[Linux-SIG] Draft PEP: Graceful cooperation between external and Python package managers

Hi Linux-SIG, At the recent PyCon US "Linux in Distros" sprint, we wrote a draft PEP about making external package managers like apt/dnf/etc. and Python-specific package managers like pip play more nicely together. This includes the "sudo pip install" problem, but it's a little more general than that. The short version is it has two recommendations: 1. When a distro indicates it's managing a Python installation, tools like pip should only install into a virtualenv, by default (with a way to override it), and show an error message that the distro can customize. 2. Distros should have two site-packages directories, one for distro-packaged files and one for local-sysadmin-installed files (e.g., /usr/lib/python3.x/site-packags vs. /usr/local/lib/python3.x/site-packages), and tools like pip should only create, delete, or modify files in the latter directory. The draft PEP is on my GitHub at https://github.com/geofft/peps branch "pip-only-in-virtualenv" in the file pep-9999.rst. I've uploaded a rendered copy to https://kerberos.club/tmp/pep-9999.html and included the text of the PEP below. Since Linux-SIG has a bunch of packagers and users of distro Python, we figured we'd run this by all of you to make sure this idea is broadly reasonable before formally bringing it to discuss.python.org. Feedback is welcome in any form - either replies here, or comments on GitHub (there's a pull request at https://github.com/geofft/peps/pull/3 if you prefer that interface). At this stage I'm mostly interested in feedback on the general idea and the high-level changes to tool semantics - we've tried to handle a good number of use cases in hopefully-reasonable ways but we definitely might have missed something. Thanks! PEP: 9999 Title: Graceful cooperation between external and Python package managers Author: Geoffrey Thomas , Matthias Klose , Filipe Laíns , Donald Stufft , Tzu-Ping Chung , Stefano Rivera , Elana Hashman , Pradyun Gedam Discussions-To: TODO discourse Status: Draft Type: Informational Content-Type: text/x-rst Created: 2021-05-18 Post-History: Abstract ======== A long-standing practical problem for Python users has been conflicts between OS package managers and Python-specific package management tools like pip. These conflicts include both Python-level API incompatibilities and conflicts over file ownership. Historically, Python-specific package management tools have defaulted to installing packages into an implicit global context. With the standardization and popularity of virtual environments, a better solution for most (but not all) use cases is to use Python-specific package management tools only within a virtual environment. This PEP proposes a mechanism for a Python installation to communicate to tools like pip that its global package installation context is managed by some means external to Python, such as an OS package manager. It specifies that Python-specific package management tools should neither install nor remove packages into the interpreter's global context, by default, and should instead guide the end user towards using a virtual environment. It also standardizes an interpretation of the ``sysconfig`` schemes so that, if a Python-specific package manager is about to install a package in an interpreter-wide context, it can do so in a manner that will avoid conflicting with the external package manager and reduces the risk of breaking software shipped by the external package manager. Terminology =========== A few terms used in this PEP have multiple meanings in the contexts that it spans. For clarity, this PEP uses the following terms in specific ways: distribution A collection of various sorts of software, ideally designed to work properly together, including (in contexts relevant to this document) the Python interpreter itself, software written in Python, and software written in other languages. That is, this is "distribution" in the sense of "Linux distro" or "Berkeley Software Distribution." A distribution can be an operating system (OS) of its own, such as Debian, Fedora, or FreeBSD. It can also be an overlay distribution that installs on top of an existing OS, such as Homebrew or MacPorts. To avoid confusion, this document does not use "distribution" in the sense of a source or binary distribution package of a single piece of Python language software, that is, in the sense of ``setuptools.dist.Distribution`` or "sdist". The provider of a distribution - the team or company that collects and publishes the software and makes any needed modifications - is its **distributor**. package A unit of software that can be installed and used within Python. That is, this refers to what Python-specific packaging tools tend to call a "distribution package" or simply a "distribution"; the colloquial abbreviation "package" is used in the sense of the Python Package Index. This document does not use "package" in the sense of an importable name that contains Python modules, though in many cases, a distribution package consists of a single importable package of the same name. This document generally does not use the term "package" to refer to units of installation by a distribution's package manager (such as ``.deb`` or ``.rpm`` files). When needed, it uses phrasing such as "a distribution's package." (Again, in many cases, a Python package is shipped inside a distribution's package named something like ``python-`` plus the Python package name.) Python-specific package manager A tool for installing, upgrading, and/or removing Python packages in a manner that conforms to Python packaging standards (such as PEP 376 [#PEP-376]_ and PEP 427 [#PEP-427]_). The most popular Python-specific package manager is pip [#pip]_; other examples include the old Easy Install command [#easy-install]_ as well as direct usage of a ``setup.py`` command. (Conda [#conda]_ is a bit of a special case, as the ``conda`` command can install much more than just Python packages, making it more like a distribution package manager in some senses. Since the ``conda`` command generally only operates on Conda-created environments, most of the concerns in this document do not apply to ``conda`` when acting as a Python-specific package manager.) distribution package manager A tool for installing, upgrading, and/or removing a distribution's packages in an installed instance of that distribution, which is capable of installing Python packages as well as non-Python packages, and therefore generally has its own database of installed software unrelated to PEP 376 [#PEP-376]. Examples include ``apt``, ``dpkg``, ``dnf``, ``rpm``, ``pacman``, and ``brew``. The salient feature is that if a package was installed by a distribution package manager, removing or upgrading it in a way that would satisfy a Python-specific package manager will generally leave a distribution package manager in an inconsistent