A pythonic file-system interface to Google Cloud Storage.

This software is beta, use at your own risk.

Please file issues and requests on github and we welcome pull requests.

This package depends on fsspec , and inherits many useful behaviours from there, including integration with Dask, and the facility for key-value dict-like objects of the type used by zarr.


The GCSFS library can be installed using conda or pip:

conda install -c conda-forge gcsfs
pip install gcsfs

or by cloning the repository:

git clone https://github.com/fsspec/gcsfs/
cd gcsfs/
pip install .


Locate and read a file:

>>> import gcsfs
>>> fs = gcsfs.GCSFileSystem(project='my-google-project')
>>> fs.ls('my-bucket')
>>> with fs.open('my-bucket/my-file.txt', 'rb') as f:
...     print(f.read())
b'Hello, world'

(see also walk and glob)

Read with delimited blocks:

>>> fs.read_block(path, offset=1000, length=10, delimiter=b'\n')
b'A whole line of text\n'

Write with blocked caching:

>>> with fs.open('mybucket/new-file', 'wb') as f:
...     f.write(2*2**20 * b'a')
...     f.write(2*2**20 * b'a') # data is flushed and file closed
>>> fs.du('mybucket/new-file')
{'mybucket/new-file': 4194304}

Because GCSFS faithfully copies the Python file interface it can be used smoothly with other projects that consume the file interface like gzip or pandas.

>>> with fs.open('mybucket/my-file.csv.gz', 'rb') as f:
...     g = gzip.GzipFile(fileobj=f)  # Decompress data with gzip
...     df = pd.read_csv(g)           # Read CSV file with Pandas


Several modes of authentication are supported:

  • if token=None (default), GCSFS will attempt to use your default gcloud credentials or, attempt to get credentials from the google metadata service, or fall back to anonymous access. This will work for most users without further action. Note that the default project may also be found, but it is often best to supply this anyway (only affects bucket- level operations).

  • if token='cloud', we assume we are running within google (compute or container engine) and fetch the credentials automatically from the metadata service.

  • you may supply a token generated by the gcloud utility; this is either a python dictionary, or the name of a file containing the JSON returned by logging in with the gcloud CLI tool (e.g., ~/.config/gcloud/application_default_credentials.json or ~/.config/gcloud/legacy_credentials/<YOUR GOOGLE USERNAME>/adc.json) or any value google Credentials object.

  • you can also generate tokens via Oauth2 in the browser using token='browser', which gcsfs then caches in a special file, ~/.gcs_tokens, and can subsequently be accessed with token='cache'.

  • anonymous only access can be selected using token='anon', e.g. to access public resources such as ‘anaconda-public-data’.

The acquired session tokens are not preserved when serializing the instances, so it is safe to pass them to worker processes on other machines if using in a distributed computation context. If credentials are given by a file path, however, then this file must exist on every machine.


The libraries intake, pandas and dask accept URLs with the prefix “gcs://”, and will use gcsfs to complete the IO operation in question. The IO functions take an argument storage_options, which will be passed to GCSFileSystem, for example:

df = pd.read_excel("gcs://bucket/path/file.xls",
                   storage_options={"token": "anon"})

This gives the chance to pass any credentials or other necessary arguments needed to s3fs.


gcsfs is implemented using aiohttp, and offers async functionality. A number of methods of GCSFileSystem are async, for for each of these, there is also a synchronous version with the same name and lack of a “_” prefix.

If you wish to call gcsfs from async code, then you should pass asynchronous=True, loop=loop to the constructor (the latter is optional, if you wish to use both async and sync methods). You must also explicitly await the client creation before making any GCS call.

loop = ...  # however you create your loop

async def run_program(loop):
    gcs = GCSFileSystem(..., asynchronous=True, loop=loop)
    await gcs.set_session()
    ...  # perform work

asyncio.run(run_program(loop))  # or call from your async code

Concurrent async operations are also used internally for bulk operations such as pipe/cat, get/put, cp/mv/rm. The async calls are hidden behind a synchronisation layer, so are designed to be called from normal code. If you are not using async-style programming, you do not need to know about how this works, but you might find the implementation interesting.



GCSFileSystem(*args, **kwargs)

Connect to Google Cloud Storage.

GCSFileSystem.cat(path[, recursive, on_error])

Fetch (potentially multiple) paths' contents

GCSFileSystem.du(path[, total, maxdepth])

Space used by files within a path

GCSFileSystem.exists(path, **kwargs)

Is there a file at the given path

GCSFileSystem.get(rpath, lpath[, recursive, ...])

