
filters,
split_row_groups,
chunksize,
aggregate_files,
ignore_metadata_file,
metadata_task_size,
require_extension=(".parq", ".parquet"),
**kwargs,
):
# Define the parquet-file (pf) object to use for metadata,
# Also, initialize `parts`. If `parts` is populated here,
# then each part will correspond to a file. Otherwise, each part will
# correspond to a row group (populated later).
#
# This logic is mostly to handle `gather_statistics=False` cases,
# because this also means we should avoid scanning every file in the
# dataset. If _metadata is available, set `gather_statistics=True`
# (if `gather_statistics=None`).
parts = []
_metadata_exists = False
if len(paths) == 1 and fs.isdir(paths[0]):
# This is a directory.
# Check if _metadata and/or _common_metadata files exists
base = paths[0]
_metadata_exists = True
if not ignore_metadata_file:
_metadata_exists = fs.isfile(fs.sep.join([base, "_metadata"]))
# Find all files if we are not using a _metadata file
if ignore_metadata_file or not _metadata_exists:
# For now, we need to discover every file under paths[0]
paths, base, fns = _sort_and_analyze_paths(fs.find(base), fs)
_update_paths = False
for fn in ["_metadata", "_common_metadata"]:
try:
fns.remove(fn)
_update_paths = True
except ValueError:
pass
if _update_paths:
paths = [fs.sep.join([base, fn]) for fn in fns]
_metadata_exists = False
if _metadata_exists:
# Using _metadata file (best-case scenario)
pf = ParquetFile(
fs.sep.join([base, "_metadata"]),
open_with=fs.open,
**kwargs,
)
if gather_statistics is None:
gather_statistics = True
else:
# Use 0th file
# Note that "_common_metadata" can cause issues for
# partitioned datasets.
if require_extension:
# Raise error if all files have been filtered by extension
len0 = len(paths)
paths = [
path for path in paths
if path.endswith(require_extension)
]
if len0 and paths == []:
raise ValueError(
"No files satisfy the `require_extension` criteria "
f"(files must end with {require_extension}).")
pf = ParquetFile(paths[:1],
open_with=fs.open,
root=base,
**kwargs)
scheme = get_file_scheme(fns)
pf.file_scheme = scheme
pf.cats = paths_to_cats(fns, scheme)
if not gather_statistics:
parts = [fs.sep.join([base, fn]) for fn in fns]
else:
# This is a list of files
paths, base, fns = _sort_and_analyze_paths(paths, fs)
# Check if _metadata is in paths, and
# remove it if ignore_metadata_file=True
_metadata_exists = "_metadata" in fns
if _metadata_exists and ignore_metadata_file:
fns.remove("_metadata")
_metadata_exists = False
paths = [fs.sep.join([base, fn]) for fn in fns]
if _metadata_exists:
# We have a _metadata file, lets use it
pf = ParquetFile(
fs.sep.join([base, "_metadata"]),
open_with=fs.open,
**kwargs,
)
else:
# Rely on metadata for 0th file.
# Will need to pass a list of paths to read_partition
scheme = get_file_scheme(fns)
pf = ParquetFile(paths[:1],
open_with=fs.open,
root=base,
**kwargs)
pf.file_scheme = scheme
pf.cats = paths_to_cats(fns, scheme)
if not gather_statistics:
parts = paths.copy()
# Check the `aggregate_files` setting
aggregation_depth = _get_aggregation_depth(
aggregate_files,
list(pf.cats),
)
# Ensure that there is no overlap between partition columns
# and explicit columns in `pf`
if pf.cats:
_partitions = [p for p in pf.cats if p not in pf.columns]
if not _partitions:
pf.cats = {}
elif len(_partitions) != len(pf.cats):
raise ValueError(
"No partition-columns should be written in the \n"
"file unless they are ALL written in the file.\n"
"columns: {} | partitions: {}".format(
pf.columns, pf.cats.keys()))
return {
"pf": pf,
"paths": paths,
"has_metadata_file": _metadata_exists,
"parts": parts,
"base": base,
"fs": fs,
"gather_statistics": gather_statistics,
"categories": categories,
"index": index,
"filters": filters,
"split_row_groups": split_row_groups,
"chunksize": chunksize,
"aggregate_files": aggregate_files,
"aggregation_depth": aggregation_depth,
"metadata_task_size": metadata_task_size,
"kwargs": kwargs,
}