The 'eu' in eucatastrophe – Why SciPy builds for Python 3.12 on Windows are a minor miracle
Published November 8, 2023
You've probably heard already that Python 3.12 was released recently. For a while already, we've been getting new feature releases every year, so perhaps this wasn't big news to you – though there's lots of interesting stuff in there!
It would appear even more "ordinary" that key packages (think
matplotlib etc.) in the ecosystem would have a release compatible with the
new Python version shortly after1, and broadly speaking, you would be right.
Not that there isn't a huge amount of dedicated maintainers (mostly
volunteers!) working tirelessly behind the scenes to make that happen,
but overall, it's a pretty routine situation.
However, behind the ordinary-seeming "SciPy released builds compatible with Python 3.12" hides an extraordinary story worth telling, because of how several unrelated, multi-decade-long timelines happened to clash in a way that could have very easily led to no Python 3.12-compatible releases for a long time.
To understand why this was such a lucky coincidence (though we tried our best to tip the scales, a lot of luck was necessary), we need to zoom out a bit and explore the big players involved in this situation. Fair warning: In the interest of brevity, the following recap is going to be incomplete and opinionated. From our perch on the shoulders of giants, let's take a quick look at the landscape.
We'll briefly shed some light on the following:
- Why is Fortran still used in so many places?
- How is that relevant to Python?
- Past struggles of NumPy/SciPy with vanilla Python packaging.
- What role conda-forge plays in this context.
If you feel you have a good-enough understanding of all this, or not much time, skip right past the background. 😉
… is almost unfathomably old in our day and age. Started in 1953, first appearing in 1958, it quickly became the most important programming language of its time, in an age when computing made several important evolutionary leaps. Most importantly for our story:
- A lot of scientific code was written in Fortran (meaning often: the accumulated math knowledge of many decades got crystallised into code), and:
- There was a huge number of Fortran compilers, and in the age before Open Source Software, all of those were proprietary.
Depending on how far behind the curtains you've dared to venture, you've probably at least heard of the concept of a compiler. Basically, computers have an exceptionally limited amount of operations at their disposal, and it's agonisingly painful to write anything in those primitives directly. The job of the compiler is to translate whatever the programmer is writing in a given compiled language, and turn it into something the computer can execute.
In many ways they're the closest to the hardware, which is why they often were hyper-specific to one CPU architecture, or one operating system – not least because you cannot write an operating system without a compiler, so the two were often tightly coupled. For example, Microsoft, from very early on, used its own compiler to create Windows and all the bits on top of it.
This coupling further compounded the tendencies towards closed source compilers, because the compiler source code might not only reveal secrets of the compiler itself, but also of the architecture or the operating system built with it. And since vendors were keen to capitalise on every advantage in an intensely competitive market, there was no incentive to share sources at all.
The pioneer to break that mould was the GNU compiler collection (GCC), which became one of the key pillars of the nascent Free & Open-Source Software (FOSS) movement in the 80s and 90s. The coupling to the operating system still stayed – in this case various flavours of Unix, most prominently Linux of course – but GCC became a truly universal compiler when it came to CPU architectures.
Let's visualise this quickly – we have a matrix of Operating Systems (OS) vs. CPU Architectures; often compilers were specific not just to one OS, but also 1-2 specific architectures (for a long time the most dominant architecture has been x86, but that has been changing in recent times). What GCC did was essentially cover an entire column of that matrix, rather than just 1-2 cells.
This matrix would be a lot larger if it included historical OSes and less common architectures, where support with the respective compiler was often in a 1:1 relationship (i.e. that combination would cover a single cell in the matrix). The matrix also does not cover which programming languages a given compiler is able to process, but for simplicity, you can picture C/C++ here.
Of course, GCC remains usable on macOS due to shared Unix roots, there are ways to make it work on Windows (through cygwin or MinGW), and the whole truth is way more complicated still. But as a first-order approximation, this shouldn't raise too many eyebrows.
