Uv is so good. I'm a curmudgeon about adopting new tooling, and tried uv with a lot of skepticism, but it was just better in every way. And even if it wasn't so polished and reliable, the raw speed makes it hard to go back to any other tool.
Uv combined with type hints reaching critical mass in the Python ecosystem, and how solid PyLance is in VSCode, feels so good it has made me consider investing in Python as my primary language for everything. But then I remember that Python is dog slow compared to other languages with comparable ergonomics and first-class support for static typing, and...idk it's a tough sell.
I know the performance meta in Python is to...not use python (bind to C, Rust, JVM) - and you can get pretty far with that (see: uv), but I'd rather spend my limited time building expertise in a language that isn't constantly hemorrhaging resources unless your code secretly calls something written in another language :/
There are so many good language options available today that compete. Python has become dominant in certain domains though, so you might not have a choice - which makes me grateful for these big steps forward in improving the tooling and ecosystem.
I wish we had a language that had the syntax of Python (notably including operator overloading, which is absolutely critical for neural networks, ML, data science and numerical computations), the performance, compile times and concurrency support of Go, the type system flexibility of Typescript, and the native platform integration of C/C++.
Wow, know you make me curious about the business processes at Microsoft. Did they see that they would earn more money if the interpreter had a 5x speedup, that they wouldn’t see with 1.5x?
Or was it trust broken?
I would surprised to see performance as good as V8, although that would be great. As I recall the v8 team performed exceptionally well in a corporate environment that badly wanted js performance to improve, and maybe inherited some Hotspot people at the right time.
I'd be quite delighted to see, say, 2x Python performance vs. 3.12. The JIT work has potential, but thus far little has come of it, but in fairness it's still the early days for the JIT. The funding is tiny compared to V8. I'm surprised someone at Google, OpenAI et al isn't sending a little more money that way. Talk about shared infrastructure!
Has something changed that allows a more relaxed refcounting / less eager "gc"? Py_DECREF was what murdered any hope of performance back when we hooked up 3.3 to OMR... Well that and the complete opacity of everything implemented in C
Am I the only one that's sad that poetry happened before pdm otherwise we might have had pdm as a standard instead of uv, addressing many of the things uv addresses without all the extra bells and whistles that make it cumbersome. I don't like the wedding between package manager and install manager.
... but then again neither pdm nor uv would have happened without poetry.
I think in Python specifically, an install manager is absolutely the right call. There's far too much breakage between Python versions.
I recently had to downgrade one of our projects to 3.12 because of a dependency we needed. With uv, I can be sure that everybody will be running the project on 3.12, it just all happens automatically. Without uv, I'd get the inevitable "but your changes crashed the code, have you even tested them?"
> But then I remember that Python is dog slow compared to other languages with comparable ergonomics and first-class support for static typing, and...idk it's a tough sell.
Post like these aptly describe why companies are downsizing in lieu of AI assistants, and they are not wrong for doing so.
Yes, Python is "slow". The thing is, compute is cheap these days and development time is expensive. $1000 per month is considered expensive as hell for an EC2 instance, but no developer would work for $12000 a year.
Furthermore, in modern software dev, most of the bottlenecks is network latency. If your total end to end operation takes 200ms mostly because of network calls, it doesn't matter if you code runs in 10 ms or 5ms as far as compute goes.
When it comes to development, the biggest uses of time are
1. Interfacing with some API or tool, for which you have to write code
2. Making a change, testing a change, fixing bugs.
Python has both covered better than any other language. Just today, it took me literally 10 mins to write code for a menu bar for my Mac using rumps python library so I have most commonly used commands available without typing into a terminal, and that is without using an LLM. Go ahead and try to do the same in Java or Rust or C++ and I promise you that unless you have experience with Mac development, its going to take you way more time. Python has additional things like just putting breakpoint() where you want the debugger, jupyter notebooks for prototyping, and things like lazy imports where you use import inside a function so large modules only get loaded when they run. No compilation step, no complex syntax. Multiprocessing is very easy to use as a replacement for threading, really dunno why people want to get rid of GIL so much. Functionally the only difference is overhead in launching a thread vs launching a process, and shared memory. But with multiprocessing API, you simply spin up a worker pool and send data over Pipes, and its pretty much just as fast as multithreading.
In the end, the things that matter are results. If LLMs can produce code that works, no matter how stringy it is, that code can run in production and start making company money, while they don't have to pay you money for multiple months to write the code yourself. Likewise, if you are able to develop things fast, and a company has to spend a bit more on compute, its a no brainer on using Python.
Meanwhile like strong typing, speed, GIL, and other popular things that get mentioned is all just echos of bullshit education that you learned in CS, and people repeat them without actually having any real world experience. So what if you have weak typing and make mistakes - code fails to run or generate correct results, you go and fix the code, and problem solved. People act like failing code makes your computer explode or something. There is no functional difference between a compilation failure and a code running failure. And as far as production goes, there has never been a case of a strong type language that gets used that gets deployed and doesn't have any bugs, because those bugs are all logic bugs within the actual code. And consequently, with Python, its way easier to fix those bugs.
Youtube, Uber, and a bunch of other well used services all run Python backends for a good reason. And now with skilled LLM usage, a single developer can write services in days that would take a team of engineers to write in weeks.
So TL:DR, if you actually want to stay competitive, use Python. The next set of LLMs are all going to be highly specialized smaller models, and being able to integrate them into services with Pytorch is going to be a very valuable skill, and nobody who is hiring will give a shit how memory safe Rust is.
I hadn't paid any attention to rust before uv, but since starting to use uv, I've switched a lot of my performance-sensitive code dev to rust (with interfaces to python). These sorts of improvements really do improve my quality of life significantly.
My hope is that conda goes away completely. I run an ML cluster and we have multi-gigabyte conda directories and researchers who can't reproduce anything because just touching an env breaks the world.
The main difference between mise and pixi is an ability to subscribe to conda channels and build (in an extremely fast way) conda environments, bypassing or eliminating most of the conda frustration (regular conda users know what I mean). mise allows to install asdf tools primarily (last I checked).
On the python front, however, I am somehow still an old faithful - poetry works just fine as far as I was every concerned. I do trust the collective wisdom that uv is great, but I just never found a good reason to try it.
Pixi is what FreeCAD is now using. (Along with Rattler).
It makes building FreeCAD pretty trivial, which is a huge deal considering FreeCAD’s really complex Python and non-python, cross-platform dependencies.
Yep, pixi is game changing. Especially for AI/ML, the ability to deal with non-python dependencies in nearly as fast a way as `uv` is huge. We have some exciting work leveraging the lower level primatives pixi uses we hope to share more about soon.
This is something that uv advocates should pay attention to, there are always contexts that need different assumptions, especially with our every growing and complex pile of libraries and systems.
The unspoken assertion that Rust and Python are interchangeable is pretty wild and needs significant defense, I think. I know a lot of scientists who would see their first borrow checker error and immediately move back to Python/C++/Matlab/Fortran/Julia and never consider rust again.
I work professionally in ML and have not had to touch conda in the last 7 years. In an ML cluster, it is hopefully containerized and there is no need for that?
At least on my cluster, few if any workloads are containerized. We also have an EKS where folks run containerized, but that's more inference and web serving, rather than training.
It would be nice indeed if there was a good solution to multi-gigabyte conda directories. Conda has been reproducible in my experience with pinned dependencies in the environment YAML... slow to build, sure, but reproducible.
I'd argue bzip compression was a mistake for Conda. There was a time when I had Conda packages made for the CUDA libraries so conda could locally install the right version of CUDA for every project, but boy it took forever for Conda to unpack 100MB+ packages.
the topic of managing large dependency chains for ML/AI workloads in a reproducible has been a deep rabbit hole for us. if you are curious, here is some of the work in open domain
As far as I get it, conda is still around because uv is focused on python while conda handles things written in other languages. Unless uv gets much more universal than expected, conda is here to stay.
Pixi is great! It doesn't purely use uv though. I just love it. It solves "creating a repo that runs natively on any developer's PC natively" problem quite well. It handles different dependency trees per OS for the same library too!
conda (and its derivatives that are also “conda” now), and conda-forge specifically, are the best ways to install things that will work across operating systems, architectures, and languages - without having to resort to compiling everything.
Want to make sure a software stack works well on a Cray with MPI+cuda+MKL, macOS, and ARM linux, with both C++ and Python libraries? It’s possible with conda-forge.
Not the OP but does this actually package CUDA and the CUDA toolchain itself or just the libraries around it? Can it work only with PyTorch or "any" other library?
Conda packaging system and the registry is capable of understanding things like ABI and binary compatibility. It can resolve not only Python dependencies but the binary dependencies too. Think more like dnf, yum, apt but OS-agnostic including Windows.
As far as I know, (apart from blindly bundling wheels), neither PyPI nor Python packaging tools have the knowledge of ABIs or purely C/C++/Rust binary dependencies.
With Conda you can even use it to just have OS-agnostic C compiler toolchains, no Python or anything. I actually use Pixi for shipping an OS-agnostic libprotobuf version for my Rust programs. It is better than containers since you can directly interact with the OS like the Windows GUI and device drivers or Linux compositors. Conda binaries are native binaries.
Until PyPI and setuptools understand the binary intricacies, I don't think it will be able to fully replace Conda. This may mean that they need to have an epoch and API break in their packaging format and the registry.
uv, poetry etc. can be very useful when the binary dependencies are shallow and do not deeply integrate or you are simply happy living behind the Linux kernel and a container and distro binaries are fulfilling your needs.
When you need complex hierarchies of package versions where half of them are not compiled with your current version of the base image and you need to bootstrap half a distro (on all OS kernels too!), Conda is a lifesaver. There is nothing like it.
If I find myself reaching a point where I would need to deal with ABIs and binary compatiblity, I pretty much stop there and say "is my workload so important that I need to recompile half the world to support it" and the answer (for me) is always no.
Well handling OS-dependent binary dependency is still unsolved because of the intricate behavior of native libraries and especially how tightly C and C++ compilers integrate with their base operating systems. vcpkg, Conan, containers, Yocto, Nix all target a limited slice of it. So there is not a fully satisfactory solution. Pixi comes very close though.
Conda ecosystem is forced to solve this problem to a point since ML libraries and their binary backends are terrible at keeping their binaries ABI-stable. Moreover different GPUs have different capabilities and support different versions of the GPGPU execution engines like CUDA. There is no easy way out without solving dependency hell.
CUDA is part of our cluster install scripts, we don't manage that with uv or conda. To me, that should be system software that only gets installed once.
Obligatory: Not only rust would be faster than python, but Rust definitely makes it easy with Cargo. Go, C, C++ should all exhibit the performance you are seeing in uv, if it had been written in one of those languages.
The curmudgeon in me feels the need to point out that fast, lightweight software has always been possible, it's just becoming easier now with package managers.
I've programmed all those languages before (learned C in '87, C++ in 93, Go in 2015 or so) and to be honest, while I still love C, I absolutely hate what C++ has become, Go never appealed to me (they really ignored numeric work for a long time). Rust feels like somebody wanted to make a better C with more standard libraries, without going the crazy path C++ took.
Also this. I liked C (don't use it now, right now it is mostly Java) but C++ didn't appeal to me.
Rust is for me similar to C just like you wrote, it is better, bigger but not the overwhelming way like C++ (and Rust has cargo, don't know if C++ has anything).
I actually got interested in Rust because its integer types and the core data structures looked sane, instead of this insanity: https://en.cppreference.com/w/cpp/types/integer.html . Fluid integer types are evil.
I stayed for the native functional programming, first class enums, good parts of C++ and the ultimate memory safety.
I'm surprised by how much I prefer prepending "uv" to everything instead of activating environments - which is still naturally an option if that's what floats your boat.
I also like how you can manage Python versions very easily with it. Everything feels very "batteries-included" and yet local to the project.
I still haven't used it long enough to tell whether it avoids the inevitable bi-yearly "debug a Python environment day" but it's shown enough promise to adopt it as a standard in all my new projects.
> how much I prefer prepending "uv" to everything instead of activating environments
You can also prepend the path to the virtual environment's bin/ (or Scripts/ on Windows). Literally all that "activating an environment" does is to manipulate a few environment variables. Generally, it puts the aforementioned directory on the path, sets $VIRTUAL_ENV to the venv root, configures the prompt (on my system that means modifying $PS1) as a reminder, and sets up whatever's necessary to undo the changes (on my system that means defining a "deactivate" function; others may have a separate explicit script for that).
I personally don't like the automatic detection of venvs, or the pressure to put them in a specific place relative to the project root.
> I also like how you can manage Python versions very easily with it.
I still don't understand why people value this so highly, but so it goes.
> the inevitable bi-yearly "debug a Python environment day"
If you're getting this because you have venvs based off the system Python and you upgrade the system Python, then no, uv can't do anything about that. Venvs aren't really designed to be relocated or to have their underlying Python modified. But uv will make it much faster to re-create the environment, and most likely that will be the practical solution for you.
Yup. I never even use activate, even though that's what you find in docs all over the place. Something about modifying my environment rubs me the wrong way. I just call ``./venv/bin/python driver.py`` (or ``./venv/bin/driver`` if you install it as a script) which is fairly self-evident, doesn't mess with your environment, and you can call into as many virtualenvs as you need to independently from one another.
``uv`` accomplishes the same thing, but it is another dependency you need to install. In some envs it's nice that you can do everything with the built-in Python tooling.
And when you control the installation, you can install multiple python versions with `make altinstall` into the same prefix, so you don't even need to pass 'project/bin/python, you can just call 'python-project' or 'project.py' or however you like.
Yep. (Although I installed into a hierarchy within /opt, and put symlinks to the binaries in /usr/local/bin. Annoyingly, I have to specify the paths to the actual executables when making venvs, so I have a little wrapper for that as well....)
I agree, once I learned (early in my programming journey) what the PATH is as a concept, I have never had an environment problem.
However, I also think many people, even many programmers, basically consider such external state "too confusing" and also don't know how they'd debug such a thing. Which I think is a shame since once you see that it's pretty simple it becomes a tool you can use everywhere. But given that people DON'T want to debug such, I can understand them liking a tool like uv.
I do think automatic compiler/interpreter version management is a pretty killer feature though, that's really annoying otherwise typically afaict, mostly because to get non-system wide installs typically seems to require compiling yourself.
Personally, I prefer prepending `uv` to my commands because they're more stateless that way. I don't need to remember which terminal my environment is sourced in, and when copying and pasting commands to people I don't need to worry about what state their terminal is it. It just works.
I use mise with uv to automatically activate a project's venv but prefixing is still useful sometimes since it would trigger a sync in case you forgot to do it.
One of the key tenets of uv is virtualenvs should be disposable. So barring any bugs with uv there should never be any debugging environments. Worst case just delete .venv and continue as normal.
The problem is, that would require support from the Python runtime itself (so that `sys.path` can be properly configured at startup) and it would have to be done in a way that doesn't degrade the experience for people who aren't using a proper "project" setup.
One of the big selling points of Python is that you can just create a .py file anywhere, willy-nilly, and execute the code with a Python interpreter, just as you would with e.g. a Bash script. And that you can incrementally build up from there, as you start out learning programming, to get a sense of importing files, and then creating meaningful "projects", and then thinking about packaging and distribution.
And how is that different from any other interpreted language? Node and PHP handle this just fine, and they don’t need a Rube Goldberg contraption to load dependencies from a relative directory or the systems library path.
I really don’t get why Python people act like that’s some kind of wicked witchcraft?
Yes, and in principle you can install each package into a separate folder (see the `--target` option for pip) and configure sys.path manually like that.
For .pth files to work, they have to be in a place where the standard library `site` module will look. You can add your own logic to `sitecustomize.py` and/or `usercustomize.py` but then you're really no better off vs. writing the sys.path manipulation logic.