Copy file(s) to local.

GCSFileSystem.glob(path, **kwargs)

Find files by glob-matching.

GCSFileSystem.info(path, **kwargs)

Give details of entry at path

GCSFileSystem.ls(path[, detail])

List objects at path.

GCSFileSystem.mkdir(path[, acl, ...])

New bucket

GCSFileSystem.mv(path1, path2[, recursive, ...])

Move file(s) from one location to another

GCSFileSystem.open(path[, mode, block_size, ...])

Return a file-like object from the filesystem

GCSFileSystem.put(lpath, rpath[, recursive, ...])

Copy file(s) from local.

GCSFileSystem.read_block(fn, offset, length)

Read a block of bytes from

GCSFileSystem.rm(path[, recursive, ...])

Delete files.

GCSFileSystem.tail(path[, size])

Get the last size bytes from file

GCSFileSystem.touch(path[, truncate])

Create empty file, or update timestamp

GCSFileSystem.get_mapper([root, check, ...])

Create key/value store based on this file-system

GCSFile(gcsfs, path[, mode, block_size, ...])



Close file


Write buffered data to backend store.


File information about this path


Return data from cache, or fetch pieces as necessary

GCSFile.seek(loc[, whence])

Set current file location


Current file location


Write data to buffer.

class gcsfs.core.GCSFileSystem(*args, **kwargs)[source]

Connect to Google Cloud Storage.

The following modes of authentication are supported:

  • token=None, GCSFS will attempt to guess your credentials in the following order: gcloud CLI default, gcsfs cached token, google compute metadata service, anonymous.

  • token='google_default', your default gcloud credentials will be used, which are typically established by doing gcloud login in a terminal.

  • token=='cache', credentials from previously successful gcsfs authentication will be used (use this after “browser” auth succeeded)

  • token='anon', no authentication is performed, and you can only access data which is accessible to allUsers (in this case, the project and access level parameters are meaningless)

  • token='browser', you get an access code with which you can authenticate via a specially provided URL

  • if token='cloud', we assume we are running within google compute or google container engine, and query the internal metadata directly for a token.

  • you may supply a token generated by the [gcloud](https://cloud.google.com/sdk/docs/) utility; this is either a python dictionary, the name of a file containing the JSON returned by logging in with the gcloud CLI tool, or a Credentials object. gcloud typically stores its tokens in locations such as ~/.config/gcloud/application_default_credentials.json, `` ~/.config/gcloud/credentials``, or ~\AppData\Roaming\gcloud\credentials, etc.

Specific methods, (eg. ls, info, …) may return object details from GCS. These detailed listings include the [object resource](https://cloud.google.com/storage/docs/json_api/v1/objects#resource)

GCS does not include “directory” objects but instead generates directories by splitting [object names](https://cloud.google.com/storage/docs/key-terms). This means that, for example, a directory does not need to exist for an object to be created within it. Creating an object implicitly creates it’s parent directories, and removing all objects from a directory implicitly deletes the empty directory.

GCSFileSystem generates listing entries for these implied directories in listing apis with the object properties:

  • “name”string

    The “{bucket}/{name}” path of the dir, used in calls to GCSFileSystem or GCSFile.

  • “bucket”string

    The name of the bucket containing this object.

  • “kind” : ‘storage#object’

  • “size” : 0

  • “storageClass” : ‘DIRECTORY’

  • type: ‘directory’ (fsspec compat)

GCSFileSystem maintains a per-implied-directory cache of object listings and fulfills all object information and listing requests from cache. This implied, for example, that objects created via other processes will not be visible to the GCSFileSystem until the cache refreshed. Calls to GCSFileSystem.open and calls to GCSFile are not effected by this cache.

In the default case the cache is never expired. This may be controlled via the cache_timeout GCSFileSystem parameter or via explicit calls to GCSFileSystem.invalidate_cache.


project_id to work under. Note that this is not the same as, but often very similar to, the project name. This is required in order to list all the buckets you have access to within a project and to create/delete buckets, or update their access policies. If token='google_default', the value is overridden by the default, if token='anon', the value is ignored.

accessone of {‘read_only’, ‘read_write’, ‘full_control’}

Full control implies read/write as well as modifying metadata, e.g., access control.

token: None, dict or string

(see description of authentication methods, above)

consistency: ‘none’, ‘size’, ‘md5’

Check method when writing files. Can be overridden in open().

cache_timeout: float, seconds

Cache expiration time in seconds for object metadata cache. Set cache_timeout <= 0 for no caching, None for no cache expiration.

