Notice: While JavaScript is not essential for this website, your interaction with the content will be limited. Please turn JavaScript on for the full experience.

PEP 665 -- A file format to list Python dependencies for reproducibility of an application

PEP:665
Title:A file format to list Python dependencies for reproducibility of an application
Author:Brett Cannon <brett at python.org>, Pradyun Gedam <pradyunsg at gmail.com>, Tzu-ping Chung <uranusjr at gmail.com>
PEP-Delegate:Paul Moore <p.f.moore at gmail.com>
Discussions-To:https://discuss.python.org/t/9911
Status:Draft
Type:Standards Track
Created:29-Jul-2021
Post-History:29-Jul-2021, 03-Nov-2021, 25-Nov-2021
Resolution:

Abstract

This PEP specifies a file format to specify the list of Python package installation requirements for an application, and the relation between the specified requirements. The list of requirements is considered exhaustive for the installation target, and thus not requiring any information beyond the platform being installed for, and the file itself. The file format is flexible enough to allow installing the requirements across different platforms, which allows for reproducibility on multiple platforms from the same file.

Terminology

There are several terms whose definition must be agreed upon in order to facilitate a discussion on the topic of this PEP.

A package is something you install as a dependency and use via the import system. The packages on PyPI are an example of this.

An application or app is an end product that other external code does not directly rely on via the import system (i.e. they are standalone). Desktop applications, command-line tools, etc. are examples of applications.

A lock file records the packages that are to be installed for an app. Traditionally, the exact version of the package to be installed is specified by a lock file, but specified packages are not always installed on a given platform (according a filtering logic described in a later section), which enables the lock file to describe reproducibility across multiple platforms. Examples of this are package-lock.json from npm [7], Poetry.lock from Poetry [13], etc.

Locking is the act of taking the input of the packages an app depends on and producting a lock file from that.

A locker is a tool which produces a lock file.

An installer consumes a lock file to install what the lock file specifies.

Motivation

Applications want reproducible installs for a few reasons (we are not worrying about package development, integration into larger systems that would handle locking dependencies external to the Python application, or other situations where flexible installation requirements are desired over strict, reproducible installations).

One, reproducibility eases development. When you and your fellow developers all end up with the same files on a specific platform, you make sure you are all developing towards the same experience for the application. You also want your users to install the same files as you expect to guarantee the experience is the same as you developed for them.

Two, you want to be able to reproduce what gets installed across multiple platforms. Thanks to Python's portability across operating systems, CPUs, etc., it is very easy and often desirable to create applications that are not restricted to a single platform. Thus, you want to be flexible enough to allow for differences in your package dependencies between platforms, while still having consistency and reproducibility on any one specific platform.

Three, reproducibility is more secure. When you control exactly what files are installed, you can make sure no malicious actor is attempting to slip nefarious code into your application (i.e. some supply chain attacks). By using a lock file which always leads to reproducible installs, we can avoid certain risks entirely.

Four, relying on the wheel file [23] format provides reproducibility without requiring build tools to support reproducibility themselves. Thanks to wheels being static and not executing code as part of installation, wheels always lead to a reproducible result. Compare this to source distributions (aka sdists) or source trees which only lead to a reproducible install if their build tool supports reproducibility due to inherent code execution. Unfortunately the vast majority of build tools do not support reproducible builds, so this PEP helps alleviate that issue by only supporting wheels as a package format.

This PEP proposes a standard for a lock file, as the current solutions don't meet the outlined goals. Today, the closest we come to a lock file standard is the requirements file format [16] from pip. Unfortunately, that format does not lead to inherently reproducible installs (it requires optional features both in the requirements file and the installer itself, to be discussed later).

The community itself has also shown a need for lock files based on the fact that multiple tools have independently created their own lock file formats:

  1. PDM [8]
  2. pip-tools [10]
  3. Pipenv [11]
  4. Poetry [13]
  5. Pyflow [14]

Unfortunately, those tools all use differing lock file formats. This means tooling around these tools much be unique. This impacts tooling such as code editors and hosting providers, which want to be as flexible as possible when it comes to accepting a user's application code, but also have a limit as to how much development resources they can spend to add support for yet another lock file format. A standardized format would allow tools to focus their work on a single target, and make sure that workflow decisions made by developers outside of the lock file format are of no concern to e.g. hosting providers.

