import io
import logging
from dataclasses import dataclass
from typing import (
TYPE_CHECKING,
Any,
Callable,
Dict,
Iterable,
Iterator,
List,
Literal,
Optional,
Union,
)
import numpy as np
import ray
from ray.data._internal.util import (
RetryingContextManager,
RetryingPyFileSystem,
_check_pyarrow_version,
_is_local_scheme,
iterate_with_retry,
make_async_gen,
)
from ray.data.block import Block, BlockAccessor
from ray.data.context import DataContext
from ray.data.datasource.datasource import Datasource, ReadTask
from ray.data.datasource.file_meta_provider import (
BaseFileMetadataProvider,
DefaultFileMetadataProvider,
)
from ray.data.datasource.partitioning import (
Partitioning,
PathPartitionFilter,
PathPartitionParser,
)
from ray.data.datasource.path_util import (
_has_file_extension,
_resolve_paths_and_filesystem,
)
from ray.util.annotations import DeveloperAPI
if TYPE_CHECKING:
import pandas as pd
import pyarrow
logger = logging.getLogger(__name__)
# We should parallelize file size fetch operations beyond this threshold.
FILE_SIZE_FETCH_PARALLELIZATION_THRESHOLD = 16
# 16 file size fetches from S3 takes ~1.5 seconds with Arrow's S3FileSystem.
PATHS_PER_FILE_SIZE_FETCH_TASK = 16
[docs]
@DeveloperAPI
@dataclass
class FileShuffleConfig:
"""Configuration for file shuffling.
This configuration object controls how files are shuffled while reading file-based
datasets.
.. note::
Even if you provided a seed, you might still observe a non-deterministic row
order. This is because tasks are executed in parallel and their completion
order might vary. If you need to preserve the order of rows, set
`DataContext.get_current().execution_options.preserve_order`.
Args:
seed: An optional integer seed for the file shuffler. If provided, Ray Data
shuffles files deterministically based on this seed.
Example:
>>> import ray
>>> from ray.data import FileShuffleConfig
>>> shuffle = FileShuffleConfig(seed=42)
>>> ds = ray.data.read_images("s3://anonymous@ray-example-data/batoidea", shuffle=shuffle)
""" # noqa: E501
seed: Optional[int] = None
def __post_init__(self):
"""Ensure that the seed is either None or an integer."""
if self.seed is not None and not isinstance(self.seed, int):
raise ValueError("Seed must be an integer or None.")
[docs]
@DeveloperAPI
class FileBasedDatasource(Datasource):
"""File-based datasource for reading files.
Don't use this class directly. Instead, subclass it and implement `_read_stream()`.
"""
# If `_WRITE_FILE_PER_ROW` is `True`, this datasource calls `_write_row` and writes
# each row to a file. Otherwise, this datasource calls `_write_block` and writes
# each block to a file.
_WRITE_FILE_PER_ROW = False
_FILE_EXTENSIONS: Optional[Union[str, List[str]]] = None
# Number of threads for concurrent reading within each read task.
# If zero or negative, reading will be performed in the main thread.
_NUM_THREADS_PER_TASK = 0
def __init__(
self,
paths: Union[str, List[str]],
*,
filesystem: Optional["pyarrow.fs.FileSystem"] = None,
schema: Optional[Union[type, "pyarrow.lib.Schema"]] = None,
open_stream_args: Optional[Dict[str, Any]] = None,
meta_provider: BaseFileMetadataProvider = DefaultFileMetadataProvider(),
partition_filter: PathPartitionFilter = None,
partitioning: Partitioning = None,
ignore_missing_paths: bool = False,
shuffle: Optional[Union[Literal["files"], FileShuffleConfig]] = None,
include_paths: bool = False,
file_extensions: Optional[List[str]] = None,
):
_check_pyarrow_version()
self._supports_distributed_reads = not _is_local_scheme(paths)
if not self._supports_distributed_reads and ray.util.client.ray.is_connected():
raise ValueError(
"Because you're using Ray Client, read tasks scheduled on the Ray "
"cluster can't access your local files. To fix this issue, store "
"files in cloud storage or a distributed filesystem like NFS."
