I appeared (was heard?) on the latest episode of The Changelog, and they very kindly sent me the embed snippet for my blog. So to celebrate, I decided to post a blog and use their embed snippet.
Those who have been following Python development on Windows recently will be aware that I’ve been actively redeveloping the installer. And if you’ve been watching closely you’ll know that there are now many more ways to install the official python.org release than in the past, not even including distributions such as WinPython or Anaconda.
In this post, I’m going to discuss each of the ways you can install the official releases of Python (since version 3.5), provide some context on when and why you would choose one over another, and discuss the positives and negatives of each approach.
Historically, there was one single MSI installer available that was intended to cover the needs of all Python users.
This installer would allow you to select a target directory and some features from its user interface or the command line (if you know the magic words), and would generally install the full distribution with all entry points (shortcuts, etc.).
Unfortunately, due to the nature of how MSIs work, there are some limitations that affect the user experience. The most major of these is the fact that MSIs cannot decide whether elevate as part of the install – it has to be hardcoded. As a result, the old installer always requires administrative privileges just in case you choose to install for all users. This prevents installation of Python on machines where you do not have full control over the system.
Secondly, while Python is often seen as one monolithic package, it is actually made up of a number of unrelated components. For example, the test suite is often not required for correct operation, nor is the documentation and often the development headers and libraries. While MSIs do support optional features, they tend to encounter issues when performing upgrades between versions (such as forgetting which options you had selected), and in general you always need to carry around the optional components even if you’re never going to install them.
Finally, some operations that are not simple file installations can be complicated. For example, when pip is installed or the standard library is precompiled, the MSI executes a background task rather than normally installing files. Without careful configuration of the MSI, these files will not be properly uninstalled or repaired, and issues in the extraction process can cause the entire task to fail. At worst, the uninstall step could fail, which can make it impossible to uninstall Python.
The issues described above have been addressed by the installers available since Python 3.5. However, there are also other uses for Python that do not lend themselves to a regular installer. For example, applications that want to include Python as a runtime dependency may not want to install a global copy of Python, build machines may require semi-public but non-conflicting installs of different versions, and platform-as-a-service web hosts may not allow normal installers.
Since Python 3.5.2, the official Python releases have been made available as executable installers, embeddable ZIP packages, nuget packages and Azure site extensions. There are also a range of third-party distributions that include the official Python binaries, along with other useful tools or libraries.
How do I know if a third-party distribution has the official binaries? Find the install directory, right-click each of the exe, dll or pyd files and select Properties, then Digital Signatures. If the signature is from the Python Software Foundation, it’s the official binary and has not been modified. If there is a different signature or no signature, it may not be the same as what is released on python.org.
The executable installers are the main way that users download Python, and are the featured downloads at python.org. I think of these as the Python Developer Kit.
These installers provide the most flexible user interface, include all dependencies such as system updates and the Python launcher, generate shortcuts for the interpreter, the manuals and the IDLE editor, and correctly support upgrades without forgetting about feature selection.
Two versions of the executable installer are available for any given release – one labelled “executable installer” and the other “web-based installer”.
The web-based installer is typically a small initial download (around 1MB), which gets you the installer UI shown above. After you have selected or deselected optional components, the minimum set of packages necessary to install Python will be downloaded and installed. This makes it easy to minimize overall download size since unused or unnecessary components are never downloaded, though it does require that you be connected to the internet at install time. (There’s also a command-line option to download all the packages you may ever need, which will then be used later instead of downloading them over and over again.)
The other installer includes the default set of features in the EXE itself. As a result, the initial download is around 30MB, but in most cases you can install without requiring any further internet access. For a single installation, your download will likely be 3-5MB larger compared to using the web-based installer, but if you use it to install on multiple machines then you’ll likely come out ahead.
Both executable installers result in identical installations and can be automated with identical command-line options. As I mentioned above, I think of this as the Python Developer Kit, which is why there are optional features to download debugging symbols or a complete debug build, which are not available in any other options. The Python Developer Kit provides everything necessary for someone to develop a complete Python application.
What about having a single MSI installer? There’s a section coming up about this. Just keep reading.
If the executable installer is the Python Developer Kit, then the embeddable package is the Python runtime redistributable. Rather than trying to be an easy-to-use installer, this package is a simple ZIP file containing the bare minimum of Python required to run applications. This includes the python[w].exe executables, the python35.dll (or later) and python3.dll modules, the standard library extension modules (*.pyd), and a precompiled copy of the standard library stored in another ZIP file.
The resulting package is about 7MB to download and around 12MB when extracted. Documentation, tools, and shortcuts are not included, and the embeddable package does not reliably build and install packages. However, once your application is ready, rather than instructing users to install Python themselves, you can include the contents of this package in your own installer. (For example, Microsoft’s command-line tools for Azure will likely do this, and installers created using pynsist can include this package automatically.)
Using the embeddable package allows you to distribute applications on Windows that use Python as a runtime without exposing it to your users. By default, a configuration file is also included to force the use of isolated mode and prevents environment variables and registry settings from affecting it (python36._pth on Python 3.6; pyvenv.cfg for Python 3.5). On Python 3.6 this file can also specify additional search paths. If your application is hosting Python, you can also choose not to distribute python.exe or any extension modules that are not used in your application.
There is no support for pip, setuptools or distutils in the embeddable package, since the idea is that you will develop against the Python Developer Kit and then lock your dependencies when you release your application. Depending on the installer technology you are using for your application, you will probably vendor any third-party packages by copying them directly into the directory with your Python code.