secure_serialize: bool (deprecated)
requester_paysbool, or str default False

Whether to use requester-pays requests. This will include your project ID project in requests as the userPorject, and you’ll be billed for accessing data from requester-pays buckets. Optionally, pass a project-id here as a string to use that as the userProject.

session_kwargs: dict

passed on to aiohttp.ClientSession; can contain, for example, proxy settings.

endpoint_url: str

If given, use this URL (format protocol://host:port , without any path part) for communication. If not given, defaults to the value of environment variable “STORAGE_EMULATOR_HOST”; if that is not set either, will use the standard Google endpoint.

default_location: str

Default location where buckets are created, like ‘US’ or ‘EUROPE-WEST3’. You can find a list of all available locations here: https://cloud.google.com/storage/docs/locations#available-locations


Return list of available project buckets.


A context within which files are committed together upon exit


cat(path[, recursive, on_error])

Fetch (potentially multiple) paths' contents

cat_file(path[, start, end])

Get the content of a file


Unique value for current version of file


Clear the cache of filesystem instances.

copy(path1, path2[, recursive, on_error])

Copy within two locations in the filesystem

cp(path1, path2, **kwargs)

Alias of AbstractFileSystem.copy.


Return the created timestamp of a file as a datetime.datetime


Return the most recently instantiated FileSystem

delete(path[, recursive, maxdepth])

Alias of AbstractFileSystem.rm.

disk_usage(path[, total, maxdepth])

Alias of AbstractFileSystem.du.

download(rpath, lpath[, recursive])

Alias of AbstractFileSystem.get.

du(path[, total, maxdepth])

Space used by files within a path


Finish write transaction, non-context version

exists(path, **kwargs)

Is there a file at the given path

expand_path(path[, recursive, maxdepth])

Turn one or more globs or directories into a list of all matching paths to files or directories.

find(path[, maxdepth, withdirs, detail])

List all files below path.


Recreate a filesystem instance from JSON representation

get(rpath, lpath[, recursive, callback])

Copy file(s) to local.

get_file(rpath, lpath[, callback, outfile])

Copy single remote file to local

get_mapper([root, check, create, ...])

Create key/value store based on this file-system

getxattr(path, attr)

Get user-defined metadata attribute

glob(path, **kwargs)

Find files by glob-matching.

head(path[, size])

Get the first size bytes from file

info(path, **kwargs)

Give details of entry at path


Invalidate listing cache for given path, it is reloaded on next use.


Is this entry directory-like?


Is this entry file-like?

lexists(path, **kwargs)

If there is a file at the given path (including broken links)

listdir(path[, detail])

Alias of AbstractFileSystem.ls.

ls(path[, detail])

List objects at path.

makedir(path[, create_parents])

Alias of AbstractFileSystem.mkdir.

makedirs(path[, exist_ok])

Recursively make directories

merge(path, paths[, acl])

Concatenate objects within a single bucket

mkdir(path[, acl, default_acl, location, ...])

New bucket

mkdirs(path[, exist_ok])

Alias of AbstractFileSystem.makedirs.


Return the modified timestamp of a file as a datetime.datetime

move(path1, path2, **kwargs)

Alias of AbstractFileSystem.mv.

mv(path1, path2[, recursive, maxdepth])

Move file(s) from one location to another

open(path[, mode, block_size, ...])

Return a file-like object from the filesystem

pipe(path[, value])

Put value into path

pipe_file(path, value, **kwargs)

Set the bytes of given file

put(lpath, rpath[, recursive, callback])

Copy file(s) from local.

put_file(lpath, rpath[, callback])

Copy single file to remote

read_block(fn, offset, length[, delimiter])

Read a block of bytes from

rename(path1, path2, **kwargs)

Alias of AbstractFileSystem.mv.


Delete a file


Delete an empty bucket

setxattrs(path[, content_type, ...])

Set/delete/add writable metadata attributes

sign(path[, expiration])

Create a signed URL representing the given path.


Size in bytes of file


Size in bytes of each file in a list of paths


Normalise GCS path string into bucket and key.


Begin write transaction for deferring files, non-context version

stat(path, **kwargs)

Alias of AbstractFileSystem.info.

tail(path[, size])

Get the last size bytes from file


JSON representation of this filesystem instance

touch(path[, truncate])

Create empty file, or update timestamp


Hash of file properties, to tell if it has changed


Format FS-specific path to generic, including protocol

upload(lpath, rpath[, recursive])

Alias of AbstractFileSystem.put.