Alright, after a full page about compilers, we can hear you thinking: "What does all this have to do with Python?!". And you're right, but just one last thing:
Once a program in C/C++/Fortran/etc. has been compiled, there are a lot of expectations built into that so-called "binary", which will expect very explicitly-sized inputs, read things from – and load things into – very specific CPU registers etc. Basically, at this point, the training wheels come off, and if you use said binary under even slightly different circumstances (different sized inputs, changed order of function arguments, different in-memory layout of some structure etc., much less another CPU type), you're going to have a bad time. This is the so-called Application Binary Interface (ABI), and any compiled program is subject to this.
We won't bore you with the history here, but the immediate thing to note is: you don't need a compiler to use Python. That's a blessing as well as a curse, because on the one hand, it dramatically lowers the difficulty of using it, and on the other hand, this makes Python slow compared to compiled languages.
Python's duck-typing means it will go through many layers of wrapping and fall-backs, before anything actually is executed. Sidenote: this process is taken care of by Python's interpreter, which – uh oh… – has itself been compiled2. Unsurprisingly this approach is slower than a compiled program that will just run with whatever instructions you've given it (and fail, if any part of the ABI is violated).
This slow-down doesn't matter so much for prototyping, or general scripting, but it becomes immensely limiting once there's medium-sized or larger amounts of data involved. The best-of-both-worlds solution to this problem was to have performant compiled code under the hood, but a nice, pythonic "wrapper" around it as the interface towards the programmer.
Remember all that scientific code that had been written in Fortran? Things like BLAS and LAPACK were already well-established standards for dealing with common problems in linear algebra, and so it was an obvious choice to reuse that when libraries like NumPy and SciPy wanted to expose such functionality to the Python world. SciPy additionally also used Fortran code to provide special mathematical functions and other mathematical tools like interpolation & integration – it's not like the basic math had changed since the 70s…
However, this came with a terrible price: Users needed a compiler to install the package. In the early Python days, when many people were on Linux distributions (which comes with GCC pre-installed), this was less of an issue. But it made it extraordinarily difficult to do this in a cross-platform way – particularly on Windows, it's not common for users to have a compiler installed, much less know how to use it.
It's even worse on the maintainer-side, because they suddenly need to become experts in the vagaries of different compilers, linkers and runtime libraries necessary on different platforms, and with sometimes wildly varying behaviour.
This situation also happens to be the fundamental problem in Python packaging. There's so much complexity hiding under the surface, that actually dealing with it in any kind of sane way is really difficult. For example, the C & C++ ecosystems never managed to standardise any sort of build & distribution mechanism, and since Python is wrapping around basically any flavour of C & C++ code, the problems in Python packaging are a superset of what's already a hugely pressing issue in C/C++.
The packaging world in Python has had to continually reinvent itself in order to solve these issues. It's instructive to check out this historical overview, to get a feeling of the many iterations things have gone through.
One of the biggest problems over the course of this evolution has been the question about the right way to distribute a package. Letting users just download the sources is problematic, because:
- packages can take a lot of time to build.
- if the packages is not just pure Python, the user needs a working compiler setup (and the source code needs to be compatible with that setup!), which is a huge usability hurdle.
- there's no reliable metadata, because even something as fundamental as the
necessary third-party dependencies used to get populated only through running
The alternative – distributing pre-compiled artefacts – is problematic for many reasons too, and especially fragile in the face of the above-mentioned ABI. Given the impact of breaking the ABI (random crashes, heisenbugs, etc.), doing "binary" distribution haphazardly is not an option.
Eventually, a feasible approach for such distribution emerged in the form of the "wheel" format, which essentially creates a bubble for each package that brings along all the required libraries, but hides them from others through a clever mechanism. While wheels still have some downsides (especially on the maintainer-side), they are a massive improvement over the situation that existed before, and nowadays, most Python packages are installable through wheels from PyPI.
But wheels were not around yet at the time when the Scientific Python community needed to solve some problems they were facing (as well as not versatile enough conceptually for solving all the key issues involved). This blog post by one of the founding contributors of NumPy and SciPy is full of relevant anecdotes:
We in the scientific python community have had difficulty and a rocky history with just waiting for the Python.org community to solve the [packaging] problem. With
distutilsfor example, we had to essentially re-write most of it (as
numpy.distutils) in order to support compilation of extensions that needed Fortran-compiled libraries.
