PyArrow comes with an abstract filesystem interface, as well as concrete implementations for various storage types.
The filesystem interface provides input and output streams as well as
directory operations. A simplified view of the underlying data
storage is exposed. Data paths are represented as abstract paths, which
/-separated, even on Windows, and shouldn’t include special path
components such as
... Symbolic links, if supported by the
underlying storage, are automatically dereferenced. Only basic
metadata about file entries, such as the file size
and modification time, is made available.
The core interface is represented by the base class
Pyarrow implements natively the following filesystem subclasses:
It is also possible to use your own fsspec-compliant filesystem with pyarrow functionalities as described in the section Using fsspec-compatible filesystems with Arrow.
Instantiating a filesystem#
A FileSystem object can be created with one of the constructors (and check the respective constructor for its options):
>>> from pyarrow import fs >>> local = fs.LocalFileSystem()
or alternatively inferred from a URI:
>>> s3, path = fs.FileSystem.from_uri("s3://my-bucket") >>> s3 <pyarrow._s3fs.S3FileSystem at 0x7f6760cbf4f0> >>> path 'my-bucket'
Reading and writing files#
Several of the IO-related functions in PyArrow accept either a URI (and infer
the filesystem) or an explicit
filesystem argument to specify the filesystem
to read or write from. For example, the
function can be used in the following ways:
import pyarrow.parquet as pq # using a URI -> filesystem is inferred pq.read_table("s3://my-bucket/data.parquet") # using a path and filesystem s3 = fs.S3FileSystem(..) pq.read_table("my-bucket/data.parquet", filesystem=s3)
The filesystem interface further allows to open files for reading (input) or writing (output) directly, which can be combined with functions that work with file-like objects. For example:
import pyarrow as pa local = fs.LocalFileSystem() with local.open_output_stream("test.arrow") as file: with pa.RecordBatchFileWriter(file, table.schema) as writer: writer.write_table(table)
>>> local.get_file_info(fs.FileSelector("dataset/", recursive=True)) [<FileInfo for 'dataset/part=B': type=FileType.Directory>, <FileInfo for 'dataset/part=B/data0.parquet': type=FileType.File, size=1564>, <FileInfo for 'dataset/part=A': type=FileType.Directory>, <FileInfo for 'dataset/part=A/data0.parquet': type=FileType.File, size=1564>]
This returns a list of
FileInfo objects, containing information about
the type (file or directory), the size, the date last modified, etc.
You can also get this information for a single explicit path (or list of paths):
>>> local.get_file_info('test.arrow') <FileInfo for 'test.arrow': type=FileType.File, size=3250> >>> local.get_file_info('non_existent') <FileInfo for 'non_existent': type=FileType.NotFound>
LocalFileSystem allows you to access files on the local machine.
Example how to write to disk and read it back:
>>> from pyarrow import fs >>> local = fs.LocalFileSystem() >>> with local.open_output_stream('/tmp/pyarrowtest.dat') as stream: stream.write(b'data') 4 >>> with local.open_input_stream('/tmp/pyarrowtest.dat') as stream: print(stream.readall()) b'data'
PyArrow implements natively a S3 filesystem for S3 compatible storage.
S3FileSystem constructor has several options to configure the S3
connection (e.g. credentials, the region, an endpoint override, etc). In
addition, the constructor will also inspect configured S3 credentials as
supported by AWS (for example the
AWS_SECRET_ACCESS_KEY environment variables).
Example how you can read contents from a S3 bucket:
>>> from pyarrow import fs >>> s3 = fs.S3FileSystem(region='eu-west-3') # List all contents in a bucket, recursively >>> s3.get_file_info(fs.FileSelector('my-test-bucket', recursive=True)) [<FileInfo for 'my-test-bucket/File1': type=FileType.File, size=10>, <FileInfo for 'my-test-bucket/File5': type=FileType.File, size=10>, <FileInfo for 'my-test-bucket/Dir1': type=FileType.Directory>, <FileInfo for 'my-test-bucket/Dir2': type=FileType.Directory>, <FileInfo for 'my-test-bucket/EmptyDir': type=FileType.Directory>, <FileInfo for 'my-test-bucket/Dir1/File2': type=FileType.File, size=11>, <FileInfo for 'my-test-bucket/Dir1/Subdir': type=FileType.Directory>, <FileInfo for 'my-test-bucket/Dir2/Subdir': type=FileType.Directory>, <FileInfo for 'my-test-bucket/Dir2/Subdir/File3': type=FileType.File, size=10>] # Open a file for reading and download its contents >>> f = s3.open_input_stream('my-test-bucket/Dir1/File2') >>> f.readall() b'some data'
See the AWS docs for the different ways to configure the AWS credentials.