)
self._schema = schema
self._data_context = DataContext.get_current()
self._open_stream_args = open_stream_args
self._meta_provider = meta_provider
self._partition_filter = partition_filter
self._partitioning = partitioning
self._ignore_missing_paths = ignore_missing_paths
self._include_paths = include_paths
self._unresolved_paths = paths
paths, self._filesystem = _resolve_paths_and_filesystem(paths, filesystem)
self._filesystem = RetryingPyFileSystem.wrap(
self._filesystem, retryable_errors=self._data_context.retried_io_errors
)
paths, file_sizes = map(
list,
zip(
*meta_provider.expand_paths(
paths,
self._filesystem,
partitioning,
ignore_missing_paths=ignore_missing_paths,
)
),
)
if ignore_missing_paths and len(paths) == 0:
raise ValueError(
"None of the provided paths exist. "
"The 'ignore_missing_paths' field is set to True."
)
if self._partition_filter is not None:
# Use partition filter to skip files which are not needed.
path_to_size = dict(zip(paths, file_sizes))
paths = self._partition_filter(paths)
file_sizes = [path_to_size[p] for p in paths]
if len(paths) == 0:
raise ValueError(
"No input files found to read. Please double check that "
"'partition_filter' field is set properly."
)
if file_extensions is not None:
path_to_size = dict(zip(paths, file_sizes))
paths = [p for p in paths if _has_file_extension(p, file_extensions)]
file_sizes = [path_to_size[p] for p in paths]
if len(paths) == 0:
raise ValueError(
"No input files found to read with the following file extensions: "
f"{file_extensions}. Please double check that "
"'file_extensions' field is set properly."
)
_validate_shuffle_arg(shuffle)
self._file_metadata_shuffler = None
if shuffle == "files":
self._file_metadata_shuffler = np.random.default_rng()
elif isinstance(shuffle, FileShuffleConfig):
# Create a NumPy random generator with a fixed seed if provided
self._file_metadata_shuffler = np.random.default_rng(shuffle.seed)
# Read tasks serialize `FileBasedDatasource` instances, and the list of paths
# can be large. To avoid slow serialization speeds, we store a reference to
# the paths rather than the paths themselves.
self._paths_ref = ray.put(paths)
self._file_sizes_ref = ray.put(file_sizes)
def _paths(self) -> List[str]:
return ray.get(self._paths_ref)
def _file_sizes(self) -> List[float]:
return ray.get(self._file_sizes_ref)
def estimate_inmemory_data_size(self) -> Optional[int]:
total_size = 0
for sz in self._file_sizes():
if sz is not None:
total_size += sz
return total_size
def get_read_tasks(self, parallelism: int) -> List[ReadTask]:
import numpy as np
open_stream_args = self._open_stream_args
partitioning = self._partitioning
paths = self._paths()
file_sizes = self._file_sizes()
if self._file_metadata_shuffler is not None:
files_metadata = list(zip(paths, file_sizes))
shuffled_files_metadata = [
files_metadata[i]
for i in self._file_metadata_shuffler.permutation(len(files_metadata))
]
paths, file_sizes = list(map(list, zip(*shuffled_files_metadata)))
filesystem = _wrap_s3_serialization_workaround(self._filesystem)
if open_stream_args is None:
open_stream_args = {}
def read_files(
read_paths: Iterable[str],
) -> Iterable[Block]:
nonlocal filesystem, open_stream_args, partitioning
fs = _unwrap_s3_serialization_workaround(filesystem)
for read_path in read_paths:
partitions: Dict[str, str] = {}
if partitioning is not None:
parse = PathPartitionParser(partitioning)
partitions = parse(read_path)
with RetryingContextManager(
self._open_input_source(fs, read_path, **open_stream_args),
context=self._data_context,
) as f:
for block in iterate_with_retry(
lambda: self._read_stream(f, read_path),
description="read stream iteratively",
match=self._data_context.retried_io_errors,
):
if partitions:
block = _add_partitions(block, partitions)
if self._include_paths:
block_accessor = BlockAccessor.for_block(block)
block = block_accessor.fill_column("path", read_path)
yield block
def create_read_task_fn(read_paths, num_threads):
def read_task_fn():
nonlocal num_threads, read_paths
# TODO: We should refactor the code so that we can get the results in
# order even when using multiple threads.
if self._data_context.execution_options.preserve_order:
num_threads = 0
if num_threads > 0:
if len(read_paths) < num_threads:
num_threads = len(read_paths)
logger.debug(
f"Reading {len(read_paths)} files with {num_threads} threads."
