PEP: 491 Title: The Wheel Binary Package Format 1.9 Version: $Revision$ Last-Modified: $Date$ Author: Daniel Holth Discussions-To: Status: Draft Type: Standards Track Content-Type: text/x-rst Created: 16 April 2015 Abstract ======== This PEP describes the second version of a built-package format for Python called "wheel". Wheel provides a Python-specific, relocatable package format that allows people to install software more quickly and predictably than re-building from source each time. A wheel is a ZIP-format archive with a specially formatted file name and the ``.whl`` extension. It contains a single distribution nearly as it would be installed according to PEP 376 with a particular installation scheme. Simple wheels can be unpacked onto ``sys.path`` and used directly but wheels are usually installed with a specialized installer. This version of the wheel specification adds support for installing distributions into many different directories, and adds a way to find those files after they have been installed. Rationale ========= Wheel 1.0 is best at installing files into ``site-packages`` and a few other locations specified by distutils, but users would like to install files from single distribution into many directories -- perhaps separate locations for docs, data, and code. Unfortunately not everyone agrees on where these install locations should be relative to the root directory. This version of the format adds many more categories, each of which can be installed to a different destination based on policy. Since it might also be important to locate the installed files at runtime, this version of the format also adds a way to record the installed paths in a way that can be read by the installed software. Details ======= Installing a wheel 'distribution-1.0-py32-none-any.whl' ------------------------------------------------------- Wheel installation notionally consists of two phases: - Unpack. a. Parse ``distribution-1.0.dist-info/WHEEL``. b. Check that installer is compatible with Wheel-Version. Warn if minor version is greater, abort if major version is greater. c. If Root-Is-Purelib == 'true', unpack archive into purelib (site-packages). d. Else unpack archive into platlib (site-packages). - Spread. a. Unpacked archive includes ``distribution-1.0.dist-info/`` and (if there is data) ``distribution-1.0.data/``. b. Move each subtree of ``distribution-1.0.data/`` onto its destination path. Each subdirectory of ``distribution-1.0.data/`` is a key into a dict of destination directories, such as ``distribution-1.0.data/(purelib|platlib|headers|scripts|data)``. c. Update scripts starting with ``#!python`` to point to the correct interpreter. (Note: Python scripts are usually handled by package metadata, and not included verbatim in wheel.) d. Update ``distribution-1.0.dist.info/RECORD`` with the installed paths. e. If empty, remove the ``distribution-1.0.data`` directory. f. Compile any installed .py to .pyc. (Uninstallers should be smart enough to remove .pyc even if it is not mentioned in RECORD.) In practice, installers will usually extract files directly from the archive to their destinations without writing a temporary ``distribution-1.0.data/`` directory. Recommended installer features '''''''''''''''''''''''''''''' Rewrite ``#!python``. In wheel, verbatim scripts are packaged in ``{distribution}-{version}.data/scripts/``. If the first line of a file in ``scripts/`` starts with exactly ``b'#!python'``, rewrite to point to the correct interpreter. Unix installers may need to add the +x bit to these files if the archive was created on Windows. The ``b'#!pythonw'`` convention is allowed. ``b'#!pythonw'`` indicates a GUI script instead of a console script. Generate script wrappers. Python scripts are more commonly represented as a ``module:callable`` string in package metadata, and are not included verbatim in the wheel archive's ``scripts`` directory. This kind of script gives the installer an opportunity to generate platform specific wrappers. Recommended archiver features ''''''''''''''''''''''''''''' Place ``.dist-info`` at the end of the archive. Archivers are encouraged to place the ``.dist-info`` files physically at the end of the archive. This enables some potentially interesting ZIP tricks including the ability to amend the metadata without rewriting the entire archive. File Format ----------- File name convention '''''''''''''''''''' The wheel filename is ``{distribution}-{version}(-{build tag})?-{python tag}-{abi tag}-{platform tag}.whl``. distribution Distribution name, e.g. 'django', 'pyramid'. version Distribution version, e.g. 1.0. build tag Optional build number. Must start with a digit. A tie breaker if two wheels have the same version. Sort as the empty string if unspecified, else sort the initial digits as a number, and the remainder lexicographically. language implementation and version tag E.g. 'py27', 'py2', 'py3'. abi tag E.g. 'cp33m', 'abi3', 'none'. platform tag E.g. 