See this blog post for more information about how to take advantage of the embeddable distribution.
Nuget is a packaging technology typically used on Windows to manage development dependencies. There are many packages available as source code or pre-built binaries, mostly for .NET assemblies, as well as build tools and extensions.
There are four Python packages available on nuget, released under my name (steve.dower) but built as part of the official python.org releases. The packages are:
- python – the latest 3.x 64-bit
- pythonx86 – the latest 3.x 32-bit
- python2 – the latest 2.x 64-bit
- python2x86 – the latest 2.x 32-bit
These may be referenced by projects in Visual Studio or directly using nuget.exe to easily install a copy of Python into a build directory. It will typically install into a directory like packages\python.3.5.2\tools\python.exe, though this can often be customised.
rem Install Python 2.7 nuget.exe install -OutputDirectory packages python2 rem Add -Prerelease to get Python 3.6 nuget.exe install -OutputDirectory packages -Prerelease python rem More options are available nuget.exe install -Help
The contents of the nuget package is somewhere between the full installation and the embeddable package. The headers, libs and pip are included so that you can install dependencies or build your own modules. The standard library is not zipped, but also does not include the CPython test suite or libraries intended for user interaction. Operating system updates are not included, so you will need to ensure your build machine is up to date before using these packages.
There is no configuration in these packages to restrict search paths or environment variables, as these are very important to control in build definitions. As a result, there is a high likelihood that a regular installation of Python may conflict with these packages. In general, it’s best to avoid installing Python on build machines where you are using these packages. If you need a full installation, avoid using the nuget packages or test for conflicts thoroughly. (Note that conflicts typically only occur within the same x.y version, so you can safely install 2.7 and use the 3.5 nuget packages.)
Azure Site Extensions
Note: This particular package is released by Microsoft and is managed by my team there. The Python Software Foundation is not responsible for this package.
Azure App Service is a platform-as-a-service offering for web services (including web apps, mobile backends, and triggered jobs). It uses site extensions to customise and enhance your web services, including a range of Python versions to simplify configuration of Python-based servers.
Because web services are sensitive to even the smallest change in a dependency, each version is available as its own package. This allows you to be confident that when your site uses one of these it is not going to change without you explicitly updating your site. The current packages available at time of writing are:
The contents of these packages is almost entirely unmodified from the official python.org releases. Some extra files for correct installation, configuration and behaviour of the web server are included, as well as copies of pip, setuptools, and certifi. Occasionally a package will include targeted patches to fix or work around issues with the platform, but we always aim to upstream fixes as soon as possible. Under the hood, these are simply nuget packages that can also be installed using nuget.exe on any copy of Windows.
C:\> nuget.exe list python -Source https://www.siteextensions.net/api/v2/ python2711x64 18.104.22.168 python2712x64 22.214.171.124 python2712x86 126.96.36.199 python351x64 188.8.131.52 python352x64 184.108.40.206 python352x86 220.127.116.11
Visit aka.ms/PythonOnAppService for the most up-to-date information about how to use these packages on Azure App Service.
While that covers the current set of available installers, there are some further use-cases that are not as well served. In this section I will briefly discuss the cases that I am currently aware of and their status. There are no promises that official installation packages for these will ever be produced (bearing in mind that Python is developed almost entirely by volunteers with limited free time), but there is also nothing preventing third-parties from producing and distributing these formats.
Are you already distributing Python in any of these formats? Let me know and I’m happy to link to you, provided I’m not concerned about the contents of your distribution.
Nuget package for source/runtime dependency
Earlier I discussed the nuget packages as build tools, but the more common use of nuget packages is for build dependencies. Normally a project (typically a Visual Studio project, but nuget can also be used independently) will specify a dependency on a source or binary package and obtain build steps or configuration from a known location within the package.
Providing a nuget package containing either the Python source code or the embeddable package may simplify projects that host the runtime. These would predominantly be C/C++ projects rather than pure Python projects, but some installer toolkits may prefer a ready-to-embed nuget package rather than a plain ZIP file.
There has not been much demand for this particular format. In general, a C/C++ project can make equally good use of the current nuget packages, and would require those for the headers and libraries anyway, while the embeddable package is not always suitable for installation completely unmodified. These reduce the value of a dependency nuget package to nearly zero, which is why we currently don’t have one.
Universal Windows Platform
The Universal Windows Platform is part of Windows 10 and specifies a common API set that is available across all Windows devices. This includes PCs, tablets, phone, IoT Core, XBox, HoloLens, and likely any new Windows hardware into the future.
Providing a UWP package of Python would allow developers to distribute Python code across all of these platforms. Indeed, the team behind IoT Core have already provided their version of this package. However, as the API set is not always compatible with the Win32 API, this task requires supporting a new platform within Python (that is, sys.platform would return a value other than 'win32'). Currently nobody has completely adapted Python for UWP, added the extensions required to access new platform APIs, or fully implemented the deployment tools needed for this to be generally useful (though the IoT Core support is a huge step towards this).
System administrators will often deploy software to some or all machines on their network using management tools such as Group Policy or System Center. While it is possible to remotely install from the executable installers, these tools often require or have enhanced functionality when the installer is a pure MSI.