Many years ago, the virtual environment model was considered saner, for whatever reasons. (I've actually heard people cite performance considerations from having an overly long `sys.path`, but I really doubt that matters.) And it's stuck.
The only word in the `source .venv/bin/activate` command that isn't a complete red flag that this was the wrong approach is probably bin. Everything else is so obviously wrong.
source - why are we using an OS level command to activate a programming language's environment
.venv - why is this hidden anyway, doesn't that just make it more confusing for people coming to the language
activate - why is this the most generic name possible as if no other element in a system might need to be called the activate command over something as far down the chain as a python environment
Feels dirty every time I've had to type it out and find it particularly annoying when Python is pushed so much as a good first language and I see people paid at a senior level not understand this command.
> why are we using an OS level command to activate a programming language's environment
Because "activating an environment" means setting environment variables in the parent process (the shell that you use to run the command), which is otherwise impossible on Linux (see for example https://stackoverflow.com/questions/6943208).
> why is this hidden anyway, doesn't that just make it more confusing for people coming to the language
It doesn't have to be. You can call it anything you want, hidden or not, and you can put it anywhere in the filesystem. It so happens that many people adopted this convention because they liked having the venv in that location and hidden; and uv gives such venvs special handling (discovering and using them by default).
> why is this the most generic name possible as if no other element in a system might need to be called the activate command over something as far down the chain as a python environment
Because the entire point is that, when you need to activate the environment, the folder in question is not on the path (the purpose of the script is to put it on the path!).
If activating virtual environments shadows e.g. /usr/bin/activate on your system (because the added path will be earlier in $PATH), you can still access that with a full absolute path; or you can forgo activation and do things like `.venv/bin/python -m foo`, `.venv/bin/my-program-wrapper`, etc.
> Feels dirty every time I've had to type it out
I use this:
$ type activate-local
activate-local is aliased to `source .local/.venv/bin/activate'
Notice that, again, you don't have to put it at .venv . I use a .local folder to store notes that I don't want to publish in my repo nor mention in my project's .gitignore; it in turn has
$ cat .local/.gitignore
# Anything found in this subdirectory will be ignored by Git.
# This is a convenient place to put unversioned files relevant to your
# working copy, without leaving any trace in the commit history.
*
> and I see people paid at a senior level not understand this command.
Fair response it's just nothing else feels like this weird duct tape'd together bunch of hacks to work around the design mistakes of the base language assuming it's a top level part of the OS.
> which is otherwise impossible on Linux
Node, Rust, etc all manage it.
> Because the entire point is that...
I just mean there is a history of Python using overly generic naming: activate, easy-install. Just feels weird and dirty to me that you'd call such a specific things names like these and I think it's indicative of this ideology that Python is deep in the OS.
Maybe if I'd aliased the activate command a decade ago I wouldn't feel this way or think about it.
Maybe it's just me, but it shouldn't be necessary to manage this and a few other things to get a python script working.
uv has increased my usage of python for production purposes because it's maintainable by a larger group of people, and beginners can become competent that much quicker.
The rest of the world handles that through PATH/PATH equivalent.
Either the package manager is invoked with a different PATH (one that contains the desired Node/Java/whatever version as a higher priority item than any other
version on the system).
Or the package manager itself has some way to figure that out through its config file.
Or there is a package manager launch tool, just like pyenv or whatever, which does that for you.
In practice it's not that a big of a deal, even for Maven, a tool created 21 years ago. As the average software dev you figure that stuff out a few weeks into using the tool, maybe you get burnt a few times early on for misconfiguring it and then you're on autopilot for the rest of your career.
Wait till you hear about Java's CLASSPATH and the idea of having a SINGLE, UNIFIED package dependency repo on your system, with no need for per-project dependency repos (node_modules), symlinks, or all of that stupidity.
CLASSPATH was introduced by Java in 1996, I think, and popularized for Java dependency management in 2004.
> The rest of the world handles that through PATH/PATH equivalent.
Activating a venv is just setting a few environment variables, including PATH, and storing the old values so that you can put them back to deactivate the environment.
Why else is this discussion getting hundreds of comments?
For any random python tool out there, I had about a 60% chance it would work out of the box. uv is the first tool in the python ecosystem that has brought that number basically to 100%. Ironically, it's written in Rust because python does not lend itself well to distributing reliable, fast tools to end users.
You need a runner for scripts that follow the PEP (actually the packaging standard established initially by the PEP, hence the note about it's historical status.)
The two main runners I am aware of are uv and pipx. (Any compliant runner can be referenced in the shebang to make a script standalone where shebangs are supported.)
I also recommend the flag for a max release date for $current_date - that basically locks all package versions to that date without a verbose lock file!
(sadly, uv cannot detect the release date of some packages. I'm looking at you, yaml!)
As long as your `/usr/bin/env` supports `-S`, yes.
It will install and use distribution packages, to use PyPA's terminology; the term "module" generally refers to a component of an import package. Which is to say: the names you write here must be the names that you would use in a `uv pip install` command, not the names you `import` in the code, although they may align.
linux core utils have supported this since 2018 (coreutils 8.3), amusingly it is the same release that added `cp --reflink`. AFAIK I know you have to opt out by having `POSIX_CORRECT=1` or `POSIX_ME_HARDER=1` or `--pedantic` set in your environment. [1]
If I were to put on my security hat, things like this give me shivers. It's one thing if you control the script and specified the dependencies. For any other use-case, you're trusting the script author to not install python dependencies that could be hiding all manner of defects or malicious intent.
This isn't a knock against UV, but more a criticism of dynamic dependency resolution. I'd feel much better about this if UV had a way to whitelist specific dependencies/dependency versions.
If you’re executing a script from an untrusted source, you should be examining it anyway. If it fails to execute because you haven’t installed the correct dependencies, that’s an inconvenience, not a lucky security benefit. You can write a reverse shell in Python with no dependencies and just a few lines of code.
it's a stretch to "executing a script with a build user" or "from a validated distro immutable package" to "allowing something to download evergreen code and install files everywhere on the system".
I've used Tiger/Saint/Satan/COPS in the distant past. But I think they're somewhat obsoleted by modern packaging and security like apparmor and selinux, not to mention docker and similar isolators.
most people like their distro to vet these things. uv et all had a reason when Python2 and 3 were a mess. i think that time is way behind us. pip is mostly to install libraries, and even that is mostly already done by the distros.
If that’s your concern you should be auditing the script and the dependencies anyway, whether they’re in a lock file or in the script. It’s just as easy to put malicious stuff in a requirements.txt
There's a completely irrational knee-jerk reaction to curl|sh. Do you trust the source or not? People who gripe about this will think nothing of downloading a tarball and running "make install", or downloading an executable and installing it in /usr/local/bin.
I will happily copy-paste this from any source I trust, for the same reason I'll happily install their software any other way.
I hate that curl $SOMETHING | sh has become normalized. One does not _have_ to blindly pipe something to a shell. It's quite possible to pull the script in a manner that allows examination. That Homebrew also endorses this behaviour doesn't make it any less of a risky abdication of administrative agency.
But then I'm a weirdo that takes personal offense at tools hijacking my rc / PATH, and keep things like homebrew at arm's length, explicitly calling shellenv when I need to use it.
> Based on the man page it doesn’t appear to be doing anything useful here
The man page tells me:
-S, --split-string=S
process and split S into separate arguments; used to pass multi‐
ple arguments on shebang lines
Without that, the system may try to treat the entirety of "uv run --script" as the program name, and fail to find it. Depending on your env implementation and/or your shell, this may not be needed.
Without -S, `uv run --script` would be treated as a binary name (including spaces) and you will get an error like "env: ‘uv run --script’: No such file or directory".
-S causes the string to be split on spaces and so the arguments are passed correctly.
If I download python project from someone on the same network as me and they have it written in a different python version to me and a requirements.txt I need all those things anyway.
I mean, if you use == constraints instead of >= you can avoid getting different versions, and if you’ve used it (or other things which combined have a superset of the requirements) you might have everything locally in your uv cache, too.
But, yes, python scripts with in-script dependencies plus uv to run them doesn't change dependency distribution, just streamlines use compared to manual setup of a venv per script.
I gotta say, I feel pretty vindicated after hearing for years how Python’s tooling was just fine and you should just use virtualenv with pip and how JS must be worse, that when Python devs finally get a taste of npm/cargo/bundler in their ecosystem, they freaking love it. Because yes, npm has its issues but lock files and consistent installs are amazing
There is nothing I dread more within the general context of software development, broadly, than trying to run other people's Python projects. Nothing. It's shocking that it has been so bad for so long.
Never underestimate cultural momentum I guess. NBA players shot long 2 pointers for decades before people realized 3 > 2. Doctors refused to wash their hands before doing procedures. There’s so many things that seem obvious in retrospect but took a long time to become accepted
They did wash their hands. Turns out that soap and water wasn't quite enough. Lister used carbolic acid (for dressing and wound cleaning) and Semmelweis used chlorinated lime (for hand washing).
And Semmelweis is a perfect case against being an asshole who's right: He was more right than wrong (he didn't fully understand why what he was doing helped, but it did) but he was such a horrible personality and such an amazing gift for pissing people off it probably cost lives by delaying the uptake of his ideas.
Or you could say it the other way around:
Even leading scientists are susceptible to letting emotions get the best of them and double-down defending their personal investments into things.
"A scientific truth does not triumph by convincing its opponents and making them see the light, but rather because its opponents eventually die and a new generation grows up that is familiar with it."
- Max Planck.
This is a fundamental problem in sports. Baseball is going the same way. Players are incentivized to win, and the league is incentivized to entertain. Turns out these incentives are not aligned.
> Players are incentivized to win, and the league is incentivized to entertain.
Players are incentivized to win due to specific decisions made by the league.
In Bananaball the league says, "practice your choreographed dance number before batting practice." And those same athletes are like, "Wait, which choreographed dance number? The seventh inning stretch, the grand finale, or the one we do in the infield when the guy on stilts is pitching?"
Edit: the grand finale dance number I saw is both teams dancing together. That should be noted.
But that's a matter of scale. When I was a child, the Harlem Globetrotters were far more more famous than any 3-4 NBA teams combined. They were in multiple Scooby Doo movies/episodes. They failed tp scale the model, but wrestling didn't.
This isn't right - the league can change the rules. NFL has done a wonderful job over the years on this.
Baseball has done a terrible job, but at least seems to have turned the corner with the pitch clock. Maybe they'll move the mound back a couple feet, make the ball 5.5oz, reduce the field by a player and then we'll get more entertainment and the players can still try their hardest to win.
People paid 100x more for their hosting when using aws cloud until they realized they never neded 99.97% uptime for their t-shirt business. Oh wait too soon. Save for post for the future.
People paid only 100x more than self hosting to use AWS until they realized that they could get a better deal by paying 200x for a service that is a wrapper over AWS but they never have to think about since it turns out that for most businesses that 100x is like 30 bucks a month.
People spent half their job figuring out self hosted infrastructure until they realized they rather just have some other company deploy their website when they make a commit.
Usually when someone comes with that argument, I ask them to pick any week date in the past year and then I take a random item on my calendar on that day; I give them the time and address of where I need to be as well as the address of my home and I ask them how long it's going to take me and how much it's going to cost. That's usually enough to bring them down a notch from "train work" to "sometimes train work". (But they tend to forget very often, they need to be reminded regularly for some reason). Do you want to play that game with me to get your reality check in order ?
> I give them the time and address of where I need to be [...] That's usually enough to bring them down a notch from "train work" to "sometimes train work" [...] Do you want to play that game with me to get your reality check in order ?
I don't think the implied claim is that there should be specifically a train to every particular address, if that's what you're counting as failure in the game, but rather that with good public transport (including trains) and pedestrian/cyclist-friendly streets it shouldn't be the case that most people need to drive.
Cars are so flexible. It's the answer to so many questions outside "how to move one or two people from A to common destination B".
Need to move 3 or 4 people? Driving the car may be cheaper.
Don't want to get rained on? Or heatstroke? Or walk through snow? Or carry a bunch of stuff, like a groceries/familyWeek or whatever else? Or go into the countryside/camping? Or move a differently-abled person? Or go somewhere outside public transport hours? Or, or .. or.
Are there many cases where people should take public transport or ride a bike instead of their car? Obviously yes. But once you have a car to cover the exigent circumstances it is easy to use them for personal comfort reasons.
They’re also a joke when it comes to moving large numbers of people. I can’t imagine the chaos if everyone leaving a concert at Wembley Stadium decided to leave by car.
Dallas would look very different if they emphasized public transport. Outside of downtown it is so sparse, many of the suburbs suffer from crumbling infrastructure because it turns out pipes made to last 30 years do poorly after 40 to 50 years when all the low density suburbs have aged out and there is no remaining land to subsidize the infrastructure ponzi scheme.
Are they crap during peak hour traffic or mass public events? Sure are! They're not some miracle device.
But people claiming that you can live a life without cars don't seem to realise the very many scenarios where cars are often easier and sometimes the only answer.
Until everyone wants to go from A to B, when a traffic jam happens. If that happens quite often, it might be more convenient to use a bicycle, an umbrella or snow boots.
The argument there is a little dishonest, given that if you only had the option of riding public transit that your schedule would indeed be well conformed to using public transit. I think everyone understands VERY well that they could get from point A to point B faster by using a dedicated vehicle which is solely concerned with getting them from point A to point B, that's not really debatable.
In the states at least if you're using public transit it's generally as an intentional time / cost tradeoff. That's not a mystery and taking a point-to-point schedule and comparing that against public transit constraints doesn't really prove much.
I live in Canada, which is similar to the US in this regard, and I can't believe how enslaved we are to the private automobile.
If you want the freedom to move across vast amounts of open nature, then yeah the private automobile is a good approximation for freedom of mobility. But designing urban areas that necessitate the use of a private vehicle (or even mass transit) for such essentials as groceries or education is enslavement. I don't buy the density argument either. Places that historically had the density to support alternative modes of transportation, densities that are lower than they are today, are only marginally accessible to alternative forms of transportation today. Then there is modern development, where the density is decreased due to infrastructure requirements.
To me, "urban planning" has a lot to answer for. They seem to have the foresight of a moth. However, they are probably constrained by politics which is similar.
Heard an anecdote about a German engineer who was in California (I think San Francisco, but if it was Los Angeles then the distances involved would be even larger) for meetings with American colleagues, and thought he would drive up to Oregon for a day trip. His American colleagues asked him to take another look at the scale on the bottom right of the map, and calculate the driving time. Once he ran the numbers, he realized that his map-reading instincts, trained in Germany, were leading him astray: the scale of maps he was used to had him thinking it was a 2- or 3-hour drive from San Francisco to Oregon. But in fact it's a 6-hour drive just to get to the Oregon border from SF, and if you want to head deeper into the interior then it's probably 9 to 10 hours depending on where you're going.
So no, I don't think Europeans who haven't been in America have quite absorbed just how vast America is. It stretches across an entire continent in the E-W direction, and N-S (its shortest border) still takes nearly a full day. (San Diego to Seattle is about 20 hours, and that's not even the full N-S breadth of the country since you can drive another 2.5 hours north of Seattle before reaching the Canadian border). In fact, I can find a route that goes nearly straight N-S the whole way, and takes 25 hours to drive, from McAllen, TX to Pembina, ND: https://maps.app.goo.gl/BpvjrzJvvdjD9vdi9
Train travel is sometimes feasible in America (I am planning Christmas travel with my family, and we are planning to take a train from Illinois to Ohio rather than fly, because the small Illinois town we'll be in has a train station but no airport; counting travel time to get to the airport, the train will be nearly as fast as flying but a lot cheaper). But there are vast stretches of the country where trains just do not make economic sense, and those whose only experience is in Europe usually don't quite realize that until they travel over here. For most people, they might have an intellectual grasp of the vastness of the United States, but it takes experiencing it before you really get it deep down. Hence why the very smart German engineer still misread the map: his instincts weren't quite lined up with the reality of America yet, and so he forgot to check the scale of the map.
