Mailman 3 October 2013 - Python-Dev

Submitting PEP 453 (pip bootstrapping) for pronouncement
by Nick Coghlan 15 Oct '13

15 Oct '13

The feedback on the previous version of PEP 453 could be summed up as "yes for 3.4, no for 2.7 and 3.3", and I've now updated the PEP accordingly. HTML: http://www.python.org/dev/peps/pep-0453/ Main diff: http://hg.python.org/peps/rev/405b80d54b7d (there are additional minor cleanup diffs after that one) Updates that directly affect the proposal: - *all* functional changes are now 3.4 only - still proposes docs changes for 2.7 & 3.3 - since ``ensurepip`` now never contacts PyPI, we tweaked the proposal to include updating source installs (via "make install" and "make altinstall") to invoke ensurepip by default, to be consistent with the Windows and Mac OS X binary installers. The proposal is explicit that there will be a mechanism to opt out of that step (we haven't decided on a exact mechanism yet, but a make variable seems likely). We had chosen to make invoking ensurepip a separate step back when the proposal still involved implicitly contacting PyPI, but with that aspect removed, we realised our original rationale for the cross-platform inconsistency was no longer valid. Other changes: - dropped rationale that was no longer relevant with the reduced scope - explicitly noted the rejection of making functional changes in 2.7 and 3.3 - noted current effort to create a Windows installer for pip - noted possibility of a future PEP to provide a combined CPython 2.7, pip and Python Launcher installer from python.org Cheers, Nick. ================================= PEP: 453 Title: Explicit bootstrapping of pip in Python installations Version: $Revision$ Last-Modified: $Date$ Author: Donald Stufft <donald(a)stufft.io>, Nick Coghlan <ncoghlan(a)gmail.com> BDFL-Delegate: Martin von Löwis Status: Draft Type: Process Content-Type: text/x-rst Created: 10-Aug-2013 Post-History: 30-Aug-2013, 15-Sep-2013, 18-Sep-2013, 19-Sep-2013, 23-Sep-2013, 29-Sep-2013, 13-Oct-2013 Abstract ======== This PEP proposes that the `Installing Python Modules <http://docs.python.org/3/install>`__ guide in Python 2.7, 3.3 and 3.4 be updated to officially recommend the use of ``pip`` as the default installer for Python packages, and that appropriate technical changes be made in Python 3.4 to provide ``pip`` by default in support of that recommendation. Rationale ========= There are two related, but distinct rationales for the proposal in this PEP. The first relates to the experience of new users, while the second relates to better enabling the evolution of the broader Python packaging ecosystem. Improving the new user experience --------------------------------- Currently, on systems without a platform package manager and repository, installing a third-party Python package into a freshly installed Python requires first identifying an appropriate package manager and then installing it. Even on systems that *do* have a platform package manager, it is unlikely to include every package that is available on the Python Package Index, and even when a desired third-party package is available, the correct name in the platform package manager may not be clear. This means that, to work effectively with the Python Package Index ecosystem, users must know which package manager to install, where to get it, and how to install it. The effect of this is that third-party Python projects are currently required to choose from a variety of undesirable alternatives: * Assume the user already has a suitable cross-platform package manager installed. * Duplicate the instructions and tell their users how to install the package manager. * Completely forgo the use of dependencies to ease installation concerns for their users. All of these available options have significant drawbacks. If a project simply assumes a user already has the tooling then beginning users may get a confusing error message when the installation command doesn't work. Some operating systems may ease this pain by providing a global hook that looks for commands that don't exist and suggest an OS package they can install to make the command work, but that only works on systems with platform package managers that include a package that provides the relevant cross-platform installer command (such as many major Linux distributions). No such assistance is available for Windows and Mac OS X users, or more conservative Linux distributions. The challenges of dealing with this problem for beginners (who are often also completely new to programming, the use of command line tools and editing system environment variables) are a regular feature of feedback the core Python developers receive from professional educators and others introducing new users to Python. If a project chooses to duplicate the installation instructions and tell their users how to install the package manager before telling them how to install their own project then whenever these instructions need updates they need updating by every project that has duplicated them. This is particular problematic when there are multiple competing installation tools available, and different projects recommend different tools. This specific problem can be partially alleviated by strongly promoting ``pip`` as the default installer and recommending that other projects reference `pip's own bootstrapping instructions <http://www.pip-installer.org/en/latest/installing.html>`__ rather than duplicating them. However the user experience created by this approach still isn't particularly good (although there is an effort under way to create a combined Windows installer for ``pip`` and its dependencies that should improve matters on that platform, and Mac OS X and \*nix platforms generally have ``wget`` and hence the ability to easily download and run the bootstrap scripts from the command line). The projects that have decided to forgo dependencies altogether are forced to either duplicate the efforts of other projects by inventing their own solutions to problems or are required to simply include the other projects in their own source trees. Both of these options present their own problems either in duplicating maintenance work across the ecosystem or potentially leaving users vulnerable to security issues because the included code or duplicated efforts are not automatically updated when upstream releases a new version. By officially recommending and providing by default a specific cross-platform package manager it will be easier for users trying to install these third-party packages as well as easier for the people distributing them as they should now be able to safely assume that most users will have the appropriate installation tools available (or access to clear instructions on how to obtain them). This is expected to become more important in the future as the Wheel_ package format (deliberately) does not have a built in "installer" in the form of ``setup.py`` so users wishing to install from a wheel file will want an installer even in the simplest cases. Reducing the burden of actually installing a third-party package should also decrease the pressure to add every useful module to the standard library. This will allow additions to the standard library to focus more on why Python should have a particular tool out of the box, and why it is reasonable for that package to adopt the standard library's 18-24 month feature release cycle, instead of using the general difficulty of installing third-party packages as justification for inclusion. Providing a standard installation system also helps with bootstrapping alternate build and installer systems, such as ``setuptools``, ``zc.buildout`` and the ``hashdist``/``conda`` combination that is aimed specifically at the scientific community. So long as ``pip install <tool>`` works, then a standard Python-specific installer provides