Connect to remote data

Dask can read data from a variety of data stores including local file systems, network file systems, cloud object stores, and Hadoop. Typically this is done by prepending a protocol like "s3://" to paths used in common data access functions like dd.read_csv:

import dask.dataframe as dd
df = dd.read_csv('s3://bucket/path/to/data-*.csv')
df = dd.read_parquet('gcs://bucket/path/to/data-*.parq')

import dask.bag as db
b = db.read_text('hdfs://path/to/*.json').map(json.loads)

Dask uses fsspec for local, cluster and remote data IO. Other file interaction, such as loading of configuration, is done using ordinary python method.

The following remote services are well supported and tested against the main codebase:

  • Local or Network File System: file:// - the local file system, default in the absence of any protocol.

  • Hadoop File System: hdfs:// - Hadoop Distributed File System, for resilient, replicated files within a cluster. This uses PyArrow as the backend.

  • Amazon S3: s3:// - Amazon S3 remote binary store, often used with Amazon EC2, using the library s3fs.

  • Google Cloud Storage: gcs:// or gs:// - Google Cloud Storage, typically used with Google Compute resource using gcsfs.

  • Microsoft Azure Storage: adl://, abfs:// or az:// - Microsoft Azure Storage using adlfs.

  • Hugging Face: hf:// - Hugging Face Hub of datasets for AI, using the huggingface_hub library.

  • HTTP(s): http:// or https:// for reading data directly from HTTP web servers.

fsspec also provides other file systems that may be of interest to Dask users, such as ssh, ftp, webhdfs and dropbox. See the documentation for more information.

When specifying a storage location, a URL should be provided using the general form protocol://path/to/data. If no protocol is provided, the local file system is assumed (same as file://).

Lower-level details on how Dask handles remote data is described below in the Internals section

Optional Parameters

Two methods exist for passing parameters to the backend file system driver: extending the URL to include username, password, server, port, etc.; and providing storage_options, a dictionary of parameters to pass on. The second form is more general, as any number of file system-specific options can be passed.


df = dd.read_csv('hdfs://user@server:port/path/*.csv')

df = dd.read_parquet('s3://bucket/path',
                     storage_options={'anon': True, 'use_ssl': False})

Details on how to provide configuration for the main back-ends are listed next, but further details can be found in the documentation pages of the relevant back-end.

Each back-end has additional installation requirements and may not be available at runtime. The dictionary fsspec.registry contains the currently imported file systems. To see which backends fsspec knows how to import, you can do

from fsspec.registry import known_implementations

Note that some backends appear twice, if they can be referenced with multiple protocol strings, like “http” and “https”.

Local File System

Local files are always accessible, and all parameters passed as part of the URL (beyond the path itself) or with the storage_options dictionary will be ignored.

This is the default back-end, and the one used if no protocol is passed at all.

We assume here that each worker has access to the same file system - either the workers are co-located on the same machine, or a network file system is mounted and referenced at the same path location for every worker node.

Locations specified relative to the current working directory will, in general, be respected (as they would be with the built-in python open), but this may fail in the case that the client and worker processes do not necessarily have the same working directory.

Hadoop File System

The Hadoop File System (HDFS) is a widely deployed, distributed, data-local file system written in Java. This file system backs many clusters running Hadoop and Spark. HDFS support can be provided by PyArrow.

By default, the back-end attempts to read the default server and port from local Hadoop configuration files on each node, so it may be that no configuration is required. However, the server, port, and user can be passed as part of the url: hdfs://user:pass@server:port/path/to/data, or using the storage_options= kwarg.

Extra Configuration for PyArrow

The following additional options may be passed to the PyArrow driver via storage_options:

  • host, port, user: Basic authentication

  • kerb_ticket: Path to kerberos ticket cache

PyArrow’s libhdfs driver can also be affected by a few environment variables. For more information on these, see the PyArrow documentation.

Amazon S3

Amazon S3 (Simple Storage Service) is a web service offered by Amazon Web Services.

The S3 back-end available to Dask is s3fs, and is importable when Dask is imported.