> there are vast stretches of the country where trains just do not make economic sense
There are plenty of city pairs where high speed trains do make economic sense and America still doesn't have them. [1] is a video "56 high speed rail links we should've built already" by CityNerd. And that's aside from providing services for the greater good instead of for profit - subsidizing public transport to make a city center more walkable and more profitable and safer and cleaner can be a worthwhile thing. The US government spends a lot subsidizing air travel.
> So no, I don't think Europeans who haven't been in America have quite absorbed just how vast America is
China had some 26,000 miles of high speed rail two years ago, almost 30,000 miles now connecting 550 cities, and adding another couple of thousand miles by 2030. A hundred plus years ago America had train networks coast to coast. Now all Americans have is excuses why the thing you used to have and tore up is impossible, infeasible, unafordable, unthinkable. You have reusable space rockets that can land on a pillar of fire. If y'all had put as much effort into it as you have into special pleading about why it's impossible, you could have had it years ago.
Personally, I'd blame California for American voters' distaste for subsidizing high-speed rail. They look at the massive budget (and time) overruns of California's celebrated high-speed rail, and say "I don't want that waste of money happening in MY state, funded with MY state taxes" and then vote against any proposed projects.
This is, of course, a massively broad generalization, and there will be plenty of voters who don't fit that generalization. But the average American voter, as best I can tell, recoils from the words "high-speed rail" like Dracula would recoil from garlic. And I do believe that California's infamous failure (multiple failures, even) to build the high-speed rail they have been working on for years has a lot to do with that "high-speed rail is a boondoggle and a waste of taxpayer dollars" knee-jerk reaction that so many voters have.
Focusing on remote spots is largely a different topic. If the majority of driving was to remote spots then we'd have 90% less driving and cars wouldn't be a problem.
Honestly people really just dont understand how far apart things are. And yeah the good remote spots are a 4 hour drive from the city (and you aren’t even half way across the state at that point).
The forests and wilderness of the PNW are much, much, much, much more remote and wild than virtually anywhere you’d go in Europe. Like not even close.
It seems like people are just talking past each other here. The fact is that 99% of driving is not done by people in the process of visiting remote nature destinations.
they can't also realize a country that ditches personal vehicles can invest in buses or more trains to "remote places". nor they realize the vehicle industry is one of the biggest pollutants on micro-plastic; which screws the "remote nature" as well our health
In the future, I hope this becomes a thing. As cars become more commodotised and self driving taxis can be ordered easily maybe there'll be competing mass fleets?
Or have a "car-cabin-without-engine-and-wheels" and treat it like a packet on a network of trains and "skateboard car platforms".
The average american mind can't comprehend this works out to a huge number of them having to commute by car 1-2 hours per day to get to work in some ungodly urban sprawl while living an alienated existence in crappy suburbs, and destroying the environment while doing so. At the same time working far more, slaving year round with laughable paid vacation time or sick day provisions, while being subjected to far worse homicide rates, and being treated as subjects by cops.
No I love being stuck in traffic every day of the week for hours, its totally worth it because I can drive to an empty patch of grassland that no one wants to go to and there's nothing there. That's why cars are so amazing and freedom granting. Trains can't take you to the middle of nowhere to do nothing for the 1% of the time you don't want to be near other civilization so cars are better
Are you arguing that trains are infeasible (due to cost or duration) for certain trips?
I'm curious how this changes (in your mind) if "trains" can be expanded to "trains, buses, bicycle", or if you consider that to be a separate discussion.
The Atlanta Metro has 6.5 million people across TWENTY THOUSAND square kilometers.
Trains just don't make sense for this. Everything is too spread out. And that's okay. Cites are allowed to have different models of transportation and living.
I like how much road infra we have. That I can visit forests, rivers, mountains, and dense city all within a relatively short amount of time with complete flexibility.
Autonomous driving is going to make this paradise. Cars will be superior to trains when they drive themselves.
The German metro area "Rheinland" has a population of 8.7 million people across 12 thousand square kilometers.
~700/sqkm vs the 240/sqkm population density of Atlanta metro. Train and metro travel in this metrk area is extremely convenient and fast. It's not that Atlanta (or anywhere else in the United States for that matter) couldn't do it because of vastness, there's just no political and societal will behind this idea. In a society that glamorizes everyone driving the biggest trucks and carrying the largest rifles, of course convenient train systems are "not feasible".
I'm not following your logic. Having nearly triple the population density in Rheinland makes trains way _more_ feasible, not _less_. That means on average you have a train 1/3 the distance away from you. That's a big difference.
I live in NYC which has 29,000/sqkm in Manhattan and 11,300/sqkm overall. Public transportation is great here and you don't need a car.
but at 240/sqkm, that's really not much public trans per person!
Having replied in good faith already, I also want to call out that your jab about trucks and rifles adds nothing to the conversation and is merely culture-war fuel.
> Please don't use Hacker News for political or ideological battle. It tramples curiosity.
> Eschew flamebait. Avoid generic tangents. Omit internet tropes.
I'm so stoked for what uv is doing for the Python ecosystem. requirements.txt and the madness around it has been a hell for over a decade. It's been so pointlessly hard to replicate what the authors of Python projects want the state of your software to be in.
uv has been much needed. It's solving the single biggest pain point for Python.
> The German metro area "Rheinland" has a population of 8.7 million people across 12 thousand square kilometers. ~700/sqkm vs the 240/sqkm population density of Atlanta metro. Train and metro travel in this metrk area is extremely convenient and fast. It's not that Atlanta (or anywhere else in the United States for that matter) couldn't do it because of vastness
Did you forget to support yourself? You're saying Rheinland has three times the population density of Atlanta, with convenient passenger rail, and that demonstrates that low population density isn't an obstacle to passenger rail in Atlanta?
Don't need one in Toronto within a ½ day or so of the snow stopping for the major bicycle routes (including the MGT).
Calgary apparently also does a good job of clearing its bike lanes.
And I do my Costco shopping by bike year-round. I think I've used the car for large purchases at Costco twice in the last year.
I _rarely_ drive my car anywhere in Toronto, and find the streets on bike safer than most of the sidewalks in January -- they get plowed sooner than most homeowners and businesses clear the ice from their sidewalks.
And in Toronto we're rank amateurs at winter biking. Look at Montreal, Oslo, or Helsinki for even better examples. Too bad we've got a addle-brained carhead who doesn't understand public safety or doing his own provincial as our premier.
Personally I've also biked to work (and everywhere, really) in sub-zero degrees many times, because the bicycle lanes are cleared and salted. It's really not too bad. It actually gets a bit too hot even, because you start out by wearing so much.
In cold weather, one should always dress for 5℃ warmer than the temperature outside when you have a bike longer than 5 km. Runners pretty much have to do the same. Your body heat and good layering will take care of everything else.
Depending how expensive is gasoline in your country, when using a car people underestimate the cost of a travel by a factor two to five, because they don't count the depreciation of their vehicle's value and the maintenance cost (and sometimes even insurance price) driven by the kilometers ridden during the trip.
By that logic cars work also turns into sometimes cars work.
Ever heard of traffic jams and have you compared the number of fatal car accidents vs fatal train accidents.
Not to mention the negative effect on air quality with many cars in dense cities.
Cars main advantage is flexibility and that’s it.
For times were the place and time usually stays the same like work, trains are a valid option.
Spaniard here. Don't lecture Southern Europeans on surviving heat when the church of the village of my parents predates America itself (and it's pretty fresh inside in Summer).
This is not people’s fault individually, but rather in aggregate (ie government). The places that have good train infrastructure that is legitimately an alternative to driving are very few and far between in the US. It’s just not an option for most people. And people can’t just all move to the places where it is an option, because housing and jobs are already strained in those places negating many of the benefits.
Have you considered that the repeated attempts to reinvent what's basically trains are not, in fact, evidence that people don't know about trains, but evidence that people like the advantages of trains but that the downsides suck so bad that people will pay literally tens of thousands of dollars a year to avoid them?
Something I think that goes underappreciated: in many parts of the world, the food supply chain is shorter and the food is fresher to begin with. You're not meant to shop for 14 days at a time; you're meant to go more frequently and get what you need, fresh.
The refrigerator is a relatively modern invention. There's always been a refrigerator for me, but as a child my mother sometimes stayed with people who didn't own one and for her mother they were a new invention many people didn't have.
Actually this idea of just buying things at "the store" is relatively new too. Historically people would make more things themselves, and more food would be purchased directly from farmers who had grown it.
So many times I have come onto a library or tool that would fix my problem, and then realized “oh crap, it’s in Python, I don’t want to spend few hours building a brittle environment for it only for that env to break next time I need to use it” - and went to look for a worse solution in better language.
If you can install it with `pip install program-name` it's usually packaged well enough to just work. But if it's a random github repository with a requirements.txt with no or very few version numbers chances are that just running `pip install -r requirements.txt` will lead you down an hour+ rabbit hole of downgrading both your venv's python version and various packages until you get a combination that is close enough to the author`s venv to actually work
Usually happens to me when I find code for some research paper. Even something that's just three months old can be a real pain to get running
I don't disagree with you, but in my experience even having a requirements.txt file is a luxury when it comes to scientific Python code: A lot of the time I end up having to figure out dependencies based purely on whatever the script is importing
Seconded. Python, even with virtualenv stuff, has never been bad. There have been a few things that have been annoying especially when you need system libraries (e.g. libav for PyAV to work, etc.), but you have the same issue with every other ecosystem unless the packages come with all batteries included.
To be fair to the GP comment, this is how I feel about Ruby software. I am not nearly as practiced at installing and upgrading in that ecosystem so if there was a way to install tools in a way that lets me easily and completely blow them away, I would be happier to use them.
This mentality is exactly what many people do wrong in Python. I mean, for a one-off, yes you can have setup instructions like that. But if you want things to work for other people, on other machines, you better include a lock file with checksums. And `pip install whatever` simply does not cut it there.
Recently (like for several years), with most packages providing wheels for most platforms, it tends to be less of a problem of things actually working, except for dependencies where the platform specifiers used by Python are insufficient to select the right build of the dependency, like PyTorch.
Recently I've been playing with Chatterbox and the setup is a nightmare. It specifically wants Python 3.11. You have 3.12? TS. Try to do pip install and you'll get an error about pkg-config calling a function that no longer exists, or something like that.
God, I hate Python. Why is it so hard to not break code?
I know, this is just how it is I guess . Those of us mystified what the big problem is with virtualenv and pip and why we all have to use a tool distributed by a for profit company and it's not even written in python will just have to start a little club or something
I guess this is mostly about data science code and maybe people who publish software in those communities are just doing very poor packaging, so this idea of a "lock file" that freezes absolutely everything with zero chance for any kind of variation is useful. Certainly the worst packaged code I've ever seen with very brittle links to certain python versions and all that is typically some ML sort of thing, so yeah.
This is all anathema to those of us who know how to package and publish software.
This comment is pithy, but I reject the sentiment.
In 2025, the overall developer experience is much better in (1) Rust compared to C++, and (2) Java/DotNet(C#) compared to Python.
I'm talking about type systems/memory safety, IDEs (incl. debuggers & compilers), package management, etc.
Recently, I came back to Python from Java (for a job). Once you take the drug of a virtual machine (Java/DotNet), it is hard to go back to native binaries.
Last, for anyone unfamiliar with this quote, the original is from Winston Churchill:
Many forms of Government have been tried, and will be tried in this world of sin and woe. No one pretends that democracy is perfect or all-wise. Indeed it has been said that democracy is the worst form of Government except for all those other forms that have been tried from time to time.
How come it's easier if the tool is in another language? What are the technical (or cultural) reasons? Do most C programs use static linking, or just not have deps?
When I need to build an established project written [mostly] in C or C++, even if I don't have the dependencies installed, it's typically just a matter of installing my distro's packages for the deps and then running configure and make, or whatever. It usually works for me. Python almost never does until I've torn half my hair out wrapping my brain around whatever new band-aid bullshit they've come up with since last time, still not having understood it fully, and muddled through to a working build via ugly shortcuts I'm sure are suboptimal at best.
I don't really know why this is, at a high level, and I don't care. All I know is that Python is, for me, with the kinds of things I tend to need to build, the absolute fucking worst. I hope uv gets adopted and drives real change.
My last dance with Python was trying to build Ardupilot, which is not written in Python but does have a build that requires a tool written in Python, for whatever reason. I think I was on my Mac, and I couldn't get this tool from Homebrew. Okay, I'll install it with Pip—but now Pip is showing me this error I've never seen before about "externally managed environments", a concept I have no knowledge of. Okay, I'll try a venv—but even with the venv activated, the Ardupilot makefile can't find the tool in its path. Okay, more googling, I'll try Pipx, as recommended broadly by the internet—I don't remember what was wrong with this approach (probably because whatever pipx does is totally incomprehensible to me) but it didn't work either. Okay, what else? I can do the thing everybody is telling me not to do, passing `--break-system-packages` to plain old Pip. Okay, now the fucking version of the tool is wrong. Back it out and install the right version. Now it's working, but at what cost?
This kind of thing always happens, even if I'm on Linux, which is where I more usually build stuff. I see errors nobody has ever posted about before in the entire history of the internet, according to Google. I run into incomprehensible changes to the already incomprehensible constellation of Python tooling, made for incomprehensible reasons, and by incomprehensible I mean I just don't care about any of it, I don't have time to care, and I shouldn't have to care. Because no other language or build system forces me to care as much, and as consistently, as Python does. And then I don't care again for 6 months, a year, 2 years, until I need to do another Python thing, and whatever I remember by then isn't exactly obsolete but it's still somehow totally fucking useless.
The universe has taught me through experience that this is what Python is, uniquely. I would welcome it teaching me otherwise.
I agree with you wholeheartedly, besides not preferring dynamic programming languages, I would in the past have given python more of a look because of its low barrier to entry...but I have been repulsed by how horrific the development ux story has been and how incredibly painful it is to then distribute the code in a portable ish way.
UV is making me give python a chance for the first time since 2015s renpy project I did for fun.
That's because many people don't pay attention to reproducibility of their developed software. If there is no lock file in a repo that nails the exact versions and checksums, then I already know it's likely gonna be a pain. That's shoddy work of course, but that doesn't stop people from not paying attention to reproducibility.
One could argue, that this is one difference between npm and such, and what many people use in the Python ecosystem. npm and cargo and so on are automatically creating lock files. Even people, who don't understand why that is important, might commit them to their repositories, while in the Python ecosystem people who don't understand it, think that committing a requirements.txt only (without checksums) is OK.
However, it is wrong, to claim, that in the Python ecosystem we didn't have the tools to do it right. We did have them, and that well before uv. It took a more care though, which is apparently too much for many people already.
The lock file shouldn't be in the repository. That forces the developers into maintenance that's more properly the responsibility of the CI/CD pipeline. Instead, the lock file should be published with the other build artifacts—the sdist and wheel(s) in Python's case. And it should be optional so that people who know what they're doing can risk breaking things by installing newer versions of locked dependencies should the need arise.
They’re weren’t that many that weren’t pre compiled for Linux in the c++ world. Python is bad, but others have issues too.
C/C++ often had to compile used “make” which I’ll admit to being better at the conda/pip.
I suspect this is because the c/c++ code was developed by people with a more comp
Sci background. Configure/make/make install..I remember compiling this one.