Unfortunately, the issues and limitations of MSI described at the start of this post still apply. It is not possible for an MSI to install all required dependencies, create an MSI that can run without administrator privileges, and robustly ensure that upgrade and remove operations behave correctly. However, it would be possible to produce a suitable MSI and installation instructions for the limited use case of administrative deployment. Such a package would likely have these characteristics:
- Fails if certain operating system updates are missing
- Always requires administrator privileges
- Only allows installation for all users
- Only allows configuration at the command line (via msiexec)
- Requires a separate task to precompile the standard library and install pip
- Requires additional cleanup task when uninstalling
- Prevents the executable installer from installing for all users
System administrators would be responsible for following the documentation associated with such an MSI, and I have no doubt that most are entirely capable of doing this. However, as this would not be a good experience for most users it cannot be the default or recommended installer. I’m aware that there are some people who are grieved by this, but interactive installs are vastly more common and so take priority when determining what to offer from python.org.
Installing Python on Windows has always been a fairly reliable process. The ability to select precisely which version you would like without fear of damaging system components allows a lot of confidence that is not always available on other platforms. Improvements in recent releases make it easier to install, upgrade and manage Python, even for non-administrator users.
We have a number of different formats in which Python may be obtained depending on your intended use. The executable installers provide the full Python Developer Kit; the embeddable package contains the runtime dependencies; nuget packages allow easy use of Python as a build tool; and site extensions for Azure App Service make it easier to manage Python as a web server dependency.
There is also potential to add new formats in the future, either through third-party distributions or as new maintainers volunteer their time towards core development. For an open-source project that is run almost entirely on volunteer time, Python is an amazing example of a robust, trustworthy product with as much flexibility as any professionally developed product.
Discussion of this post is welcome here in the comments. If you are having issues installing Python, please file an issue on bugs.python.org.
If you haven’t already, I recommend going back and reading part one up until the point where it says to come back here. That will fill you in on the background.
This is the point where I ought to make the “seriously, I’ll wait” joke, knowing full well that nobody is going to read the old post. Instead, I’m going to write one sentence, and if you understand the words in it, you have enough background to continue reading.
MSVC 14.0 (and later) and the UCRT give us independence from compiler versions, provided you build with /MT and force ucrtbase.dll to be dynamically linked.
If you scratched your head about any part of that, go and read part one. You’ll thank me in about two seconds.
The /MT Problem
Previously, I discussed some of the problems that arise when compiling with the /MT option. Mostly, because the option needs to be specified throughout all your code, it was going to cause issues with static libraries that had already been built.
Well, we found one more problem.
I want to stop for a second to thank Christoph Golhke for all his help through the 3.5.0 RCs.
For those who don’t know, Christoph maintains an epic collection of wheels for Windows. Every time you have trouble installing a package with pip, it’s worth visiting his site to see if he has a wheel available for it. After downloading the wheel, you can pass its path to pip install and you’ll get your package.
As we kept making changes to the build process, Christoph kept updating his own build steps and testing with literally hundreds of packages. His feedback has directly led to these changes to Python 3.5 that will make it much easier for everyone to have solid, future-proof builds of Python and extension modules.
Ready for the gory technical details of this new issue? Here we go.
Previously, we were statically linking functions from vcruntime140.dll into each extension module built for Python 3.5. This included, among other things, the DLL initialisation routines.
On the positive side, each extension module is now completely isolated from any others with respect to initialization, locale, debugging handlers, and other state.
On the negative side, it turns out initialization is a limited resource.
The CRT has a number of features that are thread-safe, such as errno. Each thread has its own errno, which means you do not need to perform locking around every operation that may set or use it.
To implement this per-thread state, the CRT uses fiber-local storage. On Windows, fibers are a form of cooperative multithreading that work within threads (one process may contain many threads, one thread may contain many fibers), so the CRT uses fiber-local storage to ensure the finest-grain handling. If it used thread-local storage, different fibers would see the same values, and if it used a normal global, all threads would see the same value.
Fiber-local storage slots are allocated using FlsAlloc(). If you read that doc carefully, you’ll see that a potential return value (error) is FLS_OUT_OF_INDEXES, which means you have exhausted the current process’s supply of fiber-local storage.
The CRT doesn’t like it if you’ve run out of fiber-local storage, because that means a lot of its stateful functionality will be broken. So it aborts.
How many slots do we get? It isn’t documented (so it could change at any time), but here’s how we can test it:
>>> import ctypes
>>> FlsAlloc = ctypes.windll.kernel32.FlsAlloc
>>> list(iter(lambda: FlsAlloc(None), -1))
[7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127]
On my machine, I can call FlsAlloc 121 times after starting the interpreter before it runs out of indices.
When each extension module has its own isolated initialization routine, each one will call FlsAlloc. Which means once I’ve loaded 121 extension modules, the 122nd will fail.
Going back to /MD
The only reasonable fix for this is to switch back to building both Python and extensions using /MD. So we’ve done that.
That means that vcruntime140.dll is now a runtime dependency and must be included with Python 3.5. So we’ve done that.
That means that every extension module for Python 3.5 has to depend on vcruntime140.dll and therefore must always be build with MSVC 14.0.
Well, I don’t think so.
Initialization is no longer isolated. You can load as many extension modules as you like and they’ll all use the same fiber-local storage index because vcruntime140.dll is already initialized.
But what happens when someone comes along with MSVC 15.0 (note: not a real version number) and builds an extension that depends on vcruntime150.dll? When they publish a wheel of that extension and someone installs in on Python 3.5.0? That version of Python only includes vcruntime140.dll, so there will be an unresolved dependency and the extension will fail to load.
We’ve also made a change to fix this. The hint was in that last sentence: the extension has an unresolved dependency.
