Writing Python in 2024
2024-10-26
Last updated
2024-10-26
Last updated
I have been writing code, especially python, professionally now for the past 8 years or so. A lot has changed in that time, almost exclusively for the better. Here are the tools that have made the biggest difference in my productivity over the years as a scientific software developer. (If you are an enterprise python developer writing code that scales to millions of users, not all of these will apply).
Everyone knows that the python dependency management ecosystem is fragmented and chaotic. How would you even begin to explain to someone new to python the differences between:
pip pipenv poetry conda easy_install pdm hatch flit and uv?
Luckily uv exists and I no longer think there is any need to explain any of the other tools. I think there is finally one tool to rule them all.
Some reasons why I use it over pip:
UV separates your dependencies from your development dependencies.
If your cloud function has to install a bunch of dependencies that you only use during development into your virtual environment, that means it will run slower and you'll be paying more for no reason.
UV separates your actual dependencies from all of their dependencies.
UV puts all of the libraries that you installed in a nicely organized pyproject.toml file and puts all of the libraries that your libraries need in a separate uv.lock file. No need to scan through a million dependencies in requirements.txt to find the ones you care about anymore.
UV is fast
This means your cloud functions get started faster, you pay less money and both you and your users are happy.
UV is backwards compatible with pip
Which means you can introduce it to legacy python projects without worrying at all about breaking anything.
Polars is fast
Polars uses less memory
Polars is multi-core by default (because writing multiprocessed code in python is hard)
Polars syntax is easier to remember and understand
I still use both polars and pandas since so many python tools use pandas.
As of today (2024-10-26) Claude has been a more reliable LLM programming assistant that ChatGPT for my use cases.
I still use both. I generally give Claude my hardest problems and ChatGPT the easier one to save on tokens.
I generally like Zed's minimalist aesthetic and simple settings. If they add a debugger then I will probably drop VSCode entirely from my workflow.
IDE's which allow you to write multiple languages in the same tool are better for me than single use IDE's.
FastAPI gets you up and running with an order of magnitude less code than Flask
Streamlit lets you build simple frontends for your code in pure python. It's easier to use and more powerful than dash.