I was into Python enough that I put it into my username but this is also my experience. I have had quasi-nightmares about just the bog of installing a Python project.
Same! And Python was my first, and is currently my second-highest-skill language. If someone's software's installation involves Python, I move on without trying. It used to be that it would require a Python 2 interpreter.
Honorable mention: Compiling someone else's C code. Come on; C compiles to a binary; don't make the user compile.
There's a lot more involved in distributing C (and C++) programs than just compiling them:
I'm assuming a Linux based system here, but consider the case where you have external dependencies. If you don't want to require that the user installs those, then you gotta bundle then or link them statically, which is its own can of worms.
Not to mention that a user with an older glibc may not be able to run your executable, even if they have your dependencies installed. Which you can, for example, solve by building against musl or a similar glibc alternative. But in the case of musl, the cost is a significant overhead if your program does a lot of allocations, due to it lacking many of the optimizations found in glibc's malloc. Mitigating that is yet another can of worms.
There's a reason why tools like Snap, AppImage, Docker, and many more exist, each of which are their own can of worms
Yea def. I think Linux's ABI diaspora and the way it handles dependencies is pain, and the root behind both those distro methods you mention, and why software is distributed as source instead of binaries. I contrast this with Rust. (And I know you can do this with C and C++, but it's not the norm:
- Distribute a single binary (Or zip with with a Readme, license etc) for Windows
- Distribute a single binary (or zip etc) for each broad Linux distro; you can cover the majority with 2 or 3. Make sure to compile on an older system (Or WSL edition), as you generally get forward compatibility, but not backwards.
- If someone's running a Linux distro other than what you built, they can `cargo build --release`, and it will *just work*.
this manages to be even worse. since it's setup full of holes to usable (eg reaching out on the filesystem), you get the worst of random binaries without isolation, plus the dead end for updates you get in practice when dealing with hundreds of containers outside of a professionally managed cluster.
This is very new behavior in pip. Not so long ago, imagine this:
You `pip install foo` which depends on `bar==1.0`. It installs both of those packages. Now you install `pip install baz` which depends on `bar==2.0`. It installs baz, and updates bar to 2.0. Better hope foo's compatible with the newer version!
I think pip only changed in the last year or two to resolve conflicts, or die noisily explaining why it couldn't be done.
Simple for simple cases - but you update a dependency and that updates a dependency that has a window range of dependencies because one version had a security issue which causes you to downgrade three other packages.
It can get complicated. The resolver in uv is part of its magic.
JavaScript has truly rotted the brains of software developers.
You include the security patch of whatever your dependencies are into your local vetted pypi repository. You control what you consider liabilities and you don't get shocked by breakages in what should be minor versions.
Of course you have to be able to develop software and not just snap Lego's together to manage a setup like that. Which is why uv is so popular.
You're implying that I have to run a local Pypi just to update some dependencies for a project? When other languages somehow manage without that? No way I'm doing that.
Some organizations force you to use their internal dependency repos because the "IT department" or similar has blessed only certain versions in the name of "security" (or at least security theater.)
Inevitably, these versions are out-of-date. Sometimes, they are very, very out of date. "Sorry, I can only install [version from 5 years ago.]" is always great for productivity.
I ran into this recently with a third-party. You'd think a 5 year old version would trigger alarm bells...
You can make it a language flame war, but the Python ecosystem has had no problem making this bed for themselves. That's why people are complaining about running other people's projects, not setting up their own.
Sensible defaults would completely sidestep this, that's the popularity of uv. Or you can be an ass to people online to feel superior, which I'm sure really helps.
Im wondering if people like you are getting paid to vet other people’s libraries? Because with every modern project I have ever seen, you can’t do too much the rest of the day with the amount of library updates you have to be vetting.
Open a requirements.txt and a package.lock.json next to each other and compare. Then you will know the answer to the question what npm, cargo, and others are doing better than pip. Oh, did I sneek a ".lock" in there? Damn right I did.
npm did not always do it right, and IMO still does not do it completely right (nor does pnpm, my preferred replacement for npm -- but it has `--frozen-lockfile` at least that forces it to do the right thing) because transitive dependencies can still be updated.
cargo can also update transitive dependencies (you need `--locked` to prevent that).
Ruby's Bundler does not, which is preferred and is the only correct default behaviour. Elixir's mix does not.
I don't know whether uv handles transitive dependencies correctly, but lockfiles should be absolute and strict for reproducible builds. Regardless, uv is an absolute breath of fresh air for this frequent Python tourist.
I remember advocating for running nightly tests on every project/service I worked on because inevitably one night one of the transitive dependencies would update and shit would break. And at least with the nightly test it forced it to break early vs when you needed to do something else like an emergency bug fix and ran into then..
That works, more or less. But now you have a requirements.txt file with 300 dependencies. Which ones do you actually care about, and which are just transitive things that your top-level deps brought along for the ride? And a year later, when GitHub's Dependabot is telling you have a security vulnerability in some package you've never heard of, do you remember if you even care about that package in the first place, or if it's left over cruft from that time you experimented with aiohttp instead of httpx?
I always just used pip-tools. Your requirements.in is the file that is human-readable and -writable, and sets your top-level deps and the version ranges you want. requirements.txt is your lockfile that you generate from .in with pip-compile. pip-compile writes out comments specifying from where each package in requirements.txt is being required.
uv does it a lot faster and generates requirements.txts that are cross-platform, which is a nice improvement.
As a “pip is mostly fine” person, we would direct the result to a new lock file, so you could still have your direct does and then pin transitives and update
Pips solver could still cause problems in general on changes.
UV having a better solver is nice. Being fast is also nice. Mainly tho it feeling like it is a tool that is maintained and can be improved upon without ripping one’s hair out is a godsend.
This is way less than what uv and other package managers do:
- dev dependencies (or other groups)
- distinguishing between direct and indirect dependencies (useful if you want to cut some fat from a project)
- dependencies with optional extra dependencies (if you remove the main, it will delete the orphans when relevant)
It's not unachievable with pip and virtualenvs, but verbose and prone to human error.
Like C: if you're careful enough, it can be memory safe. But teams would rather rely on memory safe languages.
I am on the same boat. I like uv for its speed and other niceties it brings and being a single tool to manage different things. But lockfile is not that big a deal. I never got Poetry as well. Tried it in a project once and the lockfile was a pain with the merges. I didn’t spend much time, so maybe I didn’t understand the tool and workflow or whatever, but pip and pip-tools were just fine working with requirements.txt.
The canonical way to do this with pip was using Constraints Files [1]. When you pollute your main requirements.txt it gets harder to see which package is an actual dependency of your project, and which ones are just sub-dependencies. Constraint files also let you not install a package if it's no longer a sub-dependency.
That being said, the uv experience is much nicer (also insanely fast).
I don’t get the hype either. Every time I’ve tried to use tools like pyenv or pipenv they fall down when I try to install anything that doesn’t provide wheels (GDAL), so I give up and stick to pip and virtualenv. Does uv let me install GDAL without hassle?
Pyenv's a different animal. It's meant for installing multiple Python versions at once so that you're not stuck with whatever dog your base OS happens to ship.
Pipenv tried to be what uv is, but it never did seem to work right, and it had too many weird corner cases ("why is it suddenly taking 3 hours to install packages? why it is literally impossible to get it to upgrade one single dependency and not all the others?") to ever be a contender.
Honestly, this feels like the difference between Cmake and cargo, sure Cmake does work and you can get to do everything you need, you just need discipline, knowledge and patience. On the other hand, you could just have a tool that does it all for you so you can get back to doing the actual work.
I've never even understood the virtual env dogma. I can see how version conflicts _could_ happen, but they never have. Admittedly, I'm surprised I never have issues installing globally, especially since others keep telling me what a terrible idea it is and how they had nightmare-scenario-X happen to them.
I write Python code for a living and no two projects I work on have the exact same dependencies. This is especially true when working with microservices, or working for multiple customers.
it's very common for different projects to have different requirements, especially for fast moving libraries like transformers. if you rarely run python stuff it might not be a big deal, but i'd rather not have to reinstall stuff (especially big stuff like pytorch builds) every time i switch projects.
That's exactly it. Imagine your company has multiple Python repos, and one depends on foo>=1.0,<2.0, and another depends on foo>=2.0. Venvs let you configure completely isolated environments for each so that they can peacefully coexist. I would not for a moment consider using Python without virtualenvs, though I'm not opinionated about which tool manages them. Uv? Great. Poetry? Fine. `python -m venv`? Whatever. They all get the job done.
Honestly, I can't think of a single good reason not to want to use a venv for Python.
I only ever had it a problem with large, poorly maintained projects from work. You know the kind that have two web frameworks required in the same project, and two orms, etc. ;-) That one I definitely put into a venv. But my stuff, no.
But then you have to m x n x o it for different combinations of Python version, OS, CPU architecture, GPU make/model... uv will solve it for you in milliseconds.
uv finds a dependency resolution that works for all platforms by default, and can do things like fork the resolution and choose different versions based on platform or python version requirements.
Webdev since 1998 here. Tabling the python vs JS/etc to comment on npm per se. PNPM is better than npm in every way. Strongest possible recommendation to use it instead of npm; it's faster, more efficient, safer, and more deterministic. See https://pnpm.io/motivation
Bun still segfaults way too often for my comfort but I’m crossing my fingers waiting for it to mature. It is definitely nice to have an alternative runtime to Node.
I'm still meeting devs who haven't heard of it and get their minds blown when they replace npm in their projects. Every day is a chance to meet one of the lucky 10000: https://xkcd.com/1053/
IME after years of using pnpm exclusively having to type `pnpm install` instead of `npm install` is easily the single biggest drawback of replacing `npm` with `pnpm`, so yes.
FWIW I use zsh with auto-auto-completion / auto-completion-as-you-type, so just hitting `p` on an empty command line will remember the most recent command starting with `p` (which was likely `pnpm`), and you can refine with further keystrokes and accept longer prefixes (like I always do that with `git add` to choose between typical ways to complete that statement). IMO people who don't use auto-completion are either people who have a magical ability to hammer text into their keyboards with the speed of light, or people who don't know about anything hence don't know about auto-completion, or terminally obsessive types who believe that only hand-crafting each line is worth while.
I don't know which type of person you are but since typing `pnpm` instead of `npm` bothers you to the degree you refuse to use `pnpm`, I assume you must be of the second type. Did you know you can alias commands? Did you know that no matter your shell it's straightforward to write shell scripts that do nothing but replace obnoxious command invocations with shorter ones? If you're a type 3 person then of course god forbid, no true hacker worth their salt will want to spoil the purity of their artisanal command line incantations with unnatural ersatz-commands, got it.
Might be worth noting that npm didn’t have lock files for quite a long time, which is the era during which I formed my mental model of npm hell. The popularity of yarn (again importing bundled/cargo-isms) seems like maybe the main reason npm isn’t as bad as it used to be.
> Maven worked fine without semantic versioning and lock files.
No, it actually has the exact same problem. You add a dependency, and that dependency specifies a sub-dependency against, say, version `[1.0,)`. Now you install your dependencies on a new machine and nothing works. Why? Because the sub-dependency released version 2.0 that's incompatible with the dependency you're directly referencing. Nobody likes helping to onboard the new guy when he goes to install dependencies on his laptop and stuff just doesn't work because the versions of sub-dependencies are silently different. Lock files completely avoid this.
Always using exact versions avoids this (your pom.xml essentially is the lock file), but it effectively meant you could never upgrade anything unless every dependency and transitive dependency also supported the new version. That could mean upgrading dozens of things for a critical patch. And it's surely one of the reasons log4j was so painful to get past.
Maven also has some terrible design where it will allow incompatible transitive dependencies to be used, one overwriting the other based on “nearest wins” rather than returning an error.
If in some supply chain attack someone switches out a version's code under your seating apparatus, then good look without lock files. I for one prefer being notified about checksums of things suddenly changing.
Yeah, python's tooling for dependency management was definitely not just fine, it was a disaster.
Coming from ruby. However, I think uv has actually now surpassed bundler and the ruby standard toolset for these things. Definitely surpassed npm, which is also not fine. Couldn't speak for cargo.
I used poetry professionally for a couple of years and hit so many bugs, it was definitely not a smooth experience. Granted that was probably 3-4 years ago.
I started using poetry abiut 4 years ago and definitely hit a lot of bugs around that time, but it seems to have improved considerably. That said, my company has largely moved to uv as it does seem easier to use (particularly for devs coming from other languages).
I've occasionally run into performance issues and bugs with dependency resolution / updates. Not so much recently, but at a previous company we had a huge monorepo and I've seen it take forever.
Exactly I jumped onto pipenv, poetry, and pyenv as soon as I heard about them, and though they provided advantages, they all had significant flaws which prevented me being able to give full-throated endorsement as the solutions to Python environments
However, I have zero reservations about uv. I have not encountered bugs, and when features are present they are ready for complete adoption. Plus there's massive speed improvements. There is zero downside to using uv in any application where it can be used and also there are advantages.
Tooling like npm, cargo, and others existed well before uv came up. I have used poetry years ago, and have had reproducible virtual environments for a long time. It's not like uv, at least in that regard, adds much. The biggest benefit I see so far, and that is also why I use it over poetry, is that it is fast. But the benefit of that is small, since usually one does not change the dependencies of a project that often, and when one does, one can also wait a few seconds longer.
Why did it take this long? Why did so many prior solutions ultimately fall flat after years and years of attempts? Was Python package/environment management such a hard problem that only VC money could have fixed it?
It didn't, though? Poetry was largely fine, it's just that uv is so much faster. I don't think uv is that much different from Poetry in the day-to-day dependency management, I'm sure there are some slight differences, but Poetry also brought all the modern stuff we expected out of a package manager.
> that when Python devs finally get a taste of npm/cargo/bundler in their ecosystem, they freaking love it. Because yes, npm has its issues but lock files and consistent installs are amazing
I think it's more like Rust devs using Python and thinking what the fuck why isn't this more like rustup+cargo?
Python might have been better at this but the community was struggling with the 2 vs 3 rift for years. Maybe new tooling will change it, but my personal opinion is that python does not scale very well beyond a homework assignment. That is its sweet spot: student-sized projects.
The community basically did reject Python 3, at first. Almost nobody used 3.0 / 3.1 / 3.2, to the point where I’ve seen them retconned as beta releases.
Even then though, the core developers made it clear that breaking everyone’s code was the only thing they were willing to do (remember Guido’s big “No 2.8” banner at PyCon?), which left the community with no choice.
Yep, working with bundler and npm for a decade plus has made me appreciate these tools more than you can know. I had just recently moved to Python for a project and was delighted to learn that Python had something similar, and indeed uv is more than just a package manager like bundler. It’s like bundler + rvenv/rvm.
Personally I never thought it was fine, but the solutions were all bad in some way that made direct venv and requirements files preferable. Poetry started to break this but I had issues with it. uv is the first one that actually feels good.
I've been using pip-tools for the best part of a decade. uv isn't the first time we got lock files. The main difference with uv is how it abstracts away the virtualenv and you run everything using `uv run` instead, like cargo. But you can still activate the virtualenv if you want. At that point the only difference is it's faster.
What weird shadow-universe do you inhabit where you found Python developers telling you the tooling was just fine? I thought everyone has agreed packaging was a trash fire since the turn of the century.
My default feeling towards using python in more ways than I did was default no because the tooling wasn't there for others to handle it, no matter how easy it was for me.
I feel uv will help python go even more mainstream.
But you are just using virtualenv with pip. It doesn't change any of the moving pieces except that uv is virtualenv aware and will set up / use them transparently.
You've been able to have the exact same setup forever with pyenv and pyenv-virtualenv except with these nothing ever has to be prefixed. Look, uv is amazing and I would recommend it over everything else but Python devs have had this flow forever.
