Limitations

Last updated on 2025-04-15 | Edit this page

Estimated time: 10 minutes

Overview

Questions

What are the limitations of venv?
What are the limitations of virtual environments more generally?

Objectives

blah

Specific limitations of `venv`

While venv is a relatively simple option for setting up and managing your virtual environments it does have some key limitations:

Python version management
Automatically keeping track of installed packages

1. Python version management

There are differences between the different versions of Python. With the most extreme example being the incompatibility between Python2 and Python 3. Depending on the complexity of your code, changes between version may not affect you, but it can be hard to tell without trying a newer version. If you want to make your code truly reproducible you should include some information about which version of Python you used to produce your results.

venv uses whichever version(s) of Python you have installed on your machine, but neither venv or pip records information about which version of Python you are using.

It is, of course, possible to include this information with the instructions on how to run your code, but as highlighted multiple times it is better to have an automated solution.

2. Automatically keeping track of installed packages

As shown, using venv with pip can create a list of dependencies for your project, but this has to be created manually and updated manually whenever you modify the dependencies for your project.

Ideally this would be updated automatically as packages are added and removed, so as to avoid any mistakes or forgetting to update it after changes have been made.

Solutions

There are quite a few different tools that have been developed to deal with these issues. Out of these the ones available, I have found most useful to be conda/ miniforge and uv.
These tools combine package, environment, and Python version management, allowing you to do everything that pip and venv do, while also being able to easily switch between different Python versions. Both will also record the version of Python used within an environment allowing you to capture both the ‘package’ and ‘language’ layers of your computational environment in one go.

There are some differences between them though:

conda/miniforge:
- Provide pre-compiled binaries for packages
- Includes other languages (notably R)
- They use their own package repositories, so may not be interoperable with other Python tools without a bit of fiddling

uv:
- Uses same package repositories as pip
- Auto updating of environment dependencies file (equivalent to requirements.txt)
- Noticably faster for installing pacakges than pip or conda
- Not reached a v1 release yet

Up until recently I have used conda to create and manage my Python environments, however having given uv a try recently I am considering using it on a more permanent basis.

General limitations of virtual environments

Beyond the specific limitations of venv, virtual environments in general have one key limitation:

They are not able to control the underlying system they are working within.

If we go back to our diagram showing the different layers of your computational environment, you can see that virtual environments only address the layer of the environment immediately below your code.

If you are using something like conda, miniforge or uv, then you may be able to extand that to the ‘language’ layer. But these tools are not able to capture the ‘operating system’ layer.

Depending on the degree of computational reproducibility you are looking for capturing the ‘package’ and ‘language’ layers may be enough. However, if you are looking for byte-for-byte reproducibility you will likely also want to capture as much of the computational environment as possible, and so you will need to turn to other tools.

Capturing the operating system layer

There are multiple different tools that can be used to capture the layer below ‘package’/ ‘language’

Virtual machines

This is one of the more commonly known tools for capturing and replicating the ‘Operating System’ layer a computational environment. Virtual machines (VMs) are effectively an emulation or virtualisation of an operating within your own “host” operating system. For example, using a virtual machine I can create and run a Windows XP operating system within my Windows 10 machine. These can also be interacted with using familiar point and click interfaces, and can be customised to access a specific portion of the host machines CPU, memory, disk, etc.
It is possible to take a snapshot of an exiting VM for distribution, or use an ‘Infrastructure as code’ tool such as Ansible to write a script to recreate a specific environment (e.g. install specific software, etc)

Containers

Containers are similar to Virtual machines in that they virtualise an operating system, but containers are typically more lightweight than VMs as they only contain software and files that are explicitly defined in order to run the project contained within them.
They also typically don’t have a GUI in the same way the OS on your computer does (although some containers will provide an interface, e.g. RStudio Server).
If you want to create an environment on your machine and then run it on something more powerful (like a HPC) then containers are a good way of achiving this.

Nix/Guix

Nix(OS) and Guix also provide a way of capturing the computational environment below the ‘packages’ and ‘language’ layers, but take a slightly different approach to this than VMs and Containers.
These are package managers based on the idea declarative configurations, that is: “specify your setup with a programmable configuration file, and then let the package manager arrange for the software available on the system to reflect that”.
This is similar to what pip is doing with the requirements.txt file described previously, but for an entire Operating system. However, this approach takes it one step further by allowing multiple versions of the same package to exist on a machine simultneously.
For a more complete outline of this approach see this post