'linux_x86_64', 'any'. For example, ``distribution-1.0-1-py27-none-any.whl`` is the first build of a package called 'distribution', and is compatible with Python 2.7 (any Python 2.7 implementation), with no ABI (pure Python), on any CPU architecture. The last three components of the filename before the extension are called "compatibility tags." The compatibility tags express the package's basic interpreter requirements and are detailed in PEP 425. Escaping and Unicode '''''''''''''''''''' Each component of the filename is escaped by replacing runs of non-alphanumeric characters with an underscore ``_``:: re.sub("[^\w\d.]+", "_", distribution, re.UNICODE) The archive filename is Unicode. The packaging tools may only support ASCII package names, but Unicode filenames are supported in this specification. The filenames *inside* the archive are encoded as UTF-8. Although some ZIP clients in common use do not properly display UTF-8 filenames, the encoding is supported by both the ZIP specification and Python's ``zipfile``. File contents ''''''''''''' The contents of a wheel file, where {distribution} is replaced with the name of the package, e.g. ``beaglevote`` and {version} is replaced with its version, e.g. ``1.0.0``, consist of: #. ``/``, the root of the archive, contains all files to be installed in ``purelib`` or ``platlib`` as specified in ``WHEEL``. ``purelib`` and ``platlib`` are usually both ``site-packages``. #. ``{distribution}-{version}.dist-info/`` contains metadata. #. ``{distribution}-{version}.data/`` contains one subdirectory for each non-empty install scheme key not already covered, where the subdirectory name is an index into a dictionary of install paths (e.g. ``data``, ``scripts``, ``include``, ``purelib``, ``platlib``). #. Python scripts must appear in ``scripts`` and begin with exactly ``b'#!python'`` in order to enjoy script wrapper generation and ``#!python`` rewriting at install time. They may have any or no extension. #. ``{distribution}-{version}.dist-info/METADATA`` is Metadata version 1.1 or greater format metadata. #. ``{distribution}-{version}.dist-info/WHEEL`` is metadata about the archive itself in the same basic key: value format:: Wheel-Version: 1.9 Generator: bdist_wheel 1.9 Root-Is-Purelib: true Tag: py2-none-any Tag: py3-none-any Build: 1 Install-Paths-To: wheel/_paths.py Install-Paths-To: wheel/_paths.json #. ``Wheel-Version`` is the version number of the Wheel specification. #. ``Generator`` is the name and optionally the version of the software that produced the archive. #. ``Root-Is-Purelib`` is true if the top level directory of the archive should be installed into purelib; otherwise the root should be installed into platlib. #. ``Tag`` is the wheel's expanded compatibility tags; in the example the filename would contain ``py2.py3-none-any``. #. ``Build`` is the build number and is omitted if there is no build number. #. ``Install-Paths-To`` is a location *relative to the archive* that will be overwritten with the install-time paths of each category in the install scheme. See the install paths section. May appear 0 or more times. #. A wheel installer should warn if Wheel-Version is greater than the version it supports, and must fail if Wheel-Version has a greater major version than the version it supports. #. Wheel, being an installation format that is intended to work across multiple versions of Python, does not generally include .pyc files. #. Wheel does not contain setup.py or setup.cfg. The .dist-info directory ^^^^^^^^^^^^^^^^^^^^^^^^ #. Wheel .dist-info directories include at a minimum METADATA, WHEEL, and RECORD. #. METADATA is the package metadata, the same format as PKG-INFO as found at the root of sdists. #. WHEEL is the wheel metadata specific to a build of the package. #. RECORD is a list of (almost) all the files in the wheel and their secure hashes. Unlike PEP 376, every file except RECORD, which cannot contain a hash of itself, must include its hash. The hash algorithm must be sha256 or better; specifically, md5 and sha1 are not permitted, as signed wheel files rely on the strong hashes in RECORD to validate the integrity of the archive. #. PEP 376's INSTALLER and REQUESTED are not included in the archive. #. RECORD.jws is used for digital signatures. It is not mentioned in RECORD. #. RECORD.p7s is allowed as a courtesy to anyone who would prefer to use S/MIME signatures to secure their wheel files. It is not mentioned in RECORD. #. During extraction, wheel installers verify all the hashes in RECORD against the file contents. Apart from RECORD and its signatures, installation will fail if any file in the archive is not both mentioned and correctly hashed in RECORD. The .data directory ^^^^^^^^^^^^^^^^^^^ Any file that is not normally installed inside site-packages goes into the .data directory, named as the .dist-info directory but with the .data/ extension:: distribution-1.0.dist-info/ distribution-1.0.data/ The .data directory contains subdirectories with the scripts, headers, documentation and so forth from the distribution. During installation the contents of these subdirectories are moved onto their destination paths. If a subdirectory is not found in the install scheme, the installer should emit a warning, and it should be installed at ``distribution-1.0.data/...