Other programming language communities have also shown the usefulness of lock files by developing their own solution to this problem. Some of those communities include:

  1. Dart [3]
  2. npm [7]/Node
  3. Go
  4. Rust [17]

The trend in programming languages in the past decade seems to have been toward providing a lock file solution.

Rationale

File Format

We wanted the file format to be easy to read as a diff when auditing a change to the lock file. As such, and thanks to PEP 518 and pyproject.toml, we decided to go with the TOML [21] file format.

Secure by Design

Viewing the requirements file format [16] as the closest we have to a lock file standard, there are a few issues with the file format when it comes to security. First is that the file format simply does not require you to specify the exact version of a package. This is why tools like pip-tools [10] exist to help manage that users of requirements files.

Second, you must opt into specifying what files are acceptable to be installed by using the --hash argument for a specific dependency. This is also optional with pip-tools as it requires specifying the --generate-hashes CLI argument.

Third, even when you control what files may be installed, it does not prevent other packages from being installed. If a dependency is not listed in the requirements file, pip will happily go searching for a file to meet that need. You must specify --no-deps as an argument to pip to prevent unintended dependency resolution outside of the requirements file.

Fourth, the format allows for installing a source distribution file [19] (aka "sdist"). By its very nature, installing an sdist requires executing arbitrary Python code, meaning that there is no control over what files may be installed. Only by specifying --only-binary :all: can you guarantee pip to only use a wheel file [23] for each package.

To recap, in order for a requirements file to be as secure as what is being proposed, a user should always do the following steps:

  1. Use pip-tools and its command pip-compile --generate-hashes
  2. Install the requirements file using pip install --require-hashes --no-deps --only-binary :all:

Critically, all of those flags, and both the specificity and exhaustion of what to install that pip-tools provides, are optional for requirements files.

As such, the proposal raised in this PEP is secure by design which combats some supply chain attacks. Hashes for files which would be used to install from are required. You can only install from wheels to unambiguously define what files will be placed in the file system. Installers must lead to an deterministic installation from a lock file for a given platform. All of this leads to a reproducible installation which you can deem trustworthy (when you have audited the lock file and what it lists).

Cross-Platform

Various projects which already have a lock file, like PDM [8] and Poetry [13], provide a lock file which is cross-platform. This allows for a single lock file to work on multiple platforms while still leading to the exact same top-level requirements to be installed everywhere with the installation being consistent/unambiguous on each platform.

As to why this is useful, let's use an example involving PyWeek [15] (a week-long game development competition). Assume you are developing on Linux, while someone you choose to partner with is using macOS. Now assume the judges are using Windows. How do you make sure everyone is using the same top-level dependencies, while allowing for any platform-specific requirements (e.g. a package requires a helper package under Windows)?

With a cross-platform lock file, you can make sure that the key requirements are met consistently across all platforms. You can then also make sure that all users on the same platform get the same reproducible installation.

Simple Installer

The separation of concerns between a locker and an installer allows for an installer to have a much simpler operation to perform. As such, it not only allows for installers to be easier to write, but facilitates in making sure installers create unambiguous, reproducible installations correctly.

The installer can also expend less computation/energy in creating the installation. This is beneficial not only for faster installs, but also from an energy consumption perspective, as installers are expected to be run more often than lockers.

This has led to a design where the locker must do more work upfront to the benefit installers. It also means the complexity of package dependencies is simpler and easier to comprehend in a lock files to avoid ambiguity.

Specification

Details

Lock files MUST use the TOML [21] file format. This not only prevents the need to have another file format in the Python packaging ecosystem thanks to its adoption by PEP 518 for pyproject.toml, but also assists in making lock files more human-readable.