state. This document also uses phrases like "external package manager" or "system's package manager" to refer to a distribution package manager in certain contexts. shadow To shadow an installed Python package is to cause some other package to be preferred for imports without removing any files from the shadowed package. This requires multiple entries on ``sys.path``: if package A 2.0 installs module ``a.py`` in one ``sys.path`` entry, and package A 1.0 installs module ``a.py`` in a later ``sys.path`` entry, then ``import a`` returns the module from the former, and we say that A 2.0 shadows A 1.0. Motivation ========== Thanks to Python's immense popularity, software distributions (by which we mean Linux and other OS distributions as well as overlay distributions like Homebrew and MacPorts) generally ship Python for two purposes: as a software package to be used in its own right by end users, and as a language dependency for other software in the distribution. For example, Fedora and Debian (and their downstream distributions, as well as many others) ship a ``/usr/bin/python3`` binary which provides the ``python3`` command available to end users as well as the ``#!/usr/bin/python3`` shebang for Python-language software included in the distribution. Because there are no official binary releases of Python for Linux/UNIX, almost all Python end users on these OSes use the Python interpreter built and shipped with their distribution. The ``python3`` executable available to the users of the distribution and the ``python3`` executable available as a dependency for other software in the distribution are typically the same binary. This means that if an end user installs a Python package using a tool like ``pip`` outside the context of a virtual environment, that package is visible to Python-language software shipped by the distribution. If the newly-installed package (or one of its dependencies) is a newer, backwards-incompatible version of a package that was installed through the distribution, it may break software shipped by the distribution. This may pose a critical problem for the integrity of distributions, which often have package-management tools that are themselves written in Python. For example, it's possible to unintentionally break Fedora's ``dnf`` command with a ``pip install`` command, making it hard to recover. This applies both to system-wide installs (``sudo pip install``) as well as user home directory installs (``pip install --user``), since packages in either location show up on the ``sys.path`` of ``/usr/bin/python3``. There is a worse problem with system-wide installs: if you attempt to recover from this situation with ``sudo pip uninstall``, you may end up removing packages that are shipped by the system's package manager. In fact, this can even happen if you simply upgrade a package - pip will try to remove the old version of the package, as shipped by the OS. At this point it may not be possible to recover the system to a consistent state using just the software remaining on the system. Over the past many years, a consensus has emerged that the best way to install Python libraries or applications (when not using a distribution's package) is to use a virtual environment. This approach was popularized by the PyPA `virtualenv`_ project, and a simple version of that approach is now available in the Python standard library as ``venv``. Installing a Python package into a virtualenv prevents it from being visible to the unqualified ``/usr/bin/python3`` interpreter and prevents breaking system software. .. _virtualenv: https://virtualenv.pypa.io/en/latest/ In some cases, however, it's useful and intentional to install a Python package from outside of the distribution that influences the behavior of distribution-shipped commands. This is common in the case of software like Sphinx or Ansible which have a mechanism for writing Python-language extensions. A user may want to use their distribution's version of the base software (for reasons of paid support or security updates) but install a small extension from PyPI, and they'd want that extension to be importable by the software in their base system. While this continues to carry the risk of installing a newer version of a dependency than the operating system expects or otherwise negatively affecting the behavior of an application, it does not need to carry the risk of removing files from the operating system. A tool like pip should be able to install packages in some directory on the default ``sys.path``, if specifically requested, without deleting files owned by the system's package manager. Therefore, this PEP proposes two things. First, it proposes **a way for distributors of a Python interpreter to mark that interpreter as having its packages managed by means external to Python**, such that Python-specific tools like pip should not change the installed packages in the interpreter's global ``sys.path`` in any way (add, upgrade/downgrade, or remove) unless specifically overridden. It also provides a means for the distributor to indicate how to use a virtual environment as an alternative. This is an opt-in mechanism: by default, the Python interpreter compiled from upstream sources will not be so marked, and so running ``pip install`` with a self-compiled interpreter, or with a distribution that has not explicitly marked its interpreter, will work as it always has worked. Second, it sets the rule that when installing packages to an interpreter's global context (either to an unmarked interpreter, or if overriding the marking), **Python-specific package managers should modify or delete files only within the directories of the sysconfig scheme in which they would create files**. This permits a distributor of a Python interpreter to set up two directories, one for its own managed packages, and one for unmanaged packages installed by the end user, and ensure that installing unmanaged packages will not delete (or overwrite) files owned by the external package manager. Rationale ========= As described in detail in the next section, the first behavior change involves creating a marker file named ``EXTERNALLY-MANAGED``, whose presence indicates that non-virtual-environment package installations are managed by some means external to Python, such as a distribution's package manager. This file is specified to live in the ``stdlib`` directory in the default ``sysconfig`` scheme, which marks the interpreter / installation as a whole, not a particular location on ``sys.path``. The reason for this is that, as identified above, there are two related problems that risk breaking an externally-managed Python: you can install an incompatible new version of a package system-wide (e.g., with ``sudo pip install``), and you can install one in your user account alone, but in a location that is on the standard Python command's ``sys.path`` (e.g., with ``pip install --user``). If the marker file were