repro-tarfile

A tiny, zero-dependency replacement for Python's tarfile standard library for creating reproducible/deterministic tar archives.

"Reproducible" or "deterministic" in this context means that the binary content of the tar archive is identical if you add files with identical binary content in the same order. It means you can reliably check equality of the contents of two tar archives by simply comparing checksums of the archive using a hash function like MD5 or SHA-256.

This Python package provides a ReproducibleTarFile class that works exactly like tarfile.TarFile from the Python standard library, except that certain archive metadata are set to fixed values. See "How does repro-tarfile work?" below for details.

You can also optionally install a command-line program, rptar. See "rptar command line program" below for more information.

Looking instead to create reproducible/deterministic ZIP archives? Check out our sister package, repro-zipfile!

Installation

repro-tarfile is available from PyPI. To install, run:

pip install repro-tarfile

It is also available from conda-forge. To install, run:

conda install repro-tarfile -c conda-forge

Usage

Simply import repro_tarfile and use it in the same way you would use regular tarfile from the Python standard library.

import repro_tarfile

with repro_tarfile.open("archive.tar.gz", "w:gz") as tar:
    tar.add("examples/data.txt", arcname="data.txt")

Note that files must be written to the archive in the same order to reproduce an identical archive. Be aware that functions that like os.listdir, os.glob, Path.iterdir, and Path.glob return files in a nondeterministic order—you should call sorted on their returned values first.

See examples/usage.py for an example script that you can run, and examples/demo_vs_tarfile.py for a demonstration in contrast with the standard library's tarfile module.

For more advanced usage, such as customizing the fixed metadata values, see the subsections under "How does repro-tarfile work?".

rptar command-line program

You can optionally install a lightweight command-line program, rptar. This includes an additional dependency on the typer CLI framework. You can install it either directly or using the cli extra with repro-tarfile. We recommend you use pipx for installing Python CLIs into isolated virtual environments. You can also install it with regular pip, too.

pipx install rptar
# or
pipx install repro-tarfile[cli]

rptar is designed to a partial drop-in replacement ubiquitous tar program. Use rptar --help to see the documentation. Here are some usage examples:

# Archive one file
rptar -czvf archive.tar.gz some_file.txt
# Archive two files
rptar -czvf archive.tar.gz file1.txt file2.txt
# Archive many files with glob
rptar -czvf archive.tar.gz some_dir/*.txt
# Archive directory recursively
rptar -czvf archive.tar.gz some_dir/

In addition to the fixed metadata values that repro-tarfile sets, rptar will also always sort all paths being archived.

How does repro-tarfile work?

Tar archives are not normally reproducible even when containing files with identical content because of metadata. In particular, the usual culprits are:

1. Last-modified timestamps of added files
2. File-system permissions (mode) of added files
3. File owner user and group of added files
4. If using gzip compression, the uncompressed filename in the gzip header
5. If using gzip compression, the last modified timestamp the gzip header

repro_tarfile.ReproducibleTarFile is a subclass of tarfile.TarFile that overrides the addfile method (which is also used interally by add) with a version that set the above file metadata to fixed values. It also overrides the gzopen method used for gzip compression to override the gzip header values. Note that repro-tarfile does not modify the original files—it simply overrides the metadata written to the archive.

You can effectively reproduce what repro-tarfile does in a .tar.gz case with something like this:

from gzip import GzipFile
from pathlib import Path
import tarfile

with Path("archive.tar.gz").open("wb") as fp:
    with GzipFile(filename="", fileobj=fp, mode="wb", mtime=315532800) as gz:
        with tarfile.open(fileobj=gz, mode="w") as tar:
            # Use write to add a file to the archive
            tarinfo = tar.gettarinfo("examples/data.txt", arcname="data.txt")
            tarinfo.mtime = 315532800
            tarinfo.mode=0o644
            tarinfo.uid=0
            tarinfo.gid=0
            tarinfo.uname=""
            tarinfo.gname=""
            with Path("examples/data.txt").open("rb") as fp2:
                tar.addfile(tarinfo, fp2)

It's kind of a pain! We believe repro-tarfile is sufficiently more convenient to justify a small package.

See the next two sections for more details about the replacement metadata values and how to customize them.

Fixed metadata values

Here's a quick reference table of the fixed metadata values. You can use the associated environment variable to override a value.

For deeper explanations, see below.

Last-modified timestamps

Tar archives store the last-modified timestamps of added files and directories. The default fixed value used by repro-tarfile is 315532800, which corresponds to 1980-01-01 00:00:00 UTC.

You can customize this value with the SOURCE_DATE_EPOCH environment variable. If set, it will be used as the fixed value instead. This should be an integer corresponding to the Unix epoch time of the timestamp you want to set, e.g., 1704067230 for 2024-01-01 00:00:00 UTC. SOURCE_DATE_EPOCH is a standard created by the Reproducible Builds project for software distributions.

File-system permissions

Tar archives store the file-system permissions of files and directories. The default permissions set for new files or directories often can be different across different systems or users without any intentional choices being made. (These default permissions are controlled by something called umask.) repro-tarfile will set these to fixed values. By default, the fixed values are 0o644 (rw-r--r--) for files and 0o755 (rwxr-xr-x) for directories, which matches the common default umask of 0o022 for root users on Unix systems. (The 0o prefix is how you can write an octal—i.e., base 8—integer literal in Python.)

You can customize these values using the environment variables REPRO_ZIPFILE_FILE_MODE and REPRO_ZIPFILE_DIR_MODE. They should be in three-digit octal Unix numeric notation, e.g., 644 for rw-r--r--.

File owner user and group

In typical file systems, every file and directory has an owner. Tar archives record the user and group information of the owner. If different users or systems are generating identical files and then archiving them, the owner information will likely be different. By default, repro-tarfile uses user and group IDs values of 0, and empty strings for the user and group names. These are the standard values recommended by the Reproducible Builds project.

You can customize the user and group IDs using the environment variables REPRO_TARFILE_UID and REPRO_TARFILE_GID. The values should be integers. You can customize the user and group names using the environment variables REPRO_TARFILE_UNAME and REPRO_TARFILE_GNAME.

Gzip header values

The gzip compression file format includes a header that contains metadata about the compressed file—in this case, the tar archive. This header includes the archive filename and the last modified timestamp of the archive. By default, repro-tarfile sets the archive filename to an empty string, and the last modified timestamp to the same default value as the added files last modified timestamp, 315532800, which corresponds to 1980-01-01 00:00:00 UTC.

The environment variable SOURCE_DATE_EPOCH used to customize the added file last modified timestamp will also be used to set the gzip header last modified timestamp. Currently, we don't support a way to customize the archive filename override.

Why care about reproducible tar archives?

Tar archives are often useful when dealing with a set of multiple files, especially if the files are large and can be compressed. Creating reproducible tar archives is often useful for:

• Building a software package. This is a development best practice to make it easier to verify distributed software packages. See the Reproducible Builds project for more explanation.
• Working with data. Verify that your data pipeline produced the same outputs, and avoid further reprocessing of identical data.
• Packaging machine learning model artifacts. Manage model artifact packages more effectively by knowing when they contain identical models.

Related Tools and Alternatives

• https://diffoscope.org/
  • Can do a rich comparison of archive files and show what specifically differs
• https://salsa.debian.org/reproducible-builds/strip-nondeterminism
  • Perl library for removing nondeterministic metadata from file archives