Despite being the product of heroic efforts by many very bright people,
numpy.distutils is generally regarded as a bandaid, and has been in minimal
maintenance mode for years.
This is largely due to fundamental limitations inherent in only having a build
setup.py), without a way to precisely control in advance the
conditions under which a package gets built.
Again from that blog post:
numpy.distutilsreplaces most of the innards of
distutilsbut is still shackled by the architecture and imperative approach to what should fundamentally be a declarative problem.
In many ways, this is what led to the creation of conda:
Therefore, you can't really address the problem of Python packaging without addressing the core problems of trying to use
distutils(at least for the NumPy stack). The problems for us in the NumPy stack started there and have to be rooted out there as well. This was confirmed for me at the first PyData meetup at Google HQ, where several of us asked Guido [van Rossum; Python's BDFL at the time] what we can do to fix Python packaging for the NumPy stack. Guido's answer was to "solve the problem ourselves".
conda & conda-forge
Conda was designed to take a more holistic view of all the things that are required for packaging – including the non-Python bits – and unsurprisingly, done in a declarative way. The basis for this is a "recipe" that explicitly defines relevant quantities (e.g. what needs to be present in the build environment), before the first line of any build script is ever run.
This introduced an unfortunate bifurcation in the Python world, because other Python tools like pip cannot install conda-packages, yet what conda had done did not easily fit into the standardisation-by-consensus model3, primarily because many things essential to conda were considered "out of scope" by the wider Python packaging community.
However, for users and maintainers of packages that are affected by some of the deep-seated issues in Python packaging, it provided a much more powerful solution and got adopted widely – especially in scientific computing, though by far not everywhere. While the recipes conda uses are the basis for many key features (e.g. having enough control to ensure that the ABI is kept consistent, or that packages are recompiled where necessary), this means in turn that integrating a given package into the conda world requires creating such a recipe in the first place.
Given the size of the ecosystem, not even a company like Anaconda4 could hope to integrate everything that users wanted, and over time, the community-driven conda-forge channel became the place to do this integration work. Anyone can submit recipes for a package that is missing, or provide fixes for those that are already being built on so-called "feedstocks".
One of the things that complicates matters is that conda-forge – as a philosophy – is aiming to support all common platforms, and even some less common ones. This multiplies the kind of problems a Linux distribution might have by at least ~three, because very often a different set of problems will appear for macOS and Windows. To have any chance at success, conda-forge uses the platform defaults (compiler, standard library, ABI, etc.) wherever possible.
Conda-forge is almost 100% volunteer-run, and dependent on public CI resources like those provided by Azure Pipelines, as well as Anaconda footing the bill for the hosting and download traffic of all the hosted artefacts. On top of that, conda-forge cannot arbitrarily package things where the licence does not allow it. There is no hidden layer for build tools; anyone can download and install the packages that the infrastructure is made of (or the underlying docker images), and that means conda-forge cannot use proprietary compilers5, but must use what's freely available.
Remember how we discussed above that all Fortran compilers are proprietary (except gfortran as part of GCC), and how GCC is not directly usable on Windows? This has been a huge issue for conda-forge, especially as the expectation of general support on Windows became more and more wide-spread. For a very long time, conda-forge was not able to build SciPy at all due to this. Instead it had to rely on packages from the Anaconda channels6, until some particularly talented contributors hacked together an unholy amalgamation of MSVC (c.f. the default ABI on Windows) and gfortran, first in SciPy itself and later on the feedstock in conda-forge.
- SciPy uses a lot of Fortran code, having incorporated existing and standard implementations (like BLAS & LAPACK) for mathematical computations.
- This did not play well with Python's native build tools, leading to the
numpy.distutils, which is however considered a bandaid and in minimal maintenance mode.
- The Python packaging world has been continually reinventing itself to overcome burning problems that are due to the immense complexity being wrapped by Python packages.
- A parallel ecosystem has arisen around conda & conda-forge, for packages and users that were underserved by the standard tools for Python packaging.
This is where our story starts in earnest.
Going back for many years already, the situation around
setuptools had been a source of great pain in the Python packaging world.