Lock files MUST end their file names with .pylock.toml. The .toml part unambiguously distinguishes the format of the file, and helps tools like code editors support the file appropriately. The .pylock part distinguishes the file from other TOML files the user has, to make the logic easier for tools to create functionality specific to Python lock files, instead of TOML files in general.

The following sections are the top-level keys of the TOML file data format. Any field not listed as required is considered optional.

version

This field is required.

The version of the lock file being used. The key MUST be a string consisting of a number that follows the same formatting as the Metadata-Version key in the core metadata spec [2].

The value MUST be set to "1.0" until a future PEP allows for a different value. The introduction of a new optional key to the file format SHOULD increase the minor version. The introduction of a new required key or changing the format MUST increase the major version. How to handle other scenarios is left as a per-PEP decision.

Installers MUST warn the user if the lock file specifies a version whose major version is supported but whose minor version is unsupported/unrecognized (e.g. the installer supports "1.0", but the lock file specifies "1.1").

Installers MUST raise an error if the lock file specifies a major version which is unsupported (e.g. the installer supports "1.9" but the lock file specifies "2.0").

created-at

This field is required.

The timestamp for when the lock file was generated (using TOML's native timestamp type). It MUST be recorded using the UTC time zone to avoid ambiguity.

If the SOURCE_DATE_EPOCH [20] environment variable is set, it MUST be used as the timestamp by the locker. This facilitates reproducibility of the lock file itself.

[tool]

Tools may create their own sub-tables under the tool table. The rules for this table match those for pyproject.toml and its [tool] table from the build system declaration spec [1].

[metadata]

This table is required.

A table containing data applying to the overall lock file.

metadata.marker

A key storing a string containing an environment marker as specified in the dependency specifier spec [4].

The locker MAY specify an environment marker which specifies any restrictions the lock file was generated under.

If the installer is installing for an environment which does not satisfy the specified environment marker, the installer MUST raise an error as the lock file does not support the target installation environment.

metadata.tag

A key storing a string specifying platform compatibility tags [12] (i.e. wheel tags). The tag MAY be a compressed tag set.

If the installer is installing for an environment which does not satisfy the specified tag (set), the installer MUST raise an error as the lock file does not support the targeted installation environment.

metadata.requires

This field is required.

An array of strings following the dependency specifier spec [4]. This array represents the top-level package dependencies of the lock file and thus the root of the dependency graph.

metadata.requires-python

A string specifying the supported version(s) of Python for this lock file. It follows the same format as that specified for the Requires-Python field in the core metadata spec [2].

[[package._name_._version_]]

This array is required.

An array per package and version containing entries for the potential (wheel) files to install (as represented by _name_ and _version_, respectively).

Lockers MUST normalize a project's name according to the simple repository API [18]. If extras are specified as part of the project to install, the extras are to be included in the key name and are to be sorted in lexicographic order.

Within the file, the tables for the projects SHOULD be sorted by:

  1. Project/key name in lexicographic order
  2. Package version, newest/highest to older/lowest according to the version specifiers spec [22]
  3. Optional dependencies (extras) via lexicographic order
  4. File name based on the filename field (discussed below)

These recommendations are to help minimize diff changes between tool executions.

package._name_._version_.filename

This field is required.

A string representing the name of the file as represented by an entry in the array. This field is required to simplify installers as the file name is required to resolve wheel tags derived from the file name. It also guarantees that the association of the array entry to the file it is meant for is always clear.

[package._name_._version_.hashes]

This table is required.

A table with keys specifying a hash algorithm and values as the hash for the file represented by this entry in the package._name_._version_ table.

Lockers SHOULD list hashes in lexicographic order. This is to help minimize diff sizes and the potential to overlook hash value changes.

An installer MUST only install a file which matches one of the specified hashes.

package._name_._version_.url

A string representing a URL where to get the file.

The installer MAY support any schemes it wants for URLs (e.g. file: as well as https:).

An installer MAY choose to not use the URL to retrieve a file if a file matching the specified hash can be found using alternative means (e.g. on the file system in a cache directory).

package._name_._version_.direct

A boolean representing whether an installer should consider the project installed "directly" as specified by the direct URL origin of installed distributions spec [5].

