An Overview of Packaging for Python¶
As a general-purpose programming language, Python is designed to be used in many ways. You can build web sites or industrial robots or a game for your friends to play, and much more, all using the same core technology.
Python’s flexibility is why the first step in every Python project must be to think about the project’s audience and the corresponding environment where the project will run. It might seem strange to think about packaging before writing code, but this process does wonders for avoiding future headaches.
This overview provides a general-purpose decision tree for reasoning about Python’s plethora of packaging options. Read on to choose the best technology for your next project.
Packages exist to be installed (or deployed), so before you package anything, you’ll want to have some answers to the deployment questions below:
Who are your software’s users? Will your software be installed by other developers doing software development, operations people in a datacenter, or a less software-savvy group?
Is your software intended to run on servers, desktops, mobile clients (phones, tablets, etc.), or embedded in dedicated devices?
Is your software installed individually, or in large deployment batches?
Packaging is all about target environment and deployment experience. There are many answers to the questions above and each combination of circumstances has its own solutions. With this information, the following overview will guide you to the packaging technologies best suited to your project.
You may have heard about PyPI,
files. These are just a few of the tools Python’s ecosystem provides
for distributing Python code to developers, which you can read about in
Packaging and distributing projects.
The following approaches to packaging are meant for libraries and tools used by technical audience in a development setting. If you’re looking for ways to package Python for a non-technical audience and/or a production setting, skip ahead to Packaging Python applications.
A Python file, provided it only relies on the standard library, can be redistributed and reused. You will also need to ensure it’s written for the right version of Python, and only relies on the standard library.
This is great for sharing simple scripts and snippets between people who both have compatible Python versions (such as via email, StackOverflow, or GitHub gists). There are even some entire Python libraries that offer this as an option, such as bottle.py and boltons.
However, this pattern won’t scale for projects that consist of multiple files, need additional libraries, or need a specific version of Python, hence the options below.
If your code consists of multiple Python files, it’s usually organized into a directory structure. Any directory containing Python files can comprise an Import Package.
Because packages consist of multiple files, they are harder to distribute. Most protocols support transferring only one file at a time (when was the last time you clicked a link and it downloaded multiple files?). It’s easier to get incomplete transfers, and harder to guarantee code integrity at the destination.
So long as your code contains nothing but pure Python code, and you know your deployment environment supports your version of Python, then you can use Python’s native packaging tools to create a source Distribution Package, or sdist for short.
Python’s sdists are compressed archives (
containing one or more packages or modules. If your code is
pure-Python, and you only depend on other Python packages, you can go
here to learn more.
If you rely on any non-Python code, or non-Python packages (such as libxml2 in the case of lxml, or BLAS libraries in the case of numpy), you will need to use the format detailed in the next section, which also has many advantages for pure-Python libraries.
Python and PyPI support multiple distributions providing different implementations of the same package. For instance the unmaintained-but-seminal PIL distribution provides the PIL package, and so does Pillow, an actively-maintained fork of PIL!
This Python packaging superpower makes it possible for Pillow to be
a drop-in replacement for PIL, just by changing your project’s
So much of Python’s practical power comes from its ability to integrate with the software ecosystem, in particular libraries written in C, C++, Fortran, Rust, and other languages.
Not all developers have the right tools or experiences to build these
components written in these compiled languages, so Python created the
Wheel, a package format designed to ship libraries with
compiled artifacts. In fact, Python’s package installer,
always prefers wheels because installation is always faster, so even
pure-Python packages work better with wheels.
Binary distributions are best when they come with source distributions to match. Even if you don’t upload wheels of your code for every operating system, by uploading the sdist, you’re enabling users of other platforms to still build it for themselves. Default to publishing both sdist and wheel archives together, unless you’re creating artifacts for a very specific use case where you know the recipient only needs one or the other.
Python and PyPI make it easy to upload both wheels and sdists together. Just follow the Packaging Python Projects tutorial.
So far we’ve only discussed Python’s native distribution tools. Based on our introduction, you would be correct to infer these built-in approaches only target environments which have Python, and an audience who knows how to install Python packages.
With the variety of operating systems, configurations, and people out there, this assumption is only safe when targeting a developer audience.
Python’s native packaging is mostly built for distributing reusable code, called libraries, between developers. You can piggyback tools, or basic applications for developers, on top of Python’s library packaging, using technologies like setuptools entry_points.
Libraries are building blocks, not complete applications. For distributing applications, there’s a whole new world of technologies out there.
The next few sections organize these application packaging options according to their dependencies on the target environment, so you can choose the right one for your project.
Some types of Python applications, like web site backends and other network services, are common enough that they have frameworks to enable their development and packaging. Other types of applications, like dynamic web frontends and mobile clients, are complex enough to target that a framework becomes more than a convenience.
In all these cases, it makes sense to work backwards, from the framework’s packaging and deployment story. Some frameworks include a deployment system which wraps the technologies outlined in the rest of the guide. In these cases, you’ll want to defer to your framework’s packaging guide for the easiest and most reliable production experience.
If you ever wonder how these platforms and frameworks work under the hood, you can always read the sections beyond.
If you’re developing for a “Platform-as-a-Service” or “PaaS” like Heroku or Google App Engine, you are going to want to follow their respective packaging guides.