a reasonably secure, cross platform mechanism to get access to these utilities. Enabling the evolution of the broader Python packaging ecosystem ---------------------------------------------------------------- As no new packaging standard can achieve widespread adoption without a transition strategy that covers the versions of Python that are in widespread *current* use (rather than merely future versions, like most language features), the change proposed in this PEP is considered a necessary step in the evolution of the Python packaging ecosystem The broader community has embraced the Python Package Index as a mechanism for distributing and installing Python software, but the different concerns of language evolution and secure software distribution mean that a faster feature release cycle that encompasses older versions is needed to properly support the latter. In addition, the core CPython development team have the luxury of dropping support for earlier Python versions well before the rest of the community, as downstream commercial redistributors pick up the task of providing support for those versions to users that still need it, while many third party libraries maintain compatibility with those versions as long as they remain in widespread use. This means that the current ``setup.py install`` based model for package installation poses serious difficulties for the development and adoption of new packaging standards, as, depending on how a project writes their ``setup.py`` file, the installation command (along with other operations) may end up invoking the standard library's ``distutils`` package. As an indicator of how this may cause problems for the broader ecosystem, consider that the feature set of ``distutils`` in Python 2.6 was frozen in June 2008 (with the release of Python 2.6b1), while the feature set of ``distutils`` in Python 2.7 was frozen in April 2010 (with the release of Python 2.7b1). By contrast, using a separate installer application like ``pip`` (which ensures that even ``setup.py`` files that invoke ``distutils`` directly still support the new packaging standards) makes it possible to support new packaging standards in older versions of Python, just by upgrading ``pip`` (which receives new feature releases roughly every 6 months). The situation on older versions of Python is further improved by making it easier for end users to install and upgrade newer build systems like ``setuptools`` or improved PyPI upload utilities like ``twine``. It is not coincidental that this proposed model of using a separate installer program with more metadata heavy and less active distribution formats matches that used by most operating systems (including Windows since the introduction of the installer service and the MSI file format), as well as many other language specific installers. For Python 2.6, this compatibility issue is largely limited to various enterprise Linux distributions (and their downstream derivatives). These distributions often have even slower update cycles than CPython, so they offer full support for versions of Python that are considered "security fix only" versions upstream (and sometimes may even be to the point where the core development team no longer support them at all - you can still get commercial support for Python 2.3 if you really need it!). In practice, the fact that tools like ``wget`` and ``curl`` are readily available on Linux systems, that most users of Python on Linux are already familiar with the command line, and that most Linux distributions ship with a default configuration that makes running Python scripts easy, means that the existing ``pip`` bootstrapping instructions for any \*nix system are already quite straightforward. Even if ``pip`` isn't provided by the system package manager, then using ``wget`` or ``curl`` to retrieve the bootstrap script from www.pip-installer.org and then running it is just a couple of shell commands that can easily be copied and pasted as necessary. Accordingly, for any version of Python on any \*nix system, the need to bootstrap ``pip`` in older versions isn't considered a major barrier to adoption of new packaging standards, since it's just one more small speedbump encountered by users of these long term stable releases. For \*nix systems, this PEP's formal endorsement of ``pip`` as the preferred default packaging tool is seen as more important than the underlying technical details involved in making ``pip`` available by default, since it shifts the nature of the conversation between the developers of ``pip`` and downstream repackagers of both ``pip`` and CPython. For Python 2.7, on the other hand, the compatibility issue for adopting new metadata standards is far more widespread, as it affects the python.org binary installers for Windows and Mac OS X, as well as even relatively fast moving \*nix platforms. Firstly, and unlike Python 2.6, Python 2.7 is still a fully supported upstream version, and will remain so until the release of Python 2.7.9 (currently scheduled for May 2015), at which time it is expected to enter the usual "security fix only" mode. That means there are at least another 19 months where Python 2.7 is a deployment target for Python applications that enjoys full upstream support. Even after the core development team switches 2.7 to security release only mode in 2015, Python 2.7 will likely remain a commercially supported legacy target out beyond 2020. While Python 3 already presents a compelling alternative over Python 2 for *new* Python applications and deployments without an existing investment in Python 2 and without a dependency on specific Python 2 only third party modules (a set which is getting ever smaller over time), it is going to take longer to create compelling business cases to update existing Python 2.7 based infrastructure to Python 3, especially in situations where the culture of automated testing is weak (or nonexistent), making it difficult to effectively use the available migration utilities. While this PEP only proposes documentation changes for Python 2.7, once ``pip`` has a Windows installer available, a separate PEP will be created and submitted proposing the creation and distribution of aggregate installers for future CPython 2.7 maintenance releases that combine the CPython, ``pip`` and Python Launcher for Windows installers into a single download (the separate downloads would still remain available - the aggregate installers would be provided as a convenience, and as a clear indication of the recommended operating environment for Python in Windows systems). Why pip? -------- ``pip`` has been chosen as the preferred default installer, as it is an already popular tool that addresses several design and user experience issues with its predecessor ``easy_install`` (these issues can't readily be fixed in ``easy_install`` itself due to backwards compatibility concerns). ``pip`` is also well suited to working within the bounds of a single Python runtime installation (including associated virtual environments), which is a desirable feature for a tool bundled with CPython. Other tools like ``zc.buildout`` and ``conda`` are more ambitious in their aims (and hence substantially better than ``pip`` at handling external binary dependencies), so it makes sense for the Python ecosystem to treat them more like platform package managers to interoperate with rather than as the default cross-platform installation tool. This relationship is similar to that between ``pip`` and platform package management systems like ``apt`` and ``yum`` (which are also designed to handle arbitrary binary dependencies). Proposal Overview ================= This PEP proposes that the `Installing Python Modules <http://docs.python.org/3/install>`__ guide be updated to officially recommend the use of ``pip`` as the default installer for Python packages, rather