Authentication for S3 is provided by the underlying library boto3. As described in the auth docs, this could be achieved by placing credentials files in one of several locations on each node: ~/.aws/credentials, ~/.aws/config, /etc/boto.cfg, and ~/.boto. Alternatively, for nodes located within Amazon EC2, IAM roles can be set up for each node, and then no further configuration is required. The final authentication option for user credentials can be passed directly in the URL (s3://keyID:keySecret/bucket/key/name) or using storage_options. In this case, however, the key/secret will be passed to all workers in-the-clear, so this method is only recommended on well-secured networks.

The following parameters may be passed to s3fs using storage_options:

  • anon: Whether access should be anonymous (default False)

  • key, secret: For user authentication

  • token: If authentication has been done with some other S3 client

  • use_ssl: Whether connections are encrypted and secure (default True)

  • client_kwargs: Dict passed to the boto3 client, with keys such as region_name or endpoint_url. Notice: do not pass the config option here, please pass it’s content to config_kwargs instead.

  • config_kwargs: Dict passed to the s3fs.S3FileSystem, which passes it to the boto3 client’s config option.

  • requester_pays: Set True if the authenticated user will assume transfer costs, which is required by some providers of bulk data

  • default_block_size, default_fill_cache: These are not of particular interest to Dask users, as they concern the behaviour of the buffer between successive reads

  • kwargs: Other parameters are passed to the boto3 Session object, such as profile_name, to pick one of the authentication sections from the configuration files referred to above (see here)

Using Other S3-Compatible Services

By using the endpoint_url option, you may use other s3-compatible services, for example, using AlibabaCloud OSS:

        "key": ...,
        "secret": ...,
        "client_kwargs": {
            "endpoint_url": "",
        # this dict goes to boto3 client's `config`
        #   `addressing_style` is required by AlibabaCloud, other services may not
        "config_kwargs": {"s3": {"addressing_style": "virtual"}},

Google Cloud Storage

Google Cloud Storage is a RESTful online file storage web service for storing and accessing data on Google’s infrastructure.

The GCS back-end is identified by the protocol identifiers gcs and gs, which are identical in their effect.

Multiple modes of authentication are supported. These options should be included in the storage_options dictionary as {'token': ..} submitted with your call to a storage-based Dask function/method. See the gcsfs documentation for further details.

General recommendations for distributed clusters, in order:

  • use anon for public data

  • use cloud if this is available

  • use gcloud to generate a JSON file, and distribute this to all workers, and supply the path to the file

  • use gcsfs directly with the browser method to generate a token cache file (~/.gcs_tokens) and distribute this to all workers, thereafter using method cache

A final suggestion is shown below, which may be the fastest and simplest for authenticated access (as opposed to anonymous), since it will not require re-authentication. However, this method is not secure since credentials will be passed directly around the cluster. This is fine if you are certain that the cluster is itself secured. You need to create a GCSFileSystem object using any method that works for you and then pass its credentials directly:

gcs = GCSFileSystem(...)
dask_function(..., storage_options={'token': gcs.session.credentials})

Microsoft Azure Storage

Microsoft Azure Storage is comprised of Data Lake Storage (Gen1) and Blob Storage (Gen2). These are identified by the protocol identifiers adl and abfs, respectively, provided by the adlfs back-end.

Authentication for adl requires tenant_id, client_id and client_secret in the storage_options dictionary.

Authentication for abfs requires storage_options to contain account_name, tenant_id, client_id and client_secret for the RBAC and ACL access models, or account_name and account_key for the shared key access model.


Direct file-like access to arbitrary URLs is available over HTTP and HTTPS. However, there is no such thing as glob functionality over HTTP, so only explicit lists of files can be used.

Server implementations differ in the information they provide - they may or may not specify the size of a file via a HEAD request or at the start of a download - and some servers may not respect byte range requests. The HTTPFileSystem therefore offers best-effort behaviour: the download is streamed but, if more data is seen than the configured block-size, an error will be raised. To be able to access such data you must read the whole file in one shot (and it must fit in memory).

Using a block size of 0 will return normal requests streaming file-like objects, which are stable, but provide no random access.