It is no longer (and technically never was) Python’s responsibility to provide dependencies required by extension modules. If you release an extension module, you are responsible for including any dependencies, or instructing your users on how to get them.
But simultaneously, we really don’t want to break python setup.py bdist_wheel upload (despite upload being insecure and you should be using twine or another uploader). If someone has MSVC 15.0 and builds a wheel using that version, somehow their users need to get vcruntime150.dll onto their machines.
So we fixed distutils.
When you build an extension using distutils (which is how most extensions are built, even if you think you’re using another package), we know which compiler version is being used. This means we know which version of the CRT you are using and which version of vcruntime###.dll your extension depends on.
We also know which versions of vcruntime###.dll shipped with the target version of Python. Currently, this is a set that contains only vcruntime140.dll (it’s also an internal implementation detail, so don’t start depending on it), and for the lifetime of Python 3.5 it will only contain that value.
When you build an extension with MSVC, we find the redistributable vcruntime###.dll in your compiler install and if it is not in the set of included files, it is copied into your build.
Currently it’s hard to see this in action, but once we have a newer MSVC to play with you’ll be able to see it. Whenever you build an extension module, it will get the new vcruntime alongside it.
Because of how DLL loading works, if that version of the DLL has not been loaded yet then the one adjacent to the extension module will be used. If it has been loaded, the one that is currently in memory will be used.
So really, we only need to include it once. The trick is making sure that it is loaded first. Ultimately, the only reliable way to do this is to include it everywhere.
Python 3.5 is always going to include vcruntime140.dll at this stage (though it may move from the top-level directory into DLLs at some point), so extensions built with MSVC 14.0 will always have it available.
Extensions built with a newer version of the compiler will include their own dependencies. As they should.
Perhaps Python 3.6 will be built with a newer version of the compiler. We can still include vcruntime140.dll, so that extensions can be built with MSVC 14.0, as well as including the newer version. We simply add the new one to the list in distutils and builds stop including it.
In this way, we can ensure that, at least for simple extensions (the majority), the easiest path will remain compiler-version independent.
Overriding the build process
One of the neat things about making the fix in the build process is that there are ways to override it. We can also make changes and improvements in bugfix releases if needed. It’s also fairly easy for developers to patch their own system if they need to customize things further. However, a few simple customisations are available without having to touch any code.
There are three stages in the build process with respect to copying vcruntime (henceforth, “the DLL”).
The first stage is locating the redistributable DLL. For regular installs of MSVC, these are in a known location, so once we’ve found vcvarsall.bat we find the DLL relative to that path.
The second stage is updating build options. If we didn’t find the redistributable DLL, we will statically link it exactly as described previously. This is not recommended, but it is still allowed and supported (and there is an opt-in described below).
The third stage is copying the DLL to your build directory.
Here are a few tricks you can use to customize your build. These are current as of Python 3.5.0rc4, but could change.
Use a different DLL
To specify your own path to the redistributable DLL, use a complete VS environment (such as the developer command prompt), set the environment variable DISTUTILS_USE_SDK=1, and then set environment variable PY_VCRUNTIME_REDIST to the full path of the DLL to use.
C:\package> "C:\Program Files\Microsoft Visual Studio 14.0\VC\vcvarsall.bat"
C:\package> set DISTUTILS_USE_SDK=1
C:\package> set PY_VCRUNTIME_REDIST=C:\Path\To\My\vcruntime140.dll
C:\package> python setup.py bdist_wheel
Note that this doesn’t override any of the following steps, so if your version of Python already includes vcruntime140.dll, then it won’t be copied. However, there’s no reason you can’t specify a different name here, and then it will be copied (but then you have to deal with a different name file, and might be missing actual dependencies… not really a great idea, but you can do it).
Always statically link
The build options are updated to statically link the DLL if PY_VCRUNTIME_REDIST is empty.
If you set DISTUTILS_USE_SDK but not PY_VCRUNTIME_REDIST, you will get statically linked vcruntime140.dll. This is probably the biggest surprise from the change.
Always dynamically link, but don’t copy
Probably the most useful customisation, you can choose to dynamically link the DLL but not copy it to your build output.
Since we check the name already, any file named vcruntime140.dll will never be copied anyway. However, with a future compiler you may already know that you are distributing the correct files and do not need distutils to help.
First, you apply the opposite of the previous customization: build options are updated to dynamically link the DLL if PY_VCRUNTIME_REDIST is not empty.
However, the DLL is only copied to the output directory if os.path.isfile is True.
The tests are deliberately different. It means you can force dynamic linking without the copy by setting PY_VCRUNTIME_REDIST to anything other than a valid file:
C:\package> "C:\Program Files\Microsoft Visual Studio 14.0\VC\vcvarsall.bat"
C:\package> set DISTUTILS_USE_SDK=1
C:\package> set PY_VCRUNTIME_REDIST=No thanks
C:\package> python setup.py bdist_wheel
And of course, if you don’t need to customise anything, you can just call setup.py directly and distutils will automatically use the latest available compiler.
It’s been a bumpy ride, especially over the last week, but we’re on track to release Python 3.5.0 in a good state.
Using MSVC 14.0 and the UCRT, we have independence from the compiler version.
When a new version of MSVC is available, extensions built for Python versions that may not include the default dependencies will bundle them in their own package. (There are still some uninstallation concerns here to resolve, but we have time for those.)
In 10 years time it should still be possible to build extensions that work with Python 3.5.0 using the latest tools available at that time. Everyone who builds for Python 2.7 today should be excited by that prospect.