than the current approach of recommending the direct invocation of the ``setup.py install`` command. However, to avoid recommending a tool that CPython does not provide, it is further proposed that the `pip`_ package manager be made available by default when installing CPython 3.4 or later and when creating virtual environments using the standard library's ``venv`` module via the ``pyvenv`` command line utility. To support that end, this PEP proposes the inclusion of an ``ensurepip`` bootstrapping module in Python 3.4, as well as automatic invocation of that module from ``pyvenv`` and changes to the way Python installed scripts are handled on Windows. Using a bootstrap module rather than providing ``pip`` directly helps to clearly demarcate development responsibilities, and to avoid inadvertently downgrading ``pip`` when updating CPython. To provide clear guidance for new users of Python that may not be starting with the latest release, this PEP also proposes that the "Installing Python Modules" guides in Python 2.7 and 3.3 be updated to recommend installing and using ``pip``, rather than invoking ``distutils`` directly. It does *not* propose backporting any of the code changes that are being proposed for Python 3.4. Finally, the PEP also strongly recommends that CPython redistributors and other Python implementations ensure that ``pip`` is available by default, or at the very least, explicitly document the fact that it is not included. This PEP does *not* propose making pip (or any dependencies) directly available as part of the standard library. Instead, pip will be a bundled application provided along with CPython for the convenience of Python users, but subject to its own development life cycle and able to be upgraded independently of the core interpreter and standard library. Explicit bootstrapping mechanism ================================ An additional module called ``ensurepip`` will be added to the standard library whose purpose is to install pip and any of its dependencies into the appropriate location (most commonly site-packages). It will expose a callable named ``bootstrap()`` as well as offer direct execution via ``python -m ensurepip``. The bootstrap will *not* contact PyPI, but instead rely on a private copy of pip stored inside the standard library. Accordingly, only options related to the installation location will be supported (``--user``, ``--root``, etc). It is considered desirable that users be strongly encouraged to use the latest available version of ``pip``, in order to take advantage of the ongoing efforts to improve the security of the PyPI based ecosystem, as well as benefiting from the efforts to improve the speed, reliability and flexibility of that ecosystem. In order to satisfy this goal of providing the most recent version of ``pip`` by default, the private copy of ``pip`` will be updated in CPython maintenance releases, which should align well with the 6-month cycle used for new ``pip`` releases. Security considerations ----------------------- The design in this PEP has been deliberately chosen to avoid making any significant changes to the trust model of the CPython installers for end users that do not subsequently make use of ``pip``. The installers will contain all the components of a fully functioning version of Python, including the ``pip`` installer. The installation process will *not* require network access, and will *not* rely on trusting the security of the network connection established between ``pip`` and the Python package index. Only users that choose to use ``pip`` directly will need to pay attention to any PyPI related security considerations. Reliability considerations -------------------------- By including the bootstrap as part of the standard library (rather than solely as a feature of the binary installers), the correct operation of the bootstrap command can be easily tested using the existing CPython buildbot infrastructure rather than adding significantly to the testing burden for the installers themselves. Implementation strategy ----------------------- To ensure there is no need for network access when installing Python or creating virtual environments, the ``ensurepip`` module will, as an implementation detail, include a complete private copy of pip and its dependencies which will be used to extract pip and install it into the target environment. It is important to stress that this private copy of pip is *only* an implementation detail and it should *not* be relied on or assumed to exist beyond the public capabilities exposed through the ``ensurepip`` module (and indirectly through ``venv``). There is not yet a reference ``ensurepip`` implementation. The existing ``get-pip.py`` bootstrap script demonstrates an earlier variation of the general concept, but the standard library version would take advantage of the improved distribution capabilities offered by the CPython installers to include private copies of ``pip`` and ``setuptools`` as wheel files (rather than as embedded base64 encoded data), and would not try to contact PyPI (instead installing directly from the private wheel files. Rather than including separate code to handle the bootstrapping, the ``ensurepip`` module will manipulate ``sys.path`` appropriately to allow the wheel files to be used to install themselves, either into the current Python installation or into a virtual environment (as determined by the options passed to the bootstrap command). It is proposed that the implementation be carried out in five separate steps (all steps after the first two are independent of each other and can be carried out in any order): * the first step would update the "Installing Python Modules" documentation to recommend the use of ``pip`` and reference the ``pip`` team's instructions for downloading and installing it. This change would be applied to Python 2.7, 3.3, and 3.4. * the ``ensurepip`` module and the private copies of the most recently released versions of pip and setuptools would be added to Python 3.4 and the 3.4 "Installing Python Modules" documentation updated accordingly. * the CPython Windows installer would be updated to offer the new ``pip`` installation option for Python 3.4. * the CPython Mac OS X installer would be updated to offer the new ``pip`` installation option for Python 3.4. * the ``venv`` module and ``pyvenv`` command would be updated to make use of ``ensurepip`` in Python 3.4 * the PATH handling and ``sysconfig`` directory layout on Windows would be updated for Python 3.4+ Proposed CLI ------------ The proposed CLI is based on a subset of the existing ``pip install`` options:: Usage: python -m ensurepip [options] General Options: -h, --help Show help. -v, --verbose Give more output. Option is additive, and can be used up to 3 times. -V, --version Show the pip version that would be extracted and exit. -q, --quiet Give less output. Installation Options: -U, --upgrade Upgrade pip and dependencies, even if already installed --user Install using the user scheme. --root <dir> Install everything relative to this alternate root directory. In most cases, end users won't need to use this CLI directly, as ``pip`` should have been installed automatically when installing Python or when creating a virtual environment. However, it is formally documented as a public interface to support at least these known use cases: * Windows and Mac OS X installations where the "Install pip" option was *not* chosen during installation * any installation where the user previously ran "pip uninstall pip" Users that want to retrieve the latest version from PyPI, or otherwise need more flexibility, can then invoke the extracted ``pip`` appropriately. Proposed module API ------------------- The proposed ``ensurepip`` module API consists of the following two functions:: def version(): """ Returns a string specifying the bundled version of