Why Hatch?¶

The high level value proposition of Hatch is that if one adopts all functionality then many other tools become unnecessary since there is support for everything one might require. Further, if one chooses to use only specific features then there are still benefits compared to alternatives.

Build backend¶

Hatchling, the build backend sister project, has many benefits compared to setuptools. Here we only compare setuptools as that is the one most people are familiar with.

• Better defaults: The default behavior for setuptools is often not desirable for the average user.
  • For source distributions, setuptools has a custom enumeration of files that get included and excluded by default. Hatchling takes the defaults from your version control system such as Git's .gitignore file.
  • For wheels, setuptools attempts to find every directory that looks like a Python package. This is often undesirable as you might ship files to the end-user unintentionally such as test or tooling directories. Hatchling defaults to very specific inclusion based on the project name and errors if no heuristic is satisfied.
• Ease of configurability: Hatchling was designed based on a history of significant challenges when configuring setuptools.
  • Hatchling uses the same glob pattern syntax as Git itself for every option which is what most users are familiar with. On the other hand, setuptools uses shell-style glob patterns for source distributions while wheels use a mix of shell-style globs and Python package syntax.
  • Configuring what gets included in source distributions requires a separate MANIFEST.in file. The custom syntax and directives must be learned and it is difficult knowing which options in the main files like setup.py influence the behavior and under what conditions. For Hatchling, everything gets configured in a single file under dedicated sections for specific targets like [tool.hatch.build.targets.wheel].
  • By default, non-Python files are excluded from wheels. Including such files requires usually verbose rules for every nested package directory. Hatchling makes no such distinction between file types and acts more like a general build system one might already be familiar with.
• Editable installations: The default behavior of Hatchling allows for proper static analysis by external tools such as IDEs. With setuptools, you must provide additional configuration which means that by default, for example, you would not get autocompletion in Visual Studio Code. This is marked as a legacy feature and may in fact be removed in future versions of setuptools.
• Reproducibility: Hatchling builds reproducible wheels and source distributions by default. setuptools does not support this for source distributions and there is no guarantee that wheels are reproducible.
• Extensibility: Although it is possible to extend setuptools, the API is quite low level. Hatchling has the concept of plugins that are separated into discrete types and only expose what is necessary, leading to an easier developer experience.

Why not?:

If building extension modules is required then it is recommended that you continue using setuptools, or even other backends that specialize in interfacing with compilers.

Environment management¶

Here we compare to both tox and nox. At a high level, there are a few common advantages:

• Python management: Hatch is able to automatically download Python distributions on the fly when specific versions that environments request cannot be found. The alternatives will raise an error, with the option to ignore unknown distributions.

• Philosophy: In the alternatives, environments are for the most part treated as executable units where a dependency set is associated with an action. If you are familiar with container ecosystems, this would be like defining a CMD at the end of a Dockerfile but without the ability to change the action at runtime. This involves significant wasted disk space usually because one often requires slight modifications to the actions and therefore will define entirely different environments inherited from a base config just to perform different logic. Additionally, this can be confusing to users not just configuration-wise but also for execution of the different environments.

  In Hatch, environments are treated as isolated areas where you can execute arbitrary commands at runtime. For example, you can define a single test environment with named scripts that runs unit vs non-unit tests, each command being potentially very long but named however you wish so you get to control the interface. Since environments are treated as places where work is performed, you can also spawn a shell into any which will execute a subprocess that automatically drops into your shell of choice. Your shell will be configured appropriately like python on PATH being updated and the prompt being changed to reflect the chosen environment.

• Configuration:

  • nox config is defined in Python which often leads to increased verbosity and makes it challenging to onboard folks compared to a standardized format with known behaviors.
• Extensibility:
  • tox allows for extending most aspects of its functionality however the API is so low-level and attached to internals that creating plugins may be challenging. For example, here is a tox plugin that was migrated to an equivalent Hatch environment collector plugin.
  • nox is configured with Python so for the local project you can do whatever you want, however there is no concept of third-party plugins per se. To achieve that, you must usually use a package that wraps nox and use that package's imports instead (example).

Why not?:

If you are using nox and you wish to migrate, and for some reason you notify sessions, then migration wouldn't be a straight translation but rather you might have to redesign that conditional step.

Python management¶

Here we compare Python management to that of pyenv.

• Cross-platform: Hatch allows for the same experience no matter the system whereas pyenv does not support Windows so you must use an entirely different project that tries to emulate the functionality.
• No build dependencies: Hatch guarantees that every available distribution is prebuilt whereas the alternative requires one to maintain a precise build environment which differs by platform and potentially Python version. Another benefit to this is extremely fast installations since the distributions are simply downloaded and unpacked.
• Optimized by default: The CPython distributions are built with profile guided optimization and link-time optimization, resulting in a 10-30% performance improvement depending on the workload. These distributions have seen wide adoption throughout the industry and are even used by the build system Bazel.
• Simplicity: Hatch treats Python installations as just another directory that one would add to PATH. It can do this for you or you can manage PATH yourself, even allowing for custom install locations. On the other hand, pyenv operates by adding shims which then act as wrappers around the actual underlying binaries. This has many unfortunate side effects:
  • It is incumbent upon the user to manage which specific Python comes first via the CLI, switch when necessary, and/or have a mental model of which versions are exposed globally and locally per-project. This can become confusing quite quickly. When working with Hatch, your global Python installations are only important insofar as they are on PATH somewhere since environments do not use them directly but rather create virtual environments from them, always using a version that is compatible with your project.
  • Configuration is required for each shell to properly set up pyenv on start, leading to inconsistencies when running processes that do not spawn a shell.
  • Debugging issues with Python search paths can be extremely difficult, especially for users of software. If you or users have ever ran into an issue where code was being executed that you did not anticipate, the issue is almost always pyenv influencing the python on PATH.

Why not?:

Currently, Hatch does not allow for the installation of specific patch release versions but rather only uses minor release granularity that tracks the latest patch release. If specific patch releases are important to you then it is best to use an alternative installation mechanism.