Developer API

The prototype for any file system back-end can be found in fsspec.spec.AbstractFileSystem. Any new implementation should provide the same API, or directly subclass, and make itself available as a protocol to Dask. For example, the following would register the protocol “myproto”, described by the implementation class MyProtoFileSystem. URLs of the form myproto:// would thereafter be dispatched to the methods of this class:

fsspec.registry['myproto'] = MyProtoFileSystem

However, it would be better to submit a PR to fsspec to include the class in the known_implementations.


Dask contains internal tools for extensible data ingestion in the dask.bytes package and uses external tools like open_files from fsspec. . These functions are developer-focused rather than for direct consumption by users. These functions power user-facing functions like dd.read_csv and db.read_text which are probably more useful for most users.

read_bytes(urlpath[, delimiter, not_zero, ...])

Given a path or paths, return delayed objects that read from those paths.

This function is extensible in its output format (bytes), its input locations (file system, S3, HDFS), line delimiters, and compression formats.

This function is lazy, returning pointers to blocks of bytes (read_bytes). It handles different storage backends by prepending protocols like s3:// or hdfs:// (see below). It handles compression formats listed in fsspec.compression, some of which may require additional packages to be installed.

This function is not used for all data sources. Some data sources like HDF5 are quite particular and receive custom treatment.


The read_bytes function takes a path (or globstring of paths) and produces a sample of the first file and a list of delayed objects for each of the other files. If passed a delimiter such as delimiter=b'\n', it will ensure that the blocks of bytes start directly after a delimiter and end directly before a delimiter. This allows other functions, like pd.read_csv, to operate on these delayed values with expected behavior.

These delimiters are useful both for typical line-based formats (log files, CSV, JSON) as well as other delimited formats like Avro, which may separate logical chunks by a complex sentinel string. Note that the delimiter finding algorithm is simple, and will not account for characters that are escaped, part of a UTF-8 code sequence or within the quote marks of a string.


These functions support widely available compression technologies like gzip, bz2, xz, snappy, and lz4. More compressions can be easily added by inserting functions into dictionaries available in the fsspec.compression module. This can be done at runtime and need not be added directly to the codebase.

However, most compression technologies like gzip do not support efficient random access, and so are useful for streaming fsspec.open_files but not useful for read_bytes which splits files at various points.


dask.bytes.read_bytes(urlpath, delimiter=None, not_zero=False, blocksize='128 MiB', sample='10 kiB', compression=None, include_path=False, **kwargs)[source]

Given a path or paths, return delayed objects that read from those paths.

The path may be a filename like '2015-01-01.csv' or a globstring like '2015-*-*.csv'.

The path may be preceded by a protocol, like s3:// or hdfs:// if those libraries are installed.

This cleanly breaks data by a delimiter if given, so that block boundaries start directly after a delimiter and end on the delimiter.

urlpathstring or list

Absolute or relative filepath(s). Prefix with a protocol like s3:// to read from alternative filesystems. To read from multiple files you can pass a globstring or a list of paths, with the caveat that they must all have the same protocol.


An optional delimiter, like b'\n' on which to split blocks of bytes.


Force seek of start-of-file delimiter, discarding header.

blocksizeint, str

Chunk size in bytes, defaults to “128 MiB”

compressionstring or None

String like ‘gzip’ or ‘xz’. Must support efficient random access.

sampleint, string, or boolean

Whether or not to return a header sample. Values can be False for “no sample requested” Or an integer or string value like 2**20 or "1 MiB"


Whether or not to include the path with the bytes representing a particular file. Default is False.


Extra options that make sense to a particular storage connection, e.g. host, port, username, password, etc.


The sample header

blockslist of lists of dask.Delayed

Each list corresponds to a file, and each delayed object computes to a block of bytes from that file.

pathslist of strings, only included if include_path is True

List of same length as blocks, where each item is the path to the file represented in the corresponding block.


>>> sample, blocks = read_bytes('2015-*-*.csv', delimiter=b'\n')  
>>> sample, blocks = read_bytes('s3://bucket/2015-*-*.csv', delimiter=b'\n')  
>>> sample, paths, blocks = read_bytes('2015-*-*.csv', include_path=True)