pip. """ def bootstrap(root=None, upgrade=False, user=False, verbosity=0): """ Bootstrap pip into the current Python installation (or the given root directory). """ Invocation from the CPython installers -------------------------------------- The CPython Windows and Mac OS X installers will each gain a new option: * Install pip (the default Python package management utility)? This option will be checked by default. If the option is checked, then the installer will invoke the following command with the just installed Python:: python -m ensurepip --upgrade This ensures that, by default, installing or updating CPython will ensure that the installed version of pip is at least as recent as the one included with that version of CPython. If a newer version of pip has already been installed then ``python -m ensurepip --upgrade`` will simply return without doing anything. Installing from source ---------------------- Just as the prebuilt binary installers will be updated to run ``python -m ensurepip`` by default, a similar change will be made to the ``make install`` and ``make altinstall`` commands of the source distribution. The directory settings in the ``sysconfig`` module should ensure the ``pip`` components are automatically installed to the expected locations. ``ensurepip`` itself (including the private copy of ``pip`` and its dependencies) will always be installed normally (as it is a regular part of the standard library), but an option will be provided to skip the invocation of ``ensurepip``. This means that even installing from source will provide ``pip`` by default, but redistributors provide ``pip`` by other means (or not providing it at all) will still be able to opt out of installing it using ``ensurepip``. Changes to virtual environments ------------------------------- Python 3.3 included a standard library approach to virtual Python environments through the ``venv`` module. Since its release it has become clear that very few users have been willing to use this feature directly, in part due to the lack of an installer present by default inside of the virtual environment. They have instead opted to continue using the ``virtualenv`` package which *does* include pip installed by default. To make the ``venv`` more useful to users it will be modified to issue the pip bootstrap by default inside of the new environment while creating it. This will allow people the same convenience inside of the virtual environment as this PEP provides outside of it as well as bringing the ``venv`` module closer to feature parity with the external ``virtualenv`` package, making it a more suitable replacement. To handle cases where a user does not wish to have pip bootstrapped into their virtual environment a ``--without-pip`` option will be added. The ``venv.EnvBuilder`` and ``venv.create`` APIs will be updated to accept one new parameter: ``with_pip`` (defaulting to ``False``). The new default for the module API is chosen for backwards compatibility with the current behaviour (as it is assumed that most invocation of the ``venv`` module happens through third part tools that likely will not want ``pip`` installed without explicitly requesting it), while the default for the command line interface is chosen to try to ensure ``pip`` is available in most virtual environments without additional action on the part of the end user. As this change will only benefit Python 3.4 and later versions, the third-party ``virtualenv`` project will still be needed to obtain a consistent cross-version experience in Python 3.3 and 2.7. Documentation ------------- The "Installing Python Modules" section of the standard library documentation in Python 2.7, 3.3 and 3.4 will be updated to recommend the use of the ``pip`` installer, either provided by default in Python 3.4 or retrieved and installed by the user in Python 2.7 or 3.3. It will give a brief description of the most common commands and options, but delegate to the externally maintained ``pip`` documentation for the full details. In Python 3.4, the ``pyvenv`` and ``venv`` documentation will also be updated to reference the revised module installation guide. The existing content of the module installation guide will be retained in all versions, but under a new "Invoking distutils directly" subsection. Bundling CA certificates with CPython ------------------------------------- The ``ensurepip`` implementation will include the ``pip`` CA bundle along with the rest of ``pip``. This means CPython effectively includes a CA bundle that is used solely by ``pip`` after it has been extracted. This is considered preferable to relying solely on the system certificate stores, as it ensures that ``pip`` will behave the same across all supported versions of Python, even those prior to Python 3.4 that cannot access the system certificate store on Windows. Automatic installation of setuptools ------------------------------------ ``pip`` currently depends on ``setuptools`` to handle metadata generation during the build process, along with some other features. While work is ongoing to reduce or eliminate this dependency, it is not clear if that work will be complete for pip 1.5 (which is the version likely to be current when Python 3.4.0 is released). This PEP proposes that, if pip still requires it as a dependency, ``ensurepip`` will include a private copy of ``setuptools`` (in addition to the private copy of ``ensurepip``). ``python -m ensurepip`` will then install the private copy in addition to installing ``pip`` itself. However, this behavior is officially considered an implementation detail. Other projects which explicitly require ``setuptools`` must still provide an appropriate dependency declaration, rather than assuming ``setuptools`` will always be installed alongside ``pip``. Once pip is able to run ``pip install --upgrade pip`` without needing ``setuptools`` installed first, then the private copy of ``setuptools`` will be removed from ``ensurepip`` in subsequent CPython releases. As long as setuptools is needed, it will be a completely unmodified copy of the latest upstream setuptools release, including the ``easy_install`` script if the upstream setuptools continues to include it. The installation of ``easy_install`` along with ``pip`` isn't considered desirable, but installing a broken setuptools would be worse. This problem will naturally resolve itself once the ``pip`` developers have managed to eliminate their dependency on ``setuptools`` and the private copy of ``setuptools`` can be removed entirely from CPython. Updating the private copy of pip -------------------------------- In order to keep up with evolutions in packaging as well as providing users with as recent version a possible the ``ensurepip`` module will be regularly updated to the latest versions of everything it bootstraps. After each new ``pip`` release, and again during the preparation for any release of Python (including feature releases), a script, provided as part of the implementation for this PEP, will be run to ensure the private copies stored in the CPython source repository have been updated to the latest versions. Updating the ensurepip module API and CLI ----------------------------------------- Like ``venv`` and ``pyvenv``, the ``ensurepip`` module API and CLI will be governed by the normal rules for the standard library: no new features are permitted in maintenance releases. However, the embedded components may be updated as noted above, so the extracted ``pip`` may offer additional functionality in maintenance releases. Uninstallation ============== No changes are proposed to the CPython uninstallation process by this PEP. The bootstrapped pip will be installed the same way as any other pip installed packages, and will be handled in the same way as any other post-install additions to the Python environment. At least on Windows, that means the bootstrapped