def place(
df: pd.DataFrame,
path: str,
filesystem: FileSystem = FileSystem(),
tmpdir: str = None,
**write_args,
) -> None:
"""Place a dataframe block onto the filesystem at the specified path
Parameters
----------
df : pd.DataFrame
The data to place
path : str
Path to the directory (possibly on GCS) in which to place the columns
write_args : dict
Any additional args to pass to the write function
filesystem : blocks.filesystem.FileSystem or similar
A filesystem object that implements the blocks.FileSystem API
"""
fname = os.path.basename(path)
tmp = os.path.join(tmpdir, fname)
write_df(df, tmp, **write_args)
filesystem.copy(tmp, path)
def _reload(data, path, **kwargs):
with open(path, 'wb') as f:
d = DataFile(path, f)
io.write_df(data, d, **kwargs)
with open(path, 'rb') as f:
d = DataFile(path, f)
df = io.read_df(d)
return df
def test_compression_parquet(randomdata, datadir_local):
pytest.importorskip("pyarrow")
pytest.importorskip("pandas", minversion="0.22.0")
path = os.path.join(datadir_local, "tmp.parquet.gz")
# write compressed with pandas
randomdata.to_parquet(path, compression="gzip")
# read compressed with blocks
d = LocalDataFile(path, path)
df = io.read_df(d)
assert np.isclose(df, randomdata).all().all()
# write compressed with blocks
d = LocalDataFile(path, path)
io.write_df(randomdata, d)
# read compressed with pandas
df = pd.read_parquet(path)
assert np.isclose(df, randomdata).all().all()
def place(df, path, filesystem=GCSFileSystem(), **write_args):
""" Place a dataframe block onto the filesystem at the specified path
Parameters
----------
df : pd.DataFrame
The data to place
path : str
Path to the directory (possibly on GCS) in which to place the columns
write_args : dict
Any additional args to pass to the write function
filesystem : blocks.filesystem.FileSystem or similar
A filesystem object that implements the blocks.FileSystem API
"""
bucket, fname = os.path.dirname(path), os.path.basename(path)
with filesystem.store(bucket, [fname]) as datafiles:
write_df(df, datafiles[0], **write_args)
def test_compression_pickle(randomdata, datadir_local):
path = os.path.join(datadir_local, 'tmp.pkl.gz')
# write compressed with pandas
randomdata.to_pickle(path, compression='gzip')
# read compressed with blocks
with open(path, 'rb') as f:
d = DataFile(path, f)
df = io.read_df(d)
assert (np.isclose(df, randomdata).all().all())
# write compressed with blocks
with open(path, 'wb') as f:
d = DataFile(path, f)
io.write_df(randomdata, d)
# read compressed with pandas
df = pd.read_pickle(path, compression='gzip')
assert (np.isclose(df, randomdata).all().all())
def test_compression_parquet(randomdata, datadir_local):
pytest.importorskip('pyarrow')
pytest.importorskip('pandas', minversion='0.22.0')
path = os.path.join(datadir_local, 'tmp.parquet.gz')
# write compressed with pandas
randomdata.to_parquet(path, compression='gzip')
# read compressed with blocks
with open(path, 'rb') as f:
d = DataFile(path, f)
df = io.read_df(d)
assert (np.isclose(df, randomdata).all().all())
# write compressed with blocks
with open(path, 'wb') as f:
d = DataFile(path, f)
io.write_df(randomdata, d)
# read compressed with pandas
df = pd.read_parquet(path)
assert (np.isclose(df, randomdata).all().all())
def divide(
df: pd.DataFrame,
path: str,
n_rgroup: int = 1,
rgroup_offset: int = 0,
cgroup_columns: Optional[Dict[Optional[cgroup], Sequence[str]]] = None,
extension: str = ".pq",
convert: bool = False,
filesystem: FileSystem = FileSystem(),
prefix=None,
tmpdir: str = None,
**write_args,
) -> None:
"""Split a dataframe into rgroups/cgroups and save to disk
Note that this splitting does not preserve the original index, so make sure
to have another column to track values
Parameters
----------
df : pd.DataFrame
The data to divide
path : str
Path to the directory (possibly on GCS) in which to place the columns
n_rgroup : int, default 1
The number of row groups to partition the data into
The rgroups will have approximately equal sizes
rgroup_offset : int, default 0
The index to start from in the name of file parts
e.g. If rgroup_offset=10 then the first file will be `part_00010.pq`
cgroup_columns : {cgroup: list of column names}
The column lists to form cgroups; if None, do not make cgroups
Each key is the name of the cgroup, and each value is the list of columns to include
To reassemble later make sure to include join keys for each cgroup
extension : str, default .pq
The file extension for the dataframe (file type inferred from this extension
convert : bool, default False
If true attempt to coerce types to numeric. This can avoid issues with ambiguous
object columns but requires additional time
filesystem : blocks.filesystem.FileSystem or similar
A filesystem object that implements the blocks.FileSystem API
prefix: str
Prefix to add to written filenames
write_args : dict
Any additional args to pass to the write function
"""
# Use a single dummy cgroup if None wanted
if cgroup_columns is None:
cgroup_columns = {None: df.columns}
# Add leading dot if not in extension
if extension[0] != ".":
extension = "." + extension
if convert:
for col in df.columns:
df[col] = pd.to_numeric(df[col], errors="ignore")
files = []
for cname, columns in cgroup_columns.items():
cgroup = df[columns]
bucket = os.path.join(path, cname) if cname else path
tmp_cgroup = os.path.join(tmpdir, cname) if cname else tmpdir
if not filesystem.isdir(tmp_cgroup):
filesystem.mkdir(tmp_cgroup)
rnames = [
"part_{:05d}{}".format(i + rgroup_offset, extension)
for i in range(n_rgroup)
]
if prefix is not None:
rnames = [prefix + "_" + rn for rn in rnames]
for rgroup, rname in zip(np.array_split(cgroup, n_rgroup), rnames):
tmp = os.path.join(tmp_cgroup, rname)
write_df(rgroup.reset_index(drop=True), tmp, **write_args)
files.append((cname, rname) if cname else (rname, ))
filesystem.copy(
[os.path.join(tmpdir, *f) for f in files],
[os.path.join(path, *f) for f in files],
)
def _reload(data, path, **kwargs):
d = LocalDataFile(path, path)
io.write_df(data, d, **kwargs)
d = LocalDataFile(path, path)
df = io.read_df(d)
return df
def divide(
df, path,
n_rgroup=1,
rgroup_offset=0,
cgroup_columns=None,
extension='.pq',
convert=False,
filesystem=GCSFileSystem(),
**write_args
):
""" Split a dataframe into rgroups/cgroups and save to disk
Note that this splitting does not preserve the original index, so make sure
to have another column to track values
Parameters
----------
df : pd.DataFrame
The data to divide
path : str
Path to the directory (possibly on GCS) in which to place the columns
n_rgroup : int, default 1
The number of row groups to partition the data into
The rgroups will have approximately equal sizes
rgroup_offset : int, default 0
The index to start from in the name of file parts
e.g. If rgroup_offset=10 then the first file will be `part_00010.pq`
cgroup_columns : {cgroup: list of column names}
The column lists to form cgroups; if None, do not make cgroups
Each key is the name of the cgroup, and each value is the list of columns to include
To reassemble later make sure to include join keys for each cgroup
extension : str, default .pq
The file extension for the dataframe (file type inferred from this extension
convert : bool, default False
If true attempt to coerce types to numeric. This can avoid issues with amiguous
object columns but requires additional time
filesystem : blocks.filesystem.FileSystem or similar
A filesystem object that implements the blocks.FileSystem API
write_args : dict
Any additional args to pass to the write function
"""
# Use a single dummy cgroup if None wanted
if cgroup_columns is None:
cgroup_columns = {None: df.columns}
# Add leading dot if not in extension
if extension[0] != '.':
extension = '.'+extension
if convert:
for col in df.columns:
df[col] = pd.to_numeric(df[col], errors='ignore')
for cname, columns in cgroup_columns.items():
cgroup = df[columns]
bucket = os.path.join(path, cname) if cname else path
rnames = ['part_{:05d}{}'.format(i+rgroup_offset, extension) for i in range(n_rgroup)]
with filesystem.store(bucket, rnames) as datafiles:
for rgroup, d in zip(np.array_split(cgroup, n_rgroup), datafiles):
write_df(rgroup.reset_index(drop=True), d, **write_args)