in the system-wide ``site-packages`` directory, it would not clearly apply to the second case. The `Alternatives`_ section has further discussion of possible locations. The second behavior change takes advantage of the existing ``sysconfig`` setup in distributions that have already encountered this class of problem, and specifically addresses the problem of a Python-specific package manager deleting or overwriting files that are owned by an external package manager. Use cases --------- The changed behavior in this PEP is intended to "do the right thing" for as many use cases as possible. In this section, we consider the changes specified by this PEP for several representative use cases / contexts. Specifically, we ask about the two behaviors that could be changed by this PEP: 1. Will a Python-specific installer tool like ``pip install`` permit installations by default, after implementation of this PEP? 2. If you do run such a tool, should it be willing to delete packages shipped by the external (non-Python-specific) package manager for that context, such as a distribution package manager? (For simplicity, this section discusses pip as the Python-specific installer tool, though the analysis should apply equally to any other Python-specific package management tool.) This table summarizes the use cases discussed in detail below: ==== ================================= =========================== =================================================== Case Description ``pip install`` permitted Deleting externally-installed packages permitted ==== ================================= =========================== =================================================== 1 Unpatched CPython Currently yes; stays yes Currently yes; stays yes 2 Distribution ``/usr/bin/python3`` Currently yes; becomes no Currently yes (except on Debian); becomes no (assuming the distribution adds a marker file) 3 Distribution Python in venv Currently yes; stays yes There are no externally-installed packages 4 Distribution Python in venv Currently yes; stays yes Currently no; stays no with ``--system-site-packages`` 5 Distribution Python in Docker Currently yes; stays yes Currently yes; becomes no (assuming the Docker image removes the marker file) 6 Conda environment Currently yes; stays yes Currently yes; stays yes 7 Dev-facing distribution Currently yes; becomes no Currently often yes; becomes no (assuming they add a (assuming they configure ``sysconfig`` as needed) marker file) 8 Distribution building packages Currently yes; can stay yes Currently yes; becomes no 9 ``PYTHONHOME`` copied from Currently yes; becomes no Currently yes; becomes no a distribution Python stdlib 10 ``PYTHONHOME`` copied from Currently yes; stays yes Currently yes; stays yes upstream Python stdlib ==== ================================= =========================== =================================================== In more detail, the use cases above are: 1. A standard unpatched CPython, without any special configuration of or patches to ``sysconfig`` and without a marker file. This PEP does not change its behavior. Such a CPython should (regardless of this PEP) not be installed in a way that that overlaps any distribution-installed Python on the same system. For instance, on an OS that ships Python in ``/usr/bin``, you should not install a custom CPython built with ``./configure --prefix=/usr``, or it will overwrite some files from the distribution and the distribution will eventually overwrite some files from your installation. Instead, your installation should be in a separate directory (perhaps ``/usr/local``, ``/opt``, or your home directory). Therefore, we can assume that such a CPython has its own ``stdlib`` directory and its own ``sysconfig`` schemes that do not overlap any distribution-installed Python. So any OS-installed packages are not visible or relevant here. If there is a concept of "externally-installed" packages in this case, it's something outside the OS and generally managed by whoever built and installed this CPython. Because the installer chose not to add a marker file or modify ``sysconfig`` schemes, they're choosing the current behavior, and ``pip install`` can remove any packages available in this CPython. 2. A distribution's ``/usr/bin/python3``, either when running ``pip install`` as root or ``pip install --user``, following our `Recommendations for distributions`_. These recommendations include shipping a marker file in the ``stdlib`` directory, to prevent ``pip install`` by default, and placing distribution-shipped packages in a location other than the default ``sysconfig`` scheme, so that ``pip`` as root does not write to that location. Many distributions (including Debian, Fedora, and their derivatives) are already doing the latter. On Debian and derivatives, ``pip install`` does not currently delete distribution-installed packages, because Debian carries a `patch to pip to prevent this`__. So, for those distributions, this PEP is not a behavior change; it simply standardizes that behavior in a way that is no longer Debian-specific and can be included into upstream pip. .. __: https://sources.debian.org/src/python-pip/20.3.4-2/debian/patches/hands-off-... (We have seen user reports of externally-installed packages being deleted on Debian or a derivative. We suspect this is because the user has previously run ``sudo pip install --upgrade pip`` and therefore now has a version of ``/usr/bin/pip`` without the Debian patch; standardizing this behavior in upstream package installers would address this problem.) 3. A distribution Python when used inside a virtual environment (either from ``venv`` or ``virtualenv``). Inside a virtual environment, all packages are owned by that environment. Even when ``pip``, ``setuptools``, etc. are installed into the environment, they are and should be managed by tools specific to that environment; they are not system-managed. 4. A distribution Python when used inside a virtual environment with ``--system-site-packages``. This is like the previous case, but worth calling out explicitly, because anything on the global ``sys.path`` is visible. Currently, the answer to "Will ``pip` delete externally-installed packages`` is no, because pip has a special case for running in a virtual environment and attempting to delete packages outside it. After this PEP, the answer remains no, but the reasoning becomes more general: system site packages will be outside any of the ``sysconfig`` schemes used for package management in the environment. 5. A distribution Python when used in a single-application container image (e.g., a Docker container). In this use case, the risk of breaking system software is lower, since generally only a single application runs in the container, and the impact is lower, since you can rebuild the container and you don't have to struggle to recover a running machine. There are also a large number of existing Dockerfiles with an unqualified ``RUN pip install ...