files will be left behind after uninstallation, since those files won't be associated with the Python MSI installer. While the case can be made for the CPython installers clearing out these directories automatically, changing that behaviour is considered outside the scope of this PEP. Script Execution on Windows =========================== While the Windows installer was updated in Python 3.3 to optionally make ``python`` available on the PATH, no such change was made to include the Scripts directory. Independently of this PEP, a proposal has also been made to rename the ``Tools\Scripts`` subdirectory to ``bin`` in order to improve consistency with the typical script installation directory names on \*nix systems. Accordingly, in addition to adding the option to extract and install ``pip`` during installation, this PEP proposes that the Windows installer (and ``sysconfig``) in Python 3.4 and later be updated to: - install scripts to PythonXY\bin rather than PythonXY\Tools\Scripts - add PythonXY\bin to the Windows PATH (in addition to PythonXY) when the PATH modification option is enabled during installation Note that these changes will only be available in Python 3.4 and later. This means that, for Python 3.3, the most reliable way to invoke pip globally on Windows (without tinkering manually with PATH) will still remain ``py -m pip`` (or ``py -3 -m pip`` to select the Python 3 version if both Python 2 and 3 are installed) rather than simply calling ``pip``. This works because Python 3.3 provides the Python Launcher for Windows (and the associated ``py`` command) by default. For Python 2.7 and 3.2, the most reliable mechanism will be to install the Python Launcher for Windows using the standalone installer and then use ``py -m pip`` as noted above. Adding the scripts directory to the system PATH will mean that ``pip`` works reliably in the "only one Python installation on the system PATH" case, with ``py -m pip``, ``pipX``, or ``pipX.Y`` needed only to select a non-default version in the parallel installation case (and outside a virtual environment). This change should also make the ``pyvenv`` command substantially easier to invoke on Windows, along with all scripts installed by ``pip``, ``easy_install`` and similar tools. While the script invocations on recent versions of Python will run through the Python launcher for Windows, this shouldn't cause any issues, as long as the Python files in the Scripts directory correctly specify a Python version in their shebang line or have an adjacent Windows executable (as ``easy_install`` and ``pip`` do). Recommendations for Downstream Distributors =========================================== A common source of Python installations are through downstream distributors such as the various Linux Distributions [#ubuntu]_ [#debian]_ [#fedora]_, OSX package managers [#homebrew]_ [#macports]_ [#fink]_, and commercial Python redistributors [#ContinuumIO]_ [#ActiveState]_ [#Enthought]_. In order to provide a consistent, user-friendly experience to all users of Python regardless of how they obtained Python this PEP recommends and asks that downstream distributors: * Ensure that whenever Python is installed ``pip`` is either installed or is otherwise made readily available to end users. * For redistributors using binary installers, this may take the form of optionally executing the ``ensurepip`` bootstrap during installation, similar to the CPython installers. * For redistributors using package management systems, it may take the form of separate packages with dependencies on each other so that installing the Python package installs the pip package and installing the pip package installs the Python package. * Another reasonable way to implement this is to package pip separately but ensure that there is some sort of global hook that will recommend installing the separate pip package when a user executes ``pip`` without it being installed. Systems that choose this option should ensure that the ``ensurepip`` module still installs pip directly when invoked inside a virtual environment, but may modify the module in the system Python installation to redirect to the platform provided mechanism when installing ``pip`` globally. * Even if pip is made available globally by other means, do not remove the ``ensurepip`` module in Python 3.4 or later. * ``ensurepip`` will be required for automatic installation of pip into virtual environments by the ``venv`` module. * This is similar to the existing ``virtualenv`` package for which many downstream distributors have already made exception to the common "debundling" policy. * This does mean that if ``pip`` needs to be updated due to a security issue, so does the private copy in the ``ensurepip`` bootstrap module * However, altering the private copy of pip to remove the embedded CA certificate bundle and rely on the system CA bundle instead is a reasonable change. * Ensure that all features of this PEP continue to work with any modifications made to the redistributed version of Python. * Checking the version of pip that will be bootstrapped using ``python -m ensurepip --version`` or ``ensurepip.version()``. * Installation of pip into a global or virtual python environment using ``python -m ensurepip`` or ``ensurepip.bootstrap()``. * ``pip install --upgrade pip`` in a global installation should not affect any already created virtual environments (but is permitted to affect future virtual environments, even though it will not do so when using the standard implementation of ``ensurepip``). * ``pip install --upgrade pip`` in a virtual environment should not affect the global installation. * Migrate build systems to utilize `pip`_ and `Wheel`_ wherever feasible and avoid directly invoking ``setup.py``. * This will help ensure a smoother and more timely migration to improved metadata formats as the Python packaging ecosystem continues to evolve. In the event that a Python redistributor chooses *not* to follow these recommendations, we request that they explicitly document this fact and provide their users with suitable guidance on translating upstream ``pip`` based installation instructions into something appropriate for the platform. Other Python implementations are also encouraged to follow these guidelines where applicable. Policies & Governance ===================== The maintainers of the bootstrapped software and the CPython core team will work together in order to address the needs of both. The bootstrapped software will still remain external to CPython and this PEP does not include CPython subsuming the development responsibilities or design decisions of the bootstrapped software. This PEP aims to decrease the burden on end users wanting to use third-party packages and the decisions inside it are pragmatic ones that represent the trust that the Python community has already placed in the Python Packaging Authority as the authors and maintainers of ``pip``, ``setuptools``, PyPI, ``virtualenv`` and other related projects. Backwards Compatibility ----------------------- The public API and CLI of the ``ensurepip`` module itself will fall under the typical backwards compatibility policy of Python for its standard library. The externally developed software that this PEP bundles does not. Most importantly, this means that the bootstrapped version of pip may gain new features in CPython maintenance releases, and pip continues to operate on its own 6 month release cycle rather than CPython's 18-24 month cycle. Security Releases ----------------- Any security update that affects the ``ensurepip`` module will be shared prior to release with the Python Security Response Team (security(a)python.org). The PSRT will then decide if the reported issue warrants a security release of CPython with an updated private copy of ``pip``. Licensing --------- ``pip`` is currently licensed as 1 Clause