Ship large files with Python packages


August 21, 2019

It is often useful to ship large data files together with a Python package, a couple of scenarios are:

If data are collectively less than 2 GB compressed and do not change very often, a simple and a bit hacky solution is to use GitHub release assets. For each packaged release on GitHub it is possible to attach one or more assets smaller than 2 GB. You can then attach data to each release, the downside is that users need to make sure to use the correct dataset for the release they are using and the first time they use the software the need to install the Python package and also download the dataset and install it in the right folder. See an example script to upload from the command line.

If data files are individually less than 10 MB and collectively less than 100 MB you can directly add them into the Python package. This is the easiest and most convenient option, for example the astropy package template automatically adds to the package any file inside the packagename/data folder.

For larger datasets I recommend to host the files externally and use the module. This module automates the process of retrieving a file from a remote server and caching it locally (in the users home folder), next time the user needs it, it is automatically retrieved from the cache:

    dataurl = ""
    with data.conf.set_temp("dataurl", dataurl), data.conf.set_temp(
        "remote_timeout", 30
        local_file_path = data.get_pkg_data_filename("myfile.jpg)

Now we need to host there files publicly, I have a few options.

Host on a dedicated GitHub repository

If files are individually less than 100MB and collectively a few GB, you can create a dedicated repository on GitHub and push there your files. Then activate GitHub Pages so that those files are published at Then use this URL as dataurl in the above script.

Host on a Supercomputer or own server

Some Supercomputers offer the feature of providing public web access from specific folders, for example NERSC allows user to publish web-pages publicly, see their documentation.

This is very useful for huge datasets because you can automatically detect if the package is being run at NERSC and then automatically access the files with their path instead of downloading them.

For example:

def get_data_from_url(filename):
    """Retrieves input templates from remote server,
    in case data is available in one of the PREDEFINED_DATA_FOLDERS defined above,
    e.g. at NERSC, those are directly returned."""

    for folder in PREDEFINED_DATA_FOLDERS:
        full_path = os.path.join(folder, filename)
        if os.path.exists(full_path):
            warnings.warn(f"Access data from {full_path}")
            return full_path
    with data.conf.set_temp("dataurl", DATAURL), data.conf.set_temp(
        "remote_timeout", 30
        warnings.warn(f"Retrieve data for {filename} (if not cached already)")
        map_out = data.get_pkg_data_filename(filename, show_progress=True)
    return map_out

Similar setup can be achieved on a GNU/Linux server, for example a powerful machine used by all members of a scientific team, where a folder is dedicated to host these data and is also published online with Apache or NGINX.

The main downside of this approach is that there is no built-in version control. One possibility is to enforce a policy where no files are ever overwritten and version control is automatically achieved with filenames. Otherwise, use git lfs in that folder to track any change in a dedicated local git repository, e.g.:

git init
git lfs track "*.fits"
git add "*.fits"
git commit -m "initial version of all FITS files"

This method tracks the checksum of all the binary files and helps managing the history, even if only locally (make sure the folder is also regularly backed up). You could push it to GitHub, that would cost $5/month for each 50GB of storage.

Host on Figshare

You can upload files to Figshare using the browser and create a dataset which also comes with a DOI and a page where you can save metadata about this object.

Once you have set the dataset public, you can find out the URL of the actual file, which is of the form, therefore we can set as the repository and use the integer defined in Figshare as filename. Using integers as filenames makes it a bit cryptic, but it has the great advantage that other people can do the uploading to Figshare and you can point to their files as easily as if the are yours. This is more convenient than alternatives where instead you need to give other people access to your file repository.

Host on Amazon S3 or other object store

A public bucket on Amazon S3 or other object store provides cheap storage and built-in version control. The cost currently is about $0.026/GB/month.

First login to the AWS console and create a new bucket, set it public by turning of “Block all public access” and under “Access Control List” set “List objects” to Yes for “Public access”.

You could upload files with the browser, but for larger files command line is better.

The files will be available at, this changes based on the chosen region.

(Advanced) Upload files from the command line

This is optional and requires some more familiarity with AWS. Go back to the AWS console to the Identity and Access Management (IAM) section, then users, create, create a policy to give access only to 1 bucket (replace bucket-name):

    "Version": "2012-10-17",
    "Statement": [
            "Sid": "ListObjectsInBucket",
            "Effect": "Allow",
            "Action": ["s3:ListBucket"],
            "Resource": ["arn:aws:s3:::bucket-name"]
            "Sid": "AllObjectActions",
            "Effect": "Allow",
            "Action": [
            "Resource": ["arn:aws:s3:::bucket-name/*"]

See the AWS documentation

Install s3cmd, then run s3cmd --configure to set it up and paste the Access and Secret keys, it will fail to test the configuration because it cannot list all the buckets, anyway choose to save the configuration.

Test it:

    s3cmd ls s3://bucket-name

Then upload your files (reduced redundancy is cheaper):

    s3cmd put --reduced-redundancy --acl-public *.fits s3://bucket-name