`` statement, etc., and it would be good not to break those. So, builders of base container images may want to ensure that the marker file is not present, even if the underlying OS ships one by default. There is a small behavior change: currently, ``pip`` run as root will delete externally-installed packages, but after this PEP it will not. We don't propose a way to override this. However, since the base image is generally minimal, there shouldn't be much of a use case for simply uninstalling packages (especially without using the distribution's own tools). The common case is when pip wants to upgrade a package, which previously would have deleted the old version (except on Debian). After this change, the old version will still be on disk, but pip will still *shadow* externally-installed packages, and we believe this to be sufficient for this not to be a breaking change in practice - a Python ``import`` statement will still get you the newly-installed package. If it becomes necessary to have a way to do this, we suggest that the distribution should document a way for the installer tool to access the ``sysconfig`` scheme used by the distribution itself. See the `Recommendations for distributions`_ section for more discussion. It is the view of the authors of this PEP that it's still a good idea to use virtual environments with distribution-installed Python interpreters, even in single-application container images. Even though they run a single *application*, that application may run commands from the OS that are implemented in Python, and if you've installed or upgraded the distribution-shipped Python packages using Python-specific tools, those commands may break. 6. Conda specifically supports the use of non-``conda`` tools like pip to install software not available in the Conda repositories. In this context, Conda acts as the external package manager / distribution and pip as the Python-specific one. In some sense, this is similar to the first case, since Conda provides its own installation of the Python interpreter. We don't believe this PEP requires any changes to Conda, and versions of pip that have implemented the changes in this PEP will continue to behave as they currently do inside Conda environments. (That said, it may be worth considering whether to use separate ``sysconfig`` schemes for pip-installed and Conda-installed software, for the same reasons it's a good idea for other distributions.) 7. By a "developer-facing distribution," we mean a specific type of distribution where direct users of Python or other languages in the distribution are expected or encouraged to make changes to the distribution itself if they wish to add libraries. Common examples include private "monorepos" at software development companies, where a single repository builds both third-party and in-house software, and the direct users of the distribution's Python interpreter are generally software developers writing said in-house software. User-level package managers like Nixpkgs_ may also count, because they encourage users of Nix who are Python developers to `package their software for Nix`__. In these cases, the distribution may want to respond to an attempted ``pip install`` with guidance encouraging use of the distribution's own facilities for adding new packages, along with a link to documentation. If the distribution supports/encourages creating a virtual environment from the distribution's Python interpreter, there may also be custom instructions for how to properly set up a virtual environment (as for example Nixpkgs does). .. _Nixpkgs: https://github.com/NixOS/nixpkgs .. __: https://nixos.wiki/wiki/Python 8. When building distribution Python packages for a distribution Python (case 2), it may be useful to have ``pip install`` be usable as part of the distribution's package build process. (Consider, for instance, building a ``python-xyz`` RPM by using ``pip install .`` inside an sdist / source tarball for ``xyz``.) The distribution may also want to use a more targeted but still Python-specific installation tool such as installer_. .. _installer: https://installer.rtfd.io/ For this case, the build process will need to find some way to suppress the marker file to allow ``pip install`` to work, and will probably need to point the Python-specific tool at the distribution's ``sysconfig`` scheme instead of the shipped default. See the `Recommendations for distributions`_ section for more discussion on how to implement this. As a result of this PEP, pip will no longer be able to remove packages already on the system. However, this behavior change is fine because a package build process should not (and generally cannot) include instructions to delete some other files on the system; it can only package up its own files. 9. A distribution Python used with ``PYTHONHOME`` to set up an alternative Python environment (as opposed to a virtual environment), where ``PYTHONHOME`` is set to some directory copied directly from the distribution Python (e.g., ``cp -a /usr/lib/python3.x pyhome/lib``). Assuming there are no modifications, then the behavior is just like the underlying distribution Python (case 2). So there are behavior changes - you can no longer ``pip install`` by default, and if you override it, it will no longer delete externally-installed packages (i.e., Python packages that were copied from the OS and live in the OS-managed ``sys.path`` entry). This behavior change seems to be defensible, in that if your ``PYTHONHOME`` is a straight copy of the distribution's Python, it should behave like the distribution's Python. 10. A distribution Python (or any Python interpreter) used with a ``PYTHONHOME`` taken from a compatible unmodified upstream Python. Because the behavior changes in this PEP are keyed off of files in the standard library (the marker file in ``stdlib`` and the behavior of the ``sysconfig`` module), the behavior is just like an unmodified upstream CPython (case 1). Specification ============= Marking an interpreter as using an external package manager ----------------------------------------------------------- Before a Python-specific package installer (that is, a tool such as pip - not an external tool such as apt) installs a package into a certain Python context, it should make the following checks by default: 1. Is it running outside of a virtual environment? It can determine this by whether ``sys.prefix == sys.base_prefix`` (but see `Backwards Compatibility`_). 