The Python Packaging Ecosystem

There have been a few recent articles reflecting on the current status of the Python packaging ecosystem from an end user perspective, so it seems worthwhile for me to write-up my perspective as one of the lead architects for that ecosystem on how I characterise the overall problem space of software publication and distribution, where I think we are at the moment, and where I'd like to see us go in the future.

For context, the specific articles I'm replying to are:

These are all excellent pieces considering the problem space from different perspectives, so if you'd like to learn more about the topics I cover here, I highly recommend reading them.

My core software ecosystem design philosophy

Since it heavily influences the way I think about packaging system design in general, it's worth stating my core design philosophy explicitly:

  • As a software consumer, I should be able to consume libraries, frameworks, and applications in the binary format of my choice, regardless of whether or not the relevant software publishers directly publish in that format

  • As a software publisher working in the Python ecosystem, I should be able to publish my software once, in a single source-based format, and have it be automatically consumable in any binary format my users care to use

This is emphatically not the way many software packaging systems work - for a great many systems, the publication format and the consumption format are tightly coupled, and the folks managing the publication format or the consumption format actively seek to use it as a lever of control over a commercial market (think operating system vendor controlled application stores, especially for mobile devices).

While we're unlikely to ever pursue the specific design documented in the rest of the PEP (hence the "Deferred" status), the "Development, Distribution, and Deployment of Python Software" section of PEP 426 provides additional details on how this philosophy applies in practice.

I'll also note that while I now work on software supply chain management tooling at Red Hat, that wasn't the case when I first started actively participating in the upstream Python packaging ecosystem design process. Back then I was working on Red Hat's main hardware integration testing system, and growing increasingly frustrated with the level of effort involved in integrating new Python level dependencies into Beaker's RPM based development and deployment model. Getting actively involved in tackling these problems on the Python upstream side of things then led to also getting more actively involved in addressing them on the Red Hat downstream side.

The key conundrum

When talking about the design of software packaging ecosystems, it's very easy to fall into the trap of only considering the "direct to peer developers" use case, where the software consumer we're attempting to reach is another developer working in the same problem domain that we are, using a similar set of development tools. Common examples of this include:

  • Linux distro developers publishing software for use by other contributors to the same Linux distro ecosystem

  • Web service developers publishing software for use by other web service developers

  • Data scientists publishing software for use by other data scientists

In these more constrained contexts, you can frequently get away with using a single toolchain for both publication and consumption:

  • Linux: just use the system package manager for the relevant distro

  • Web services: just use the Python Packaging Authority's twine for publication and pip for consumption

  • Data science: just use conda for everything

For newer languages that start in one particular domain with a preferred package manager and expand outwards from there, the apparent simplicity arising from this homogeneity of use cases may frequently be attributed as an essential property of the design of the package manager, but that perception of inherent simplicity will typically fade if the language is able to successfully expand beyond the original niche its default package manager was designed to handle.

In the case of Python, for example, distutils was designed as a consistent build interface for Linux distro package management, setuptools for plugin management in the Open Source Application Foundation's Chandler project, pip for dependency management in web service development, and conda for local language-independent environment management in data science. distutils and setuptools haven't fared especially well from a usability perspective when pushed beyond their original design parameters (hence the current efforts to make it easier to use full-fledged build systems like Scons and Meson as an alternative when publishing Python packages), while pip and conda both seem to be doing a better job of accommodating increases in their scope of application.

This history helps illustrate that where things really have the potential to get complicated (even beyond the inherent challenges of domain-specific software distribution) is when you start needing to cross domain boundaries. For example, as the lead maintainer of contextlib in the Python standard library, I'm also the maintainer of the contextlib2 backport project on PyPI. That's not a domain specific utility - folks may need it regardless of whether they're using a self-built Python runtime, a pre-built Windows or Mac OS X binary they downloaded from, a pre-built binary from a Linux distribution, a CPython runtime from some other redistributor (homebrew, pyenv, Enthought Canopy, ActiveState, Continuum Analytics, AWS Lambda, Azure Machine Learning, etc), or perhaps even a different Python runtime entirely (PyPy, PyPy.js, Jython, IronPython, MicroPython, VOC, Batavia, etc).