It literally does, though iyt maintains a mostly-parallel low-level interface, the implementation is replaced with improved (in speed, in dependency solving, and in other areas.) You are using virtual environments (but not venv/virtualenv) and the same sources that pip uses (but not pip).
> You've been able to have the exact same setup forever with pyenv and pyenv-virtualenv except with these nothing ever has to be prefixed.
Yes, you can do a subset of what uv does with those without prefixes, and if you add pipx and hatch (though with hatch you’ll be prefixing for much the same reason as in uv) you’ll get closer to uv’s functionality.
> Look, uv is amazing and I would recommend it over everything else but Python devs have had this flow forever.
If you ignore the parts of the flow built around modern Python packaging standards like pyproject.toml, sure, pieces of the flow have been around and supported by the right constellation of other standard and nonstandard tools for a while.
same here; I now prefer uv but conda served us very well, and allowed us to maintain stable reproducible environments; being able to have multiple environments for a given project is also sometimes handy vs a single pyproject.toml
As someone who moved from Python to NodeJS/npm ~10yrs ago I can fully support that statement. Dissatisfaction with Python's refusal to get its dependency/package-management act together and seeing how reasonably the task is being dealt with by `npm`—notably with all its flaws—made me firmly stay with NodeJS. Actually virtualenv was for me another reason to keep my fingers out of whatever they're doing now over there in Python-land, but maybe `uv` can change that.
This is the most insulting take in the ongoing ruination of Python. You used to be able to avoid virtualenvs and install scripts and dependencies directly runnable from any shell. Now you get endlessly chastised for trying to use Python as a general purpose utility. Debian was a bastion of sanity with the split between dist_packages and site_packages but that's ruined now too.
Unless all python dependencies you ever used were available in your distro (and then at that point, you're no longer using pip, you're using dpkg...), this never worked well. What solves this well is PEP 723 and tooling around it.
With PEP 723 and confortable tooling (like uv), now you get scripts, that are "actually directly runnable", not just "fake directly runnable oops forgot to apt-get install something sorta runnable", and work reliably even when stuff around you is updated.
> You used to be able to avoid virtualenvs and install scripts and dependencies directly runnable from any shell.
This wasn't really the case; in principle anything you installed in the system Python environment, even "at user level", had the potential to pollute that environment and thus interfere with system tools written in Python. And if you did install it at system level, that became files within the environment your system package manager is managing, that it doesn't know how to deal with, because they didn't come from a system package.
But it's worse now because of how many system tools are written in Python — i.e., a mark of Python's success.
Notably, these tools commonly include the system package manager itself. Since you mentioned Debian (actually this is Mint, but ya know):
$ file `which apt`
/usr/local/bin/apt: Python script, ASCII text executable
> Now you get endlessly chastised for trying to use Python as a general purpose utility.
No, you don't. Nothing prevents you from running scripts with the system Python that make use of system-provided libraries (including ones that you install later with the system package manager).
If you need something that isn't packaged by your distro, then of course you shouldn't expect your distro to be able to help with it, and of course you should expect to use an environment isolated from the distro's environment. In Python, virtual environments are the method of isolation. All reasonable tooling uses them, including uv.
> Debian was a bastion of sanity with the split between dist_packages and site_packages but that's ruined now too.
It's not "ruined". If you choose to install the system package for pip and to use it with --break-system-packages, the consequences are on you, but you get the legacy behaviour back. And the system packages still put files separately in dist-packages. It's just that... doing this doesn't actually solve all the problems, fundamentally because of how the Python import system works.
Nowadays pip also defaults to installing to the users home folder if you don't run it as root.
Basically the only thing missing from pip install being a smooth experience is something like npx to cleanly run modules/binary files that were installed to that directory. It's still futzing with the PATH variable to run those scripts correctly.
> Nowadays pip also defaults to installing to the users home folder if you don't run it as root.
This could still cause problems if you run system tools as that user.
I haven't checked (because I didn't install my distro's system package for pip, and because I use virtual environments properly) but I'm pretty sure that the same marker-file protection would apply to that folder (there's no folder there, on my system).
This ideology is what caused all the problems to begin with, the base python is built as if it's the only thing in the entire operating systems environment when it's entire packaging system is also built in a way that makes that impossible to do without manually having to juggle package conflicts/incompatibilities.
This is very true! I was highly surprised when I installed Python from source and found out, that the entire problem is fixed since decades. You can have different Python versions in the same prefix just fine, you just need to pick a default one you install with `make install` and install all the others with `make altinstall`.
it's because so many essential system tools now rely on python, and if you install arbitrary code outside of a venv it can clobber the global namespace and break the core OS' guarantees.
I do agree it is annoying, and what they need to do is just provide an automatic "userspace" virtualenv for anything a user installs themselves... but that is a pandoras box tbh. (Do you do it per user? How does the user become aware of this?)
What they needed to do is allow side-by-side installs of different versions of the same distribution package and allow specifying or constraining versions at import time, then you wouldn't have the problem at all.
But that's probably not practical to retrofit given the ecosystem as it is now.
For "applications" (which are distributed on PyPI but include specified entry points for command-line use), yes. For development — installing libraries that your own code will use — you'll still generally need something else (although the restriction is really quite arbitrary).
Agreed! Sorry my read was for apps. You can use --user with pip to install into the user site rather than the system site, however it still causes overlap which can be problematic
These rust based tools really change the idea of what's possible (when you can get feedback in milliseconds). But I'm trying to figure out what Astral as a company does for revenue. I don't see any paid products on their website. They even have investors.
So far it seems like they have a bunch of these high performance tools. Is this part of an upcoming product suite for python or something? Just curious. I'm not a full-time python developer.
"What I want to do is build software that vertically integrates with our open source tools, and sell that software to companies that are already using Ruff, uv, etc. Alternatives to things that companies already pay for today. An example of what this might look like [...] would be something like an enterprise-focused private package registry."
I'd put type annotations and GIL removal above UV without a second thought. UV is still young and I hit some of those growing pains. While it is very nice, I'm not going to put it up there with sliced bread, it's just another package manager among many
For that matter, IMX much of what people praise uv for is simply stuff that pip (and venv) can now do that it couldn't back when they gave up on pip. Which in turn has become possible because of several ecosystem standards (defined across many PEPs) and increasing awareness and adoption of those standards.
The "install things that have complex non-Python dependencies using pip" story is much better than several years ago, because of things like pip gaining a new resolver in 2020, but in large part simply because it's now much more likely that the package you want offers a pre-built wheel (and that its dependencies also do). A decade ago, it was common enough that you'd be stuck with source packages even for pure-Python projects, which forced pip to build a wheel locally first (https://pradyunsg.me/blog/2022/12/31/wheels-are-faster-pure-...).
Another important change is that for wheels on PyPI the installer can now obtain separate .metadata files, so it can learn what the transitive dependencies are for a given version of a given project from a small plain-text file rather than having to speculatively download the entire wheel and unpack the METADATA file from it. (This is also possible for source distributions that include PKG-INFO, but they aren't forced to do so, and a source distribution's metadata is allowed to have "dynamic" dependencies that aren't known until the wheel is built (worst case) or a special metadata-only build hook is run (requires additional effort for the build system to support and the developer to implement)).
Things uv does better by pip by default:
- really hard to install a package globally by accident (pip: forgetting to activate venv)
- really easy to distinguish de and main dependencies (pip: create different files for different groups and set up their relationship)
- distinguish direct dependencies from indirect dependencies, making it easy to find when a package is not needed anymore (pip: I bet most devs are either not tracking sub dependencies or mixing all together with pip freeze)
- easily use different python versions for different projects (pip: not really)
With uv it just works. With pip, technically you can make it work, and I bet you'll screw something up along the way.
> - really hard to install a package globally by accident (pip: forgetting to activate venv)
This is different as of Python 3.11. Please see https://peps.python.org/pep-0668/ for details. Nowadays, to install a package globally, you first have to have a global copy of pip (Debian makes you install that separately), then you have to intentionally bypass a security marker using --break-system-packages.
Also, you don't have to activate the venv to use it. You can specify the path to the venv's pip explicitly; or you can use a different copy of pip (e.g. a globally-installed one) passing it the `--python` argument (you have been able to do this for about 3 years now).
(Pedantically, yes, you could use a venv-installed copy of pip to install into the system environment, passing both --python and --break-system-packages. I can't prove that anyone has ever done this, and I can't fathom a reason beyond bragging rights.)
> - really easy to distinguish [dev] and main dependencies
As of 25.1, pip can install from dependency groups described in pyproject.toml, which is the standard way to group your dependencies in metadata.
> distinguish direct dependencies from indirect dependencies, making it easy to find when a package is not needed anymore
As of 25.1, pip can create PEP 751 standard lockfiles.
> easily use different python versions for different projects
If you want something to install Python for you, yes, that was never in pip's purview, by design.
If you want to use an environment based off an existing Python, that's what venv is for.
For sure, we see the same thing in the JS ecosystem. New tooling adds some feature, other options implement feature, convergence to a larger common set.
The things you list may be a reason for some, but in all discussions I’ve had and read about on uv, the reason is that it behaves as a package manger should. It can just install dependencies from an automatically generated lockfile. It can update outdated minor versions. It can tell me about outdated versions of my dependencies. It can reproduce a build on another machine. The lock file can be put into version control. A coworker can run a single command to install everything. It abstracts the stupidity that is virtual environments away so much you don’t even have to touch them anymore. And also, it’s fast.
The overwhelming majority of developers seem to agree with me though.
If anything, pip is a dependency installer, while working with even trivial projects requires a dependency manager. Parent's point was that pip is actually good enough that you don’t even need uv anymore, but as long as pip doesn’t satisfy 80% of the requirements, that’s just plain false.
I'm not sure an overwhelming majority of Python developers care one way or the other. Like, I'm sure uv is nice, but I've somehow never had an issue with pip or conda, so there's just no reason to futz with uv. Same deal with Jujutsu. It's probably great, but git isn't a problem, so jj isn't a priority.
A majority of HN users might agree with you, but I'd guess that a majority of developers, to paraphrase Don Draper, don't think about it at all.
As far as impact on the ecosystem I’d say uv is up there. For the language itself you are right. Curious if you’ve come across any real use cases for Gil-less python. I haven’t yet. Seems like everything that would benefit from it is already written in highly optimized native modules.
I'm pretty ignorant about this stuff but I think asyncio is for exactly that, asynchronus I/O. Whereas GIL-less Python would be beneficial for CPU bound programs. My day job is boring so I'm never CPU bound, always IO bound on the database or network. If there is CPU heavy code, it's in Numpy. So I'm not sure if Gil-less actually helps there.
I second and third this. I HATE python but uv was what made it usable to me. No other language had such a confusing obnoxious setup to do anything with outside of js land. uv made it sane for me.
Currently they are a bit pointless. Sure they aid in documentation, but they are effort and cause you pain when making modifications (mind you with halfarse agentic coding its probably less of a problem. )
What would be better is to have a strict mode where instead of duck typing its pre-declared. It would also make a bunch of things faster (along with breaking everything and the spirit of the language)
I still don't get the appeal of UV, but thats possibly because I'm old and have been using pyenv and venv for many many years. This means that anything new is an attack on my very being.
however if it means that conda fucks off and dies, then I'm willing to move to UV.
It's the python version of fink vs macports vs homebrew. Or apt vs deb. or pkgsrc vs ports.
But I don't think "its just another" gets the value proposition here. It's significantly simpler to deploy in practice for people like me, writing ad hoc scripts and running git downloaded scripts and codelets.
Yes, virtualenv and pip existed. No, they turned out to be a lot more fiddly to run in practice than UV.
That UV is rust is funny, but not in a terrible way. The llvm compiler toolchain is written in C but compiles other languages. Using one language to do things for another language isn't such a terrible outcome.
I hope UV supplants the others. Not to disrespect their authors, but UV is better for end users. If its worse for package maintainers I think the UV authors should be told.
If you've ever used python on a project above a certain size (both lines of code and people who contribute to it), type annotations quickly become something you find useful.
I must be the odd man out but I am not a fan of uv.
1. It tries to do too many things. Please just do one thing and do it well. It's simultaneously trying to replace pip, pyenv, virtualenv, and ruff in one command.
2. You end up needing to use `uv pip` so it's not even a full replacement for pip.
3. It does not play well with Docker.
4. It adds more complexity. You end up needing to understand all of these new environmental variables: `UV_TOOL_BIN_DIR`, `UV_SYSTEM_PYTHON`, `UV_LINK_MODE`, etc.
Your implication is that pyenv, virtualenv, and pip should be 3 different tools. But for the average developer, these tools are all related to managing the python environment and versions which in my head sounds like one thing. Other languages don't have 3 different tools for this.
pip and virtualenv also add a ton of complexity and when they break (which happens quite often) debugging it is even harder despite them being "battle tested" tools.
I don't think that's true, most projects using uv don't rely on those tools at all, and you don't need to understand them. You just `uv sync` and do your work.
On the other hand, Nix and Bazel and friends are a lot of pain. I'm sure the tradeoff makes sense in a lot of situations, but not needing to bring in Nix or Bazel just to manage dependencies is a pretty big boon. It would be great to see some of the all-in-one build tools become more usable though. Maybe one day it will seem insane that every language ecosystem has its own build tool because there's some all-in-one tool that is just as easy to use as `(car)go build`!
Well Nix is the only sane way I know to manage fully reproducible envs that incorporate programs/scripts spanning multiple ecosystems. Very common situation in applied data analysis.
Nix is a 10x force multiplier for managing Linux systems. The fact that I can write python, go, bash, jq, any tool that is right for the job of managing and configuring the system is amazing. And on top of that I can patch any part of the entire system with just that, a patch from my fork on GitHub or anywhere else.
Top that off with first class programming capabilities and modularization and I can share common configuration and packages across systems. And add that those same customized packages can be directly included in a dev shell making all of the amazing software out there available for tooling and support. Really has changed my outlook and I have so much fun now not EVER dealing with tooling issues except when I have explicitly upgrade my shell and nixpkgs version.
I just rebuilt our CI infrastructure with nix and was a able to configure multiple dockerd isolated daemons per host, calculate the subnet spread for all the networks, write scripts configuring the env so you can run docker1 and hit daemon 1. Now we can saturate our CI machines with more parallel work without them fighting over docker system resources like ports. Never would have attempting doing this without nix, being able to generate the entire system config tree and inspect systemd service configs befor even applying to a host reduced my iteration loop to an all time low in the infrastructure land where 10-15mins lead times of building images to find out I misspelling Kafka and kakfa somewhere and now need to rebuild again for 15mins. Now I get almost instant feedback for most of these types of errors.
> Maybe one day it will seem insane that every language ecosystem has its own build tool because there's some all-in-one tool that is just as easy to use as `(car)go build`!
Yeah, I agree. In particular it seems insane to me that virtualenv should have to exist. I can't see any valid use case for a machine-global pool of dependencies. Why would anyone think it should be a separate tool rather than just the obvious thing that a dependency manager does? I say this as someone with nearly 20 years of Python experience.
It's the same sort of deal with pyenv--the Python version is itself a dependency of most libraries, so it's a little silly to have a dependency manager that only manages some dependencies.
And in practice it usually ends up being 6 different machine-global pools that all weirdly intersect, and some are python2.
I started using NodeJS more after lots of Python experience. Packages make so much more sense there. Even imports. You know how hard it is to do the equivalent of "require '../foo.js'" in Python?
`virtualenv` is a heavy-duty third-party library that adds functionality to the standard library venv. Or rather, venv was created as a subset of virtualenv in Python 3.3, and the projects have diverged since.
The standard library `venv` provides "obvious thing that a dependency manager does" functionality, so that every dependency manager has the opportunity to use it, and so that developers can also choose to work at a lower level. And the virtual-environment standard needs to exist so that Python can know about the pool of dependencies thus stored. Otherwise you would be forced to... depend on the dependency manager to start Python and tell it where its dependency pool is.
Fundamentally, the only things a venv needs are the `pyvenv.cfg` config file, the appropriate folder hierarchy, and some symlinks to Python (stub executables on Windows). All it's doing is providing a place for that "pool of dependencies" to exist, and providing configuration info so that Python can understand the dependency path at startup. The venvs created by the standard library module — and by uv — also provide "activation" scripts to manipulate some environment variables for ease of use; but these are completely unnecessary to making the system work.
Fundamentally, tools like uv create the same kind of virtual environment that the standard library does — because there is only one kind. Uv doesn't bootstrap pip into its environments (since that's slow and would be pointless), but you can equally well disable that with the standard library: `python -m venv --without-pip`.
> the Python version is itself a dependency of most libraries
This is a strange way of thinking about it IMO. If you're trying to obtain Python libraries, it's normally because you already have Python, and want to obtain libraries that are compatible with the Python you already have, so that you can write Python code that uses the libraries and works under that Python.
If you're trying to solve the problem of deploying an application to people who don't have Python (or to people who don't understand what Python is), you need another layer of wrapping anyway. You aren't going to get end users to install uv first.
I don’t think people consider things from a first principles perspective these days.
“…I can't see any valid use case for a machine-global pool of dependencies…” - Rhetorical question for OP but how do you run an operating system without having said operating systems dependencies available to everything else?
That quote is mine, so I think you’re meaning to address me?
> how do you run an operating system without having said operating systems dependencies available to everything else?
I’m not sure if I understand your question, but I’ll answer based on what I think you mean. The OS gets compiled into an artifact, so the dependencies aren’t available to the system itself unless they are explicitly added.
I agree with all of that context about virtualenv and venv, but it all seems orthogonal to my point. I still can’t see a case where you would want the default Python behavior (global dependencies).
> This is a strange way of thinking about it IMO. If you're trying to obtain Python libraries, it's normally because you already have Python, and want to obtain libraries that are compatible with the Python you already have, so that you can write Python code that uses the libraries and works under that Python.
“normally” is biased by what the tooling supports. If Python tooling supported pinning to an interpreter by default then perhaps it would seem more normal?
I write a lot of Go these days, and the libs pin to a version of Go. When you build a project, the toolchain will resolve and (if necessary) install the necessary Go dependency just like all of the other dependencies. It’s a very natural and pleasant workflow.
> 1. It tries to do too many things. Please just do one thing and do it well. It's simultaneously trying to replace pip, pyenv, virtualenv, and ruff in one command.
I think there are more cases where pip, pyenv, and virtualenv are used together than not. It makes sense to bundle the features of the three into one. uv does not replace ruff.
> 2. You end up needing to use `uv pip` so it's not even a full replacement for pip.
uv pip is there for compatibility and to facilitate migration but once you are full on the uv workflow you rarely need `uv pip` if ever
> 3. It does not play well with Docker.
In what sense?
> 4. It adds more complexity. You end up needing to understand all of these new environmental variables: `UV_TOOL_BIN_DIR`, `UV_SYSTEM_PYTHON`, `UV_LINK_MODE`, etc.
> It tries to do too many things. Please just do one thing and do it well. It's simultaneously trying to replace pip, pyenv, virtualenv, and ruff in one command.
uv doesn’t try to replace ruff.
> You end up needing to use `uv pip` so it's not even a full replacement for pip.
"uv pip" doesn't use pip, it provides a low-level pip-compatible interface for uv, so it is, in fact, still uv replacing pip, with the speed and other advantages of uv when using that interface.
Also, while I’ve used uv pip and uv venv as part of familiarizing myself with the tool, I’ve never run into a situation where I need either of those low-level interfaces rather than the normal high-level interface.
> 1. It tries to do too many things. Please just do one thing and do it well. It's simultaneously trying to replace pip, pyenv, virtualenv, and ruff in one command.
In my experience it generally does all of those well. Are you running into issues with the uv replacements?
> 2. You end up needing to use `uv pip` so it's not even a full replacement for pip.
So I have been doing Python for far too long and have all sort of tooling I've accreted to make Python work well for me across projects and computers and I never quite made the leap to Poetry and was suspicious of uv.
Happened to buy a new machine and decided to jump in the deep end and it's been glorious. I think the difference from your comment (and others in this chain) and my experience is that you're trying to make uv fit how you have done things. Jumping all the way in, I just . . . never needed virtualenvs. Don't really think about them once I sorted out a mistake I was making. uv init and you're pretty much there.
>You end up needing to use `uv pip` so it's not even a full replacement for pip
The only time I've used uv pip is on a project at work that isn't a uv-powered project. uv add should be doing what you need and it really fights you if you're trying to add something to global because it assumes that's an accident, which it probably is (but you can drop back to uv pip for that).
>`UV_TOOL_BIN_DIR`, `UV_SYSTEM_PYTHON`, `UV_LINK_MODE`, etc.
I've been using it for six months and didn't know those existed. I would suggest this is a symptom of trying to make it be what you're used to. I would also gently suggest those of us who have decades of Python experience may have a bit of Stockholm Syndrome around package management, packaging, etc.
> It tries to do too many things. Please just do one thing and do it well.
I disagree with this principle. Sometimes what I need is a kitset. I don't want to go shopping for things, or browse multiple docs. I just want it taken care of for me. I don't use uv so I don't know if the pieces fit together well but the kitset can work well and so can a la carte.
uv has played well with Docker in my experience, from dev containers to CI/CD to production image builds. Would be interested to hear what is not working for you.
> 2. You end up needing to use `uv pip` so it's not even a full replacement for pip.
Needing pip and virtualenvs was enough to make me realize uv wasn't what I was looking for. If I still need to manage virtualenvs and call pip I'm just going to do so with both of these directly.
I had been hoping someone would introduce the non-virtualenv package management solution that every single other language has where there's a dependency list and version requirements (including of the language itself) in a manifest file (go.mod, package.json, etc) and everything happens in the context of that directory alone without shell shenanigans.
> I had been hoping someone would introduce the non-virtualenv package management solution that every single other language has where there's a dependency list and version requirements (including of the language itself) in a manifest file (go.mod, package.json, etc) and everything happens in the context of that directory alone without shell shenanigans.
Isn't that exactly a pyproject.toml via the the uv add/sync/run interface? What is that missing that you need?
What are you needing to use `uv pip` for? I don't think I ever call into pip from uv for anything nowadays. I typically just need to do `uv sync` and `uv run`, maybe sometimes `uvx` if I want to run some random 3rd party python script
> I had been hoping someone would introduce the non-virtualenv package management solution that every single other language has where there's a dependency list and version requirements (including of the language itself) in a manifest file (go.mod, package.json, etc) and everything happens in the context of that directory alone without shell shenanigans.
If you are using uv, you don’t need to do shell shenanigans, you just use uv run. So I'm not sure how uv with pyproject.toml doesn't meet this description (yes, the venv is still there, it is used exactly as you describe.)
In most cases, you don't really need to manage virtual envs though ? uv commands that need a venv will just create one for you or install to the existing one automatically.
So here uv installs the Python version wanted. But it's just a venv. And we pip install using requirements.txt, like normal, within that venv.
Someone, please tell me what's wrong with this. To me, this seems much less complicated that some uv-centric .toml config file, plus some uv-centric commands for more kinds of actions.
This. I was researching uv to replace my pipenv+pyenv setup, but after reading up a bit I decided to just give up. Pipenv is just straightforward and “just works”. Aside from being slow, not much is wrong with it. I’m not in the mood to start configuring uv, a tool that should take me 2 minutes and a “uv —-help” to learn.
I have worked on numerous projects that started with pipenv and it has never "just works" ever. Either there's some trivial dependency conflict that it can't resolve or it's slow as molasses or something or the other. pipenv has been horrible to use. I started switching projects to pip-tools and now I recommend using uv
What doesn’t just work about uv in particular? You basically need three commands - uv add, uv sync, and uv run. Forget about virtual environments, and get back to working. No configuration necessary.
Slow doesn't really begin to do justice, I'd have to wait for >5 minutes for pipenv to finish figuring out our lock file. uv does it in less than a second.
Yeah, I'm with you. I'm forcing myself to learn it because it looks like that's the way PyWorld is going. I don't dislike uv as much as poetry. But I guess I never really ran into issues using pyenv and pip. shrug Maybe I wasn't working on complex enough projects.
What's your use case for UV_NO_SYNC? I assume the option exists for a reason, but aside from maybe a modest performance improvement when working with a massive complex package environment, I'm not sure what problem it solves.
> You end up needing to use `uv pip` so it's not even a full replacement for pip.
No you don't. That's just a set of compatibility approaches for people who can't let go of pip/venv. Move to uv/PEP723, world's your oyster.
> It does not play well with Docker.
Huh? I use uv both during container build and container runtime, and it works just fine?
> You end up needing to understand all of these new environmental variables
Not encountered the need for any of these yet. Your comments on uv are so far out of line of all the uses I've seen, I'd love to hear what you're specifically doing that these become breaking points.
Idk, for me ruff was more of a game changer. No more explaining why we need both flake8 and pylint (and isort), no more flake8 plugins... Just one command that does it all.
UV is great but I use it as a more convenient pip+venv. Maybe I'm not using it to it's full potential.
I agree flake8 -> ruff was more of a game changer for me than pip+venv -> uv. I use flake8/ruff for more often than pip/venv.
uv is probably much more of a game changer for beginner python users who just need to install stuff and don't need to lint. So it's a bigger deal for the broader python ecosystem.
You aren't, but that's fine. Everyone has their own idea about how tooling should work and come together, and I happen to be in your camp (from what I can tell). I actively don't want an all-in-one tool to do "project management".
As someone who generally prefers not to use python in a production context (I think it's excellent for one-off scripts or cron jobs that require more features then what bash provides), I agree with this sentiment. I recently wrote some python (using uv) and found it to be pleasant and well-integrated with a variety of LSPs.
Can't agree more. We were using pyenv+poetry before and regularly had to pin our poetry version to a specific one, because new poetry releases would stall trying to resolve dependencies.
pyenv was problematic because you needed the right concoction of system packages to ensure it compiled python with the right features, and we have a mix of MacOS and Linux devs so this was often non-trivial.
uv is much faster than both of these tools, has a more ergonomic CLI, and solves both of the issues I just mentioned.
I'm hoping astral's type checker is suitably good once released, because we're on mypy right now and it's a constant source of frustration (slow and buggy).
> because new poetry releases would stall trying to resolve dependencies.
> uv is much faster than both of these tools
conda is also (in)famous for being slow at this, although the new mamba solver is much faster. What does uv do in order to resolve dependencies much faster?
> What does uv do in order to resolve dependencies much faster?
- Representing version numbers as single integer for fast comparison.
- Being implemented in rust rather than Python (compared to Poetry)
- Parallel downloads
- Caching individual files rather than zipped wheel, so installation is just hard-linking files, zero copy (on unix at least). Also makes it very storage efficient.
> Instead of
>
> source .venv/bin/activate
> python myscript.py
>
> you can just do
>
> > uv run myscript
>
This is by far the biggest turn off for me. The whole point of an environment manager is set the environment so that the commands I run work. They need to run natively how they are supposed to when the environment is set, not put through a translation layer.
Side rant: yes I get triggered whenever someone tells me "you can just" do this thing that is actually longer and worse than the original.
> The whole point of an environment manager is set the environment so that the commands I run work. They need to run natively how they are supposed to when the environment is set, not put through a translation layer.
The `uv run` command is an optional shortcut for avoiding needing to activate the virtual environment. I personally don't like the whole "needing to activate an environment" before I can run commands "natively", so I like `uv run`. (Actually for the last 10 years I've had my `./manage.py` auto-set up the virtual environment for me.)
The `uv add` / `uv lock` / `uv sync` commands are still useful without `uv run`.
> They need to run natively how they are supposed to when the environment is set, not put through a translation layer.
There is a new standard mechanism for specifying the same things you would specify when setting up a venv with a python version and dependencies in the header of a single file script, so that tooling can setup up the environment and run the script using only the script file itself as a spec.
uv (and PyPA’s own pipx) support this standard.
> yes I get triggered whenever someone tells me "you can just" do this thing that is actually longer and worse than the original.
"uv run myscript" is neither longer nor worse than separately manually building a venv, activating it, installing dependencies into it, and then running the script.
> I get triggered whenever someone tells me "you can just" do this thing that is actually longer and worse than the original.
Apologies for triggering you in advance, but in case you or others find it useful, here’s how to do the equivalent env-activation commands with uv: https://news.ycombinator.com/item?id=44360892
Yes, you still have the option of manually activating a venv, and that makes sense if the amortized cost of that is lower than several instances of typing `uv run `. Though sometimes when working in one project with its venv activated, I end up needing to run a tool from another project with a separate vent, so uv still ends up being useful.
Unless I'm an AI, I'm pretty sure "uv run" is the same number of characters as "python". So it's shorter. Also venvs are a translation layer already, changing path.
it's not really the number of characters so much as the cognitive load of having to do something different here vs there and anything I run successfully on the command line can't be directly lifted over into scripts etc. Along with training a team of people to do that.
Before uv, I was fairly happy with pyenv + venv + pip for development and pipx for running "tools". IMO, the specific things uv improves upon are:
- Faster dependency resolution. In fact, everything uv does is extremely fast.
- Better ergonomics in a dozen ways (`uv run` instead of activating the virtual env, support for script metadata to run scripts with dependencies, uv add to modify the pyproject.toml (that it created for you), etc.)
- Stack of one tool instead of four+
- Easier Python installation (although I usually use both pyenv and uv on my machine)
UV and the crew at Astral really moved the Python packaging community forward.
I would love to see them compete with the likes of Conda and try to handle the Python C extension story.
But in the interim, I agree with everyone else who has already commented, Pixi which is partly built atop of UV’s solver is an even bigger deal and I think the longer term winner here.
Having a topologically complete package manager who can speak Conda and PyPi, is amazing.
This article appears to be NOT about someone who discovered uv after using venv/pip, but rather an article about someone who discovered uv after not using virtual environments at all, and is mostly excited about the cleanliness of virtual environments.
The article shows some advantages compared to plain virtual environments:
In principle, you can ‘activate’ this new virtual environment like any typical virtual environment that you may have seen in other tools, but the most ‘uv-onic’ way to use uv is simply to prepend any command with uv run. This command automatically picks up the correct virtual environment for you and runs your command with it. For instance, to run a script — instead of
> The article shows some advantages compared to plain virtual environments
No; they are plain virtual environments. There is no special kind of virtual environment. Uv simply offers its own command structure for managing those environments. In particular, `uv run` just ensures a venv in a specific location, then uses it.
There is no requirement to activate virtual environments in order to use them (unless you have some other tooling that specifically depends on the environment variables being set). You can, similarly, "just do"
.venv/bin/python myscript.py
without uv installed.
> This command automatically picks up the correct virtual environment for you
Some people dislike such magic, especially since it involves uv having an opinion about where the virtual environment is located.
This blog very strongly echoes my own experiental sense of the field of play.
It's just simpler to use, and better overall. It's reduced friction significantly.
I think the Python community should put it as a first preference vehicle, and be respectful to the prior arts, and their developers, but not insist they have primacy.
But what was wrong with pip, venv and pyproject.toml in the first place? I just keep a system installation of python for my personal things and an environment for every project I'm working on. I'd get suspicious if a developer is picky about python versions or library versions like what crazy programs are you writing?
What's wrong? Having modify the shell environment, no lockfile, slow download/installation, lack of a standard dependency dir, ...
> I'd get suspicious if a developer is picky about python versions or library versions
Certain library versions only support certain python versions. And they also break API. So moving up/down the python versions also means moving library versions which means stuff no longer works.
You don't have to modify the environment (this is provided as an option for convenience). The alternatives are to use higher-level management like uv does, or to specify the path to executables in the virtual environment directly. But uv works by creating virtual environments that are essentially the same as what you get with `python -m venv --without-pip` (although they reimplemented the venv creation logic).
Pip can install from dependency groups in a pyproject.toml file, and can write PEP 751 lockfiles, and work is under way to allow it to install from those lockfiles as well.
I don't know what you mean about a "standard dependency dir". When you make a venv yourself, you can call it what you want, and put it where you want. If you want to put it in a "standard" place, you can trivially make a shell alias to do so. (You can also trivially make a shell alias for "activate the venv at a hard-coded relative path", and use that from your project root.)
Yes, pip installation is needlessly slow for a variety of reasons (that mostly do not have to do with being implemented in Python rather than Rust). Resolving dependencies is also slow (and Rust may be more relevant here; I haven't done detailed testing). But your download speed is still going to be primarily limited by your internet connection to PyPI.
> The alternatives are to use higher-level management like uv does,
The question was specifically what's wrong with pip, venv and pyproject toml, i.e. what issues uv is trying to address. Well of course the thing trying to address the problem addresses the problem....
> I don't know what you mean about a "standard dependency dir".
like node's node_modules, or cargo's ~/.cargo/registry. You shouldn't have to manually create and manage that. installing/building should just create it. Which is what uv does and pip doesn't.
> the same as what you get with `python -m venv --without-pip`
The thing that should be automatic. And even if it is not it should at least be less arcane. An important command like that should have been streamlined long ago. One of the many improvements uv brings to the table.
> and work is under way to allow it to install from those lockfiles as well.
Yeah well, the lack up until now is one of those "what is wrong" things.
> But your download speed is still going to be primarily limited by your internet connection to PyPI.
Downloading lots of small packages dependencies serially leaves a lot of performance on the table due to latency and non-instantaneous response from congestion controllers. Downloading and installing concurrently reduces walltime further.
> Well of course the thing trying to address the problem addresses the problem....
The point is that it is a thing trying to address the "problem", and that not everyone considers it a problem.
> Which is what uv does and pip doesn't.
The point is that you might want to install something not for use in a "project", and that you might want to explicitly hand-craft the full contents of the environment. Pip is fundamentally a lower-level tool than uv.
> The thing that should be automatic.
Bootstrapping pip is the default so that people who have barely learned what Python is don't ask where pip is, or why pip isn't installing into the (right) virtual environment.
Yes, there are lots of flaws in pip. The problem is not virtual environments. Uv uses the same virtual environments. Neither is the problem "being a low-level tool that directly installs packages and their dependencies". I actively want to have that tool, and actively don't want a tool that tries to take over my entire project workflow.
As mostly a Python outsider, in the infrequent times that I do use python package management, uv just works. When I use pip I’d get all sorts of obscure error messages that I’d have to go track down, probably because I got some obscure environment detail wrong. With uv I never run into that nonsense.
Design-wise, nothing, IMO. But I don't fault people who prefer the uv workflow, either. Chacun a son gout.
Implementation-wise, there's nothing wrong in my view with venv. Or rather, everything is compelled to use virtual environments, including uv, and venv is just a simple tool for doing so manually. Pip, on the other hand, is slow and bulky due to poor architecture, a problem made worse by the expectation (you can work around it, but it requires additional understanding and setup, and isn't a perfect solution) of re-installing it into each virtual environment.
(The standard library venv defaults to such installation; you can disable this, but then you have to have a global pip set up, and you have to direct it to install into the necessary environment. One sneaky way to do this is to install Pipx, and then set up some script wrappers that use Pipx's vendored copy of pip. I describe my techniques for this in https://zahlman.github.io/posts/2025/01/07/python-packaging-....)
Edit: by "design" above I meant the broad strokes of how you use pip, installing single packages with their transitive dependencies etc. There's a lot I would change about the CLI syntax, and other design issues like that.
venv is lower-level tooling. Literally all it does is create a virtual environment — the same kind that uv creates and manages. There's nothing to "integrate".
I love uv. But the post starts with a simple install using a oneliner curl piping to sh, which is such a big attack surface area… I would much rather have a much longer one liner that increases safety.
If you look at the script, this is indeed more or less what happens. Except the folks over there are very clever about ergonomics, so the script is quite long so it can detect your architecture, OS, and even libc to give you an appropriate binary. There’s a tool that they use (which they wrote) which generates such install scripts for you
I love how uv allows me to not think of all the options anymore.
virtualenv, venv, pyenv, pipenv... I think at one point the recommended option changed because it was integrated into Python, but I can't even remember which is which anymore.
Such a pleasure to finally have just one, for maybe... ~99% of my needs.
I'm weird, this opinion Is even weirder, but how about start removing what makes Python slow? Sometimes you need yo remove features. It's hard, a pain but i dont see any other way of fixing really big mistakes...
Ty is still under very active development, so it either works or very much doesn't. I run it occasionally to see if it works on my codebases, and while it is getting closer, it isn't quite there yet.
Seems like a commercial blog. And imho hatch is better from a Foss perspective.
UV means getting more strings attached with VC funded companies and leaning on their infrastructure. This is a high risk for any FOSS community and history tells us how this ends….
This is going to sound harsh, but the problem with hatch is that it's pypa. And look at all the people that equate python-the-language with problems in pypa-managed solutions already. Pypa does not make good stuff or make good decisions.
Speaking of history, I was very sympathetic to the "we are open-source volunteers, give us a break" kind of stuff for the first N years.. but pypa has a pattern of creating problems, ignoring them, ignoring criticism, ignoring people who are trying to help, and pushing talent+interest elsewhere. This has fragmented the packaging ecosystem in a way that confuses newcomers, forces constant maintenance and training burden on experts, and damages the credibility of the language and its users. Hatch is frankly too little too late, and even if it becomes a wonderful standard, it would just force more maintenance, more confusion for a "temporary" period that lasts many, many years. Confidence is too far gone.
As mentioned elsewhere in the thread, there are tons of conflicting tools in the space already, and due to the fragmentation, poetry etc could never get critical mass. That's partly because pypa stuff felt most "official" and a safer long term bet than anything else, but partly because 33% better was never good enough to encourage widespread adoption until it was closer to 200% better. But uv actually IS that much better. Just let it win.
And let pypa be a case-study in how to NOT do FOSS. Fragmentation is fine up to a point, but you know what? If it wasn't for KDE / Gnome reinventing the wheel for every single kind of individual GUI then we'd have already seen the glorious "year of the linux desktop" by now.
curl|sh and iwr|iex chills my spine, no one should recommend these methods of installation in 2025. I'm against closed computers but I'm also against reckless install. Even without the security concerns these way of installation tends to put files in a whole random places making it hard to manage and cleanup.
While I do share the sentiment, I firmly believe that for opensource, no one should require the author to distribute their software, or even ask them to provide os-specific installation methods. They wrote it for free, use it or don't. They provide a handy install script - don't like it? sure, grab the source and build it yourself. Oops, you don't know what the software does? Gotta read every line of it, right?
Maybe if you trust the software, then trusting the install script isn't that big of a stretch?
For small project open source with a CLI audience, why bother with an install script at all and not just provide tarballs/ZIP files and assume that the CLI audience is smart enough to untarball/unzip it to somewhere on their PATH?
Also, many of the "distribution" tools like brew, scoop, winget, and more are just "PR a YAML file with your zip file URL, name of your EXE to add to a PATH, and a checksum hash of the zip to this git repository". We're about at a minimum effort needed to generate a "distribution" point in software history, so seems interesting shell scripts to install things seem to have picked up instead.
Part of writing software involves writing a way to deploy that software to a computer. Piping a web URL to a bash interpreter is not good enough. if that's the best installer you can do the rest of your code is probably trash.
A package is at least a signable, checksummable artefact. The curl | sh thing could have been anything and after running it you have no record of what it was you did.
There have also been PoCs on serving malicious content only when piped to sh rather than saved to file.
If you want to execute shell code from the internet, at the very least store it in a file first and store that file somewhere persistent before executing it. It will make forensics easier
If you're going to run code without inspecting it though, the methods are similar. One case has https, the other a signature (which you're trusting due to obtaining it over https). You can't inspect it reliably only after getting hypothetically compromised.
That iwr|iex example is especially egregious because it hardcodes the PowerShell <7.0 EXE name to include `-ExecutionPolicy Bypass`. So it'll fail on Linux or macOS, but more importantly iwr|iex is already an execution bypass, so including a second one seems a red flag to me. (What else is it downloading?)
Also, most reasonable developers should already be running with the ExecutionPolicy RemoteSigned, it would be nice if code signing these install script was a little more common, too. (There was even a proposal for icm [Invoke-Command] to take signed script URLs directly for a much safer alternative code-golfed version of iwr|iex. Maybe that proposal should be picked back up.)
A malicious server could detect whether the user is actually running "curl | sh" instead of just "curl" and only serve a malicious shell script when the code is executed blindly. See this thread for reference: https://news.ycombinator.com/item?id=17636032
well you still have to execute the shell script at some point. You could do curl > install.sh, open it up to inspect, and then run the install script which would still trigger the callback to the server mentioned in the link you posted. I guess it's really up to the user to decide what programs to run and not run.
That doesn't fix the core issue. You can put anything inside a .deb file, even preinstall script can send your ~/.aws/credentials to China. The core concern is getting a package that's verified by a volunteer human to not contain anything malicious, and then getting that package into Debian repository or equivalent.
It's surprising to me how long it can take for some languages to get decent package management solutions. There are no silver bullets because it's tricky to "encode" compatibility in a version number. I personally think semver helped a little and damaged a lot more by selling a pseudo solution that stands no chance to solve the real problem it needs to.
Maven has always been a very good solution. I think Bazel is too, but haven't had much experience with it.
I think ruff is the best thing to happen to the python ecosystem in a decade, it really sold the entire community on the difference fast native tooling could make.
I still feel bitten by diving into poetry when starting some projects. Has the ecosystem fully moved on to uv, now? Do they have good influence on what python's main ecosystem is moving to?
I for now prefer to stick to whatever the default is from the python packaging crew and standard library i.e. `python -m venv` and `pip install` inside of it.
Python for me is great when things can remain as simple to wrap your head around as possible.
Managing environments with `python -m venv` and all of the easy ways that goes wrong is exactly what I don't want to deal with. Is almost enough to make me never want to use python.
I completely agree. Deploying Python packages like MCP servers has been a real game changer. I'm so glad the days of wrestling with conda environments and Jupyter kernels are behind us. I used to start personal projects, decide to clean up my Python setup first, and inevitably give up on the project after getting lost in the cleanup.
What about Pixi[1]? It has become an irreplaceable part of my dev stack. Fantastic for tool + library version management. It has replaced a number of tools for me and greatly simplified bootstrapping in a new environment (like lxc containers when I am experimenting with stuff) or creating a lightweight sandbox for AI agents.
uv is great. I am a Ruby developer and I always loathed having to work with Python libraries because of how bad the tooling was. It was too complex to learn for the one-off times that I needed it and nothing worked properly.
Now with uv everything just works and I can play around easily with all the great Python projects that exist.
Been using pip and venv for long. I get uv is faster, but I don't get why people drool over it this much. What is wrong with pip + venv? I build webapps so perhaps I don't see issues in ML world
Using Python on and off for OS scripting since version 1.6.
It has always been enough to place installations in separate directories, and use the same bash scripts for environment variables configuration for all these years.
I’m surprised not to see a discussion of the biggest drawback: despite being fewer characters, “uv” is harder to type than “pip”. It requires two different hands to participate and a longer reach with my left index finger. pip is convenient – just a little rattle off with my right hand.
Can I just start using python if I've already got a bunch of projects manage with venv / pyenv / virtualenv ( and tbh I've kinda got into a confused mess with all these venv things, and at this point just hope they all keep working...)
Does uv handle CUDA versioning? This is the big reason I'm still on conda -- I can save a whole environment with `conda list --explicit`, including CUDA stuff, and I can set up a new machine with the same environment just from that file.
> uv is an incredibly powerful simplification for us that we use across our entire tech stack. As developers, we can all work with identical Python installations, which is especially important given a number of semi-experimental dependencies that we use that have breaking changes with every version. On GitHub Actions, we’re planning to use uv to quickly build a Python environment and run our unit tests. In production, uv already manages Python for all of our servers.
> It’s just so nice to always know that Python and package installation will always be handled consistently and correctly across all of our machines. That’s why uv is the best thing to happen to the Python ecosystem in a decade.
I can only conclude, that the author of the article, and perhaps even the organization they work in, is unaware of other tools that did the job long before uv. If they really value reproducibility that much, how come they didn't look into the matter before? Things much have been really hastily stitched together, if no one ever looked at existing tooling before, and only now they make things reproducible.
I guess reproducibility is still very much a huge problem, especially in jobs, where it should be one of the most important things to take care of: Research. ("Astronomer & Science Communicator" it says on the website). My recommendation is: Get an actual software developer (at least mid-level) to support your research team. A capable and responsibly acting developer would have sorted this problem out right from the beginning.
I am glad they improved their project setups to the level they should be at, if they want to call it research.
> I can only conclude, that the author of the article, and perhaps even the organization they work in, is unaware of other tools that did the job long before uv. If they really value reproducibility that much, how come they didn't look into the matter before? Things much have been really hastily stitched together, if no one ever looked at existing tooling before, and only now they make things reproducible.
Yes, Poetry has had lock files for years, and pyenv has been able to manage installations, but uv is "an incredibly powerful simplification" that makes it easy to do everything really well with just one tool.
Also, I kinda feel dirty criticizing an open source project but Poetry seems to be struggling with technical debt. I hit bugs which have been open for years or stuff which is WONTFIXed, and while they truly do not owe me anything, it’s a lot more rewarding to use uv where I hit fewer issues in general and the stuff I do hit is usually fixed quickly.
There’s a bigger conversation about open source maintenance there, but if I have to get my job done it’s increasingly tempting to take the simplifications and speed.
Doesn't really explain, how their organization apparently ran around without proper lock files before, when they are a researcher. If anything, then this article is shining light on the previously bad state of project setup in the organization.
My biggest frustration is the lack of a good universal REPL to just play around with. It's frustrating how I have to run `uvx --with x,y,z ipython` every single time I just want to spin up some python code which may or may not use packages. (Hard to overstate how annoying it is to type out the modules list).
To me, Python's best feature is the ability to quickly experiment without a second thought. Conda is nice since it keeps everything installed globally so I can just run `python` or iPython/Jupyter anywhere and know I won't have to reinstall everything every single time.
Would creating a `main.py` with the dependencies installed either as a uv project or inline work for you?
One thing I did recently was create a one-off script with functions to exercise a piece of equipment connected to the PC via USB, and pass that to my coworkers. I created a `main.py` and uv add'ed the library. Then when I wanted to use the script in the REPL, I just did `uv run python -i main.py`.
This let me just call functions I defined in there, like `set_led_on_equipment(led='green', on=True)` directly in the REPL, rather than having to modify the script body and re-run it every time.
Edit: another idea that I just had is to use just[0] and modify your justfile accordingly, e.g. `just pything` and in your justfile, `pything` target is actually `uv run --with x,y,z ipython`
Edit edit: I guess the above doesn't even require just, it could be a command alias or something, I probably am overengineering that lol.
Python is not my first language but I've always liked it. But project and dependency
management was always a bit meh and an afterthought.