2. Is there a ``EXTERNALLY-MANAGED`` file in the directory identified by ``sysconfig.get_path("stdlib", sysconfig.get_default_scheme())`` If both of these conditions are true, the installer should exit with an error message indicating that package installation into this Python interpreter's directory are disabled outside of a virtual environment. The installer should have a way for the user to override these rules, such as a command-line flag ``--break-system-packages``. This option should not be enabled by default and should carry some connotation that its use is risky. The ``EXTERNALLY-MANAGED`` file is a metadata file in the `packaging core metadata format`_, which is an email-message-like format with headers and a body. (At current writing, that format is defined exactly as what the standard library ``email.parser`` module can parse using ``policy=email.policy.compat32``.) If the file can be parsed as a core metadata file, then the installer should output an error message from that file as part of its error. If ``locale.getlocale(locale.LC_MESSAGES)`` returns non-``None`` and the first element is a string of the form ``xx_YY``, and the file contains a header variable ``Error-xx_YY`` or failing that ``Error-xx``, then the installer should use the value of that header as the error. Otherwise, it should use the body of the message as an error. .. _`packaging core metadata format`: https://packaging.python.org/specifications/core-metadata/ If the file does not parse as a core metadata file, then the installer should ignore the parse failure and instead just use a pre-defined error message of its own, which should suggest that the user create a virtual environment to install packages. Software distributors who have a non-Python-specific package manager that manages libraries in the ``sys.path`` of their Python package should, in general, ship a ``EXTERNALLY-MANAGED`` file in their standard library directory. For instance, Debian may ship a file in ``/usr/lib/python3.9/EXTERNALLY-MANAGED`` consisting of something like :: To install Python packages system-wide, try apt install python3-xyz, where xyz is the package you are trying to install. If you wish to install a non-Debian-packaged Python package, create a virtual environment using python3 -m venv path/to/venv. Then use path/to/venv/bin/python and path/to/venv/bin/pip. Make sure you have python3-full installed. If you wish to install a non-Debian packaged Python application, it may be easiest to use pipx install xyz, which will manage a virtual environment for you. Make sure you have pipx installed. See /usr/share/doc/python3.9/README.venv for more information. which provides useful and distribution-relevant information to a user trying to install a package. In certain contexts, such as single-application container images that aren't updated after creation, a distributor may choose not to ship an ``EXTERNALLY-MANAGED`` file, so that users can install whatever they like (as they can today) without having to manually override this rule. Writing to only the target ``sysconfig`` scheme ----------------------------------------------- Usually, a Python package installer installs to directories in a scheme returned by the ``sysconfig`` standard library package. Ordinarily, this is the scheme returned by ``sysconfig.get_default_scheme()``, but based on configuration (e.g. ``pip install --user``), it may use a different scheme. Whenever the installer is installing to a ``sysconfig`` scheme, this PEP specifies that the installer should never modify or delete files outside of that scheme. For instance, if it's upgrading a package, and the package is already installed in a directory outside that scheme (perhaps in a directory from another scheme), it should leave the existing files alone. If the installer does end up shadowing an existing installation during an upgrade, we recommend that it produces a warning at the end of its run. If the installer is installing to a location outside of a ``sysconfig`` scheme (e.g., ``pip install --target``), then this subsection does not apply. Recommendations for distributions ================================= This section is non-normative. It provides best practices we believe distributions should follow unless they have a specific reason otherwise. Mark the installation as externally managed ------------------------------------------- Distributions should create an ``EXTERNALLY-MANAGED`` file in their ``stdlib`` directory. Guide users towards virtual environments ---------------------------------------- The file should contain a useful and distribution-relevant error message indicating both how to install system-wide packages via the distribution's package manager and how to set up a virtual environment. If your distribution is often used by users in a state where the ``python3`` command is available (and especially where ``pip`` or ``get-pip`` is available) but ``python3 -m venv`` does not work, the message should indicate clearly how to make ``python3 -m venv`` work properly. Consider packaging pipx_, a tool for installing Python-language applications, and suggesting it in the error. pipx automatically creates a virtual environment for that application alone, which is a much better default for end users who want to install some Python-language software (which isn't available in the distribution) but are not themselves Python users. Packaging pipx in the distribution avoids the irony of instructing users to ``pip install --user --break-system-packages pipx`` to *avoid* breaking system packages. Consider arranging things so your distribution's package / environment for Python for end users (e.g., ``python3`` on Fedora or ``python3-full`` on Debian) depends on pipx. .. _pipx: https://github.com/pypa/pipx Remove the marker file in container images ------------------------------------------ Distributions that produce official images for single-application containers (e.g., Docker container images) should remove the ``EXTERNALLY-MANAGED`` file, preferably in a way that makes it not come back if a user of that image installs package updates inside their image (think ``RUN apt-get dist-upgrade``). On dpkg-based systems, using ``dpkg-divert --local`` to persistently rename the file would work. On other systems, there may need to be some configuration flag available to a post-install script to re-remove the ``EXTERNALLY-MANAGED`` file. Create separate distribution and local directories -------------------------------------------------- Distributions should place two separate paths on the system interpreter's ``sys.path``, one for distribution-installed packages and one for packages installed by the local system administrator, and configure ``sysconfig.get_default_scheme()`` to point at the latter path. This ensures that tools like pip will not modify distribution-installed packages. The path for the local system administrator should come before the distribution path on ``sys.path`` so that local installs take preference over distribution packages. For example, Fedora and Debian (and their derivatives) both implement this split by using ``/usr/local`` for locally-installed packages and ``/usr`` for distribution-installed packages. Fedora uses ``/usr/local/lib/python3.x/site-packages`` vs. ``/usr/lib/python3.x/site-packages``. (Debian uses ``/usr/local/lib/python3/dist-packages`` vs. ``/usr/lib/python3/dist-packages`` as