Fortunately for me, I don't need to worry about all that complexity in the wider ecosystem when I'm specifically wearing my contextlib2 maintainer hat - I just publish an sdist and a universal wheel file to PyPI, and the rest of the ecosystem has everything it needs to take care of redistribution and end user consumption without any further input from me.

However, contextlib2 is a pure Python project that only depends on the standard library, so it's pretty much the simplest possible case from a tooling perspective (the only reason I needed to upgrade from distutils to setuptools was so I could publish my own wheel files, and the only reason I haven't switched to using the much simpler pure-Python-only flit instead of either of them is that that doesn't yet easily support publishing backwards compatible based sdists).

This means that things get significantly more complex once we start wanting to use and depend on components written in languages other than Python, so that's the broader context I'll consider next.

Platform management or plugin management?

When it comes to handling the software distribution problem in general, there are two main ways of approaching it:

  • design a plugin management system that doesn't concern itself with the management of the application framework that runs the plugins

  • design a platform component manager that not only manages the plugins themselves, but also the application frameworks that run them

This "plugin manager or platform component manager?" question shows up over and over again in software distribution architecture designs, but the case of most relevance to Python developers is in the contrasting approaches that pip and conda have adopted to handling the problem of external dependencies for Python projects:

  • pip is a plugin manager for Python runtimes. Once you have a Python runtime (any Python runtime), pip can help you add pieces to it. However, by design, it won't help you manage the underlying Python runtime (just as it wouldn't make any sense to try to install Mozilla Firefox as a Firefox Add-On, or Google Chrome as a Chrome Extension)

  • conda, by contrast, is a component manager for a cross-platform platform that provides its own Python runtimes (as well as runtimes for other languages). This means that you can get pre-integrated components, rather than having to do your own integration between plugins obtained via pip and language runtimes obtained via other means

What this means is that pip, on its own, is not in any way a direct alternative to conda. To get comparable capabilities to those offered by conda, you have to add in a mechanism for obtaining the underlying language runtimes, which means the alternatives are combinations like:

  • apt-get + pip

  • dnf + pip

  • yum + pip

  • pyenv + pip

  • homebrew (Mac OS X) + pip

  • Windows installer + pip

  • Enthought Canopy

  • ActiveState's Python runtime + PyPM

This is the main reason why "just use conda" is excellent advice to any prospective Pythonista that isn't already using one of the platform component managers mentioned above: giving that answer replaces an otherwise operating system dependent or Python specific answer to the runtime management problem with a cross-platform and (at least somewhat) language neutral one.

It's an especially good answer for Windows users, as chocalatey/OneGet/Windows Package Management isn't remotely comparable to pyenv or homebrew at this point in time, other runtime managers don't work on Windows, and getting folks bootstrapped with MinGW, Cygwin or the new (still experimental) Windows Subsystem for Linux is just another hurdle to place between them and whatever goal they're learning Python for in the first place.

However, conda's pre-integration based approach to tackling the external dependency problem is also why "just use conda for everything" isn't a sufficient answer for the Python software ecosystem as a whole.

If you're working on an operating system component for Fedora, Debian, or any other distro, you actually want to be using the system provided Python runtime, and hence need to be able to readily convert your upstream Python dependencies into policy compliant system dependencies.

Similarly, if you're wanting to support folks that deploy to a preconfigured Python environment in services like AWS Lambda, Azure Cloud Functions, Heroku, OpenShift or Cloud Foundry, or that use alternative Python runtimes like PyPy or MicroPython, then you need a publication technology that doesn't tightly couple your releases to a specific version of the underlying language runtime.

As a result, pip and conda end up existing at slightly different points in the system integration pipeline:

  • Publishing and consuming Python software with pip is a matter of "bring your own Python runtime". This has the benefit that you can readily bring your own runtime (and manage it using whichever tools make sense for your use case), but also has the downside that you must supply your own runtime (which can sometimes prove to be a significant barrier to entry for new Python users, as well as being a pain for cross-platform environment management).