`` as if the package was unpacked by a standard unzip tool. Install paths ^^^^^^^^^^^^^ In addition to the distutils install paths, wheel now includes the listed categories based on GNU autotools. This expanded scheme should help installers to implement system policy, but installers may root each category at any location. A UNIX install scheme might map the categories to their installation paths like this:: { 'bindir': '$eprefix/bin', 'sbindir': '$eprefix/sbin', 'libexecdir': '$eprefix/libexec', 'sysconfdir': '$prefix/etc', 'sharedstatedir': '$prefix/com', 'localstatedir': '$prefix/var', 'libdir': '$eprefix/lib', 'static_libdir': r'$prefix/lib', 'includedir': '$prefix/include', 'datarootdir': '$prefix/share', 'datadir': '$datarootdir', 'mandir': '$datarootdir/man', 'infodir': '$datarootdir/info', 'localedir': '$datarootdir/locale', 'docdir': '$datarootdir/doc/$dist_name', 'htmldir': '$docdir', 'dvidir': '$docdir', 'psdir': '$docdir', 'pdfdir': '$docdir', 'pkgdatadir': '$datadir/$dist_name' } If a package needs to find its files at runtime, it can request they be written to a specified file or files by the installer *and* included in those same files inside the archive itself, relative to their location within the archive (so a wheel is still installed correctly if unpacked with a standard unzip tool, or perhaps not unpacked at all). If the ``WHEEL`` metadata contains these fields:: Install-Paths-To: wheel/_paths.py Install-Paths-To: wheel/_paths.json Then the wheel installer, when it is about to unpack ``wheel/_paths.py`` from the archive, replaces it with the actual paths used at install time. The paths may be absolute or relative to the generated file. If the filename ends with ``.py`` then a Python script is written. The script MUST be executed to get the paths, but it will probably look like this:: data='../wheel-0.26.0.dev1.data/data' headers='../wheel-0.26.0.dev1.data/headers' platlib='../wheel-0.26.0.dev1.data/platlib' purelib='../wheel-0.26.0.dev1.data/purelib' scripts='../wheel-0.26.0.dev1.data/scripts' # ... If the filename ends with ``.json`` then a JSON document is written:: { "data": "../wheel-0.26.0.dev1.data/data", ... } Only the categories actually used by a particular wheel must be written to this file. These files are designed to be written to a location that can be found by the installed package without introducing any dependency on a packaging library. Signed wheel files ------------------ Wheel files include an extended RECORD that enables digital signatures. PEP 376's RECORD is altered to include a secure hash ``digestname=urlsafe_b64encode_nopad(digest)`` (urlsafe base64 encoding with no trailing = characters) as the second column instead of an md5sum. All possible entries are hashed, including any generated files such as .pyc files, but not RECORD which cannot contain its own hash. For example:: file.py,sha256=AVTFPZpEKzuHr7OvQZmhaU3LvwKz06AJw8mT\_pNh2yI,3144 distribution-1.0.dist-info/RECORD,, The signature file(s) RECORD.jws and RECORD.p7s are not mentioned in RECORD at all since they can only be added after RECORD is generated. Every other file in the archive must have a correct hash in RECORD or the installation will fail. If JSON web signatures are used, one or more JSON Web Signature JSON Serialization (JWS-JS) signatures is stored in a file RECORD.jws adjacent to RECORD. JWS is used to sign RECORD by including the SHA-256 hash of RECORD as the signature's JSON payload:: { "hash": "sha256=ADD-r2urObZHcxBW3Cr-vDCu5RJwT4CaRTHiFmbcIYY" } (The hash value is the same format used in RECORD.) If RECORD.p7s is used, it must contain a detached S/MIME format signature of RECORD. A wheel installer is not required to understand digital signatures but MUST verify the hashes in RECORD against the extracted file contents. When the installer checks file hashes against RECORD, a separate signature checker only needs to establish that RECORD matches the signature. See - http://self-issued.info/docs/draft-ietf-jose-json-web-signature.html - http://self-issued.info/docs/draft-jones-jose-jws-json-serialization.html - http://self-issued.info/docs/draft-ietf-jose-json-web-key.html - http://self-issued.info/docs/draft-jones-jose-json-private-key.html Comparison to .egg ------------------ #. Wheel is an installation format; egg is importable. Wheel archives do not need to include .pyc and are less tied to a specific Python version or implementation. Wheel can install (pure Python) packages built with previous versions of Python so you don't always have to wait for the packager to catch up. #. Wheel uses .dist-info directories; egg uses .egg-info. Wheel is compatible with the new world of Python packaging and the new concepts it brings. #. Wheel has a richer file naming convention for today's multi-implementation world. A single wheel archive can indicate its compatibility with a number of Python language versions and implementations, ABIs, and system architectures. Historically the ABI has been specific to a CPython release, wheel is ready for the stable ABI. #. Wheel is lossless. The first wheel implementation bdist_wheel always generates egg-info, and then converts it to a .whl. It is also possible to convert existing eggs and bdist_wininst distributions. #. Wheel is versioned. Every wheel file contains the version of the wheel specification and the implementation that packaged it. Hopefully the next migration can simply be to Wheel 2.0. #. Wheel is a reference to the other Python. FAQ === Wheel defines a .data directory. Should