Hadoop Distributed File System (HDFS)#
PyArrow comes with bindings to the Hadoop File System (based on C++ bindings
libhdfs, a JNI-based interface to the Java Hadoop client). You connect
from pyarrow import fs hdfs = fs.HadoopFileSystem(host, port, user=user, kerb_ticket=ticket_cache_path)
libhdfs library is loaded at runtime (rather than at link / library
load time, since the library may not be in your LD_LIBRARY_PATH), and relies on
some environment variables.
HADOOP_HOME: the root of your installed Hadoop distribution. Often has lib/native/libhdfs.so.
JAVA_HOME: the location of your Java SDK installation.
ARROW_LIBHDFS_DIR(optional): explicit location of
libhdfs.soif it is installed somewhere other than
CLASSPATH: must contain the Hadoop jars. You can set these using:
export CLASSPATH=`$HADOOP_HOME/bin/hadoop classpath --glob` # or on Windows %HADOOP_HOME%/bin/hadoop classpath --glob > %CLASSPATH%
In contrast to the legacy HDFS filesystem with
CLASSPATHis not optional (pyarrow will not attempt to infer it).
Using fsspec-compatible filesystems with Arrow#
The filesystems mentioned above are natively supported by Arrow C++ / PyArrow. The Python ecosystem, however, also has several filesystem packages. Those packages following the fsspec interface can be used in PyArrow as well.
Functions accepting a filesystem object will also accept an fsspec subclass. For example:
# creating an fsspec-based filesystem object for Google Cloud Storage import gcsfs fs = gcsfs.GCSFileSystem(project='my-google-project') # using this to read a partitioned dataset import pyarrow.dataset as ds ds.dataset("data/", filesystem=fs)
Similarly for Azure Blob Storage:
import adlfs # ... load your credentials and configure the filesystem fs = adlfs.AzureBlobFileSystem(account_name=account_name, account_key=account_key) import pyarrow.dataset as ds ds.dataset("mycontainer/data/", filesystem=fs)
Under the hood, the fsspec filesystem object is wrapped into a python-based
PyArrow filesystem (
You can also manually do this to get an object with the PyArrow FileSystem
from pyarrow.fs import PyFileSystem, FSSpecHandler pa_fs = PyFileSystem(FSSpecHandler(fs))
Then all the functionalities of
FileSystem are accessible:
# write data with pa_fs.open_output_stream('mycontainer/pyarrowtest.dat') as stream: stream.write(b'data') # read data with pa_fs.open_input_stream('mycontainer/pyarrowtest.dat') as stream: print(stream.readall()) #b'data' # read a partitioned dataset ds.dataset("data/", filesystem=pa_fs)
Using Arrow filesystems with fsspec#
The Arrow FileSystem interface has a limited, developer-oriented API surface. This is sufficient for basic interactions and for using this with Arrow’s IO functionality. On the other hand, the fsspec interface provides a very large API with many helper methods. If you want to use those, or if you need to interact with a package that expects fsspec-compatible filesystem objects, you can wrap an Arrow FileSystem object with fsspec.
fsspec version 2021.09, the
ArrowFSWrapper can be used
>>> from pyarrow import fs >>> local = fs.LocalFileSystem() >>> from fsspec.implementations.arrow import ArrowFSWrapper >>> local_fsspec = ArrowFSWrapper(local)
The resulting object now has an fsspec-compatible interface, while being backed by the Arrow FileSystem under the hood. Example usage to create a directory and file, and list the content:
>>> local_fsspec.mkdir("./test") >>> local_fsspec.touch("./test/file.txt") >>> local_fsspec.ls("./test/") ['./test/file.txt']
For more information, see the fsspec documentation.