Get HTTP URL of the given path

walk(path[, maxdepth])

Return all files belows path








property buckets

Return list of available project buckets.

getxattr(path, attr)

Get user-defined metadata attribute


Invalidate listing cache for given path, it is reloaded on next use.

path: string or None

If None, clear all listings cached else listings at or under given path.

merge(path, paths, acl=None)

Concatenate objects within a single bucket

mkdir(path, acl='projectPrivate', default_acl='bucketOwnerFullControl', location=None, create_parents=True, **kwargs)

New bucket

If path is more than just a bucket, will create bucket if create_parents=True; otherwise is a noop. If create_parents is False and bucket does not exist, will produce FileNotFFoundError.

path: str

bucket name. If contains ‘/’ (i.e., looks like subdir), will have no effect because GCS doesn’t have real directories.

acl: string, one of bACLs

access for the bucket itself

default_acl: str, one of ACLs

default ACL for objects created in this bucket

location: Optional[str]

Location where buckets are created, like ‘US’ or ‘EUROPE-WEST3’. If not provided, defaults to self.default_location. You can find a list of all available locations here: https://cloud.google.com/storage/docs/locations#available-locations

create_parents: bool

If True, creates the bucket in question, if it doesn’t already exist

rm(path, recursive=False, maxdepth=None, batchsize=20)

Delete files.

path: str or list of str

File(s) to delete.

recursive: bool

If file(s) are directories, recursively delete contents and then also remove the directory

maxdepth: int or None

Depth to pass to walk for finding files to delete, if recursive. If None, there will be no limit and infinite recursion may be possible.


Delete an empty bucket

bucket: str

bucket name. If contains ‘/’ (i.e., looks like subdir), will have no effect because GCS doesn’t have real directories.

setxattrs(path, content_type=None, content_encoding=None, fixed_key_metadata=None, **kwargs)

Set/delete/add writable metadata attributes

content_type: str

If not None, set the content-type to this value

content_encoding: str

This parameter is deprecated, you may use fixed_key_metadata instead. If not None, set the content-encoding. See https://cloud.google.com/storage/docs/transcoding

fixed_key_metadata: dict
Google metadata, in key/value pairs, supported keys:
  • cache_control

  • content_disposition

  • content_encoding

  • content_language

  • custom_time

More info: https://cloud.google.com/storage/docs/metadata#mutable

kw_args: key-value pairs like field=”value” or field=None

value must be string to add or modify, or None to delete

Entire metadata after update (even if only path is passed)
sign(path, expiration=100, **kwargs)[source]

Create a signed URL representing the given path.


The path on the filesystem


Number of seconds to enable the URL for


The signed URL

classmethod split_path(path)[source]

Normalise GCS path string into bucket and key.


Input path, like gcs://mybucket/path/to/file. Path is of the form: ‘[gs|gcs://]bucket[/key]’

(bucket, key) tuple

Get HTTP URL of the given path

class gcsfs.core.GCSFile(gcsfs, path, mode='rb', block_size=5242880, autocommit=True, cache_type='readahead', cache_options=None, acl=None, consistency='md5', metadata=None, content_type=None, timeout=None, fixed_key_metadata=None, **kwargs)[source]



Close file


If not auto-committing, finalize file


Cancel in-progress multi-upload


Returns underlying file descriptor if one exists.


Write buffered data to backend store.


File information about this path


Return whether this is an 'interactive' stream.


Return data from cache, or fetch pieces as necessary


Whether opened for reading


mirrors builtin file's readinto method


Read until first occurrence of newline character


Return all data, split by the newline character

readuntil([char, blocks])

Return data between current position and first occurrence of char

seek(loc[, whence])

Set current file location


Whether is seekable (only in read mode)


Current file location


Truncate file to size bytes.


HTTP link to this file's data


Whether opened for writing


Write data to buffer.

writelines(lines, /)

Write a list of lines to stream.



If not auto-committing, finalize file


Cancel in-progress multi-upload

Should only happen during discarding this write-mode file


File information about this path


HTTP link to this file’s data

For Developers

We welcome contributions to gcsfs!

Please file issues and requests on github and we welcome pull requests.


The testing framework supports using your own GCS-compliant endpoint, by setting the “STORAGE_EMULATOR_HOST” environment variable. If this is not set, then an emulator will be spun up using docker and fake-gcs-server. This emulator has almost all the functionality of real GCS. A small number of tests run differently or are skipped.