If the key is true, then the installer MUST follow the direct URL origin of installed distributions spec [5] for recording the installation as "direct".

package._name_._version_.requires-python

A string specifying the support version(s) of Python for this file. It follows the same format as that specified for the Requires-Python field in the core metadata spec [2].

package._name_._version_.requires

An array of strings following the dependency specifier spec [4] which represent the dependencies of this file.

Example

version = "1.0"
created-at = 2021-10-19T22:33:45.520739+00:00

[tool]
# Tool-specific table.

[metadata]
requires = ["mousebender"]
requires-python = ">=3.6"

[[package.attrs."21.2.0"]]
filename = "attrs-21.2.0-py2.py3-none-any.whl"
hashes.sha256 = "149e90d6d8ac20db7a955ad60cf0e6881a3f20d37096140088356da6c716b0b1"
url = "https://files.pythonhosted.org/packages/20/a9/ba6f1cd1a1517ff022b35acd6a7e4246371dfab08b8e42b829b6d07913cc/attrs-21.2.0-py2.py3-none-any.whl"
requires-python = ">=2.7, !=3.0.*, !=3.1.*, !=3.2.*, !=3.3.*, !=3.4.*"

[[package.attrs."21.2.0"]]
# If attrs had another wheel file (e.g. that was platform-specific),
# it could be listed here.

[[package.mousebender."2.0.0"]]
filename = "mousebender-2.0.0-py3-none-any.whl"
hashes.sha256 = "a6f9adfbd17bfb0e6bb5de9a27083e01dfb86ed9c3861e04143d9fd6db373f7c"
url = "https://files.pythonhosted.org/packages/f4/b3/f6fdbff6395e9b77b5619160180489410fb2f42f41272994353e7ecf5bdf/mousebender-2.0.0-py3-none-any.whl"
required-python = ">=3.6"
requires = ["attrs", "packaging"]

[[package.packaging."20.9"]]
filename = "packaging-20.9-py2.py3-none-any.whl"
hashes.blake-256 = "3e897ea760b4daa42653ece2380531c90f64788d979110a2ab51049d92f408af"
hashes.sha256 = "67714da7f7bc052e064859c05c595155bd1ee9f69f76557e21f051443c20947a"
url = "https://files.pythonhosted.org/packages/3e/89/7ea760b4daa42653ece2380531c90f64788d979110a2ab51049d92f408af/packaging-20.9-py2.py3-none-any.whl"
requires-python = ">=3.6"
requires = ["pyparsing"]

[[package.pyparsing."2.4.7"]]
filename = "pyparsing-2.4.7-py2.py3-none-any.whl"
hashes.sha256="ef9d7589ef3c200abe66653d3f1ab1033c3c419ae9b9bdb1240a85b024efc88b"
url = "https://files.pythonhosted.org/packages/8a/bb/488841f56197b13700afd5658fc279a2025a39e22449b7cf29864669b15d/pyparsing-2.4.7-py2.py3-none-any.whl"
direct = true  # For demonstration purposes.
requires-python = ">=2.6, !=3.0.*, !=3.1.*, !=3.2.*"

Expectations for Lockers

Lockers MUST create lock files for which a topological sort of the packages which qualify for installation on the specified platform results in a graph for which only a single version of any package qualifies for installation and there is at least one compatible file to install for each package. This leads to a lock file for any supported platform where the only decision an installer can make is what the "best-fitting" wheel is to install (which is discussed below).

Lockers are expected to utilize metadata.marker, metadata.tag, and metadata.requires-python as appropriate as well as environment markers specified via requires and Python version requirements via requires-python to enforce this result for installers. Put another way, the information used in the lock file is not expected to be pristine/raw from the locker's input and instead is to be changed as necessary to the benefit of the locker's goals.