BSD, and it contains code taken from other projects. Additionally this PEP will include setuptools until such time as pip no longer requires it. The licenses for these appear in the table below. ================= ============ Project License ================= ============ requests Apache 2.0 six 1 Clause BSD html5lib 1 Clause BSD distlib PSF colorama 3 Clause BSD Mozilla CA Bundle LGPL setuptools PSF ================= ============ All of these licenses should be compatible with the PSF license. Additionally it is unclear if a CA Bundle is copyrightable material and thus if it needs or can be licensed at all. Appendix: Rejected Proposals ============================ Including ensurepip in Python 2.7, and 3.3 ------------------------------------------ Earlier versions of this PEP made the case that the challenges of getting ``pip`` bootstrapped for new users posed a significant enough barrier to Python's future growth that it justified adding ``ensurepip`` as a new feature in the upcoming Python 2.7 and 3.3 maintenance releases. While the proposal to provide ``pip`` with Python 3.4 was universally popular, this part of the proposal was highly controversial and ultimately `rejected by MvL as BDFL-Delegate <https://mail.python.org/pipermail/python-dev/2013-September/129091.html>`__. Accordingly, the proposal to backport ``ensurepip`` to Python 2.7 and 3.3 has been removed from this PEP in favour of creating a Windows installer for ``pip`` and a possible future PEP suggesting creation of an aggregate installer for Python 2.7 that combines CPython 2.7, ``pip`` and the Python Launcher for Windows. Automatically contacting PyPI when bootstrapping pip ---------------------------------------------------- Earlier versions of this PEP called the bootstrapping module ``getpip`` and defaulted to downloading and installing ``pip`` from PyPI, with the private copy used only as a fallback option or when explicitly requested. This resulted in several complex edge cases, along with difficulties in defining a clean API and CLI for the bootstrap module. It also significantly altered the default trust model for the binary installers published on python.org, as end users would need to explicitly *opt-out* of trusting the security of the PyPI ecosystem (rather than opting in to it by explicitly invoking ``pip`` following installation). As a result, the PEP was simplified to the current design, where the bootstrapping *always* uses the private copy of ``pip``. Contacting PyPI is now always an explicit separate step, with direct access to the full pip interface. Removing the implicit attempt to access PyPI also made it feasible to invoke ``ensurepip`` by default when installing from a custom source build. Implicit bootstrap ------------------ `PEP439`_, the predecessor for this PEP, proposes its own solution. Its solution involves shipping a fake ``pip`` command that when executed would implicitly bootstrap and install pip if it does not already exist. This has been rejected because it is too "magical". It hides from the end user when exactly the pip command will be installed or that it is being installed at all. It also does not provide any recommendations or considerations towards downstream packagers who wish to manage the globally installed pip through the mechanisms typical for their system. The implicit bootstrap mechanism also ran into possible permissions issues, if a user inadvertently attempted to bootstrap pip without write access to the appropriate installation directories. Including pip directly in the standard library ---------------------------------------------- Similar to this PEP is the proposal of just including pip in the standard library. This would ensure that Python always includes pip and fixes all of the end user facing problems with not having pip present by default. This has been rejected because we've learned, through the inclusion and history of ``distutils`` in the standard library, that losing the ability to update the packaging tools independently can leave the tooling in a state of constant limbo. Making it unable to ever reasonably evolve in a time frame that actually affects users as any new features will not be available to the general population for *years*. Allowing the packaging tools to progress separately from the Python release and adoption schedules allows the improvements to be used by *all* members of the Python community and not just those able to live on the bleeding edge of Python releases. There have also been issues in the past with the "dual maintenance" problem if a project continues to be maintained externally while *also* having a fork maintained in the standard library. Since external maintenance of ``pip`` will always be needed to support earlier Python versions, the proposed bootstrapping mechanism will becoming the explicit responsibility of the CPython core developers (assisted by the pip developers), while pip issues reported to the CPython tracker will be migrated to the pip issue tracker. There will no doubt still be some user confusion over which tracker to use, but hopefully less than has been seen historically when including complete public copies of third-party projects in the standard library. The approach described in this PEP also avoids some technical issues related to handling CPython maintenance updates when pip has been independently updated to a more recent version. The proposed pip-based bootstrapping mechanism handles that automatically, since pip and the system installer never get into a fight about who owns the pip installation (it is always managed through pip, either directly, or indirectly via the ``ensurepip`` bootstrap module). Finally, the separate bootstrapping step means it is also easy to avoid installing ``pip`` at all if end users so desire. This is often the case if integrators are using system packages to handle installation of components written in multiple languages using a common set of tools. Defaulting to --user installation --------------------------------- Some consideration was given to bootstrapping pip into the per-user site-packages directory by default. However, this behavior would be surprising (as it differs from the default behavior of pip itself) and is also not currently considered reliable (there are some edge cases which are not handled correctly when pip is installed into the user site-packages directory rather than the system site-packages). .. _Wheel: http://www.python.org/dev/peps/pep-0427/ .. _pip: http://www.pip-installer.org .. _setuptools: https://pypi.python.org/pypi/setuptools .. _PEP439: http://www.python.org/dev/peps/pep-0439/ References ========== .. [1] Discussion thread 1 (distutils-sig) (https://mail.python.org/pipermail/distutils-sig/2013-August/022529.html) .. [2] Discussion thread 2 (distutils-sig) (https://mail.python.org/pipermail/distutils-sig/2013-September/022702.html) .. [3] Discussion thread 3 (python-dev) (https://mail.python.org/pipermail/python-dev/2013-September/128723.html) .. [4] Discussion thread 4 (python-dev) (https://mail.python.org/pipermail/python-dev/2013-September/128780.html) .. [5] Discussion thread 5 (python-dev) (https://mail.python.org/pipermail/python-dev/2013-September/128894.html) .. [#ubuntu] `Ubuntu <http://www.ubuntu.com/>` .. [#debian] `Debian <http://www.debian.org>` .. [#fedora] `Fedora <https://fedoraproject.org/>` .. [#homebrew] `Homebrew <http://brew.sh/>` .. [#macports] `MacPorts <http://macports.org>` .. [#fink] `Fink <http://finkproject.org>` .. [#ContinuumIO] `Anaconda <https://store.continuum.io/cshop/anaconda/>` .. [#ActiveState] `ActivePython <http://www.activestate.com/activepython>` .. [#Enthought] `Enthought Canopy <https://www.enthought.com/products/canopy/>` Copyright ========= This document has been placed in the public domain. -- Nick Coghlan | ncoghlan(a)gmail.com | Brisbane, Australia