)
yield from make_async_gen(
iter(read_paths),
read_files,
num_workers=num_threads,
preserve_ordering=True,
buffer_size=max(len(read_paths) // num_threads, 1),
)
else:
logger.debug(f"Reading {len(read_paths)} files.")
yield from read_files(read_paths)
return read_task_fn
# fix https://github.com/ray-project/ray/issues/24296
parallelism = min(parallelism, len(paths))
read_tasks = []
split_paths = np.array_split(paths, parallelism)
split_file_sizes = np.array_split(file_sizes, parallelism)
for read_paths, file_sizes in zip(split_paths, split_file_sizes):
if len(read_paths) <= 0:
continue
meta = self._meta_provider(
read_paths,
self._schema,
rows_per_file=self._rows_per_file(),
file_sizes=file_sizes,
)
read_task_fn = create_read_task_fn(read_paths, self._NUM_THREADS_PER_TASK)
read_task = ReadTask(read_task_fn, meta)
read_tasks.append(read_task)
return read_tasks
def _open_input_source(
self,
filesystem: "RetryingPyFileSystem",
path: str,
**open_args,
) -> "pyarrow.NativeFile":
"""Opens a source path for reading and returns the associated Arrow NativeFile.
The default implementation opens the source path as a sequential input stream,
using self._data_context.streaming_read_buffer_size as the buffer size if none
is given by the caller.
Implementations that do not support streaming reads (e.g. that require random
access) should override this method.
"""
import pyarrow as pa
from pyarrow.fs import HadoopFileSystem
compression = open_args.get("compression", None)
if compression is None:
try:
# If no compression manually given, try to detect
# compression codec from path.
compression = pa.Codec.detect(path).name
except (ValueError, TypeError):
# Arrow's compression inference on the file path
# doesn't work for Snappy, so we double-check ourselves.
import pathlib
suffix = pathlib.Path(path).suffix
if suffix and suffix[1:] == "snappy":
compression = "snappy"
else:
compression = None
buffer_size = open_args.pop("buffer_size", None)
if buffer_size is None:
buffer_size = self._data_context.streaming_read_buffer_size
if compression == "snappy":
# Arrow doesn't support streaming Snappy decompression since the canonical
# C++ Snappy library doesn't natively support streaming decompression. We
# works around this by manually decompressing the file with python-snappy.
open_args["compression"] = None
else:
open_args["compression"] = compression
file = filesystem.open_input_stream(path, buffer_size=buffer_size, **open_args)
if compression == "snappy":
import snappy
stream = io.BytesIO()
if isinstance(filesystem.unwrap(), HadoopFileSystem):
snappy.hadoop_snappy.stream_decompress(src=file, dst=stream)
else:
snappy.stream_decompress(src=file, dst=stream)
stream.seek(0)
file = pa.PythonFile(stream, mode="r")
return file
def _rows_per_file(self):
"""Returns the number of rows per file, or None if unknown."""
return None
def _read_stream(self, f: "pyarrow.NativeFile", path: str) -> Iterator[Block]:
"""Streaming read a single file.
This method should be implemented by subclasses.
"""
raise NotImplementedError(
"Subclasses of FileBasedDatasource must implement _read_stream()."
)
@property
def supports_distributed_reads(self) -> bool:
return self._supports_distributed_reads
def _add_partitions(
data: Union["pyarrow.Table", "pd.DataFrame"], partitions: Dict[str, Any]
) -> Union["pyarrow.Table", "pd.DataFrame"]:
import pandas as pd
import pyarrow as pa
assert isinstance(data, (pa.Table, pd.DataFrame))
if isinstance(data, pa.Table):
return _add_partitions_to_table(data, partitions)
if isinstance(data, pd.DataFrame):
return _add_partitions_to_dataframe(data, partitions)
def _add_partitions_to_table(
table: "pyarrow.Table", partitions: Dict[str, Any]
) -> "pyarrow.Table":
import pyarrow as pa
import pyarrow.compute as pc
column_names = set(table.column_names)
for field, value in partitions.items():
column = pa.array([value] * len(table))
if field in column_names:
# TODO: Handle cast error.
column_type = table.schema.field(field).type
column = column.cast(column_type)
values_are_equal = pc.all(pc.equal(column, table[field]))
values_are_equal = values_are_equal.as_py()
if not values_are_equal:
raise ValueError(
f"Partition column {field} exists in table data, but partition "
f"value '{value}' is different from in-data values: "
f"{table[field].unique().to_pylist()}."
)
i = table.schema.get_field_index(field)
table = table.set_column(i, field, column)
else:
table = table.append_column(field, column)
return table
def _add_partitions_to_dataframe(
df: "pd.DataFrame", partitions: Dict[str, Any]
) -> "pd.DataFrame":
import pandas as pd
for field, value in partitions.items():
column = pd.Series(data=[value] * len(df), name=field)
if field in df:
column = column.astype(df[field].dtype)
mask = df[field].notna()
if not df[field][mask].equals(column[mask]):
raise ValueError(
f"Partition column {field} exists in table data, but partition "
f"value '{value}' is different from in-data values: "
f"{list(df[field].unique())}."
)
df[field] = column
return df
def _wrap_s3_serialization_workaround(filesystem: "pyarrow.fs.FileSystem"):
# This is needed because pa.fs.S3FileSystem assumes pa.fs is already
# imported before deserialization. See #17085.
import pyarrow as pa
import pyarrow.fs
base_fs = filesystem
if isinstance(filesystem, RetryingPyFileSystem):
base_fs = filesystem.unwrap()
if isinstance(base_fs, pa.fs.S3FileSystem):
return _S3FileSystemWrapper(filesystem)
return filesystem
def _unwrap_s3_serialization_workaround(
filesystem: Union["pyarrow.fs.FileSystem", "_S3FileSystemWrapper"],
):
if isinstance(filesystem, _S3FileSystemWrapper):
filesystem = filesystem.unwrap()
return filesystem
class _S3FileSystemWrapper:
"""pyarrow.fs.S3FileSystem wrapper that can be deserialized safely.
Importing pyarrow.fs during reconstruction triggers the pyarrow
S3 subsystem initialization.
NOTE: This is only needed for pyarrow<14.0.0 and should be removed
once the minimum supported pyarrow version exceeds that.
See https://github.com/apache/arrow/pull/38375 for context.
"""
def __init__(self, fs: "pyarrow.fs.FileSystem"):
self._fs = fs
def unwrap(self):
return self._fs
@classmethod
def _reconstruct(cls, fs_reconstruct, fs_args):
# Implicitly trigger S3 subsystem initialization by importing
# pyarrow.fs.
import pyarrow.fs # noqa: F401
return cls(fs_reconstruct(*fs_args))
def __reduce__(self):
return _S3FileSystemWrapper._reconstruct, self._fs.__reduce__()
def _resolve_kwargs(
kwargs_fn: Callable[[], Dict[str, Any]], **kwargs
) -> Dict[str, Any]:
if kwargs_fn:
kwarg_overrides = kwargs_fn()
kwargs.update(kwarg_overrides)
return kwargs
def _validate_shuffle_arg(
shuffle: Union[Literal["files"], FileShuffleConfig, None],
) -> None:
if not (
shuffle is None or shuffle == "files" or isinstance(shuffle, FileShuffleConfig)
):
raise ValueError(
f"Invalid value for 'shuffle': {shuffle}. "
"Valid values are None, 'files', `FileShuffleConfig`."
)