Thank you all again for everyone who has contributed testing, feedback, fixes and suggestions. We hope you will love Python 3.5.
Some parts of this post have been superseded. If you are interested in the background, continue reading (I have marked the parts that are now incorrect). If you simply want to see how Python 3.5 and later are avoiding CRT compatibility issues, you can jump straight to part two.
Windows builds now use Microsoft Visual C++ 14.0, and extension modules should use the same.
For most Python users, this will not (and should not) mean anything. It doesn’t affect how you use or interact with Python, it isn’t new syntax or a new library, and generally you won’t notice any difference other than a small performance improvement (yay!).
However, Python extenders and embedders care a lot about this change, because it directly affects how they build their code. Since you are almost certainly using their work (numpy, anyone? Blender? MATLAB?) you should really hope that they care. However, this is a change, and no matter how good the end result is, change takes time. Please be patient with project maintainers, as they will have to spend more time supporting Python 3.5 than previous versions.
So while the most obvious benefit for most people may be a performance improvement, we haven’t even bothered benchmarking this precisely. Why not? Because the long-term benefits of the change are so good it would be worth sacrificing performance to get them. And we know it’s going to hurt some project maintainers in the short term, and again, the long-term benefits for the entire ecosystem – and those same maintainers – are worth it.
As I’m largely responsible for the compiler change (with the full support of the CPython developers, of course), this post is my attempt to help our ecosystem catch up and set the context so everyone can see just how the benefits are worth the pain. Python 3.5.0rc2 is already available with all of these changes, so project maintainers can be testing their builds now.
First, some definitions
While this post is intended for advanced audiences who probably know all of these terms, I’ll set out some definitions along the way just to make sure we’re all talking about the same thing.
MSVC is Microsoft Visual C++, the compiler used to build CPython on Windows. It’s often specified with a version number, and in this post I’ll refer to MSVC 9.0, MSVC 10.0 and MSVC 14.0.
CRT refers to the C Runtime library, which for MSVC is provided and supported by Microsoft and contains all of the standard functions your C programs can call. This is a heavily overloaded term, so I’ll be more specific and refer to DLL names (like msvcr90.dll) or import library names (like libucrt.lib) where it matters.
MSVCRT refers specifically to the CRT required by MSVC. Other compilers like gcc have their own CRT, typically known as libc, and even MSVCRT is made up of parts with their own distinct names.
VCRedist will come up later, and it refers to the redistributable package provided by Microsoft that installs the CRT, as well as extra files required for C++ programs and Microsoft extensions such as C++ AMP and the Concurrency Runtime.
MSVCRT’s Little Problem
The problem is rooted in a design decision made many years ago as part of MSVC (I don’t even know which version – probably the very earliest). While we can view it differently today, at the time it was clearly a good design. However, the long-term ramifications were not obvious without the rise of the internet.
Each version of MSVC (the compiler) comes with a matched version of the CRT (the library), and the compiler has intimate knowledge of the library. This allows for some cool optimizations, like choosing different implementations of (say) memcpy automatically based on what the compiler knows about the variables involved – if it can prove the ranges never overlap, why bother checking for overlap at runtime?
However, it does mean that when you use a different compiler, you also have to use the matched CRT version or everything breaks down very quickly. Generally this is okay, since when a developer upgrades to a newer compiler they can rebuild all of their code. The reason the internet causes this to break down is the rise of plugins and the ease of updates.
Many applications support plugins that are loadable shared libraries (DLLs, often with a special extension such as .pyd). While the application may not consider or describe these as plugins – Python prefers “extensions” or “native modules” – it is still a plugin architecture. And with the internet, we have easier access than ever to download and install many such plugins, and also to update the host application.
The CRT comes into play because it is shared between the host application and every plugin. Or rather, it assumes that it is shared. Because of the way Windows loads DLLs, if the host application and all its plugins are built with the same MSVC version and hence use the same CRT version, the state kept within that CRT would be shared.
Shared state includes things such as file descriptors, standard input/output buffering, locale, memory allocators and more. These features can be used equally by the host and its plugins without conflicts resulting in data corruption or crashes.
However, when a plugin is built with a different CRT, this state is no longer shared. File descriptors opened by the plugin do not exist (or worse, refer to a different file) in the host, file and console buffering gets confused, error handling is no longer synchronised, memory allocated in one cannot be freed in the other and so on. It is possible to safely use a plugin built with a different CRT, but it takes care. A lot of care.
This is the situation that Python 2.7 currently suffers from, and will continue to suffer from until it is completely retired. Python 2.7 is built with MSVC 9.0, and because of compatibility requirements, will always be built with MSVC 9.0 – otherwise a minor upgrade would break all of your extensions simultaneously, including the ones that nobody is able to build anymore.
Unfortunately, MSVC 9.0 is no longer supported by Microsoft and all the free downloads were removed, making it essentially impossible to build extensions for Python 2.7. The easiest mitigation was to keep making the compilers available in an unsupported manner, so we did that, but it still leaves projects in a place where they are using old tools, likely with unpatched bugs and vulnerabilities. Not ideal.
Python 3.3 and 3.4 were built with MSVC 10.0, which is in essentially the same position. The compiler is no longer supported and the tools are no longer easily available. Building extensions with later versions of MSVC results in CRT conflicts, and building with the older tools misses out on security fixes and other improvements.
One example of an improvement in MSVC 14 that is not in MSVC 10 or earlier is support for the C99 standard. I’m not claiming it’s 100% supported (it’s not), but even 90% support is much more useful than what was previously available.