2 3

C extension import time
by Stefan Krah 14 Oct '13

14 Oct '13

Hi, recently there has been some talk about reducing import times. It seems that the current import strategy for C extensions (i.e. importing the extension at the bottom of the .py file) is quite slow: ==================== import sys for i in range(10000): import decimal del sys.modules('decimal') del sys.modules('_decimal') ==================== With the regular decimal.py the script is about 9 times slower than running it with a modified decimal.py that just contains "from _decimal import *". I don't know enough about the import mechanism to say whether this is a particularly good benchmark. If it is, perhaps we should devise another strategy for loading C extensions. Stefan Krah

10 24

GC pauses in CPython
by Maciej Fijalkowski 14 Oct '13

14 Oct '13

HI I'm working on an incremental GC for PyPy. How do I measure GC pauses in CPython? (that is, the circular reference searching stuff) Cheers, fijal

4 8

Make str/bytes hash algorithm pluggable?
by Christian Heimes 14 Oct '13

14 Oct '13

Hi, some of you may have seen that I'm working on a PEP for a new hash API and new algorithms for hashing of bytes and str. The PEP has three major aspects. It introduces DJB's SipHash as secure hash algorithm, chances the hash API to process blocks of data instead characters and it adds an API to make the algorithm pluggable. A draft is already available [1]. Now I got some negative feedback on the 'pluggable' aspect of the new PEP on Twitter [2]. I like to get feedback from you before I finalize the PEP. The PEP proposes a pluggable hash API for a couple of reasons. I like to give users of Python a chance to replace a secure hash algorithm with a faster hash algorithm. SipHash is about as fast as FNV for common cases as our implementation of FNV process only 8 to 32 bits per cycle instead of 32 or 64. I haven't actually benchmarked how a faster hash algorithm affects the a real program, though ... I also like to make it easier to replace the hash algorithm with a different one in case a vulnerability is found. With the new API vendors and embedders have an easy and clean way to use their own hash implementation or an optimized version that is more suitable for their platform, too. For example a mobile phone vendor could provide an optimized implementation with ARM NEON intrinsics. On which level should Python support a pluggable hash algorithm? 1) Compile time option: The hash code is compiled into Python's core. Embedders have to recompile Python with different options to replace the function. 2) Library option: A hash algorithm can be added and one avaible hash algorithm can be set before Py_Initialize() is called for the first time. The approach gives embedders the chance the set their own algorithm without recompiling Python. 3) Startup options: Like 2) plus an additional environment variable and command line argument to select an algorithm. With a startup option users can select a different algorithm themselves. Christian [1] http://www.python.org/dev/peps/pep-0456/ [2] https://twitter.com/EDEADLK/status/385572395777818624

13 31

Right place for PBKDF2 wrapper
by Christian Heimes 14 Oct '13

14 Oct '13

Hi, I have written a interface to OpenSSL's PKCS5_PBKDF2_HMAC() function. It implements PKCS#5's password based key derivation function 2 with HMAC as pseudo-random function. It supports any digest that is supported by OpenSSL, e.g. SHA-1, SHA-256 and SHA-512. It's a low level inteface that takes the digest as unicode name, password and salt as bytes/buffer, keylen and rounds as int. I'd like to add the feature to Python 3.4. Now I'm looking for a good place to put it and some high level functions. In the future I like to add scrypt and bcrypt key stretching and key derivation functions, too. What's a good place for them?? * add a new ``kdf`` module (key derivation function) * add PBKDF2 to ``hashlib`` * make ``hashlib`` a package and add PBKDF2 to a new ``hashlib.kdf`` module * make ``hashlib`` a package and add PBKDF2 to a new ``hashlib.pbkdf2`` module * make ``crypt`` work under Windows and add PKBDF2 to it The patch is available at http://bugs.python.org/issue18582 Regards, Christian

7 12

__objclass__ documentation
by Ethan Furman 13 Oct '13

13 Oct '13

It was pointed in Issue16938[1] that __objclass__ is not documented anywhere. Is the following an appropriate description? (in Doc/reference/datamodel.rst in user-defined functions) +-------------------------+-------------------------------+-----------+ | :attr:`__objclass__` | The class this object belongs | | | | to; useful when the object is | | | | a descriptor, or a virtual or | | | | dynamic class attribute, and | | | | it's __class__ attribute does | | | | not match the class it is | | | | associated with, or it is not | | | | in that class' ``__dict__``. | | +-------------------------+-------------------------------+-----------+ [1] http://bugs.python.org/issue16938 -- ~Ethan~

3 4

inspect() and dir()
by Ethan Furman 12 Oct '13

12 Oct '13

Greetings, Currently, inspect() is dependent on dir(). Now that we can override what dir() returns on a class by class basis, we are seeing the side-effect of (possibly) incomplete inspect results, with also leads to (possibly) incomplete help(). I would think we want inspect to be more thorough, and return whatever it finds on the object, as well as whatever an abbreviated dir() might return. Thoughts? -- ~Ethan~

4 10

Support for marking limited API elements in C API docs
by Georg Brandl 12 Oct '13

12 Oct '13

Hi, in light of the recent thread about PEPs not forming part of the docs, I've just pushed a change that allows to document C API elements not part of the limited API as such. It is done like this: .. c:function:: int _PyTuple_Resize(PyObject **p, Py_ssize_t newsize) :notlimited: I have not yet begun adding these to the documents; if someone wants to help with this I am glad for volunteers. cheers, Georg

5 13

PEPs shouldn't be considered docs
by Ned Batchelder 12 Oct '13

12 Oct '13

I wanted to teach a co-worker about "from __future__ import absolute_import" today, so I thought I'd point them at the docs. The page for "__future__" starts with a bunch of internal details that almost no one needs to know. There's a table at the end that mentions the actual importable names, including "absolute_import", but says nothing about then except to link to a PEP. The absolute imports PEP has no simple description of what it does. Like many PEPs, it's mostly a summary of the debate around the design of the feature. The closest the PEP comes to describing the behavior of "absolute_import" is this parenthetical: For the second problem, it is proposed that all import statements be absolute by default (searching sys.path only) with special syntax (leading dots) for accessing package-relative imports. And notice: that sentence describes it as a "proposal." I'd like to suggest that we not consider PEPs to be documentation. If a PEP has a good succinct description of how something works, then let's copy that text into the documentation at an appropriate place. If a PEP doesn't have such a description, then all the more reason not to send readers there. A link to the PEP is appropriate as a "see also" in the docs, but we shouldn't pretend that a PEP addresses the needs of people new to the feature. --Ned.