I put all my data there? ----------------------------------------------------------------- This specification does not have an opinion on how you should organize your code. The .data directory is just a place for any files that are not normally installed inside ``site-packages`` or on the PYTHONPATH. In other words, you may continue to use ``pkgutil.get_data(package, resource)`` even though *those* files will usually not be distributed in *wheel's* ``.data`` directory. Why does wheel include attached signatures? ------------------------------------------- Attached signatures are more convenient than detached signatures because they travel with the archive. Since only the individual files are signed, the archive can be recompressed without invalidating the signature or individual files can be verified without having to download the whole archive. Why does wheel allow JWS signatures? ------------------------------------ The JOSE specifications of which JWS is a part are designed to be easy to implement, a feature that is also one of wheel's primary design goals. JWS yields a useful, concise pure-Python implementation. Why does wheel also allow S/MIME signatures? -------------------------------------------- S/MIME signatures are allowed for users who need or want to use existing public key infrastructure with wheel. Signed packages are only a basic building block in a secure package update system. Wheel only provides the building block. What's the deal with "purelib" vs. "platlib"? --------------------------------------------- Wheel preserves the "purelib" vs. "platlib" distinction, which is significant on some platforms. For example, Fedora installs pure Python packages to '/usr/lib/pythonX.Y/site-packages' and platform dependent packages to '/usr/lib64/pythonX.Y/site-packages'. A wheel with "Root-Is-Purelib: false" with all its files in ``{name}-{version}.data/purelib`` is equivalent to a wheel with "Root-Is-Purelib: true" with those same files in the root, and it is legal to have files in both the "purelib" and "platlib" categories. In practice a wheel should have only one of "purelib" or "platlib" depending on whether it is pure Python or not and those files should be at the root with the appropriate setting given for "Root-is-purelib". Is it possible to import Python code directly from a wheel file? ---------------------------------------------------------------- Technically, due to the combination of supporting installation via simple extraction and using an archive format that is compatible with ``zipimport``, a subset of wheel files *do* support being placed directly on ``sys.path``. However, while this behaviour is a natural consequence of the format design, actually relying on it is generally discouraged. Firstly, wheel *is* designed primarily as a distribution format, so skipping the installation step also means deliberately avoiding any reliance on features that assume full installation (such as being able to use standard tools like ``pip`` and ``virtualenv`` to capture and manage dependencies in a way that can be properly tracked for auditing and security update purposes, or integrating fully with the standard build machinery for C extensions by publishing header files in the appropriate place). Secondly, while some Python software is written to support running directly from a zip archive, it is still common for code to be written assuming it has been fully installed. When that assumption is broken by trying to run the software from a zip archive, the failures can often be obscure and hard to diagnose (especially when they occur in third party libraries). The two most common sources of problems with this are the fact that importing C extensions from a zip archive is *not* supported by CPython (since doing so is not supported directly by the dynamic loading machinery on any platform) and that when running from a zip archive the ``__file__`` attribute no longer refers to an ordinary filesystem path, but to a combination path that includes both the location of the zip archive on the filesystem and the relative path to the module inside the archive. Even when software correctly uses the abstract resource APIs internally, interfacing with external components may still require the availability of an actual on-disk file. Like metaclasses, monkeypatching and metapath importers, if you're not already sure you need to take advantage of this feature, you almost certainly don't need it. If you *do* decide to use it anyway, be aware that many projects will require a failure to be reproduced with a fully installed package before accepting it as a genuine bug. References ========== .. [1] PEP acceptance (https://mail.python.org/pipermail/python-dev/2013-February/124103.html) Appendix ======== Example urlsafe-base64-nopad implementation:: # urlsafe-base64-nopad for Python 3 import base64 def urlsafe_b64encode_nopad(data): return base64.urlsafe_b64encode(data).rstrip(b'=') def urlsafe_b64decode_nopad(data): pad = b'=' * (4 - (len(data) & 3)) return base64.urlsafe_b64decode(data + pad) Copyright ========= This document has been placed into the public domain. .. Local Variables: mode: indented-text indent-tabs-mode: nil sentence-end-double-space: t fill-column: 70 coding: utf-8 End: