Following system colour scheme Selected dark colour scheme Selected light colour scheme

Python Enhancement Proposals

PEP 426 – Metadata for Python Software Packages 2.0

Author:
Alyssa Coghlan <ncoghlan at gmail.com>, Daniel Holth <dholth at gmail.com>, Donald Stufft <donald at stufft.io>
BDFL-Delegate:
Donald Stufft <donald at stufft.io>
Discussions-To:
Distutils-SIG list
Status:
Withdrawn
Type:
Informational
Topic:
Packaging
Requires:
440, 508, 518
Created:
30-Aug-2012
Post-History:
14-Nov-2012, 05-Feb-2013, 07-Feb-2013, 09-Feb-2013, 27-May-2013, 20-Jun-2013, 23-Jun-2013, 14-Jul-2013, 21-Dec-2013
Replaces:
345

Table of Contents

PEP Withdrawal

The ground-up metadata redesign proposed in this PEP has been withdrawn in favour of the more modest proposal in PEP 566, which retains the basic Key:Value format of previous metadata versions, but also defines a standardised mechanism for translating that format to nested JSON-compatible data structures.

Some of the ideas in this PEP (or the related PEP 459) may still be considered as part of later proposals, but they will be handled in a more incremental fashion, rather than as a single large proposed change with no feasible migration plan.

Abstract

This PEP describes a mechanism for publishing and exchanging metadata related to Python distributions. It includes specifics of the field names, and their semantics and usage.

This document specifies the never released version 2.0 of the metadata format.

Version 1.0 is specified in PEP 241. Version 1.1 is specified in PEP 314. Version 1.2 is specified in PEP 345.

Version 2.0 of the metadata format proposed migrating from directly defining a custom key-value file format to instead defining a JSON-compatible in-memory representation that may be used to define metadata representation in other contexts (such as API and archive format definitions).

This version also defines a formal extension mechanism, allowing new fields to be added for particular purposes without requiring updates to the core metadata format.

Note on PEP History

This PEP was initially deferred for an extended period, from December 2013 through to March 2017, as distutils-sig worked through a number of other changes. These changes included:

  • defining a binary compatibility tagging format in PEP 425
  • defining a binary archive format (wheel) in PEP 427
  • explicitly defining versioning and version comparison in PEP 440
  • explicitly defining the PyPI “simple” API in PEP 503
  • explicitly defining dependency specifiers and the extras system in PEP 508
  • declaring static build system dependencies (pyproject.toml) in PEP 518
  • migrating PyPI hosting to Rackspace, and placing it behind the Fastly CDN
  • shipping pip with CPython by default in PEP 453, and backporting that addition to Python 2.7 in PEP 477
  • establishing packaging.python.org as the common access point for Python packaging ecosystem documentation
  • migrating to using the specifications section of packaging.python.org as the central location for tracking packaging related PEPs

The time spent pursuing these changes provided additional perspective on which metadata format changes were genuinely desirable, and which could be omitted from the revised specification as merely being “change for change’s sake”.

It also allowed a number of features that aren’t critical to the core activity of publishing and distributing software to be moved out to PEP 459, a separate proposal for a number of standard metadata extensions that provide additional optional information about a release.

As of September 2017, it was deferred again, on the grounds that it doesn’t actually help solve any particularly pressing problems:

  • JSON representation would be better handled through defining a transformation of the existing metadata 1.2 fields
  • clarification of the additional fields defined in the past few years and related changes to the spec management process would be better covered in a minor spec version update

Finally, the PEP was withdrawn in February 2018 in favour of PEP 566 (which pursues that more incremental strategy).

Purpose

The purpose of this PEP is to define a common metadata interchange format for communication between software publication tools and software integration tools in the Python ecosystem. One key aim is to support full dependency analysis in that ecosystem without requiring the execution of arbitrary Python code by those doing the analysis. Another aim is to encourage good software distribution practices by default, while continuing to support the current practices of almost all existing users of the Python Package Index (both publishers and integrators). Finally, the aim is to support an upgrade path from the currently in use metadata formats that is transparent to end users.

The design draws on the Python community’s nearly 20 years of experience with distutils based software distribution, and incorporates ideas and concepts from other distribution systems, including Python’s setuptools, pip and other projects, Ruby’s gems, Perl’s CPAN, Node.js’s npm, PHP’s composer and Linux packaging systems such as RPM and APT.

While the specifics of this format are aimed at the Python ecosystem, some of the ideas may also be useful in the future evolution of other dependency management ecosystems.

Development, Distribution and Deployment of Python Software

The metadata design in this PEP is based on a particular conceptual model of the software development and distribution process. This model consists of the following phases:

  • Software development: this phase involves working with a source checkout for a particular application to add features and fix bugs. It is expected that developers in this phase will need to be able to build the software, run the software’s automated test suite, run project specific utility scripts and publish the software.
  • Software publication: this phase involves taking the developed software and making it available for use by software integrators. This includes creating the descriptive metadata defined in this PEP, as well as making the software available (typically by uploading it to an index server).
  • Software integration: this phase involves taking published software components and combining them into a coherent, integrated system. This may be done directly using Python specific cross-platform tools, or it may be handled through conversion to development language neutral platform specific packaging systems.
  • Software deployment: this phase involves taking integrated software components and deploying them on to the target system where the software will actually execute.

The publication and integration phases are collectively referred to as the distribution phase, and the individual software components distributed in that phase are formally referred to as “distribution packages”, but are more colloquially known as just “packages” (relying on context to disambiguate them from the “module with submodules” kind of Python package).

The exact details of these phases will vary greatly for particular use cases. Deploying a web application to a public Platform-as-a-Service provider, publishing a new release of a web framework or scientific library, creating an integrated Linux distribution, or upgrading a custom application running in a secure enclave are all situations this metadata design should be able to handle.

The complexity of the metadata described in this PEP thus arises directly from the actual complexities associated with software development, distribution and deployment in a wide range of scenarios.

Supporting definitions

The key words “MUST”, “MUST NOT”, “REQUIRED”, “SHALL”, “SHALL NOT”, “SHOULD”, “SHOULD NOT”, “RECOMMENDED”, “MAY”, and “OPTIONAL” in this document are to be interpreted as described in RFC 2119.

“Projects” are software components that are made available for integration. Projects include Python libraries, frameworks, scripts, plugins, applications, collections of data or other resources, and various combinations thereof. Public Python projects are typically registered on the Python Package Index.

“Releases” are uniquely identified snapshots of a project.

“Distribution packages” are the packaged files which are used to publish and distribute a release.

Depending on context, “package” may refer to either a distribution, or to an importable Python module that has a __path__ attribute and hence may also have importable submodules.

“Source archive” and “VCS checkout” both refer to the raw source code for a release, prior to creation of an sdist or binary archive.

An “sdist” is a publication format providing the distribution metadata and any source files that are essential to creating a binary archive for the distribution. Creating a binary archive from an sdist requires that the appropriate build tools be available on the system.

“Binary archives” only require that prebuilt files be moved to the correct location on the target system. As Python is a dynamically bound cross-platform language, many so-called “binary” archives will contain only pure Python source code.

“Contributors” are individuals and organizations that work together to develop a software component.

“Publishers” are individuals and organizations that make software components available for integration (typically by uploading distributions to an index server)

“Integrators” are individuals and organizations that incorporate published distributions as components of an application or larger system.

“Build tools” are automated tools intended to run on development systems, producing source and binary distribution archives. Build tools may also be invoked by integration tools in order to build software distributed as sdists rather than prebuilt binary archives.

“Index servers” are active distribution registries which publish version and dependency metadata and place constraints on the permitted metadata.

“Public index servers” are index servers which allow distribution uploads from untrusted third parties. The Python Package Index is a public index server.

“Publication tools” are automated tools intended to run on development systems and upload source and binary distribution archives to index servers.

“Integration tools” are automated tools that consume the metadata and distribution archives published by an index server or other designated source, and make use of them in some fashion, such as installing them or converting them to a platform specific packaging format.

“Installation tools” are integration tools specifically intended to run on deployment targets, consuming source and binary distribution archives from an index server or other designated location and deploying them to the target system.

“Automated tools” is a collective term covering build tools, index servers, publication tools, integration tools and any other software that produces or consumes distribution version and dependency metadata.

“Legacy metadata” refers to earlier versions of this metadata specification, along with the supporting metadata file formats defined by the setuptools project.

“Distro” is used as the preferred term for Linux distributions, to help avoid confusion with the Python-specific use of the term “distribution package”.

“Qualified name” is a dotted Python identifier. For imported modules and packages, the qualified name is available as the __name__ attribute, while for functions and classes it is available as the __qualname__ attribute.

A “fully qualified name” uniquely locates an object in the Python module namespace. For imported modules and packages, it is the same as the qualified name. For other Python objects, the fully qualified name consists of the qualified name of the containing module or package, a colon (:) and the qualified name of the object relative to the containing module or package.

A “prefixed name” starts with a qualified name, but is not necessarily a qualified name - it may contain additional dot separated segments which are not valid identifiers.

Integration and deployment of distributions

The primary purpose of the distribution metadata is to support integration and deployment of distributions as part of larger applications and systems.

Integration and deployment can in turn be broken down into further substeps.

  • Build: the build step is the process of turning a VCS checkout, source archive or sdist into a binary archive. Dependencies must be available in order to build and create a binary archive of the distribution (including any documentation that is installed on target systems).
  • Installation: the installation step involves getting the distribution and all of its runtime dependencies onto the target system. In this step, the distribution may already be on the system (when upgrading or reinstalling) or else it may be a completely new installation.
  • Runtime: this is normal usage of a distribution after it has been installed on the target system.

These three steps may all occur directly on the target system. Alternatively the build step may be separated out by using binary archives provided by the publisher of the distribution, or by creating the binary archives on a separate system prior to deployment. The advantage of the latter approach is that it minimizes the dependencies that need to be installed on deployment targets (as the build dependencies will be needed only on the build systems).

The published metadata for distribution packages SHOULD allow integrators, with the aid of build and integration tools, to:

  • obtain the original source code that was used to create a distribution
  • identify and retrieve the dependencies (if any) required to use a distribution
  • identify and retrieve the dependencies (if any) required to build a distribution from source
  • identify and retrieve the dependencies (if any) required to run a distribution’s test suite

Development and publication of distributions

The secondary purpose of the distribution metadata is to support effective collaboration amongst software contributors and publishers during the development phase.

The published metadata for distributions SHOULD allow contributors and publishers, with the aid of build and publication tools, to:

  • perform all the same activities needed to effectively integrate and deploy the distribution
  • identify and retrieve the additional dependencies needed to develop and publish the distribution
  • specify the dependencies (if any) required to use the distribution
  • specify the dependencies (if any) required to build the distribution from source
  • specify the dependencies (if any) required to run the distribution’s test suite
  • specify the additional dependencies (if any) required to develop and publish the distribution

Metadata format

The format defined in this PEP is an in-memory representation of Python distribution metadata as a string-keyed dictionary. Permitted values for individual entries are strings, lists of strings, and additional nested string-keyed dictionaries.

Except where otherwise noted, dictionary keys in distribution metadata MUST be valid Python identifiers in order to support attribute based metadata access APIs.

The individual field descriptions show examples of the key name and value as they would be serialised as part of a JSON mapping.

Unless otherwise indicated, the fields identified as core metadata are required. Automated tools MUST NOT accept distributions with missing core metadata as valid Python distributions.

All other fields are optional. Automated tools MUST operate correctly if a distribution does not provide them, except for those operations which specifically require the omitted fields.

Automated tools MUST NOT insert dummy data for missing fields. If a valid value is not provided for a required field then the metadata and the associated distribution MUST be rejected as invalid. If a valid value is not provided for an optional field, that field MUST be omitted entirely. Automated tools MAY automatically derive valid values from other information sources (such as a version control system).

Automated tools, especially public index servers, MAY impose additional length restrictions on metadata beyond those enumerated in this PEP. Such limits SHOULD be imposed where necessary to protect the integrity of a service, based on the available resources and the service provider’s judgment of reasonable metadata capacity requirements.

Metadata files

The information defined in this PEP is serialised to pysdist.json files for some use cases. These are files containing UTF-8 encoded JSON metadata.

Each metadata file consists of a single serialised mapping, with fields as described in this PEP. When serialising metadata, automated tools SHOULD lexically sort any keys and list elements in order to simplify reviews of any changes.

There are expected to be three standard locations for these metadata files:

  • as a {distribution}-{version}.dist-info/pysdist.json file in an sdist source distribution archive
  • as a {distribution}-{version}.dist-info/pysdist.json file in a wheel binary distribution archive
  • as a {distribution}-{version}.dist-info/pysdist.json file in a local Python installation database

This file is expected to be identical in all three locations - it is generated when creating a source archive or binary archive from a source tree, and then preserved unchanged on installation, or when building a binary archive from a source archive.

Note

These locations are to be confirmed, since they depend on the definition of sdist 2.0 and the revised installation database standard. There will also be a wheel 1.1 format update after this PEP is approved that mandates provision of 2.0+ metadata.

Note that these metadata files MAY be processed even if the version of the containing location is too low to indicate that they are valid. Specifically, unversioned sdist archives, unversioned installation database directories and version 1.0 of the wheel specification may still provide pysdist.json files.

Note

Until this specification is formally marked as Active, it is recommended that tools following the draft format use an alternative filename like metadata.json or pep426-20131213.json to avoid colliding with the eventually standardised files.

Other tools involved in Python distribution MAY also use this format.

Note that these metadata files are generated by build tools based on other input formats (such as setup.py and pyproject.toml) rather than being used directly as a data input format. Generating the metadata as part of the publication process also helps to deal with version specific fields (including the source URL and the version field itself).

For backwards compatibility with older installation tools, metadata 2.0 files MAY be distributed alongside legacy metadata.

Index servers MAY allow distributions to be uploaded and installation tools MAY allow distributions to be installed with only legacy metadata.

Automated tools MAY attempt to automatically translate legacy metadata to the format described in this PEP. Advice for doing so effectively is given in Appendix A.

Metadata validation

A jsonschema description of the distribution metadata is available.

This schema does NOT currently handle validation of some of the more complex string fields (instead treating them as opaque strings).

Except where otherwise noted, all URL fields in the metadata MUST comply with RFC 3986.

Note

The current version of the schema file covers the previous draft of the PEP, and has not yet been updated for the split into the essential dependency resolution metadata and multiple standard extensions, and nor has it been updated for the various other differences between the current draft and the earlier drafts.

Core metadata

This section specifies the core metadata fields that are required for every Python distribution.

Publication tools MUST ensure at least these fields are present when publishing a distribution.

Index servers MUST ensure at least these fields are present in the metadata when distributions are uploaded.

Installation tools MUST refuse to install distributions with one or more of these fields missing by default, but MAY allow users to force such an installation to occur.

Metadata version

Version of the file format; "2.0" is the only legal value.

Automated tools consuming metadata SHOULD warn if metadata_version is greater than the highest version they support, and MUST fail if metadata_version has a greater major version than the highest version they support (as described in PEP 440, the major version is the value before the first dot).

For broader compatibility, build tools MAY choose to produce distribution metadata using the lowest metadata version that includes all of the needed fields.

Example:

"metadata_version": "2.0"

Generator

Name (and optional version) of the program that generated the file, if any. A manually produced file would omit this field.

Examples:

"generator": "flit"
"generator": "setuptools (34.3.1)"

Name

The name of the distribution, as defined in PEP 508.

As distribution names are used as part of URLs, filenames, command line parameters and must also interoperate with other packaging systems, the permitted characters are constrained to:

  • ASCII letters ([a-zA-Z])
  • ASCII digits ([0-9])
  • underscores (_)
  • hyphens (-)
  • periods (.)

Distribution names MUST start and end with an ASCII letter or digit.

Automated tools MUST reject non-compliant names. A regular expression to enforce these constraints (when run with re.IGNORECASE) is:

^([A-Z0-9]|[A-Z0-9][A-Z0-9._-]*[A-Z0-9])$

All comparisons of distribution names MUST be case insensitive, and MUST consider hyphens and underscores to be equivalent.

Index servers MAY consider “confusable” characters (as defined by the Unicode Consortium in TR39: Unicode Security Mechanisms) to be equivalent.

Index servers that permit arbitrary distribution name registrations from untrusted sources SHOULD consider confusable characters to be equivalent when registering new distributions (and hence reject them as duplicates).

Integration tools MUST NOT silently accept a confusable alternate spelling as matching a requested distribution name.

At time of writing, the characters in the ASCII subset designated as confusables by the Unicode Consortium are:

  • 1 (DIGIT ONE), l (LATIN SMALL LETTER L), and I (LATIN CAPITAL LETTER I)
  • 0 (DIGIT ZERO), and O (LATIN CAPITAL LETTER O)

Example:

"name": "ComfyChair"

Version

The distribution’s public or local version identifier, as defined in PEP 440. Version identifiers are designed for consumption by automated tools and support a variety of flexible version specification mechanisms (see PEP 440 for details).

Version identifiers MUST comply with the format defined in PEP 440.

Version identifiers MUST be unique within each project.

Index servers MAY place restrictions on the use of local version identifiers as described in PEP 440.

Example:

"version": "1.0a2"

Summary

A short summary of what the distribution does.

This field SHOULD contain fewer than 512 characters and MUST contain fewer than 2048.

This field SHOULD NOT contain any line breaks.

A more complete description SHOULD be included as a separate file in the sdist for the distribution. Refer to the python-details extension in PEP 459 for more information.

Example:

"summary": "A module that is more fiendish than soft cushions."

Source code metadata

This section specifies fields that provide identifying details for the source code used to produce this distribution.

All of these fields are optional. Automated tools MUST operate correctly if a distribution does not provide them, including failing cleanly when an operation depending on one of these fields is requested.

Source labels

Source labels are text strings with minimal defined semantics. They are intended to allow the original source code to be unambiguously identified, even if an integrator has applied additional local modifications to a particular distribution.

To ensure source labels can be readily incorporated as part of file names and URLs, and to avoid formatting inconsistencies in hexadecimal hash representations they MUST be limited to the following set of permitted characters:

  • Lowercase ASCII letters ([a-z])
  • ASCII digits ([0-9])
  • underscores (_)
  • hyphens (-)
  • periods (.)
  • plus signs (+)

Source labels MUST start and end with an ASCII letter or digit.

A regular expression to rnforce these constraints (when run with re.IGNORECASE) is:

^([A-Z0-9]|[A-Z0-9][A-Z0-9._-+]*[A-Z0-9])$

A source label for a project MUST NOT match any defined version for that project. This restriction ensures that there is no ambiguity between version identifiers and source labels.

Examples:

"source_label": "1.0.0-alpha.1"

"source_label": "1.3.7+build.11.e0f985a"

"source_label": "v1.8.1.301.ga0df26f"

"source_label": "2013.02.17.dev123"

Source URL

A string containing a full URL where the source for this specific version of the distribution can be downloaded.

Source URLs MUST be unique within each project. This means that the URL can’t be something like "https://github.com/pypa/pip/archive/main.zip", but instead must be "https://github.com/pypa/pip/archive/1.3.1.zip".

The source URL MUST reference either a source archive or a tag or specific commit in an online version control system that permits creation of a suitable VCS checkout. It is intended primarily for integrators that wish to recreate the distribution from the original source form.

All source URL references SHOULD specify a secure transport mechanism (such as https) AND include an expected hash value in the URL for verification purposes. If a source URL is specified without any hash information, with hash information that the tool doesn’t understand, or with a selected hash algorithm that the tool considers too weak to trust, automated tools SHOULD at least emit a warning and MAY refuse to rely on the URL. If such a source URL also uses an insecure transport, automated tools SHOULD NOT rely on the URL.

For source archive references, an expected hash value may be specified by including a <hash-algorithm>=<expected-hash> entry as part of the URL fragment.

As of 2017, it is RECOMMENDED that 'sha256' hashes be used for source URLs, as this hash is not yet known to be vulnerable to generation of malicious collisions, while also being widely available on client systems.

For version control references, the VCS+protocol scheme SHOULD be used to identify both the version control system and the secure transport, and a version control system with hash based commit identifiers SHOULD be used. Automated tools MAY omit warnings about missing hashes for version control systems that do not provide hash based commit identifiers.

To handle version control systems that do not support including commit or tag references directly in the URL, that information may be appended to the end of the URL using the @<commit-hash> or the @<tag>#<commit-hash> notation.

Note

This isn’t quite the same as the existing VCS reference notation supported by pip. Firstly, the distribution name is a separate field rather than embedded as part of the URL. Secondly, the commit hash is included even when retrieving based on a tag, in order to meet the requirement above that every link should include a hash to make things harder to forge (creating a malicious repo with a particular tag is easy, creating one with a specific hash, less so).

Example:

"source_url": "https://github.com/pypa/pip/archive/1.3.1.zip#sha256=2dc6b5a470a1bde68946f263f1af1515a2574a150a30d6ce02c6ff742fcc0db8
"source_url": "git+https://github.com/pypa/pip.git@1.3.1#7921be1537eac1e97bc40179a57f0349c2aee67d"
"source_url": "git+https://github.com/pypa/pip.git@7921be1537eac1e97bc40179a57f0349c2aee67d"

Semantic dependencies

Dependency metadata allows published projects to make use of functionality provided by other published projects, without needing to bundle copies of particular releases of those projects.

Semantic dependencies allow publishers to indicate not only which other projects are needed, but also why they’re needed. This additional information allows integrators to install just the dependencies they need for specific activities, making it easier to minimise installation footprints in constrained environments (regardless of the reasons for those constraints).

By default, dependency declarations are assumed to be for “runtime dependencies”: other releases that are needed to actually use the published release.

There are also four different kinds of optional dependency that releases may declare:

  • test dependencies: other releases that are needed to run the automated test suite for this release, but are not needed just to use it (e.g. nose2 or pytest)
  • build dependencies: other releases that are needed to build this a deployable binary version of this release from source (e.g. flit or setuptools)
  • doc dependencies: other releases that are needed to build the documentation for this distribution (e.g. the sphinx build tool)
  • dev dependencies: other releases that are needed when working on this distribution, but do not fit into exactly one of the other optional dependency categories (e.g. pylint, flake8). dev dependencies are also effectively considered as combined test, build, and doc dependencies, without needing to be listed three times

These optional categories are known as Extras. In addition to the four standard categories, projects may also declare their own custom categories in the Extras field.

There are also two standard extra categories that imply dependencies on other extras:

  • alldev: implies the test, build, doc, dev extras
  • all: if not otherwise defined, implies all declared extras

Dependency management is heavily dependent on the version identification and specification scheme defined in PEP 440 and the dependency specification, extra, and environment marker schemes defined in PEP 508.

All of these fields are optional. Automated tools MUST operate correctly if a distribution does not provide them, by assuming that a missing field indicates “Not applicable for this distribution”.

Mapping dependencies to development and distribution activities

The different categories of dependency are based on the various distribution and development activities identified above, and govern which dependencies should be installed for the specified activities:

  • Required runtime dependencies:
    • unconditional dependencies
  • Required build dependencies:
    • the build extra
    • the dev extra
    • If running the distribution’s test suite as part of the build process, also install the unconditional dependencies and test extra
  • Required development and publication dependencies:
    • unconditional dependencies
    • the test extra
    • the build extra
    • the doc extra
    • the dev extra

The notation described in Extras (optional dependencies) SHOULD be used to determine exactly what gets installed for various operations.

Installation tools SHOULD report an error if dependencies cannot be satisfied, MUST at least emit a warning, and MAY allow the user to force the installation to proceed regardless.

See Appendix B for an overview of mapping these dependencies to an RPM spec file.

Extras

A list of optional sets of dependencies that may be used to define conditional dependencies in dependency fields. See Extras (optional dependencies) for details.

The names of extras MUST abide by the same restrictions as those for distribution names.

The following extra names are available by default and MUST NOT be declared explicitly in this field:

  • all
  • alldev
  • build
  • dev
  • doc
  • test

Example:

"extras": ["warmup", "tea"]

Dependencies

A list of release requirements needed to actually run this release.

Public index servers MAY prohibit strict version matching clauses or direct references in this field.

Example:

"dependencies":
  {
    "requires": ["SciPy", "PasteDeploy", "zope.interface > 3.5.0"]
  },
  {
    "requires": ["pywin32 > 1.0"],
    "environment": "sys_platform == 'win32'"
  },
  {
    "requires": ["SoftCushions"],
    "extra": "warmup"
  }
]

While many dependencies will be needed to use a project release at all, others are needed only on particular platforms or only when particular optional features of the release are needed.

To handle this, release dependency specifiers are mappings with the following subfields:

  • requires: a list of requirements needed to satisfy the dependency
  • extra: the name of a set of optional dependencies that are requested and installed together. See Extras (optional dependencies) for details
  • environment: an environment marker defining the environment that needs these dependencies. The syntax and capabilities of environment markers are defined in PEP 508

Individual entries in the requires lists are strings using the dependency declaration format defined in PEP 508, with the exception that environment markers MUST NOT be included in the individual dependency declarations, and are instead supplied in the separate environment field.

requires is the only required subfield. When it is the only subfield, the dependencies are said to be unconditional. If extra or environment is specified, then the dependencies are conditional.

All three fields may be supplied, indicating that the dependencies are needed only when the named extra is requested in a particular environment.

Automated tools MUST combine related dependency specifiers (those with common values for extra and environment) into a single specifier listing multiple requirements when serialising metadata.

Despite this required normalisation, the same extra name or environment marker MAY appear in multiple conditional dependencies. This may happen, for example, if an extra itself only needs some of its dependencies in specific environments. It is only the combination of extras and environment markers that is required to be unique in a list of dependency specifiers.

Aside from the six standard extra categories, any extras referenced from a dependency specifier MUST be named in the Extras field for this distribution. This helps avoid typographical errors and also makes it straightforward to identify the available extras without scanning the full set of dependencies.

To reuse an extra definition as part of another extra, project releases MAY declare dependencies on themselves. To avoid infinite recursion in these cases, automated tools MUST special case dependencies from a project back onto itself.

Metadata Extensions

Extensions to the metadata MAY be present in a mapping under the extensions key. The keys MUST be valid prefixed names, while the values MUST themselves be nested mappings.

Two key names are reserved and MUST NOT be used by extensions, except as described below:

  • extension_version
  • installer_must_handle

The following example shows the python.details and python.commands standard extensions from PEP 459:

"extensions" : {
  "python.details": {
    "license": "GPL version 3, excluding DRM provisions",
    "keywords": [
      "comfy", "chair", "cushions", "too silly", "monty python"
    ],
    "classifiers": [
      "Development Status :: 4 - Beta",
      "Environment :: Console (Text Based)",
      "License :: OSI Approved :: GNU General Public License v3 (GPLv3)"
    ],
    "document_names": {
        "description": "README.rst",
        "license": "LICENSE.rst",
        "changelog": "NEWS"
    }
  },
  "python.commands": {
    "wrap_console": [{"chair": "chair:run_cli"}],
    "wrap_gui": [{"chair-gui": "chair:run_gui"}],
    "prebuilt": ["reduniforms"]
  },
}

Extension names are defined by distributions that will then make use of the additional published metadata in some way.

To reduce the chance of name conflicts, extension names SHOULD use a prefix that corresponds to a module name in the distribution that defines the meaning of the extension. This practice will also make it easier to find authoritative documentation for metadata extensions.

Metadata extensions allow development tools to record information in the metadata that may be useful during later phases of distribution, but is not essential for dependency resolution or building the software.

Extension versioning

Extensions MUST be versioned, using the extension_version key. However, if this key is omitted, then the implied version is 1.0.

Automated tools consuming extension metadata SHOULD warn if extension_version is greater than the highest version they support, and MUST fail if extension_version has a greater major version than the highest version they support (as described in PEP 440, the major version is the value before the first dot).

For broader compatibility, build tools MAY choose to produce extension metadata using the lowest metadata version that includes all of the needed fields.

Required extension handling

A project may consider correct handling of some extensions to be essential to correct installation of the software. This is indicated by setting the installer_must_handle field to true. Setting it to false or omitting it altogether indicates that processing the extension when installing the distribution is not considered mandatory by the developers.

Installation tools MUST fail if installer_must_handle is set to true for an extension and the tool does not have any ability to process that particular extension (whether directly or through a tool-specific plugin system).

If an installation tool encounters a required extension it doesn’t understand when attempting to install from a wheel archive, it MAY fall back on attempting to install from source rather than failing entirely.

Extras (optional dependencies)

As defined in PEP 508, extras are additional dependencies that enable an optional aspect of a project release, often corresponding to a try: import optional_dependency ... block in the code. They are also used to indicate semantic dependencies for activities other than normal runtime using (such as testing, building, or working on the component).

To support the use of the release with or without the optional dependencies, they are listed separately from the release’s core runtime dependencies and must be requested explicitly, either in the dependency specifications of another project, or else when issuing a command to an installation tool.

Example of a distribution with optional dependencies:

"name": "ComfyChair",
"extras": ["warmup"]
"dependencies": [
  {
    "requires": ["SoftCushions"],
    "extra": "warmup"
  },
  {
    "requires": ["cython"],
    "extra": "build"
  }
]

Other distributions require the additional dependencies by placing the relevant extra names inside square brackets after the distribution name when specifying the dependency. Multiple extras from a dependency can be requested by placing to

If the standard all extra has no explicitly declared entries, then integration tools SHOULD implicitly define it as a dependency on all of the extras explicitly declared by the project.

If the standard alldev extra has no explicitly declared entries, then integration tools SHOULD implicitly define it as a dependency on the standard test, build, doc, and dev extras.

The full set of dependency requirements is then based on the unconditional dependencies, along with those of any requested extras.

Dependency examples (showing just the requires subfield):

"requires": ["ComfyChair"]
    -> requires ``ComfyChair`` only

"requires": ["ComfyChair[warmup]"]
    -> requires ``ComfyChair`` and ``SoftCushions``

"requires": ["ComfyChair[all]"]
    -> requires ``ComfyChair`` and ``SoftCushions``, but will also
       pick up any new extras defined in later versions

Updating the metadata specification

The metadata specification may be updated with clarifications without requiring a new PEP or a change to the metadata version.

Changing the meaning of existing fields or adding new features (other than through the extension mechanism) requires a new metadata version defined in a new PEP.

Appendix A: Conversion notes for legacy metadata

The reference implementations for converting from legacy metadata to metadata 2.0 are:

  • the wheel project, which adds the bdist_wheel command to setuptools
  • the Warehouse project, which will eventually be migrated to the Python Packaging Authority as the next generation Python Package Index implementation
  • the distlib project which is derived from the core packaging infrastructure created for the distutils2 project

Note

These tools have yet to be updated for the switch to standard extensions for several fields.

While it is expected that there may be some edge cases where manual intervention is needed for clean conversion, the specification has been designed to allow fully automated conversion of almost all projects on PyPI.

Metadata conversion (especially on the part of the index server) is a necessary step to allow installation and analysis tools to start benefiting from the new metadata format, without having to wait for developers to upgrade to newer build systems.

Appendix B: Mapping dependency declarations to an RPM SPEC file

As an example of mapping this PEP to Linux distro packages, assume an example project without any extras defined is split into 2 RPMs in a SPEC file: example and example-devel.

The unconditional dependencies would be mapped to the Requires dependencies for the “example” RPM (a mapping from environment markers relevant to Linux to SPEC file conditions would also allow those to be handled correctly).

The build and dev extra dependencies would be mapped to the BuildRequires dependencies for the “example” RPM. Depending on how the %check section in the RPM was defined, the test extra may also be mapped to the BuildRequires declaration for the RPM.

All defined dependencies relevant to Linux in the dev, test, build, and doc extras would become Requires dependencies for the “example-devel” RPM.

A documentation toolchain dependency like Sphinx would either go in the build extra (for example, if man pages were included in the built distribution) or in the doc extra (for example, if the documentation is published solely through Read the Docs or the project website). This would be enough to allow an automated converter to map it to an appropriate dependency in the spec file.

If the project did define any extras, those could be mapped to additional virtual RPMs with appropriate BuildRequires and Requires entries based on the details of the dependency specifications. Alternatively, they could be mapped to other system package manager features (such as weak dependencies).

The metadata extension format should also provide a way for distribution specific hints to be included in the upstream project metadata without needing to manually duplicate any of the upstream metadata in a distribution specific format.

Appendix C: Summary of differences from PEP 345

  • Metadata-Version is now 2.0, with semantics specified for handling version changes
  • The increasingly complex ad hoc “Key: Value” format has been replaced by a more structured JSON compatible format that is easily represented as Python dictionaries, strings, lists.
  • Most fields are now optional and filling in dummy data for omitted fields is explicitly disallowed
  • Explicit permission for in-place clarifications without releasing a new version of the specification
  • The PEP now attempts to provide more of an explanation of why the fields exist and how they are intended to be used, rather than being a simple description of the permitted contents
  • Changed the version scheme to be based on PEP 440 rather than PEP 386
  • Added the source label mechanism as described in PEP 440
  • Formally defined dependency declarations, extras, and environment markers in PEP 508
  • Support for different kinds of dependencies through additional reserved extra names
  • Updated obsolescence mechanism
  • A well-defined metadata extension mechanism, and migration of any fields not needed for dependency resolution to standard extensions
  • With all due respect to Charles Schulz and Peanuts, many of the examples have been updated to be more thematically appropriate for Python ;)

The rationale for major changes is given in the following sections.

Metadata-Version semantics

The semantics of major and minor version increments are now specified, and follow the same model as the format version semantics specified for the wheel format in PEP 427: minor version increments must behave reasonably when processed by a tool that only understand earlier metadata versions with the same major version, while major version increments may include changes that are not compatible with existing tools.

The major version number of the specification has been incremented accordingly, as interpreting PEP 426 metadata obviously cannot be interpreted in accordance with earlier metadata specifications.

Whenever the major version number of the specification is incremented, it is expected that deployment will take some time, as either metadata consuming tools must be updated before other tools can safely start producing the new format, or else the sdist and wheel formats, along with the installation database definition, will need to be updated to support provision of multiple versions of the metadata in parallel.

Existing tools won’t abide by this guideline until they’re updated to support the new metadata standard, so the new semantics will first take effect for a hypothetical 2.x -> 3.0 transition. For the 1.x -> 2.x transition, we will use the approach where tools continue to produce the existing supplementary files (such as entry_points.txt) in addition to any equivalents specified using the new features of the standard metadata format (including the formal extension mechanism).

Switching to a JSON compatible format

The old “Key:Value” format was becoming increasingly limiting, with various complexities like parsers needing to know which fields were permitted to occur more than once, which fields supported the environment marker syntax (with an optional ";" to separate the value from the marker) and eventually even the option to embed arbitrary JSON inside particular subfields.

The old serialisation format also wasn’t amenable to easy conversion to standard Python data structures for use in any new install hook APIs, or in future extensions to the runtime importer APIs to allow them to provide information for inclusion in the installation database.

Accordingly, we’ve taken the step of switching to a JSON-compatible metadata format. This works better for APIs and is much easier for tools to parse and generate correctly. Changing the name of the metadata file also makes it easy to distribute 1.x and 2.x metadata in parallel, greatly simplifying several aspects of the migration to the new metadata format.

The specific choice of pydist.json as the preferred file name relates to the fact that the metadata described in these files applies to the distribution as a whole, rather than to any particular build. Additional metadata formats may be defined in the future to hold information that can only be determined after building a binary distribution for a particular target environment.

Changing the version scheme

See PEP 440 for a detailed rationale for the various changes made to the versioning scheme.

Source labels

The new source label support is intended to make it clearer that the constraints on public version identifiers are there primarily to aid in the creation of reliable automated dependency analysis tools. Projects are free to use whatever versioning scheme they like internally, so long as they are able to translate it to something the dependency analysis tools will understand.

Source labels also make it straightforward to record specific details of a version, like a hash or tag name that allows the release to be reconstructed from the project version control system.

Support for optional dependencies for distributions

The new extras system allows distributions to declare optional behaviour, and to use the dependency fields to indicate when particular dependencies are needed only to support that behaviour. It is derived from the equivalent system that is already in widespread use as part of setuptools and allows that aspect of the legacy setuptools metadata to be accurately represented in the new metadata format.

The additions to the extras syntax relative to setuptools are defined to make it easier to express the various possible combinations of dependencies, in particular those associated with build systems (with optional support for running the test suite) and development systems.

Support for different kinds of semantic dependencies

The separation of the five different kinds of dependency through the Extras system allows a project to optionally indicate whether a dependency is needed specifically to develop, build, test or use the distribution.

The advantage of having these distinctions supported in the upstream Python specific metadata is that even if a project doesn’t care about these distinction themselves, they may be more amenable to patches from downstream redistributors that separate the fields appropriately. Over time, this should allow much greater control over where and when particular dependencies end up being installed.

Support for metadata extensions

The new extension effectively allows sections of the metadata namespace to be delegated to other projects, while preserving a standard overall format metadata format for easy of processing by distribution tools that do not support a particular extension.

It also works well in combination with the new build extra to allow a distribution to depend on tools which do know how to handle the chosen extension, and the new extras mechanism in general, allowing support for particular extensions to be provided as optional features.

Possible future uses for extensions include declaration of plugins for other projects and hints for automatic conversion to Linux system packages.

The ability to declare an extension as required is included primarily to allow the definition of the metadata hooks extension to be deferred until some time after the initial adoption of the metadata 2.0 specification. If a release needs a postinstall hook to run in order to complete the installation successfully, then earlier versions of tools should fall back to installing from source rather than installing from a wheel file and then failing to run the expected postinstall hook.

Appendix D: Deferred features

Several potentially useful features have been deliberately deferred in order to better prioritise our efforts in migrating to the new metadata standard. These all reflect information that may be nice to have in the new metadata, but which can be readily added through metadata extensions or in metadata 2.1 without breaking any use cases already supported by metadata 2.0.

Once the pypi, setuptools, pip, wheel and distlib projects support creation and consumption of metadata 2.0, then we may revisit the creation of metadata 2.1 with some or all of these additional features.

Standard extensions

Some of the information provided by the legacy metadata system has been moved out to standard extensions defined in PEP 459.

This allows publication of the core dependency metadata in a more readily consumable format to proceed even before the full details of those extensions have been resolved.

Improved handling of project obsolescence, renames and mergers

Earlier drafts of this PEP included new Provides and Obsoleted-By fields for more robust automated notifications and tracking of project obsolescence, renames and mergers.

This isn’t an essential feature of a dependency management system, and has been deferred indefinitely as a possible future metadata extension.

MIME type registration

At some point after acceptance of the PEP, we may submit the following MIME type registration request to IANA:

  • application/vnd.python.pydist+json

It’s even possible we may be able to just register the vnd.python namespace under the banner of the PSF rather than having to register the individual subformats.

String methods in environment markers

Supporting at least “.startswith” and “.endswith” string methods in environment markers would allow some conditions to be written more naturally. For example, "sys.platform.startswith('win')" is a somewhat more intuitive way to mark Windows specific dependencies, since "'win' in sys.platform" is incorrect thanks to cygwin and the fact that 64-bit Windows still shows up as win32 is more than a little strange.

Appendix E: Rejected features

The following features have been explicitly considered and rejected as introducing too much additional complexity for too small a gain in expressiveness.

Separate lists for conditional and unconditional dependencies

Earlier versions of this PEP used separate lists for conditional and unconditional dependencies. This turned out to be annoying to handle in automated tools and removing it also made the PEP and metadata schema substantially shorter, suggesting it was actually harder to explain as well.

Separate lists for semantic dependencies

Earlier versions of this PEP used separate fields rather than the extras system for test, build, documentation, and development dependencies. This turned out to be annoying to handle in automated tools and removing it also made the PEP and metadata schema substantially shorter, suggesting it was actually harder to explain as well.

Introducing friction for overly precise dependency declarations

Earlier versions of this PEP attempted to introduce friction into the inappropriate use of overly strict dependency declarations in published releases. Discussion on distutils-sig came to the conclusion that wasn’t a serious enough problem to tackle directly at the interoperability specification layer, and if it does become a problem in the future, it would be better tackled at the point where projects are uploaded to the public Python Package Index.

Disallowing underscores in distribution names

Debian doesn’t actually permit underscores in names, but that seems unduly restrictive for this spec given the common practice of using valid Python identifiers as Python distribution names. A Debian side policy of converting underscores to hyphens seems easy enough to implement (and the requirement to consider hyphens and underscores as equivalent ensures that doing so won’t introduce any name conflicts).

Allowing the use of Unicode in distribution names

This PEP deliberately avoids following Python 3 down the path of arbitrary Unicode identifiers, as the security implications of doing so are substantially worse in the software distribution use case (it opens up far more interesting attack vectors than mere code obfuscation).

In addition, the existing tools really only work properly if you restrict names to ASCII and changing that would require a lot of work for all the automated tools in the chain.

It may be reasonable to revisit this question at some point in the (distant) future, but setting up a more reliable software distribution system is challenging enough without adding more general Unicode identifier support into the mix.

Depending on source labels

There is no mechanism to express a dependency on a source label - they are included in the metadata for internal project reference only. Instead, dependencies must be expressed in terms of either public versions or else direct URL references.

Alternative dependencies

An earlier draft of this PEP considered allowing lists in place of the usual strings in dependency specifications to indicate that there are multiple ways to satisfy a dependency.

If at least one of the individual dependencies was already available, then the entire dependency would be considered satisfied, otherwise the first entry would be added to the dependency set.

Alternative dependency specification example:

["Pillow", "PIL"]
["mysql", "psycopg2 >= 4", "sqlite3"]

However, neither of the given examples is particularly compelling, since Pillow/PIL style forks aren’t common, and the database driver use case would arguably be better served by an SQL Alchemy defined “supported database driver” metadata extension where a project depends on SQL Alchemy, and then declares in the extension which database drivers are checked for compatibility by the upstream project.

Compatible release comparisons in environment markers

PEP 440 defines a rich syntax for version comparisons that could potentially be useful with python_version and python_full_version in environment markers. However, allowing the full syntax would mean environment markers are no longer a Python subset, while allowing only some of the comparisons would introduce yet another special case to handle.

Given that environment markers are only used in cases where a higher level “or” is implied by the metadata structure, it seems easier to require the use of multiple comparisons against specific Python versions for the rare cases where this would be useful.

Conditional provides

Under the revised metadata design, conditional “provides” based on runtime features or the environment would go in a separate “may_provide” field. However, it isn’t clear there’s any use case for doing that, so the idea is rejected unless someone can present a compelling use case (and even then the idea won’t be reconsidered until metadata 2.1 at the earliest).

References

This document specifies version 2.0 of the metadata format. Version 1.0 is specified in PEP 241. Version 1.1 is specified in PEP 314. Version 1.2 is specified in PEP 345.

The initial attempt at a standardised version scheme, along with the justifications for needing such a standard can be found in PEP 386.


Source: https://github.com/python/peps/blob/main/peps/pep-0426.rst

Last modified: 2023-12-14 21:37:56 GMT