7 8

Re: [Python-Dev] [Python-checkins] cpython: Issue #19209: Remove import of copyreg from the os module to speed up
by Nick Coghlan 12 Oct '13

12 Oct '13

On 12 Oct 2013 09:27, "christian.heimes" <python-checkins(a)python.org> wrote: > > http://hg.python.org/cpython/rev/29c4a6a11e76 > changeset: 86216:29c4a6a11e76 > user: Christian Heimes <christian(a)cheimes.de> > date: Sat Oct 12 01:27:08 2013 +0200 > summary: > Issue #19209: Remove import of copyreg from the os module to speed up > interpreter startup. stat_result and statvfs_result are now hard-coded to > reside in the os module. > The patch is based on Victor Stinner's patch. > > files: > Lib/os.py | 27 --------------------------- > Lib/test/test_os.py | 22 ++++++++++++++++++++++ > Lib/test/test_site.py | 5 ++++- > Misc/NEWS | 4 ++++ > Modules/posixmodule.c | 4 ++-- > 5 files changed, 32 insertions(+), 30 deletions(-) > > > diff --git a/Lib/os.py b/Lib/os.py > --- a/Lib/os.py > +++ b/Lib/os.py > @@ -945,33 +945,6 @@ > __all__.extend(["spawnlp", "spawnlpe"]) > > > -import copyreg as _copyreg > - > -def _make_stat_result(tup, dict): > - return stat_result(tup, dict) > - > -def _pickle_stat_result(sr): > - (type, args) = sr.__reduce__() > - return (_make_stat_result, args) > - > -try: > - _copyreg.pickle(stat_result, _pickle_stat_result, _make_stat_result) > -except NameError: # stat_result may not exist > - pass > - > -def _make_statvfs_result(tup, dict): > - return statvfs_result(tup, dict) > - > -def _pickle_statvfs_result(sr): > - (type, args) = sr.__reduce__() > - return (_make_statvfs_result, args) > - > -try: > - _copyreg.pickle(statvfs_result, _pickle_statvfs_result, > - _make_statvfs_result) > -except NameError: # statvfs_result may not exist > - pass > - > # Supply os.popen() > def popen(cmd, mode="r", buffering=-1): > if not isinstance(cmd, str): > diff --git a/Lib/test/test_os.py b/Lib/test/test_os.py > --- a/Lib/test/test_os.py > +++ b/Lib/test/test_os.py > @@ -26,6 +26,7 @@ > import codecs > import decimal > import fractions > +import pickle > try: > import threading > except ImportError: > @@ -264,6 +265,13 @@ > warnings.simplefilter("ignore", DeprecationWarning) > self.check_stat_attributes(fname) > > + def test_stat_result_pickle(self): > + result = os.stat(self.fname) > + p = pickle.dumps(result) > + self.assertIn(b'\x03cos\nstat_result\n', p) > + unpickled = pickle.loads(p) > + self.assertEqual(result, unpickled) > + > def test_statvfs_attributes(self): > if not hasattr(os, "statvfs"): > return > @@ -310,6 +318,20 @@ > except TypeError: > pass > > + @unittest.skipUnless(hasattr(os, 'statvfs'), > + "need os.statvfs()") > + def test_statvfs_result_pickle(self): > + try: > + result = os.statvfs(self.fname) > + except OSError as e: > + # On AtheOS, glibc always returns ENOSYS > + if e.errno == errno.ENOSYS: > + return > + p = pickle.dumps(result) > + self.assertIn(b'\x03cos\nstatvfs_result\n', p) > + unpickled = pickle.loads(p) > + self.assertEqual(result, unpickled) > + > def test_utime_dir(self): > delta = 1000000 > st = os.stat(support.TESTFN) > diff --git a/Lib/test/test_site.py b/Lib/test/test_site.py > --- a/Lib/test/test_site.py > +++ b/Lib/test/test_site.py > @@ -431,10 +431,13 @@ > modules = eval(stdout.decode('utf-8')) > self.assertIn('site', modules) > > + # http://bugs.python.org/issue19205 > re_mods = {'re', '_sre', 'sre_compile', 'sre_constants', 'sre_parse'} > self.assertFalse(modules.intersection(re_mods)) > - > + # http://bugs.python.org/issue9548 > self.assertNotIn('locale', modules) This looks like it snuck in from a separate patch. Cheers, Nick. > + # http://bugs.python.org/issue19209 > + self.assertNotIn('copyreg', modules) > > > if __name__ == "__main__": > diff --git a/Misc/NEWS b/Misc/NEWS > --- a/Misc/NEWS > +++ b/Misc/NEWS > @@ -36,6 +36,10 @@ > Library > ------- > > +- Issue #19209: Remove import of copyreg from the os module to speed up > + interpreter startup. stat_result and statvfs_result are now hard-coded to > + reside in the os module. > + > - Issue #19205: Don't import the 're' module in site and sysconfig module to > to speed up interpreter start. > > diff --git a/Modules/posixmodule.c b/Modules/posixmodule.c > --- a/Modules/posixmodule.c > +++ b/Modules/posixmodule.c > @@ -11974,7 +11974,7 @@ > return NULL; > #endif > > - stat_result_desc.name = MODNAME ".stat_result"; > + stat_result_desc.name = "os.stat_result"; /* see issue #19209 */ > stat_result_desc.fields[7].name = PyStructSequence_UnnamedField; > stat_result_desc.fields[8].name = PyStructSequence_UnnamedField; > stat_result_desc.fields[9].name = PyStructSequence_UnnamedField; > @@ -11983,7 +11983,7 @@ > structseq_new = StatResultType.tp_new; > StatResultType.tp_new = statresult_new; > > - statvfs_result_desc.name = MODNAME ".statvfs_result"; > + statvfs_result_desc.name = "os.statvfs_result"; /* see issue #19209 */ > if (PyStructSequence_InitType2(&StatVFSResultType, > &statvfs_result_desc) < 0) > return NULL; > > -- > Repository URL: http://hg.python.org/cpython > > _______________________________________________ > Python-checkins mailing list > Python-checkins(a)python.org > https://mail.python.org/mailman/listinfo/python-checkins >

2 1

Python-Dev October 2013