  • Like Linux system package managers before it, conda takes away the requirement to supply your own Python runtime by providing one for you. This is great if you don't have any particular preference as to which runtime you want to use, but if you do need to use a different runtime for some reason, you're likely to end up fighting against the tooling, rather than having it help you. (If you're tempted to answer "Just add another interpreter to the pre-integrated set!" here, keep in mind that doing so without the aid of a runtime independent plugin manager like pip acts as a multiplier on the platform level integration testing needed, which can be a significant cost even when it's automated)

Where do we go next?

In case it isn't already clear from the above, I'm largely happy with the respective niches that pip and conda are carving out for themselves as a plugin manager for Python runtimes and as a cross-platform platform focused on (but not limited to) data analysis use cases.

However, there's still plenty of scope to improve the effectiveness of the collaboration between the upstream Python Packaging Authority and downstream Python redistributors, as well as to reduce barriers to entry for participation in the ecosystem in general, so I'll go over some of the key areas I see for potential improvement.

Sustainability and the bystander effect

It's not a secret that the core PyPA infrastructure (PyPI, pip, twine, setuptools) is nowhere near as well-funded as you might expect given its criticality to the operations of some truly enormous organisations.

The biggest impact of this is that even when volunteers show up ready and willing to work, there may not be anybody in a position to effectively wrangle those volunteers, and help keep them collaborating effectively and moving in a productive direction.

To secure long term sustainability for the core Python packaging infrastructure, we're only talking amounts on the order of a few hundred thousand dollars a year - enough to cover some dedicated operations and publisher support staff for PyPI (freeing up the volunteers currently handling those tasks to help work on ecosystem improvements), as well as to fund targeted development directed at some of the other problems described below.

However, rather than being a true "tragedy of the commons", I personally chalk this situation up to a different human cognitive bias: the bystander effect.

The reason I think that is that we have so many potential sources of the necessary funding that even folks that agree there's a problem that needs to be solved are assuming that someone else will take care of it, without actually checking whether or not that assumption is entirely valid.

The primary responsibility for correcting that oversight falls squarely on the Python Software Foundation, which is why the Packaging Working Group was formed in order to investigate possible sources of additional funding, as well as to determine how any such funding can be spent most effectively.

However, a secondary responsibility also falls on customers and staff of commercial Python redistributors, as this is exactly the kind of ecosystem level risk that commercial redistributors are being paid to manage on behalf of their customers, and they're currently not handling this particular situation very well. Accordingly, anyone that's actually paying for CPython, pip, and related tools (either directly or as a component of a larger offering), and expecting them to be supported properly as a result, really needs to be asking some very pointed question of their suppliers right about now. (Here's a sample question: "We pay you X dollars a year, and the upstream Python ecosystem is one of the things we expect you to support with that revenue. How much of what we pay you goes towards maintenance of the upstream Python packaging infrastructure that we rely on every day?").

One key point to note about the current situation is that as a 501(c)(3) public interest charity, any work the PSF funds will be directed towards better fulfilling that public interest mission, and that means focusing primarily on the needs of educators and non-profit organisations, rather than those of private for-profit entities.

Commercial redistributors are thus far better positioned to properly represent their customers interests in areas where their priorities may diverge from those of the wider community (closing the "insider threat" loophole in PyPI's current security model is a particular case that comes to mind - see Making PyPI security independent of SSL/TLS).

Migrating PyPI to

An instance of the new PyPI implementation (Warehouse) is up and running at and connected directly to the production PyPI database, so folks can already explicitly opt-in to using it over the legacy implementation if they prefer to do so.

However, there's still a non-trivial amount of design, development and QA work needed on the new version before all existing traffic can be transparently switched over to using it.

Getting at least this step appropriately funded and a clear project management plan in place is the main current focus of the PSF's Packaging Working Group.

Making the presence of a compiler on end user systems optional

Between the wheel format and the manylinux1 usefully-distro-independent ABI definition, this is largely handled now, with conda available as an option to handle the relatively small number of cases that are still a problem for pip.

The main unsolved problem is to allow projects to properly express the constraints they place on target environments so that issues can be detected at install time or repackaging time, rather than only being detected as runtime failures. Such a feature will also greatly expand the ability to correctly generate platform level dependencies when converting Python projects to downstream package formats like those used by conda and Linux system package managers.