an additional layer of separation from a locally-compiled Python interpreter: if you build and install upstream CPython in ``/usr/local/bin``, it will look at ``/usr/local/lib/python3/site-packages``, and Debian wishes to make sure that packages installed via the locally-built interpreter don't show up on ``sys.path`` for the distribution interpreter.) Note that the ``/usr/local`` vs. ``/usr`` split is analogous to how the ``PATH`` environment variable typically includes ``/usr/local/bin:/usr/bin`` and non-distribution software installs to ``/usr/local`` by default. This split is `recommended by the Filesystem Hierarchy Standard`__. .. __: https://refspecs.linuxfoundation.org/FHS_3.0/fhs/ch04s09.html There are two ways you could do this. One is, if you are building and packaging Python libraries directly (e.g., your packaging helpers unpack a PEP 517-built wheel or call ``setup.py install``), arrange for those tools to use a directory that is not in a ``sysconfig`` scheme but is still on ``sys.path``. The other is to arrange for the default ``sysconfig`` scheme to change when running inside a package build versus when running on an installed system. The ``sysconfig`` customization hooks from bpo-43976_ should make this easy: make your packaging tool set an environment variable or some other detectable configuration, and define a ``get_preferred_schemes`` function to return a different scheme when called from inside a package build. Then you can use ``pip install`` as part of your distribution packaging. .. _bpo-43976: https://bugs.python.org/issue43976 We propose adding a ``--scheme=...`` option to instruct pip to run against a specific scheme. (See `Implementation Notes`_ below for how pip currently determines schemes.) Once that's available, for local testing and possibly for actual packaging, you would be able to run something like ``pip install --scheme=posix_distro`` to explicitly install a package into your distribution's location (bypassing ``get_preferred_schemes``). One could also, if absolutely needed, use ``pip uninstall --scheme=posix_distro`` to use pip to remove packages from the system-managed directory, which addresses the (hopefully theoretical) regression in use case 5 in Rationale_. To install packages with pip, you would also need to either suppress the ``EXTERNALLY-MANAGED`` marker file to allow pip to run or to override it on the command line. You may want to use the same means for suppressing the marker file in build chroots as you do in container images. The advantage of setting these up to be automatic (suppressing the marker file in your build environment and having ``get_preferred_schemes`` automatically return your distribution's scheme) is that an unadorned ``pip install`` will work inside a package build, which generally means that an unmodified upstream build script that happens to internally call ``pip install`` will do the right thing. You can, of course, just ensure that your packaging process always calls ``pip install --scheme=posix_distro --break-system-packages``, which would work too. The best approach here depends a lot on your distribution's conventions and mechanisms for packaging. Similarly, the ``sysconfig`` paths that are not for importable Python code - that is, ``include``, ``platinclude``, ``scripts``, and ``data`` - should also have two variants, one for use by distribution-packaged software and one for use for locally-installed software, and the distribution should be set up such that both are usable. For instance, a typical FHS-compliant distribution will use ``/usr/local/include`` for the default scheme's ``include`` and ``/usr/include`` for distribution-packaged headers and place both on the compiler's search path, and it will use ``/usr/local/bin`` for the default scheme's ``scripts`` and ``/usr/bin`` for distribution-packaged entry points and place both on ``$PATH``. Backwards Compatibility ======================= All of these mechanisms are proposed for new distribution releases and new versions of tools like pip only. In particular, we strongly recommend that distributions with a concept of major versions only add the marker file or change ``sysconfig`` schemes in a new major version; otherwise there is a risk that, on an existing system, software installed via a Python-specific package manager now becomes unmanageable (without an override option). For a rolling-release distribution, if possible, only add the marker file or change ``sysconfig`` schemes in a new Python minor version. One particular backwards-compatibility difficulty for package installation tools is likely to be managing environments created by old versions of ``virtualenv`` which have the latest version of the tool installed. A "virtual environment" now has a fairly precise definition: it uses the ``pyvenv.cfg`` mechanism, which causes ``sys.base_prefix != sys.prefix``. It is possible, however, that a user may have an old virtual environment created by an older version of ``virtualenv``; as of this writing, pip supports Python 3.6 onwards, which is in turn supported by ``virtualenv`` 15.1.0 onwards, so this scenario is possible. In older versions of ``virtualenv``, the mechanism is instead to set a new attribute, ``sys.real_prefix``, and it does not use the standard library support for virtual environments, so ``sys.base_prefix`` is the same as ``sys.prefix``. So the logic for robustly detecting a virtual environment is something like:: def is_virtual_environment(): return sys.base_prefix != sys.prefix or hasattr(sys, "real_prefix") Security Implications ===================== The purpose of this feature is not to implement a security boundary; it is to discourage well-intended changes from unexpectedly breaking a user's environment. That is to say, the reason this PEP restricts ``pip install`` outside a virtual environment is not that it's a security risk to be able to do so; it's that "There should be one-- and preferably only one --obvious way to do it," and that way should be using a virtual environment. ``pip install`` outside a virtual environment is rather too obvious for what is almost always the wrong way to do it. If there is a case where a user should not be able to ``sudo pip install`` or ``pip install --user`` and add files to ``sys.path`` *for security reasons*, that needs to be implemented either via access control rules on what files the user can write to or an explicitly secured ``sys.path`` for the program in question. Neither of the mechanisms in this PEP should be interpreted as a way to address such a scenario. For those reasons, an attempted install with a marker file present is not a security incident, and there is no need to raise an auditing event for it. If the calling user legitimately has access to ``sudo pip install`` or ``pip install --user``, they can accomplish the same installation entirely outside of Python; if they do not legitimately have such