Expectations for Installers

The expected algorithm for resolving what to install is:

  1. Construct a dependency graph based on the data in the lock file with metadata.requires as the starting/root point.
  2. Eliminate all files that are unsupported by the specified platform.
  3. Eliminate all irrelevant edges between packages based on marker evaluation for requires.
  4. Raise an error if a package version is still reachable from the root of the dependency graph but lacks any compatible file.
  5. Verify that all packages left only have one version to install, raising an error otherwise.
  6. Install the best-fitting wheel file for each package which remains.

Installers MUST follow a deterministic algorithm determine what the "best-fitting wheel file" is. A simple solution for this is to rely upon the packaging project and its packaging.tags module to determine wheel file precedence.

Installers MUST support installing into an empty environment. Installers MAY support installing into an environment that already contains installed packages (and whatever that would entail to be supported).

(Potential) Tool Support

The pip [9] team has said they are interested in supporting this PEP if accepted. The current proposal for pip may even supplant the need for pip-tools [10].

PDM [8] has also said they would support the PEP if accepted.

Pyflow [14] has said they "like the idea" of the PEP.

Poetry [13] has said they would not support the PEP as-is because "Poetry supports sdists files, directory and VCS dependencies which are not supported". Recording requirements at the file level, which is on purpose to better reflect what can occur when it comes to dependencies, "is contradictory to the design of Poetry". This also excludes export support to a this PEP's lock file as "Poetry exports the information present in the poetry.lock file into another format" and sdists and source trees are included in Poetry.lock files. Thus it is not a clean translation from Poetry's lock file to this PEP's lock file format.

Backwards Compatibility

As there is no pre-existing specification regarding lock files, there are no explicit backwards compatibility concerns.

As for pre-existing tools that have their own lock file, some updating will be required. Most document the lock file name, but not its contents. For projects which do not commit their lock file to version control, they will need to update the equivalent of their .gitignore file. For projects that do commit their lock file to version control, what file(s) get committed will need an update.

For projects which do document their lock file format like pipenv [11], they will very likely need a major version release which changes the lock file format.

Security Implications

A lock file should not introduce security issues but instead help solve them. By requiring the recording of hashes for files, a lock file is able to help prevent tampering with code since the hash details were recorded. Relying on only wheel files means what files will be installed can be known ahead of time and is reproducible. A lock file also helps prevent unexpected package updates being installed which may in turn be malicious.

How to Teach This

Teaching of this PEP will very much be dependent on the lockers and installers being used for day-to-day use. Conceptually, though, users could be taught that a lock file specifies what should be installed for a project to work. The benefits of consistency and security should be emphasized to help users realize why they should care about lock files.

Reference Implementation

No proof-of-concept or reference implementation currently exists. An example locker and installer will be provided before this PEP is fully accepted (although this is not a necessarily a requirement for conditional acceptance).

Rejected Ideas

File Formats Other Than TOML

JSON [6] was briefly considered, but due to:

  1. TOML already being used for pyproject.toml
  2. TOML being more human-readable
  3. TOML leading to better diffs

the decision was made to go with TOML. There was some concern over Python's standard library lacking a TOML parser, but most packaging tools already use a TOML parser thanks to pyproject.toml so this issue did not seem to be a showstopper. Some have also argued against this concern in the past by the fact that if packaging tools abhor installing dependencies and feel they can't vendor a package then the packaging ecosystem has much bigger issues to rectify than the need to depend on a third-party TOML parser.

Alternative Naming Schemes

Specifying a directory to install file to was considered, but ultimately rejected due to people's distaste for the idea.

It was also suggested to not have a special file name suffix, but it was decided that hurt discoverability by tools too much.

Supporting a Single Lock File

At one point the idea of only supporting single lock file which contained all possible lock information was considered. But it quickly became apparent that trying to devise a data format which could encompass both a lock file format which could support multiple environments as well as strict lock outcomes for reproducible builds would become quite complex and cumbersome.

The idea of supporting a directory of lock files as well as a single lock file named pyproject-lock.toml was also considered. But any possible simplicity from skipping the directory in the case of a single lock file seemed unnecessary. Trying to define appropriate logic for what should be the pyproject-lock.toml file and what should go into pyproject-lock.d seemed unnecessarily complicated.

