def serialize(vineyard_socket, object_id):
'''Serialize a vineyard object as a stream.
The serialization executes in the following steps:
1. glob all blobs in the meta
2. build a stream for each blob
3. generate a hierarchical `StreamCollection` object as the result
'''
client = vineyard.connect(vineyard_socket)
meta = client.get_meta(object_id)
queue: "ConcurrentQueue[Tuple[ByteStream, memoryview]]" = ConcurrentQueue()
serialized_id = traverse_to_serialize(client, meta, queue, '')
# object id done
client.persist(serialized_id)
report_success(serialized_id)
# start transfer
#
# easy to be implemented as a threaded executor in a future
executor = ThreadStreamExecutor(SerializeExecutor,
parallism=1,
task_queue=queue)
results = executor.execute()
logger.info('finish serialization: %s', results)
def deserialize(vineyard_socket, object_id, proc_num, proc_index):
client = vineyard.connect(vineyard_socket)
streams = client.get(object_id)
if len(streams) != proc_num:
report_error("Expected: %s stream partitions" % proc_num)
sys.exit(-1)
queue: "ConcurrentQueue[Tuple[ByteStream, Union[BlobBuilder, Blob]]]" = (
ConcurrentQueue())
traverse_to_prepare(client, streams[proc_index].id, queue)
# serve as a stream id -> blob id mapping
rqueue: "ConcurrentQueue[Tuple[ObjectID, str, Blob]]" = ConcurrentQueue()
# copy blobs
executor = ThreadStreamExecutor(
ReconstructExecututor,
parallism=1,
client=client,
task_queue=queue,
result_queue=rqueue,
)
executor.execute()
blobs: Dict[ObjectID, Blob] = dict()
while not rqueue.empty():
bs, memberpath, blob = rqueue.get(block=False)
blobs[bs] = (memberpath, blob)
_, result = traverse_to_rebuild(client, streams[proc_index].id, blobs)
client.persist(result.id)
report_success(result.id)
def read_vineyard_dataframe(vineyard_socket, path, storage_options,
read_options, proc_num, proc_index):
client = vineyard.connect(vineyard_socket)
params = dict()
if storage_options:
raise ValueError("Read vineyard current not support storage options")
params["header_row"] = "1" if read_options.get("header_row",
False) else "0"
params["delimiter"] = bytes(read_options.get("delimiter", ","),
"utf-8").decode("unicode_escape")
stream = DataframeStream.new(client, params)
client.persist(stream.id)
report_success(stream.id)
name = urlparse(path).netloc
# the "name" part in URL can be a name, or an ObjectID for convenience.
try:
df_id = client.get_name(name)
except Exception:
df_id = vineyard.ObjectID(name)
dataframes = client.get(df_id)
writer: DataframeStream.Writer = stream.open_writer(client)
try:
for df in dataframes:
batch = pa.RecordBatch.from_pandas(df)
writer.write(batch)
writer.finish()
except Exception:
report_exception()
writer.fail()
sys.exit(-1)
def read_bytes_collection(
vineyard_socket, prefix, storage_options, proc_num, proc_index
):
"""Read a set of files as a collection of ByteStreams."""
client = vineyard.connect(vineyard_socket)
protocol, prefix_path = split_protocol(prefix)
fs = fsspec.filesystem(protocol, **storage_options)
worker_prefix = os.path.join(prefix_path, '%s-%s' % (proc_num, proc_index))
logger.info("start creating blobs ...")
queue: "ConcurrentQueue[Tuple[ByteStream, str]]" = ConcurrentQueue()
stream_id = read_stream_collections(client, fs, queue, worker_prefix, worker_prefix)
client.persist(stream_id)
report_success(stream_id)
logger.info("start reading blobs ...")
executor = ThreadStreamExecutor(
ReadToByteStreamExecutor,
parallism=1,
client=client,
fs=fs,
task_queue=queue,
chunk_size=CHUNK_SIZE,
)
executor.execute()
def read_orc(
vineyard_socket,
path,
storage_options: Dict,
read_options: Dict,
proc_num,
proc_index,
):
# This method is to read the data files of a specific hive table
# that is stored as orc format in HDFS.
#
# In general, the data files of a hive table are stored at the hive
# space in the HDFS with the table name as the directory,
# e.g.,
#
# .. code:: python
#
# '/user/hive/warehouse/sometable'
#
# To read the entire table, simply use 'hive://user/hive/warehouse/sometable'
# as the path.
#
# In case the table is partitioned, use the sub-directory of a specific partition
# to read only the data from that partition. For example, sometable is partitioned
# by column date, we can read the data in a given date by giving path as
#
# .. code:: python
#
# 'hive://user/hive/warehouse/sometable/date=20201112'
#
if proc_index:
raise ValueError("Parallel reading ORC hasn't been supported yet")
if read_options:
raise ValueError("Reading ORC doesn't support read options.")
client = vineyard.connect(vineyard_socket)
stream = DataframeStream.new(client)
client.persist(stream.id)
report_success(stream.id)
writer = stream.open_writer(client)
parsed = urlparse(path)
fs = fsspec.filesystem(parsed.scheme, **storage_options)
if fs.isfile(parsed.path):
files = [parsed.path]
else:
files = [f for f in fs.ls(parsed.path, detail=False) if fs.isfile(f)]
for file_path in files:
read_single_orc(file_path, fs, writer)
# hdfs = HDFileSystem(
# host=host, port=int(port), pars={"dfs.client.read.shortcircuit": "false"}
# )
writer.finish()
def parse_dataframe(vineyard_socket, stream_id, proc_num, proc_index):
client = vineyard.connect(vineyard_socket)
streams = client.get(stream_id)
if len(streams) != proc_num or streams[proc_index] is None:
raise ValueError(
f"Fetch stream error with proc_num={proc_num},proc_index={proc_index}"
)
instream: DataframeStream = streams[proc_index]
stream_reader = instream.open_reader(client)
generate_header_row = instream.params.get("header_row", None) == "1"
delimiter = instream.params.get("delimiter", ",")
stream = ByteStream.new(client, params=instream.params)
client.persist(stream.id)
report_success(stream.id)
stream_writer = stream.open_writer(client)
first_write = generate_header_row
try:
while True:
try:
batch = stream_reader.next() # pa.RecordBatch
except (StopIteration, vineyard.StreamDrainedException):
stream_writer.finish()
break
df = batch.to_pandas()
csv_content = df.to_csv(header=first_write,
index=False,
sep=delimiter).encode('utf-8')
# write to byte stream
first_write = False
chunk = stream_writer.next(len(csv_content))
vineyard.memory_copy(chunk, 0, csv_content)
except Exception:
report_exception()
stream_writer.fail()
sys.exit(-1)
def write_vineyard_dataframe(vineyard_socket, stream_id, proc_num, proc_index):
client = vineyard.connect(vineyard_socket)
streams = client.get(stream_id)
if len(streams) != proc_num or streams[proc_index] is None:
raise ValueError(
f"Fetch stream error with proc_num={proc_num},proc_index={proc_index}"
)
instream: DataframeStream = streams[proc_index]
stream_reader = instream.open_reader(client)
batch_index = 0
while True:
try:
batch = stream_reader.next()
except Exception:
break
df = batch.to_pandas()
df_id = client.put(df,
partition_index=[proc_index, 0],
row_batch_index=batch_index)
batch_index += 1
client.persist(df_id)
report_success(df_id)
def read_bytes( # noqa: C901
vineyard_socket: str,
path: str,
storage_options: Dict,
read_options: Dict,
proc_num: int,
proc_index: int,
):
"""Read bytes from external storage and produce a ByteStream,
which will later be assembled into a ParallelStream.
Args:
vineyard_socket (str): Ipc socket
path (str): External storage path to write to
storage_options (dict): Configurations of external storage
read_options (dict): Additional options that could control the behavior of read
proc_num (int): Total amount of process
proc_index (int): The sequence of this process
Raises:
ValueError: If the stream is invalid.
"""
client = vineyard.connect(vineyard_socket)
params = dict()
read_block_delimiter = read_options.pop('read_block_delimiter', '\n')
if read_block_delimiter is not None:
read_block_delimiter = read_block_delimiter.encode('utf-8')
# Used when reading tables from external storage.
# Usually for load a property graph
header_row = read_options.get("header_row", False)
for k, v in read_options.items():
if k in ("header_row", "include_all_columns"):
params[k] = "1" if v else "0"
elif k == "delimiter":
params[k] = bytes(v, "utf-8").decode("unicode_escape")
else:
params[k] = v
try:
protocol = split_protocol(path)[0]
fs = fsspec.filesystem(protocol, **storage_options)
except Exception:
report_error(
f"Cannot initialize such filesystem for '{path}', "
f"exception is:\n{traceback.format_exc()}"
)
sys.exit(-1)
if fs.isfile(path):
files = [path]
else:
try:
files = fs.glob(path + '*')
assert files, f"Cannot find such files: {path}"
except Exception:
report_error(f"Cannot find such files for '{path}'")
sys.exit(-1)
''' Note [Semantic of read_block with delimiter]:
read_block(fp, begin, size, delimiter) will:
- find the first `delimiter` from `begin`, then starts read
- after `size`, go through util the next `delimiter` or EOF, then finishes read.
Note that the returned size may exceed `size`.
'''
stream, writer = None, None
if 'chunk_size' in storage_options:
chunk_size = parse_readable_size(storage_options['chunk_size'])
else:
chunk_size = 1024 * 1024 * 64 # default: 64MB
try:
for index, file_path in enumerate(files):
with fs.open(file_path, mode="rb") as f:
offset = 0
# Only process header line when processing first file
# And open the writer when processing first file
if index == 0:
if header_row:
header_line = read_block(f, 0, 1, read_block_delimiter)
params["header_line"] = header_line.decode("unicode_escape")
offset = len(header_line)
stream = ByteStream.new(client, params)
client.persist(stream.id)
report_success(stream.id)
writer = stream.open_writer(client)
try:
total_size = f.size()
except TypeError:
total_size = f.size
part_size = (total_size - offset) // proc_num
begin = part_size * proc_index + offset
end = min(begin + part_size, total_size)
# See Note [Semantic of read_block with delimiter].
if index == 0 and proc_index == 0:
begin -= int(header_row)
while begin < end:
buffer = read_block(
f,
begin,
min(chunk_size, end - begin),
delimiter=read_block_delimiter,
)
size = len(buffer)
if size <= 0:
break
begin += size - 1
chunk = writer.next(size)
vineyard.memory_copy(chunk, 0, buffer)
writer.finish()
except Exception:
report_exception()
if writer is not None:
writer.fail()
sys.exit(-1)
def parse_bytes(vineyard_socket, stream_id, proc_num,
proc_index): # noqa: C901
client = vineyard.connect(vineyard_socket)
# get input streams
streams = client.get(stream_id)
if len(streams) != proc_num or streams[proc_index] is None:
raise ValueError(
f"Fetch stream error with proc_num={proc_num},proc_index={proc_index}"
)
instream: ByteStream = streams[proc_index]
stream_reader = instream.open_reader(client)
use_header_row = instream.params.get("header_row", None) == "1"
delimiter = instream.params.get("delimiter", ",")
# process parsing and coverting options
columns = []
column_types = []
original_columns = []
header_line = None
if use_header_row:
header_line: str = instream.params.get('header_line', None)
if not header_line:
report_error(
'Header line not found while header_row is set to True')
sys.exit(-1)
original_columns = header_line.strip().split(delimiter)
schema = instream.params.get('schema', None)
if schema:
columns = schema.split(',')
column_types = instream.params.get('column_types', [])
if column_types:
column_types = column_types.split(',')
include_all_columns = instream.params.get('include_all_columns',
None) == '1'
read_options = pa.csv.ReadOptions()
parse_options = pa.csv.ParseOptions()
convert_options = pa.csv.ConvertOptions()
if original_columns:
read_options.column_names = original_columns
else:
read_options.autogenerate_column_names = True
parse_options.delimiter = delimiter
indices = []
for i, column in enumerate(columns):
if original_columns:
if column.isdigit():
column_index = int(column)
if column_index >= len(original_columns):
raise IndexError('Column index out of range: %s of %s' %
(column_index, original_columns))
indices.append(i)
columns[i] = original_columns[column_index]
else:
columns[
i] = 'f%s' % i # arrow auto generates column names in that way.
if include_all_columns:
for column in original_columns:
if column not in columns:
columns.append(column)
if columns:
convert_options.include_columns = columns
if len(column_types) > len(columns):
raise ValueError(
"Format of column type schema is incorrect: too many columns")
arrow_column_types = dict()
for i, column_type in enumerate(column_types):
if column_type:
arrow_column_types[columns[i]] = normalize_arrow_dtype(column_type)
convert_options.column_types = arrow_column_types
stream = DataframeStream.new(client, params=instream.params)
client.persist(stream.id)
report_success(stream.id)
stream_writer = stream.open_writer(client)
try:
while True:
try:
content = stream_reader.next()
except (StopIteration, vineyard.StreamDrainedException):
stream_writer.finish()
break
# parse csv
table = parse_dataframe_blocks(content, read_options,
parse_options, convert_options)
# write recordbatches
stream_writer.write_table(table)
except Exception:
report_exception()
stream_writer.fail()
sys.exit(-1)