access, that's a problem outside the scope of this PEP. The marker file itself is located in the standard library directory, which is a trusted location (i.e., anyone who can write to the marker file used by a particular installer could, presumably, run arbitrary code inside the installer). Therefore, there is generally no need to filter out terminal escape sequences or other potentially-malicious content in the error message. Alternatives ============== There are a number of similar proposals we considered that this PEP rejects or defers, largely to preserve the behavior in the case-by-case analysis in Rationale_. Marker file ----------- Should the marker file be in ``sys.path``, marking a particular directory as not to be written to by a Python-specific package manager? This would help with the second problem addressed by this PEP (not overwriting deleting distribution-owned files) but not the first (incompatible installs). A directory-specific marker in ``/usr/lib/python3.x/site-packages`` would not discourage installations into either ``/usr/local/lib/python3.x/site-packages`` or ``~/.local/lib/python3.x/site-packages``, both of which are on ``sys.path`` for ``/usr/bin/python3``. In other words, the marker file should not be interpreted as marking a single *directory* as externally managed (even though it happens to be in a directory on ``sys.path``); it marks the entire *Python installation* as externally managed. Another variant of the above: should the marker file be in ``sys.path``, where if it can be found in any directory in ``sys.path``, it marks the installation as externally managed? An apparent advantage of this approach is that it automatically disables itself in virtual environments. Unfortunately, This has the wrong behavior with a ``--system-site-packages`` virtual environment, where the system-wide ``sys.path`` is visible but package installations are allowed. (It could work if the rule of exempting virtual environments is preserved, but that seems to have no advantage over the current scheme.) Should the marker just be a new attribute of a ``sysconfig`` scheme? There is some conceptual cleanliness to this, except that it's hard to override. We want to make it easy for container images, package build environments, etc. to suppress the marker file. A file that you can remove is easy; code in ``sysconfig`` is much harder to modify. Should the file be in ``/etc``? No, because again, it refers to a specific Python installation. A user who installs their own Python may well want to install packages within the global context of that interpreter. Should the configuration setting be in ``pip.conf`` or ``distutils.cfg``? Apart from the above objections about marking an installation, this mechanism isn't specific to either of those tools. (It seems reasonable for pip to *also* implement a configuration flag for users to prevent themselves from performing accidental non-virtual-environment installs in any Python installation, but that is outside the scope of this PEP.) Should the file be TOML? TOML is gaining popularity for packaging (see e.g. PEP-517) but does not yet have an implementation in the standard library. Strictly speaking, this isn't a blocker - distributions need only write the file, not read it, so they don't need a TOML library (the file will probably be written by hand, regardless of format), and packaging tools likely have a TOML reader already. However, the ``email.message`` format is currently used for various other forms of packaging metadata, meets our needs, and is parseable by the standard library, and the pip maintainers expressed a preference to avoid using TOML for this yet. Should the marker file be executable Python code that evaluates whether installation should be allowed or not? Apart from the concerns above about having the file in ``sys.path``, we have a concern that making it executable is committing to too powerful of an API and risks making behavior harder to understand. (Note that the ``get_default_scheme`` hook of bpo-43976_ is in fact executable, but that code needs to be supplied when the interpreter builds; it isn't intended to be supplied post-build.) When overriding the marker, should a Python-specific package manager be disallowed from shadowing a package installed by the external package manager (i.e., installing modules of the same name)? This would minimize the risk of breaking system software, but it's not clear it's worth the additional user experience complexity. There are legitimate use cases for shadowing system packages, and an additional command-line option to permit it would be more confusing. Meanwhile, not passing that option wouldn't eliminate the risk of breaking system software, which may be relying on a ``try: import xyz`` failing, finding a limited set of entry points, etc. Communicating this distinction seems difficult. We think it's a good idea for Python-specific package managers to print a warning if they shadow a package, but we think it's not worth disabling it by default. Why not use the ``INSTALLER`` file from PEP 376 to determine who installed a package and whether it can be removed? First, it's specific to a particular package (it's in the package's ``dist-info`` directory), so like some of the alternatives above, it doesn't provide information on an entire environment and whether package installations are permissible. PEP 627 also updates PEP 376 to prevent programmatic use of ``INSTALLER``, specifying that the file is "to be used for informational purposes only. [...] Our goal is supporting interoperating tools, and basing any action on which tool happened to install a package runs counter to that goal." Finally, as PEP 627 envisions, there are legitimate use cases for one tool knowing how to handle packages installed by another tool; for instance, ``conda`` can safely remove a package installed by ``pip`` into a Conda environment. Why does the specification give no means for disabling package installations inside a virtual environment? We can't see a particularly strong use case for it (at least not one related to the purposes of this PEP). If you need it, it's simple enough to ``pip uninstall pip`` inside that environment, which should discourage at least unintentional changes to the environment (and this specification makes no provision to disable *intentional* changes, since after all the marker file can be easily removed). System Python ------------- Shouldn't distribution software just run with the distribution ``site-packages`` directory alone on ``sys.path`` and ignore the local system administrator's ``site-packages`` as well as the user-specific one? This is a worthwhile idea, and various versions of it have been circulating for a while under the name of "system Python" or "platform Python" (with a separate "user Python" for end users writing Python or installing Python software separate from the system). However, it's much more involved of a