The two (in close interaction) represented the de-facto standard way of
building Python packages, but changing anything about them was extremely hard
because of how many use-cases had to be considered, how much everything had
grown organically, and so on.
The next evolutionary step turned out to be the combination of PEP517 / PEP518,
which promised a level playing field between various possible build tools, and
a way to specify the required information through a new interface
This eventually gained enough traction that
distutils got scheduled for
removal from the Python standard library per version 3.12.
However, this had far-reaching consequences for all the things in Python-land
which had grown around
distutils, and even more so for things layered on top.
The maintainers of
numpy.distutils took the decision that they're not going
to maintain this layer going forward – even though
setuptools ingested much
distutils interface, it did so in ways that were subtly incompatible
numpy.distutils did, and ultimately no-one wanted to keep
maintaining this old pile of hacks any longer than absolutely necessary.
This kicked off a flurry of work – already years ago, because NumPy and SciPy are under excellent stewardship – to find a better solution. SciPy in particular took a long, hard look at the landscape, and decided to go with Meson as a build tool, which had a nice Python-style feel to it, and less of the historic baggage or cumbersome DSL of CMake. This was a far-reaching and forward-looking decision at the time, not least because Meson did not yet support all the things that would be necessary to pull things off.
To put this effort into context, all this required many, many engineering months of highly talented people over the last ~2.5 years. Unbeknownst to most, we are benefitting from the great foresight of the SciPy and NumPy developers in tackling the issue this far in advance, rather than having an "uh oh" moment once Python 3.12 is released.
It's necessary to note that Meson as a build tool has a much broader audience than Python developers – it is used widely for C & C++ projects for example7. As part of an overall design ideology to bring some much-needed sanity to this space (and especially the world of C/C++ where no uniform solution exists), Meson will not let you do things that are sure to go wrong in all but the most expert hands.
In particular, Meson was going to refuse to accept the MSVC+gfortran combination that was in use in conda-forge. In fact, it looked like we were going to be completely boxed in. As of Python 3.12:
distutilswould be gone, and contemporary
- Meson would refuse to work with the hack conda-forge was using.
- There were no free (and ABI-compatible) Fortran compilers on Windows at all.
Fortran, the revival
While Fortran has long been the butt of the joke in IT departments the world over, in a curious twist of fate, it has seen a dramatic resurgence over the last few years. While the reasons for this are not exactly obvious (at least to us!), a few possible explanations are:
- It has relatively simple syntax, and maps well to the structure of mathematical formulae.
- It produces performant code, especially in combination with parallelisation frameworks like OpenMP.
- Due to the large amount of Fortran code in scientific computing, and the promised performance gains through GPUs, it became an attractive target for supporting GPU computations.
As such, there was renewed vigour in the Fortran compiler space, and several important developments happened in short succession. For example, PGI / NVIDIA open sourced a version of their compiler called pgfortran with a new backend based on LLVM (more on that below), which later turned into what's now known as "classic" Flang. In the process of trying to upstream this into LLVM itself, the compiler got rewritten completely under the name f18, which later turned into "new" Flang that eventually got merged into LLVM itself. Pretty much at the same time, another group started developing a Fortran compiler based on LLVM: LFortran.
LLVM & Clang
Further up we had looked at a matrix of Operating Systems versus CPU architectures, and how most compilers tended to only cover 1-2 "cells", whereas GCC covered essentially all possible architectures on Linux – i.e. a whole column of the matrix.
As it happened, the last dimension of universality was tackled by another large project that grew from a research project into a fully cross-architecture and cross-platform compiler infrastructure8: LLVM, and its flagship compiler Clang. This was an absolutely massive undertaking, but due to its cross-platform nature, permissive licensing, and modular infrastructure, has become the focal point of efforts around all kinds of compiler development.
Without going into the technical details, LLVM provides several so-called "Intermediate Representations" (IRs), which are somewhere in between machine code and the code that most programmers write, but in a way that is language-agnostic. In particular, it would be possible to target the LLVM IRs from Fortran, and then automatically benefit from all the "backends" that do most of the heavy optimization work between the IRs and the actual CPU architecture.