The best mitigation we have for MSVC 10.0 builds of Python is to migrate to Python 3.5. Luckily, doing so does not require the same porting effort as moving from Python 2.7 would require, but it raises the question: why is Python 3.5 any better?
The answer is: UCRT.
The UCRT Solution
As part of Visual Studio 2015, MSVCRT was significantly refactored. Rather than being a single msvcr140.dll file, as would be expected based on previous versions, it is now separated into a few separate DLLs.
The most exciting one of these is ucrtbase.dll. Look carefully – there is no version number in the filename! This DLL contains the bulk of the C Runtime and is not tied to a particular compiler version, so plugins that reference ucrtbase.dll will share all the state we discussed above, even if they were built with different compilers.
Another great benefit is that ucrtbase.dll is an operating system component, installed by default on Windows 10 and coming as a recommended update for earlier versions of Windows. This means that soon every Windows PC will include the CRT and we will not need to distribute it ourselves (though the Python 3.5 installer will install the update if necessary).
It’s very important to clarify here that the compatibility guarantees only hold when linked through ucrt.lib. The public exports of ucrtbase.dll may change at any time, but linking through ucrt.lib uses API Sets to provide cross-version compatibility. Using the exports of ucrtbase.dll directly is not supported.
So the major issue faced by earlier versions of Python no longer exist. The next version of MSVC will be able to build extensions for Python 3.5, and it may even be possible for later version of Python 3.5 to be built with newer compilers without affecting users. But while this is the start of the story, it isn’t the end and the rest is not so pretty.
The UCRT Problems
While ucrt.lib is a great improvement over earlier versions, if you followed the link above or just read my comment carefully, you’ll see the rest of the problem. Besides ucrtbase.dll, there are other libraries we need to link with.
For pure C applications, the other DLL we need is vcruntime140.dll. Notice how this one includes a version number? Yeah, it depends on the version of the compiler that was used. Applications using C++ will likely depend on msvcp140.dll, which is also versioned. We have not yet completely escaped DLL hell.
Why weren’t these libraries also made version independent? Unfortunately, there are places where the compiler still needs intimate knowledge of the CRT. They are very few, and vcruntime140.dll in particular exports almost no functions that are both documented and have no preferred alternative in ucrtbase.dll (for example, memcpy may be used from vcruntime140.dll, but memcpy_s from ucrtbase.dll should be preferred). However, much of the critical startup code is part of vcruntime140.dll, and this is so closely tied to what the compiler generates that it cannot reasonably be made compatible across versions.
Ultimately, depending on any version-specific DLL takes us right back to the earlier issues. Extensions for Python 3.5 will need to use MSVC 14.0 or else include the version-specific DLLs – Python 3.5 could include vcruntime140.dll, but if an extension depends on vcruntime150.dll then it is not easily distributable.
Luckily, this concern was raised as the UCRT was being developed, and so there is a semi-official solution for this that happens to work well for Python’s needs.
The End (of part one)
Remember how I said at the start that some of this blog is no longer valid for Python 3.5? Yeah, that’s from here to the end. To see what we’ve actually done, stop reading here and read part two instead.
The Partially-Static Solution
To avoid having a runtime dependency on vcruntime140.dll, it is possible to statically link just that part of the CRT. Effectively, the required functions, which tend to be a very small subset of the complete DLL, are compiled into the final binary. However, the functions from ucrtbase.dll are still loaded from the DLL on the user’s machine, so many of the issues associated with static-linking are avoided.
There are many downsides to static linking, especially of the core runtime, ranging from larger binaries through to not automatically receiving security updates from the operating system. Previously, applications including Python have avoided static linking by distributing the CRT as part of the application (“app local”), but while this avoids some of the bloat concerns, the application distributor is still responsible for providing updates to the CRT. Statically linking vcruntime140.dll also leaves responsibility with the distributor for some updates, but significantly fewer.
Warning: This is where things get technical. Skip to the next section if you just want to know what you’ll need to fix.
The difference between dynamic linking and static linking is based on a few options passed to both the compiler (cl.exe) and the linker (link.exe). Most people are familiar with the compiler option, one of /MD (dynamic link), /MDd (debug dynamic link), /MT (static link) and /MTd (debug static link). As well as automatically filling out the remaining settings, these also control some code generation at compile time – different code needs to be compiled for static linking versus dynamic linking, and this is how that option is selected at compile time.
For the linker, there are separate libraries to link with. If the compiler option is provided, these are selected automatically, but can be overridden with the /nodefaultlib option. This table is adapted from the VC Blog post I linked above:
Release DLLs (/MD): msvcrt.lib vcruntime.lib ucrt.lib Debug DLLs (/MDd): msvcrtd.lib vcruntimed.lib ucrtd.lib Release Static (/MT): libcmt.lib libvcruntime.lib libucrt.lib Debug Static (/MTd): libcmtd.lib libvcruntimed.lib libucrtd.lib
I will ignore the debug options for the rest of this post, as debug builds should generally not be redistributed and can therefore reliably assume all the DLLs they need are available. This is why the Python 3.5 debug binaries option requires Visual Studio 2015 – to make sure you have the debug DLLs.
For a fully dynamic release build, we’ve built with /MD. This enables codepaths in the CRT header files that decorate CRT functions with declspec(dllimport) and so code is generated for calls to go through an import stub. Linking in vcruntime.lib and ucrt.lib provides the stubs that will be corrected at load time to refer to the actual DLLs.