change. First, it would be a backwards-incompatible change. As mentioned in the Motivation_ section, there are valid use cases for running distribution-installed Python applications like Sphinx or Ansible with locally-installed Python libraries available on their ``sys.path``. A wholesale switch to ignoring local packages would break these use cases, and a distribution would have to make a case-by-case analysis of whether an application ought to see locally-installed libraries or not. Furthermore, `Fedora attempted this change and reverted it`_, finding, ironically, that their implementation of the change `broke their package manager`_. Given that experience, there are clearly details to be worked out before distributions can reliably implement that approach, and a PEP recommending it would be premature. .. _`Fedora attempted this change and reverted it`: https://lists.fedoraproject.org/archives/list/devel@lists.fedoraproject.org/... .. _`broke their package manager`: https://bugzilla.redhat.com/show_bug.cgi?id=1483342 This PEP is intended to be a complete and self-contained change that is independent of a distributor's decision for or against "system Python" or similar proposals. It is not incompatible with a distribution implementing "system Python" in the future, and even though both proposals address the same class of problems, there are still arguments in favor of implementing something like "system Python" even after implementing this PEP. At the same time, though, this PEP specifically tries to make a more targeted and minimal change, such that it can be implemented by distributors who don't expect to adopt "system Python" (or don't expect to implement it immediately). The changes in this PEP stand on their own merits and are not an intermediate step for some future proposal. This PEP reduces (but does not eliminate) the risk of breaking system software while minimizing (but not completely avoiding) breaking changes, which should therefore be much easier to implement than the full "system Python" idea, which comes with the downsides mentioned above. We expect that the guidance in this PEP - that users should use virtual environments whenever possible and that distributions should have separate ``sys.path`` directories for distribution-managed and locally-managed modules - should make further experiments easier in the future. These may include distributing wholly separate "system" and "user" Python interpreters, running system software out of a distribution-owned virtual environment or ``PYTHONHOME`` (but shipping a single interpreter), or modifying the entry points for certain software (such as the distribution's package manager) to use a ``sys.path`` that only sees distribution-managed directories. Those ideas themselves, however, remain outside the scope of this PEP. Implementation Notes ==================== This section is non-normative and contains notes relevant to both the specification and potential implementations. Currently, pip does not directly expose a way to choose a target ``sysconfig`` scheme, but it has three ways of looking up schemes when installing: ``pip install`` Calls ``sysconfig.get_default_scheme()``, which is usually (in upstream CPython and most current distributions) the same as ``get_preferred_scheme('prefix')``. ``pip install --prefix=/some/path`` Calls ``sysconfig.get_preferred_scheme('prefix')``. ``pip install --user`` Calls ``sysconfig.get_preferred_scheme('user')``. Finally, ``pip install --target=/some/path`` writes directly to ``/some/path`` without looking up any schemes. Debian currently carries a `patch to change the default install location inside a virtual environment`__, using a few heuristics (including checking for the ``VIRTUAL_ENV`` environment variable), largely so that the directory used in a virtual environment remains ``site-packages`` and not ``dist-packages``. This does not particularly affect this proposal, because the implementation of that patch does not actually change the default ``sysconfig`` scheme, and notably does not change the result of ``sysconfig.get_path("stdlib")``. .. __: https://sources.debian.org/src/python3.7/3.7.3-2+deb10u3/debian/patches/dist... Fedora currently carries a `patch to change the default install location when not running inside rpmbuild`__, which they use to implement the two-system-wide-directories approach. This is conceptually the sort of hook envisioned by bpo-43976_, except implemented as a code patch to ``distutils`` instead of as a changed ``sysconfig`` scheme. .. __: https://src.fedoraproject.org/rpms/python3.9/blob/f34/f/00251-change-user-in... The implementation of ``is_virtual_environment`` above, as well as the logic to load the ``EXTERNALLY-MANAGED`` file and find the error message from it, may as well get added to the standard library (``sys`` and ``sysconfig``, respectively), to centralize their implementations, but they don't need to be added yet. References ========== For additional background on these problems and previous attempts to solve them, see `Debian bug 771794`_ "pip silently removes/updates system provided python packages` from 2014, Fedora's 2018 article `Making sudo pip safe`_ about pointing ``sudo pip`` at /usr/local (which acknowledges that the changes still do not make ``sudo pip`` completely safe), pip issues 5605_ ("Disable upgrades to existing python modules which were not installed via pip") and 5722_ ("pip should respect /usr/local") from 2018, and the post-PyCon US 2019 discussion thread `Playing nice with external package managers`_. .. _`Debian bug 771794`: https://bugs.debian.org/771794 .. _`Making sudo pip safe`: https://fedoraproject.org/wiki/Changes/Making_sudo_pip_safe .. _5605: https://github.com/pypa/pip/issues/5605 .. _5722: https://github.com/pypa/pip/issues/5722 .. _`Playing nice with external package managers`: https://discuss.python.org/t/playing-nice-with-external-package-managers/196... TODO: We can open these before the PEP is accepted and should link to these: * PR to pip for EXTERNALLY-MANAGED + ``--break-system-packages`` * PR to pip for hands-off-system-packages.patch v2 * MR to Debian Python to create the EXTERNALLY-MANAGED file * PR to upstream Python for ``is_virtual_env``/``is_externally_managed`` maybe? .. [#PEP-376] PEP 376, Database of Installed Python Distributions Ziadé (http://www.python.org/dev/peps/pep-0376) .. [#PEP-427] PEP 427, The Wheel Binary Package Format 1.0, Holth (http://www.python.org/dev/peps/pep-0427) .. [#pip] https://pip.pypa.io/en/stable/ .. [#easy-install] https://setuptools.readthedocs.io/en/latest/deprecated/easy_install.html (Note that the ``easy_install`` command was removed in setuptools version 52, released 23 January 2021.) .. [#Conda] https://conda.io Copyright ========= This document is placed in the public domain or under the CC0-1.0-Universal license, whichever is more permissive. -- Geoffrey Thomas https://ldpreload.com geofft@ldpreload.com