In all these setups, the platform takes care of packaging and deployment, as long as you follow their patterns. Most software does not fit one of these templates, hence the existence of all the other options below.
If you’re developing software that will be deployed to machines you own, users’ personal computers, or any other arrangement, read on.
Python’s steady advances are leading it into new spaces. These days you can write a mobile app or web application frontend in Python. While the language may be familiar, the packaging and deployment practices are brand new.
If you’re planning on releasing to these new frontiers, you’ll want to check out the following frameworks, and refer to their packaging guides:
If you are not interested in using a framework or platform, or just wonder about some of the technologies and techniques utilized by the frameworks above, continue reading below.
Pick an arbitrary computer, and depending on the context, there’s a very good chance Python is already installed. Included by default in most Linux and Mac operating systems for many years now, you can reasonably depend on Python preexisting in your data centers or on the personal machines of developers and data scientists.
Technologies which support this model:
PEX (Python EXecutable)
zipapp (does not help manage dependencies, requires Python 3.5+)
shiv (requires Python 3)
Of all the approaches here, depending on a pre-installed Python relies the most on the target environment. Of course, this also makes for the smallest package, as small as single-digit megabytes, or even kilobytes.
In general, decreasing the dependency on the target system increases the size of our package, so the solutions here are roughly arranged by increasing size of output.
For a long time many operating systems, including Mac and Windows, lacked built-in package management. Only recently did these OSes gain so-called “app stores”, but even those focus on consumer applications and offer little for developers.
Developers long sought remedies, and in this struggle, emerged with their own package management solutions, such as Homebrew. The most relevant alternative for Python developers is a package ecosystem called Anaconda. Anaconda is built around Python and is increasingly common in academic, analytical, and other data-oriented environments, even making its way into server-oriented environments.
Instructions on building and publishing for the Anaconda ecosystem:
A similar model involves installing an alternative Python distribution, but does not support arbitrary operating system-level packages:
Computing as we know it is defined by the ability to execute programs. Every operating system natively supports one or more formats of program they can natively execute.
There are many techniques and technologies which turn your Python program into one of these formats, most of which involve embedding the Python interpreter and any other dependencies into a single executable file.
This approach, called freezing, offers wide compatibility and seamless user experience, though often requires multiple technologies, and a good amount of effort.
A selection of Python freezers:
pyInstaller - Cross-platform
cx_Freeze - Cross-platform
constructor - For command-line installers
py2exe - Windows only
py2app - Mac only
bbFreeze - Windows, Linux, Python 2 only
osnap - Windows and Mac
pynsist - Windows only
Most of the above imply single-user deployments. For multi-component server applications, see Chef Omnibus.
An increasing number of operating systems – including Linux, Mac OS, and Windows – can be set up to run applications packaged as lightweight images, using a relatively modern arrangement often referred to as operating-system-level virtualization, or containerization.
These techniques are mostly Python agnostic, because they package whole OS filesystems, not just Python or Python packages.
Adoption is most extensive among Linux servers, where the technology originated and where the technologies below work best:
Most operating systems support some form of classical virtualization, running applications packaged as images containing a full operating system of their own. Running these virtual machines, or VMs, is a mature approach, widespread in data center environments.
These techniques are mostly reserved for larger scale deployments in data centers, though certain complex applications can benefit from this packaging. Technologies are Python agnostic, and include:
The most all-encompassing way to ship your software would be to ship it already-installed on some hardware. This way, your software’s user would require only electricity.
Whereas the virtual machines described above are primarily reserved for the tech-savvy, you can find hardware appliances being used by everyone from the most advanced data centers to the youngest children.
The sections above can only summarize so much, and you might be wondering about some of the more conspicuous gaps.
As mentioned in Depending on a separate software distribution ecosystem above, some operating systems have package managers of their own. If you’re very sure of the operating system you’re targeting, you can depend directly on a format like deb (for Debian, Ubuntu, etc.) or RPM (for Red Hat, Fedora, etc.), and use that built-in package manager to take care of installation, and even deployment. You can even use FPM to generate both deb and RPMs from the same source.
In most deployment pipelines, the OS package manager is just one piece of the puzzle.
Virtualenvs have been an indispensable tool for multiple generations of Python developer, but are slowly fading from view, as they are being wrapped by higher-level tools. With packaging in particular, virtualenvs are used as a primitive in the dh-virtualenv tool and osnap, both of which wrap virtualenvs in a self-contained way.
For production deployments, do not rely on running
from the Internet into a virtualenv, as one might do in a development
environment. The overview above is full of much better solutions.
The further down the gradient you come, the harder it gets to update components of your package. Everything is more tightly bound together.
For example, if a kernel security issue emerges, and you’re deploying containers, the host system’s kernel can be updated without requiring a new build on behalf of the application. If you deploy VM images, you’ll need a new build. Whether or not this dynamic makes one option more secure is still a bit of an old debate, going back to the still-unsettled matter of static versus dynamic linking.
Packaging in Python has a bit of a reputation for being a bumpy ride. This impression is mostly a byproduct of Python’s versatility. Once you understand the natural boundaries between each packaging solution, you begin to realize that the varied landscape is a small price Python programmers pay for using one of the most balanced, flexible languages available.