Bootstrapping dependency management tools on end user systems

With pip being bundled with recent versions of CPython (including CPython 2.7 maintenance releases), and pip (or a variant like upip) also being bundled with most other Python runtimes, the ecosystem bootstrapping problem has largely been addressed for new Python users.

There are still a few usability challenges to be addressed (like defaulting to per-user installations when outside a virtual environment, interoperating more effectively with platform component managers like conda, and providing an officially supported installation interface that works at the Python prompt rather than via the operating system command line), but those don't require the same level of political coordination across multiple groups that was needed to establish pip as the lowest common denominator approach to dependency management for Python applications.

Making the use of distutils and setuptools optional

As mentioned above, distutils was designed ~18 years ago as a common interface for Linux distributions to build Python projects, while setuptools was designed ~12 years ago as a plugin management system for an open source Microsoft Exchange replacement. While both projects have given admirable service in their original target niches, and quite a few more besides, their age and original purpose means they're significantly more complex than what a user needs if all they want to do is to publish their pure Python library or framework to the Python Package index.

Their underlying complexity also makes it incredibly difficult to improve the problematic state of their documentation, which is split between the legacy distutils documentation in the CPython standard library and the additional setuptools specific documentation in the setuptools project.

Accordingly, what we want to do is to change the way build toolchains for Python projects are organised to have 3 clearly distinct tiers:

  • toolchains for pure Python projects

  • toolchains for Python projects with simple C extensions

  • toolchains for C/C++/other projects with Python bindings

This allows folks to be introduced to simpler tools like flit first, better enables the development of potential alternatives to setuptools at the second tier, and supports the use of full-fledged pip-installable build systems like Scons and Meson at the third tier.

The first step in this project, defining the pyproject.toml format to allow declarative specification of the dependencies needed to launch, has been implemented, and Daniel Holth's enscons project demonstrates that that is already sufficient to bootstrap an external build system even without the later stages of the project.

Future steps include providing native support for pyproject.toml in pip and easy_install, as well as defining a declarative approach to invoking the build system rather than having to run with the relevant distutils & setuptools flags.

Making PyPI security independent of SSL/TLS

PyPI currently relies entirely on SSL/TLS to protect the integrity of the link between software publishers and PyPI, and between PyPI and software consumers. The only protections against insider threats from within the PyPI administration team are ad hoc usage of GPG artifact signing by some projects, personal vetting of new team members by existing team members and 3rd party checks against previously published artifact hashes unexpectedly changing.

A credible design for end-to-end package signing that adequately accounts for the significant usability issues that can arise around publisher and consumer key management has been available for almost 3 years at this point (see Surviving a Compromise of PyPI and Surviving a Compromise of PyPI: the Maximum Security Edition).

However, implementing that solution has been gated not only on being able to first retire the legacy infrastructure, but also the PyPI administators being able to credibly commit to the key management obligations of operating the signing system, as well as to ensuring that the system-as-implemented actually provides the security guarantees of the system-as-designed.

Accordingly, this isn't a project that can realistically be pursued until the underlying sustainability problems have been suitably addressed.

Automating wheel creation

While redistributors will generally take care of converting upstream Python packages into their own preferred formats, the Python-specific wheel format is currently a case where it is left up to publishers to decide whether or not to create them, and if they do decide to create them, how to automate that process.

Having PyPI take care of this process automatically is an obviously desirable feature, but it's also an incredibly expensive one to build and operate.

Thus, it currently makes sense to defer this cost to individual projects, as there are quite a few commercial continuous integration and continuous deployment service providers willing to offer free accounts to open source projects, and these can also be used for the task of producing release artifacts. Projects also remain free to only publish source artifacts, relying on pip's implicit wheel creation and caching and the appropriate use of private PyPI mirrors and caches to meet the needs of end users.

For downstream platform communities already offering shared build infrastructure to their members (such as Linux distributions and conda-forge), it may make sense to offer Python wheel generation as a supported output option for cross-platform development use cases, in addition to the platform's native binary packaging format.


Comments powered by Disqus