Ask HN: Python developers at big companies what is your setup? I am trying to transition from .net to python, but finding conflicting information online about what people Actually Use. People recommend their favorite tool, but when you look into it, it is barely supported or already abandoned. I will focus on working with LLMs. Edit: By tools I meant everything related to python development starting from python version, package manager, environment manager, IDE and ending with deployment tools |
- uv (went from pipenv to poetry to uv), which was the right decision imo.
- ruff (replaced black)
- PyCharm, but some people use vscode or neovim.
- mypy for type checking, which I am not 100% happy with. One experienced developer swears by mypy, thats the reason we haven't explored alternatives.
- 'just' as task runner
Ruff is my goto for linting and formatting. replacing black, isort, flake8, etc.
There's lsps for both, so you can use them in vscode or wherever you're doing your editing.
- poetry for dependency management
- black for formatting
- flake for linting (you can use ruff now)
- PyCharm as IDE
- pyenv for python version management -> but I switched to devbox a year ago and never go back
- docker as end artifact
- deployed to kubernetes (at the end k3s) with ansible and terraform
With that out of the way: mise for managing versions of all the runtimes (asdf is fine too), uv for packages (previously poetry which is ok), mypy type checking for things I make. Otherwise I switch between projects a lot so just use whatever the projects use by default.
Deployment tools aren't really language specific (when they are they're rarely good). The editing, it's just Cursor, although I did like the jetbrains IDE previously.
But overall - just give things a go. It's usually quite easy to swap those tools as needed.
Packages managed with pip and versions pinned with pipcompile. One ML production environment forces us to use conda we inherit from the pip environment and try to have as much done by pip as possible.
Formatted with black and flake8
Unit tests with unittest
Azure Dev Ops for running tests, security screening, and CI
The happy path for deployment is build docker images in AzureDevOps unfortunately we can’t do that for good reasons
I’m not a fan of typed Python, it doesn’t work as smoothly as you expect if you’re coming from a statically typed language.
Be careful enforcing pylint quality. It can be a great way to kill your teams velocity.
I’m describing a boring stack that works. All of these tools are 5+ years old and very well supported.
I think language servers and autocomplete are much more useful for .net (I used to do a lot of C# as a freelancer in college) than they are for python. Personally I just open a REPL and tab around in that if I need it but I do that maybe once a month. Most software is well organized such that you can keep enough of the theory in your head for the default nvim config to work easily. A small FITM model with a ring context like Qwen can help a bit with boilerplate if you really want autocomplete.
Python just isn't that verbose.
If I had my choice, I'd switch environment and package management to uv.
- ruff for linting and all the other stuff it does.
- mypy for type hint checking
- VS Code (although I'm playing with zed.dev too)
Vim + PythonMode
`python -m venv` for local environment separation
Git + https://github.com/dandavison/delta for change management
I only use it for local models and don't have any hosted ones that I'm willing to pay for so YMMV.
We authenticate through a GitHub via SSO, which has a whole RBAC model behind it.
Maybe it would work with something else, but it’s usually more trouble than it’s worth, and depending on the source of the plugin for the other editors, could set off alarms with the security organization, and maybe legal as well. We have a lot of rules around our use of AI.
Find defs and refs. Master a little bit of regex and you will reduce false positives.
Grep serves as a rudimentary autocomplete. find the definition, open in a buffer, observe fields. This is analogous to an autocomplete popup displayed inline. The buffer can now power your contextual completions, similar to an inline popup.
Can you use a dev container with a dockerfile and use UV for package management, ruff for formatting and lint, mypy for type hint checking, and pytest.
And put all of that on your CI as well so if they fail you don’t merge failing code.
That said, Neovim + basedpyright + pipenv works well for me.
A lot of people here are using uv and pyenv. Can anyone hint at why I might want to switch?
And when a new enterprise tool is purchased it may or may not change.
(I only believed it once it was lived through.)
* poetry for package manager
* vscode + copilot