For a fully static build, we use /MT which omits the declspec‘s and generates normal extern definitions. Linking with libvcruntime.lib and libucrt.lib provides the actual function implementation and the linker resolves the symbols normally, just as if you were calling your own function in a separate .c file.
What we want to achieve is linking with libvcruntime.lib for the static definitions, but ucrt.lib for the import stubs. Unfortunately, the compiler does not know how to generate code for this case, so it will either assume import stubs for all functions, or none of them, which results in linker errors later on.
There is one case that works: if we compile with /MT so the CRT will be statically linked, the generated code assumes everything can be resolved through it’s regular name. When linking, if we then substitute ucrt.lib instead of libucrt.lib, the linker can generate the import stubs needed to call into the DLL.
The build commands look like this:
cl.exe /MT /GL file.c link.exe /LTCG /NODEFAULTLIB:libucrt.lib ucrt.lib file.obj
We use /MT to select the static CRT. The /GL and /LTCG options enable link-time code generation, and the /NODEFAULTLIB:libucrt.lib ucrt.lib arguments ignore the static library and replace it with the import library. The linker then generates the code needed for this to work, and we end up with a DLL or an executable that only depends on ucrtbase.dll (via the API sets).
Unfortunately, there are some follow-on effects because of this change.
What else does this break?
In case you skipped the last warning, the rest of this post is now invalid. To see what we’ve actually done, stop reading here and read part two instead.
With Python 3.5, distutils has been updated to build extensions in a portable manner by default. Most simple extensions will build fine, and your wheels can be distributed to and will work on any machine with Python 3.5 installed. However, in some cases, your extension may fail to build, may produce a significantly different .pyd file from previously, or may need extra dependencies when distributed.
The first likely problem is linking static libraries. Because of the compiler change, you will probably need to rebuild other static libraries anyway, and it is important that when you do you select the static CRT option (/MT). As discussed above, we don’t actually link the entire CRT statically, but if your library expects to dynamically load the CRT DLL then it will fail to link.
If your library requires C++, your resulting .pyd will statically link any parts of the C++ runtime, and so it may be significantly larger than the same extension for Python 3.4. This is unfortunate, but not a critical issue, and it actually has the benefit that your extension will not be interfered with by other extensions that also use C++.
Of course, in some cases you really do not want to do this. In that case, I would strongly discourage you from uploading your wheels to PyPI, since you will also need to get your users to install the VCRedist version that matches the compiler you used. Currently, there is no way to check or enforce this through tools like pip.
Since it is so strongly discouraged, I’m not even going to show you how to do it, though I’ll give basic directions. In your setup.py file, you’ll want to monkeypatch distutils._msvccompiler.MSVCCompiler.initialize() (yes, the underscore in _msvccompiler means this is not supported and we may break it at any point), call the original implementation and then replace the '/MT' element in self.compile_options with '/MD'.
Ugly? Yep. By going down this path, you are making it near impossible for non-administrative users to use your extension. Please be very certain your users will be okay with this.
If you have binary dependencies that you can’t recompile but have to include, then your best option is to include the redistributable DLLs alongside them. Test thoroughly for CRT incompatibilities, especially if your dependencies use a different version of MSVCRT, and generally assume that only ucrtbase.dll will be available on your user’s machines. Dependency Walker is an amazing tool for checking binary dependencies.
The third likely issue that will be faced is code that no longer compiles. There has been an entire deprecation cycle between MSVC 10.0 and MSVC 14.0, which means some functions may simply disappear without warning (because the warning was in MSVC 11.0 and MSVC 12.0, which Python never used). There have also been changes to unsupported names and a number of non-standard names are now indicated correctly with a leading underscore.
Also, as with every release, the graph of header files may have changed, and so names that were implicitly #included previously may now require the correct header file to be specified. (This is not necessarily names moving into different header files, rather, one header file may have included another and the name was available that way. Dependencies within header files are not guaranteed stable – you should always include all headers directly when you require their definitions.)
A lot of code tries to fill gaps in various compilers and runtimes by defining functions under #ifdef directives. With the range of changes that have occurred, most of these should be checked and updated – _MSC_VER is defined as at least 1900 now, and because of the switch to /MT some defines of CRT exports may need to have the declspec(dllimport) removed (or remove the entire declaration and use the official headers).
Finally, extensions that are built with gcc under MinGW are likely to have compatibility issues for some time yet, since the UCRT is not a supported target for those tools. Again, this pain is unfortunate, but long term it should be entirely feasible for the MinGW toolchain to support the Universal CRT better than the MSVC 10.0 CRT.
If you made it this far, technically this part is still mostly correct. But then, it doesn’t really add anything new. To see what we’ve actually done, read part two.
By moving Python 3.5 to use MSVC 14.0 and the Universal CRT, we have (hopefully) removed the restriction to build extensions with matched compilers.
Extensions built with distutils can be distributed easily, though they may be larger or have more build errors as a result of different build settings.
Long term, we believe this change will avoid the problems currently faced by those building for Python 2.7 as toolchains are deprecated and retired.
The short-term pain we are going to experience would have occurred for any compiler change, but after this we should be largely insulated against the next.
Finally, please show some respect and grace towards the maintainers of projects you depend upon. It may take some time to see fully compatible releases for Python 3.5, and shouting at or abusing people online is not necessary or even helpful.
I personally want to thank everyone who distributes builds of their packages for Windows, which they don’t strictly need to do, and I apologise for the pain of transition. This change is meant to help you all, not to hurt you, though the benefits won’t be seen for some time. Thank you for your work, and for making the Python ecosystem exist for millions of users.