This made LLVM an attractive foundation for these new Fortran compiler efforts, and both incarnations of Flang as well as LFortran followed this approach, though with different aims in what they built on top.
All of these FOSS Fortran compilers attracted great hope to unblock the general situation regarding the lack of free Fortran compilers on Windows. Classic Flang already had preliminary support for Windows, but this never fully took off, not least as the lion's share of resources behind Flang was refocussed on getting the rewritten version merged into LLVM itself.
But these were not the only changes happening in this space. For example, Intel did a major overhaul of their Fortran compiler (also basing it on LLVM), and started making it freely available, though not open source. Unfortunately this precludes conda-forge from using it, unless Intel themselves packages the compilers for us; we asked, but there was no timeline.
Furthermore, through the Mingw-w64 project, it slowly started to become possible to use gfortran with Microsoft's "Universal C Runtime" (UCRT), which is essentially the biggest hurdle in achieving ABI-compatibility on Windows. However, switching out the underlying C standard library would represent a major overhaul in the MinGW stack for windows within conda-forge, and was also not something we could count on to be ready in time.
For people on the sidelines like ourselves, this turned into something akin to "Waiting for Godot" – a seemingly endless period as things kept stretching and stretching and stretching out further into the future. To give some perspective, things already seemed "not too far out" 3-4 years ago.
At this point you may be asking yourself: "wait a second… how did SciPy actually switch to Meson, if there are no Fortran compilers for Windows?"
How upstream SciPy does it (for now)
While SciPy has some amount of funding through grants and such, and could conceivably pay for a compiler licence to use for producing wheels, the requirements of building things in CI regularly – to avoid the otherwise inevitable bitrot – would cost a lot and not be a judicious use of the limited funds available.
What ended up happening is that one lone developer managed to custom-build a MinGW-based toolchain, carefully adapted to use the right-sized integers and to conform to the MSVC ABI. This is enough to pass through Meson's sanity checks (after all, GCC is a consistent toolchain), however it is again subtly-yet-crucially different from the requirements that would be necessary for wide-spread use in conda-forge.
conda-forge and the migration problem
One final aspect in the whole saga is that conda-forge has some additional constraints in terms of how large-scale maintenance efforts work, e.g. something like rolling out a new Python version. Essentially, this means rebuilding all Python-related feedstocks in the order how they depend on each other, and the only reasonable way to do it at scale is to do all OSes and architectures at the same time, as they're all built simultaneously per feedstock.
So it looked conceivable that we'd end up in a situation where:
- We have no Windows builds for SciPy, because there's no Fortran compiler conda-forge can use.
- Not being able to rebuild SciPy for Python 3.12 on Windows would either hold up the migration for at least ~1000 packages that depend on SciPy on all platforms, or we would have to drop Windows from the Python migration completely, which would be similarly disruptive and a lot of effort to fix later.
A pious hope
So while the
distutils removal was coming down the line like an oncoming
freight train, all we could try to do was cross our fingers that one Fortran
compiler would be ready in time (most likely either llvm-flang or LFortran), and
be as prepared as possible in terms of having all the other pieces ready to go.
LLVM is a big baby, and keeping all the pieces involved up-to-date is a
non-trivial exercise already9, but we put extra effort into trying to
identify potential problems with Flang early on, both in our own
as well as
But no matter how motivated we were to pursue this, the iron-clad constraints of FOSS remain at play: no-one can take care of everything, much less keep all the required knowledge in their head. That means we need other people's input & feedback – and when everyone's a volunteer10, things can take a while.
As it happened, the release of LLVM 17.0 was coming up, which would be the first
release where Flang had matured enough to remove the
flag, and was estimated
to be roughly "0.8 level maturity".
As soon as we had the release candidates built in conda-forge, we tried to throw
it at SciPy, only to realise that Meson did not support the new Flang yet.
Even worse, it wasn't clear which ABI Flang was using (as mentioned above, conda-forge uses the default ABI on Windows everywhere), not least because the Clang driver which Flang uses under the hood can target both modes. Unfortunately, it wasn't clear which mode was being chosen, and generally it looked like Windows support was (as often happens in FOSS) the least mature platform.