If you want to actually test against real GCS, then you should set STORAGE_EMULATOR_HOST to “https://storage.googleapis.com” and also provide appropriate GCSFS_TEST_BUCKET and GCSFS_TEST_PROJECT, as well as setting your default google credentials (or providing them via the fsspec config).


Warning, this functionality is experimental

FUSE is a mechanism to mount user-level filesystems in unix-like systems (linux, osx, etc.). GCSFS is able to use FUSE to present remote data/keys as if they were a directory on your local file-system. This allows for standard shell command manipulation, and loading of data by libraries that can only handle local file-paths (e.g., netCDF/HDF5).


In addition to a standard installation of GCSFS, you also need:

  • libfuse as a system install. The way to install this will depend on your OS. Examples include sudo apt-get install fuse, sudo yum install fuse and download from osxfuse.

  • fusepy, which can be installed via conda or pip

  • pandas, which can also be installed via conda or pip (this library is used only for its timestring parsing.


FUSE functionality is available via the fsspec.fuse module. See the docstrings for further details.

gcs = gcsfs.GCSFileSystem(..)
from fsspec.fuse import run
run(gcs, "bucket/path", "local/path", foreground=True, threads=False)


This functionality is experimental. The command usage may change, and you should expect exceptions.


  • although mutation operations tentatively work, you should not at the moment depend on gcsfuse as a reliable system that won’t loose your data.

  • permissions on GCS are complicated, so all files will be shown as fully-open 0o777, regardless of state. If a read fails, you likely don’t have the right permissions.



  • invalidate listings cache for simple put/pipe (#474)

  • conform _mkdir and _cat_file to upstream (#471)


(note that this release happened in 2022.4, but we label as 2022.3 to match fsspec)

  • bucket exists workaround (#464)

  • dirmarkers (#459)

  • check connection (#457)

  • browser connection now uses local server (#456)

  • bucket location (#455)

  • ensure auth is closed (#452)


  • fix list_buckets without cache (#449)

  • drop py36 (#445)


  • update refname for versions (#442)


  • don’t touch cache when doing find with a prefix (#437)


  • move to fsspec org

  • add support for google fixed_key_metadata (#429)

  • deprecate content_encoding parameter of setxattrs method (#429)

  • use emulator for resting instead of vcrpy (#424)


  • url signing (#411)

  • default callback (#422)


  • min version for decorator

  • default callback in get (#422)


  • correctly recognise 404 (#419)

  • fix for .details due to upstream (#417)

  • callbacks in get/put (#416)

  • “%” in paths (#415)


  • don’t retry 404s (#406)


  • fix find/glob with a prefix (#399)


  • kwargs to aiohttpClient session

  • graceful timeout when disconnecting at finalise (#397)


  • negative ranges in cat_file (#394)


  • no credentials bug fix (#390)

  • use googleapis.com (#388)

  • more retries (#387, 385, 380)

  • Code cleanup (#381)

  • license to match stated one (#378)

  • deps updated (#376)

Version 2021.04.0

  • switch to calver and fsspec pin

Version 0.8.0

  • keep up with fsspec 0.9.0 async

  • one-shot find

  • consistency checkers

  • retries for intermittent issues

  • timeouts

  • partial cat

  • http error status

  • CI to GHA

Version 0.7.0

  • async operations via aiohttp

Version 0.6.0

  • API-breaking: Changed requester-pays handling for GCSFileSystem.

    The user_project keyword has been removed, and has been replaced with the requester_pays keyword. If you’re working with a requester_pays bucket you will need to explicitly pass requester_pays-True. This will include your project ID in requests made to GCS.

Version 0.5.3

  • GCSFileSystem now validates that the project provided, if any, matches the Google default project when using token-'google_default' to authenticate (PR #219).

  • Fixed bug in GCSFileSystem.cat on objects in requester-pays buckets (PR #217).

Version 0.5.2

  • Fixed bug in user_project fallback for default Google authentication (PR #213)

Version 0.5.1

  • user_project now falls back to the project if provided (PR #208)

Version 0.5.0

  • Added the ability to make requester-pays requests with the user_project parameter (PR #206)

Version 0.4.0

  • Improved performance when serializing filesystem objects (PR #182)

  • Fixed authorization errors when using gcsfs within multithreaded code (PR #183, PR #192)

  • Added contributing instructions (PR #185)

  • Improved performance for gcsfs.GCSFileSystem.info() (PR #187)

  • Fixed bug in gcsfs.GCSFileSystem.info() raising an error (PR #190)

Indices and tables