Last year at PyCon US, I volunteered to take over maintenance and development of the Python installers for Windows. My stated plan was to keep building the installer for Python 2.7 without modification, and to develop a new installer for Python 3.5. In this post, I’m going to show some of the changes I’ve been working on. Nothing is settled yet, and more changes are basically guaranteed before the first releases occur, but I’m happy that we’ll soon have a more powerful and flexible installer.
The installer will first be available for Python 3.5.0 alpha 1, due to be released in February.
Changes You Will Notice
The most dramatic change (and the most likely to be removed before the final release) is new default installation locations.
Installing a copy of Python for all users on a machine and allowing everyone to modify it (the default under Python 3.4 and earlier) is a massive security hole. When installed into Program Files, only administrators can modify the shared files, and so users are better protected from malicious or accidental modifications.
Those who have used the Just for Me option in previous versions of Python are likely to have been surprised when it did not work as expected. For Python 3.5, this is now a true per-user installation. All files are installed into a directory than can only be accessed by the current user and the installation will work without administrative privileges.
The first two buttons on this page are single-click installs, meaning you’ll get all the default features and options, including pip and IDLE. For most users, these will dramatically simplify the process of installing Python.
However, many of us (myself included) like to be a bit more selective when we install Python. The third button, Customize installation, is for us.
There are two pages of options. The first is a list of features that can be added or removed independently of the rest of the installation. Compared to the old-style tree view, the simple list of checkboxes makes it easier to see what each feature provides. This is also the screen you’ll see when you choose to modify an existing installation.
The second page is advanced options, including the install location which (currently) defaults to the legacy directory, allowing you to install Python 3.5 identically to the older versions with the same amount of clicking. Right now, the options are basically identical to previous versions, but they are no longer mixed up with installable features. The way they are implemented has also been improved to be more reliable.
From here, the rest of the installation proceeds as you’d expect. The final page retains the familiar message (thanks, Mark!) and also adds some links into the online documentation.
Changes You Will Not Notice
One interesting option you may have spotted on the Advanced Options page is a checkbox to install debugging symbols (.pdb files). These are handy if you work on or debug C extensions (for example, Python Tools for Visual Studio‘s mixed-mode C/Python debugging feature requires Python’s PDB files), and this is an easy way to install them. Previously the symbol files were available in a separate ZIP, but now they are just a checkbox away.
But wait, doesn’t this make the installer a larger download? Yes, or at least it would if the installer included the debugging symbols.
The biggest change to the installer is its architecture. Previously, it was a single MSI with a single embedded CAB that contained all the files. The new installer is a collection of MSIs (currently 12, though this could change), CABs (currently 16) and a single EXE. The EXE is the setup program shown above, while the CABs contain the install files and the MSIs have the install logic.
With this change, it is possible to lazily download MSIs and CABs as needed. Although it’s not marked in the screenshot above, the “Install debugging symbols” option will require an active internet connection and will download symbols on demand. In fact, it’s trivially easy to download all the components on demand, which reduces the initial download to less than 1MB.
My initial plan is to release four downloadable installers for Python 3.5.0 alpha1: two “web” installers (32-bit and 64-bit) and two “offline” installers that include the default components (download size is around 20MB, and it includes everything that was included in earlier versions). Depending on feedback and usage, this may change by the final release, but initially I want to offer useful alternatively without being too disruptive.
Another change that is part of the build process is code signing. Previously, only the installer was signed, which meant that undetectable changes could be made to python.exe or pythonXY.dll after installation. As part of reworking the installer, I’ve also moved to signing every binary that is part of a Python installation. This improves the level of trust for those who choose to validate signatures, as well as using the signed UAC dialog rather than the unsigned one when running Python as an administrator.
Changes For Administrators
For those who have scripted or automated Python installation from the old MSIs, things are going to change a bit. I believe these are for the better, as we never previously really documented and supported command-line installation, and I’ll be interested in the feedback from early adopters.
The first concern likely to arise is the web installers – how do I avoid downloading from the Python servers every time I install? What if I have to install on two hundred machines? Two thousand? The easiest way is to simply download everything once with the “/layout” option:
python-3.5.0a1.exe /layout \\shared\python\3.5.0a1
This will not install Python, but it will create a folder on a shared machine (or a local path) and download all the components into that folder. You can then run `python-3.5.0a1.exe` from that location and it will not need to download anything. Currently the entire layout is around 26MB for each of the 32-bit and 64-bit versions.
To silently install, you can run the executable with `/quiet` or `/passive`, and customisation options can be provided as properties on the command-line:
python-3.5.0a1.exe /quiet TargetDir=C:\Python35 InstallAllUsers=1 Include_pip=0 AssociateFiles=0 CompileAll=1
I’m not going to document the full list yet, as they may change up until the final release, but there will be a documentation page dedicated to installing and configuring Python on Windows.
How Can I Try This Out Early?
I’m still very actively working on this, but you can get my changes from hg.python.org/sandbox/steve.dower on the Installer branch. The build files are in Tools/msi and will (should) work with either Visual Studio 2013 or Visual Studio 2015 Preview.
Where Do I Complain About This?
I am keen to hear constructive feedback and concerns, so come and find the threads at python-dev. Nothing is unchangeable, and the Python community gets to have its say, though right now I’m looking to stabilise things up until alpha so please don’t be too upset if your suggestion doesn’t appear in the first release.
If you’re at all angry or upset, please make sure you’ve read the entire post before sharing that anger with everyone else. (That’s just general good advice.)