To top it off, there are also two different standard libraries, two different runtime libraries and two different linkers to consider in all this – all of which had the potential to lead to subtle breakage. At this point, we still hadn't compiled more than a "Hello World!" example with Flang, so it wasn't even clear that it could compile all of SciPy, much less pass the test suite...
The final push came once SciPy got ready
to remove support for
setup.py, the last link to the "old world" before the
With some kind hints from the Meson devs and a bit of persistence, we managed to
teach Meson to be able to handle llvm-flang just enough to compile something.
We eventually passed the first build, and after some more wrangling with Meson,
the first installation too.
What happened next felt like nothing short of magical – while we were expecting test failures, hangs or even crashes, the test suite… just passed 100%?!
........................................................................ [ 99%] .................................................s....... [100%] = 54987 passed, 2866 skipped, 245 xfailed, 11 xpassed, 1 warning in 979.88s (0:16:19) = Resource usage statistics from testing scipy-tests: Process count: 43 CPU time: Sys=0:01:20.6, User=0:29:59.3 Memory: 1.7G Disk usage: 1.8K Time elapsed: 0:16:24.6
55'000 tests, and not a single failure? When does that ever happen on the first try after major surgery?
Somehow, armed only with a pre-release of a still-experimental compiler and a hastily patched-up build system, we managed to dodge the freight train, and the migration in conda-forge could progress past SciPy with no more than a few days delay, rather than being stuck for months. In the end, the respective SciPy builds in conda-forge were available two days after the release of Python 3.12.0.
At the end of this journey, we can only marvel at how long and winding the ways have been that lead us here. Depending on where you draw the line, hundreds or even thousands of person years of effort went into the ingredients that were necessary for us to achieve the final result (SciPy, Meson, Flang, LLVM, LAPACK etc.). We cannot realistically thank everyone, but a few call-outs nevertheless:
- Ralf Gommers, for tirelessly positive support and for outstanding stewardship of NumPy and SciPy, and driving the massive effort to move SciPy to Meson.
- Isuru Fernando, for building & maintaining some of the trickiest pieces in conda-forge, and consistently helping out with keeping them running or building on top of them, as well as reviewing with incredible expertise.
- The Meson developers, for a clean build system with a solid design philosophy.
- The Flang developers, for providing the world with an open source Fortran compiler that also works on Windows.
We are also very grateful that this work was supported in part by a grant from NASA to SciPy, scikit-learn, pandas and NumPy under the NASA ROSES 2020 program.
PS. In case you're still wondering what the "eu" in eucatastrophe means, it's a neologism coined by J. R. R. Tolkien from Greek ευ- "good" and καταστροφή "sudden turn". Quoting from the link above: "In essence, a eucatastrophe is a massive turn in fortune from a seemingly unconquerable situation to an unforeseen victory, usually brought by grace rather than heroic effort. Such a turn is catastrophic in the sense of its breadth and surprise and positive in that a great evil or misfortune is averted."
or even before, based on Python release candidates! ↩
and thus also has an ABI, if you want to speak to Python from another compiled language. ↩
coupled with fears (at least initially) that a single company would "take over" such a critical piece of the ecosystem. ↩
which grew around the needs conda addresses, and is the main driver behind the tool. ↩
because their licenses forbid redistribution – which conda-forge wouldn't be able to prevent from happening. ↩
Anaconda can afford to have their own build infrastructure, and is able to enter into contracts with compiler vendors. ↩
if you're on Linux, much of your graphics stack is being built with Meson nowadays. ↩
not least because Apple hired its founder, with the ostensible goal of having a compiler that was not subject to GCC's license. ↩
aside from the fact that conda-forge cannot build the whole enchilada in one go within the 6h time limit on the relatively modest CI agents that Azure provides for free, any substantial change in our LLVM setup needs a lot of care due to the sheer number of packages that can be affected by something so deep in the stack. ↩
i.e. people who have lives and their own interests, or who're just plain busy with other urgencies or things they care about more. And even though there are people employed to work on some of these projects, they almost always aren't paid to solve your problem. ↩