Using a Flat List Instead of a Dependency Graph

The first version of this PEP proposed that the lock file have no concept of a dependency graph. Instead, the lock file would list exactly what should be installed for a specific platform such that installers did not have to make any decisions about what to install, only validating that the lock file would work for the target platform.

This idea was eventually rejected due to the number of combinations of potential PEP 508 environment markers. The decision was made that trying to have lockers generate all possible combinations as individual lock files when a project wants to be cross-platform would be too much.

Use Wheel Tags in the File Name

Instead of having the metadata.tag field there was a suggestion of encoding the tags into the file name. But due to the addition of the metadata.marker field and what to do when no tags were needed, the idea was dropped.

Alternative Names for requires

Some other names for what became requires were installs, needs, and dependencies. Initially this PEP chose needs after asking a Python beginner which term they preferred. But based on feedback on an earlier draft of this PEP, requires was chosen as the term.

Accepting PEP 650

PEP 650 was an earlier attempt at trying to tackle this problem by specifying an API for installers instead of standardizing on a lock file format (ala PEP 517). The initial response to PEP 650 could be considered mild/lukewarm. People seemed to be consistently confused over which tools should provide what functionality to implement the PEP. It also potentially incurred more overhead as it would require executing Python APIs to perform any actions involving packaging.

This PEP chooses to standardize around an artifact instead of an API (ala PEP 621). This would allow for more tool integrations as it removes the need to specifically use Python to do things such as create a lock file, update it, or even install packages listed in a lock file. It also allows for easier introspection by forcing dependency graph details to be written in a human-readable format. It also allows for easier sharing of knowledge by standardizing what people need to know more (e.g. tutorials become more portable between tools when it comes to understanding the artifact they produce). It's also simply the approach other language communities have taken and seem to be happy with.

Acceptance of this PEP would mean PEP 650 gets rejected.

Specifying Requirements per Package Instead of per File

An earlier draft of this PEP specified dependencies at the package level instead of per file. While this has traditionally been how packaging systems work, it actually did not reflect accurately how things are specified. As such, this PEP was subsequently updated to reflect the granularity that dependencies can truly be specified at.

Specify Where Lockers Gather Input

This PEP does not specify how a locker gets its input. An initial suggestion was to partially reuse PEP 621, but due to disagreements on how flexible the potential input should be in terms of specifying things such as indexes, etc., it was decided this would best be left to a separate PEP.

Allowing Source Distributions and Source Trees to be an Opt-In, Supported File Format

After extensive discussion, it was decided that this PEP would not support source distributions (aka sdists) or source trees as an acceptable format for code. Introducing sdists and source trees to this PEP would immediately undo the reproducibility and security goals due to needing to execute code to build the sdist or source tree. It would also greatly increase the complexity for (at least) installers as the dynamic build nature of sdists and source trees means the installer would need to handle fully resolving whatever requirements the sdists produced dynamically, both from a building and installation perspective.

Due to all of this, it was decided it was best to have a separate discussion about what supporting sdists and source trees after this PEP is accepted/rejected. As the proposed file format is versioned, introducing sdists and source tree support in a later PEP is doable.

It should be noted, though, that this PEP is not stop an out-of-band solution from being developed to be used in conjunction with this PEP. Building wheel files from sdists and shipping them with code upon deployment so they can be included in the lock file is one option. Another is to use a requirements file just for sdists and source trees, then relying on a lock file for all wheels.

Open Issues

None.

Acknowledgments

Thanks to Frost Ming of PDM [8] and S├ębastien Eustace of Poetry [13] for providing input around dynamic install-time resolution of PEP 508 requirements.

Thanks to Kushal Das for making sure reproducible builds stayed a concern for this PEP.

Thanks to Andrea McInnes for initially settling the bikeshedding and choosing the paint colour of needs (at which point people ralled around the requires colour instead).

Source: https://github.com/python/peps/blob/master/pep-0665.rst