Over the years, I've tried venv, conda, pipenv, petry, plain pip with requirements.txt. I've played with uv on some recent projects and it's a definite step up. I like it.
Uv actually fixes most of the issues with what came before and actually builds on existing things. Which is not a small compliment because the state of the art before uv was pretty bad. Venv, pip, etc. are fine. They are just not enough by themselves. Uv embraces both. Without that, all we had was just a lot of puzzle pieces that barely worked together and didn't really fit together that well. I tried making conda + pipenv work at some point. Pipenv shell just makes using your shell state-full just adds a lot of complexity. None of the IDEs I tried figured that out properly. I had high hopes for poetry but it ended up a bit underwhelming and still left a lot of stuff to solve. Uv succeeds in providing a bit more of an end to end solution. Everything from having project specific python installation, venv by default without hassle, dependency management, etc.
My basic needs are simple. I don't want to pollute my system python with random crap I need for some project. So, like uv, I need to have whatever solution deal with installing the right python version. Besides, the system python is usually out of date and behind the current stable version of python which is what I would use for new projects.
God yes. I got dragged into the uv when I started using copyparty and I am a fanatical admirer ever since. I also use pipx to install tools often. I really don't understand why you can't just pip install something globally. I want this package to be available to me EVERYWHERE, why can't I do it? I only use python recreationally because everyone uses python everywhere and you can't escape it. So there is a massive possibility I am simply wrong and pip-installing something globally is a huge risk. I'm just not understanding it.
> I really don't understand why you can't just pip install something globally. I want this package to be available to me EVERYWHERE, why can't I do it? I only use python recreationally because everyone uses python everywhere and you can't escape it. So there is a massive possibility I am simply wrong and pip-installing something globally is a huge risk. I'm just not understanding it.
You may have a library that's been globally installed, and you have multiple projects that rely on it. One day you may need to upgrade the library for use in one project, but there are backward incompatibile changes in the upgrade, so now all of your other projects break when you upgrade the global library.
In general, when projects are used by multiple people across multiple computers, it's best to have the specific dependencies and versions specified in the project itself so that everyone using that project is using the exact same version of each dependency.
For recreational projects it's not as big of a deal. It's just harder to do a recreation of your environment.
> I want this package to be available to me EVERYWHERE, why can't I do it?
Because it being available in the system environment could cause problems for system tools, which are expecting to find something else with the same name.
And because those tools could include your system's package manager (like Apt).
> So there is a massive possibility I am simply wrong and pip-installing something globally is a huge risk. I'm just not understanding it.
I assume you're referring to the new protections created by the EXTERNALLY-MANAGED marker file, which will throw up a large boilerplate warning if you try to use pip to install packages in the system environment (even with --user, where they can still cause problems when you run the system tools without sudo).
* the Python forum discussion (https://discuss.python.org/t/_/56900) where it was originally noticed that the Stack Overflow Q&A was advising people to circumvent the protection without understanding it, and a coordinated attempt was made to remedy that problem.
The Machine-Learning world, especially "Google Brain" research team figured out that NumPy was an awesome piece of software for dealing with large arrays of numbers and matrix multiplication. They built "TensorFlow" on top of it around 2015 which became very popular. Facebook followed suit and released PyTorch in 2016.
IPython/Jupiter notebooks (for Julia, Python and R) from 2015 were another factor, also adopted by the AI/ML community.
The alternative data-science languages at the time were Mathematica, MATLAB, SAS, Fortran, Julia, R, etc, but Python probably won because it was general purpose and open source.
I suspect Python would not have survived the 2/3 split very well if it wasn't for AI/ML adopting Python as its main language.
> when the tooling was so inferior
Since 2012, Conda/Anaconda has been the go-to installer in the SciPy/NumPy world which also solves a lot of problems that uv solves.
I either want one universal tool that can manage this sort of thing across multiple languages (eg. devenv) or a native, built-in tool (eg. go's tooling). I don't see how this is any different from all the previous incarnations of Python's project/package management tools. The constant churning of 3rd party tooling for Python was one of the main reasons I mostly stopped using it for anything but smaller scripts.
The difference is that this one is actually good. So good, in fact, that there is considerable momentum and thus adoption with this tool, and I wouldn’t be surprised if it reaches a similar state like npm is for node eventually.
Every time I see one of these comment threads it seems like uv desperately needs a better home page that doesn’t start with a long list of technical stuff. It’s really simple to use, in fact so simple that it confuses people!
There is literally no easy way to also have a configuration for CUDA, you have to have a second config, and, the worse, manually copy/symlink them into the hardcoded pyproject.toml file
Or is it a corporate grab to gain more influence in the ecosystem? I like the idea, but for profit backing is out of the question. This lesson has been learned countless times.
no, it's a python library, get a grip. Also "This lesson has been learned countless times"? No it hasn't, since when has a package manager developed by a for-profit company hurt the ecosystem?
uv is finally an all-in-one tool that finally takes all of the good ideas from previous projects and combines them together to work well as one (and unbelievably fast).
The fact that it's a binary, not written in python, also simplifies bootstrapping. So you don't need python+dependencies installed in order to install your python+dependencies.
One helpful element that has changed over the years compared to the old wild west days is the large number of PEPs that have quietly in the background bit by bit standardized packaging formats and requirements.
Some foundations have moved into the stdlib. This means that newer tools are much more compatible with each other and mainly just differ in implementation rather than doing different things altogether. The new stuff is working on a much more standard base and can leave behind many dark crufty corners.
Unravelling the legacy stuff and putting the standards in place seems to have taken 15+ years?
All of these tools are third-party and the Python core development team can't do anything to prevent people from inventing new ones. Even pip is technically at arms length; it has special support in the standard library (Python releases will vendor a wheel for it, which is designed to be able to bootstrap itself for installation[0]), but is developed separately.
Standards are developed to allow existing tools to inter-operate; this entails allowing new tools to appear (and inter-operate), too.
This system was in some regards deliberate, specifically to support competition in "build backends". The background here is that many popular Python projects must interface to non-Python code provided with the project; in many cases this is code in compiled languages (typically C, Fortran or Rust) and it's not always possible to pre-build for the user's system. This can get really, really complicated, and people need to connect to heavyweight build systems in some cases. The Python ecosystem standards are designed with the idea that installers can automatically obtain and use those systems when necessary.
And by doing all of this, Python core developers get to focus on Python itself.
Another important concern is that some bad choices were made initially with Setuptools, and we have been seeing a very long transition because of a very careful attitude towards backwards compatibility (even if it doesn't seem that way!) which in turn is motivated by the battle scars of the 2->3 transition. In particular, it used to be normal and expected that your project would use arbitrary Python code (in `setup.py` at the project root) simply to specify metadata. Further, `setup.py` generally expects to `import setuptools`, and might require a specific version of Setuptools; but it can't express its build-time Setuptools version requirement until the file is already running - a chicken-and-egg scenario.
Modern projects use a declarative TOML file for "abstract" metadata instead (which is the source for concrete metadata included in the actual build artifacts), but the whole ecosystem still has to support a lot of really outdated ways of doing things, because in part of how much abandonware is out there.
[0]: Wheels are zip-compressed, and Python can run code from a zip file, with some restrictions. The pip project is designed to make sure that this will work. The standard library provides a module "ensurepip" which locates this wheel and runs a bootstrap script from that wheel, which will then install into the current environment. Further, the standard library "venv", used to create virtual environments, defaults to using this bootstrap in the newly created environment.
It's helpful context but still seems like a lost opportunity for python to provide the UI. It feels like every couple years we are reworking the wheel and redefining how to publish software.
With python over the years i can think of pip, pipx, setuptools, easy_install, distutils, venv, conda, wheel, .egg, wheel (formats) , now uv.
PHP stabilized with composer, perl with cpan , go with `go mod` and `go get` (builtin).
Java and Swift had some competition with Gradle/maven and swiftPM / cocoapods, but nothing as egregious.
file tree, dep tree, task DAG. how many ways can they be written?
It's not just a matter of how they're written. For Python specifically, build orchestration is a big deal. But also, you know, there are all the architecture ideas that make uv faster than pip. Smarter (and more generous) caching; hard-linking files where possible rather than copying them; parallel downloads (I tend to write this off but it probably does help a bit, even though the downloading process is intermingled with resolution); using multiple cores for precompiling bytecode (the one real CPU-intensive task for a large pure-Python installation).
It sounds great and I’m not against Uv . It probably is the best . I’m wondering what’s wrong with the Python community that 25 years sees 10 package managers. I’m not being cynical it’s a clinical / empirical question
Yes, uv is probably the best thing to happen to the Py ecosystem in the last decade. That is mainly because the rest of the ecosystem is somewhere between garbage fire and mediocre at best.
uv in itself is a great tool, I have no complaints about it whatsoever! But we have to remember just how bad the rest of things are and never forget that everything's still in a pretty bad state even after more than 3 ** DECADES ** of constant evolution.
They have different use cases. uv is meant to be the singular tool for managing Python packages and dependencies, replacing pip, virtualenv, and pip-tools. Conda is for more general-purpose environment management, not just Python. If you're doing something with Node or R, uv won't work at all because it's only for Python.
uv's biggest advantage is speed. It claims a 10-100x performance speedup over pip and Conda [1]. uv can also manage python versions and supports using Python scripts as executables via inline dependencies [2].
But Conda is better for non-Python usage and is more mature, especially for data science related uses.
I can only agree. I'm not an python expert, but I always struggled when installing a new package and got the warning, that it could break the system packages, or when cloning an existing repo on a new installed system.
Always wondered, why it became so "complicated" over the years.
I have one problem with uv as of now, and it's more of an annoyance. It doesn't seem to understand the concept of >= when it's trying to resolve a local wheel I built and use. If I have 6.4.1 published on GitLab and the pyproject says $WHEEL_NAME>=6.2.0, it still goes to look for 6.2.0 (which I deleted) and errors out.
I haven't tried uv yet, but I did use it's precursor - rye.
I had to update some messy python code and I was looking for a tool that could handle python versions, package updates, etc. with the least amount of documentation needing be read and troubleshooting.
Rye was that for me! Next time I write python I'm definitely going to use uv.
Indeed rye is great and switching to uv is pretty straight forward. I still think rye's use of shims was pretty cool but probably uv's approach is more sane
uv is the best tool out there as long as you have python only dependencies. It's really fast, and you can avoid using poetry, pipenv, etc. The only reason for conda to still exist is non pythonic dependencies, but that's another beast to tackle in itself.
This resonates so much. As someone who's more on the builder/product side than engineering, I've always felt that barrier with Python tooling. The learning curve for environment management has been one of those silent productivity killers.
What strikes me about uv is that it seems to understand that not everyone launching a Python-based project has a CS degree. That accessibility matters—especially in the era where more non-engineers are building products.
Curious: for those who've switched to uv, did you notice any friction when collaborating with team members who were still on traditional setups? I'm thinking about adoption challenges when you're not a solo builder.
For years I've avoided using Python tools because I've always struggled to get them working properly. Will uv solve this pain for me? Can I install a Python app globally with it?
I'm not a pythonista, and the most recent time I've been playing with python has been using octodns. origninally I was using a pip setup, and honestly wow UV was so much faster.
I'm very happy the python community has better tooling.
I don't like that it defaults to putting the virtual environment right there, I much prefer how pipenv does it with a shared one in the users home directory, but it's a small price to pay for how fast it is.
I am still learning and I have the same feeling as someone who don't consider myself good with python. At least I can keep my venv in control now is all I can feel with Uv approach.
Python venv's is the #1 reason I've avoided working with it more. It used to be #2 behind strong typing, but now that Linux OSes' take up the default python install and block it from being used for quick scripts, it jumped to #1.
I've always wondered why Linux OSes that rely on python scripts don't make their own default venv and instead clobber the user's default python environment...
Hmpf. I am using uv now, but I have been doing fine before using poetry. For me it is not a huge revolution, as I always value reproducibility, which means lock file and checksums, and that, I was able to have before using poetry. Yes, yes, ... uv is faster. I grant them that. And yes, it's pleasant, when it runs so quickly. But I am not changing dependencies that often, that this really impacts my productivity. A venv is created, it stays. Until at some point I update pyproject.toml and the lock file.
Since I am mostly avoiding non-reproducible use-cases, like for example stating dependencies inside the python scripts themselves, without checksums, only with versions, and stuff like that, I am not really benefiting that much. I guess, I am just not writing enough throwaway code, to benefit from those use-cases.
Some people here act, like uv is the first tool ever to install dependencies like npm and cargo and so on. Well, I guess they didn't use poetry before, which did just that.
I've used it in various work projects/services, and in my free time in various projects. Never had anything "flaky" about it happening. Care to elaborate what you mean by that?
uv has made working with different python versions and environments much, much nicer for me. Most of my colleagues in computational genomics use conda, but I've yet to encounter a scenario where I've been unable to just use uv instead.
I find the python tooling so confusing now. There’s pip, virtualenv, pipx, uv, probably half a dozen others I’m missing. I like node, npm isolates by default, npx is easy to understand, and the ecosystem is much less fragmented. I see a python app on GitHub and they’re all listing different package management tools. Reminds me of that competing standards xkcd.
Node has at least bun, and probably other tools, that attempt to speed things up in similar ways. New tooling is always coming for our languages of choice, even if we aren't paying attention.
Yes, for years I've sat on the sidelines avoiding the fragmented Poetry, ppyenv, pipenv, pipx, pip-tools/pip-compile, rye, etc, but uv does now finally seem to be the all-in-one solution that seems to be succeeding where other tools have failed.
I had a recent period in my programming career where I started to actually believe that the "worse is better" philosophy is true in practice. It was a dark period and thankfully the existence of tools like uv save me from that abyss.
Rust is the best thing to happen to the Python (and JS) ecosystem in a decade. Once people realized that the tooling doesn't need to be written in the same language as the target language, it opens up all sorts of performance possibilities.
Am I the only one who feels like this is obviated by Docker?
uv is a clear improvement over pip and venv, for sure.
But I do everything in dev containers these days. Very few things get to install on my laptop itself outside a container. I've gotten so used to this that tools that uninstall/install packages on my box on the fly give me the heebie-jeebies.
> Am I the only one who feels like this is obviated by Docker?
This whole discussion has the same vibes like digital photography 15 years ago. Back then some people spent more time on discussing the tech spec their cameras than takin photos. Now some people spend more time on discussing the pros and cons of different Python environment management solutions than building real things.
The last time I had to touch one of my dockerized environments was when Miniconda and Miniforge were merged. I said the agent "fix the dockerfile", and the third attempt worked. Another time, one dependency was updated and I had to switch to Poetry. Once again, I said the agent "refactor the repository to Poetry" and it worked. Maybe because all my Python package versions are frozen and I only update them when they break or when I need the functionality of the new version.
Whenever this topic pops up in real life, I always ask back what was the longest time they managed the same Python service in the cloud. In the most cases, the answer is never. The last time someone said one year. After a while this service was turned into two .py files.
I don't know. Maybe I'm just too far away from FAANG level sorcery. Everything is a hammer if all you have to deal with are nails.
Speed matters everywhere. How much compute is spent on things that could easily be 100x faster than they are? Compare using VMware with pip to run a battery of unit tests with firecracker plus uv. It’s orders of magnitude quicker, and avoids a whole suite of issues related to persistent state on the machine
Possibly for some workflows, though personally I find the emphasis on speed baffling and a big part of the reason I don’t find most of these uv testimonials credible. I’m a regular python user across multiple environments and I’ve never considered waiting for pip to be a material part of my time, it’s trivial to the point of being irrelevant. The fact that so many people come out of the woodwork to talk about how fast it is, means either there’s some big group somewhere with a niche use case that gets them bogged down in pip dependency resolving or whatever gets sped up (obviously the actual downloading can’t be faster) or it’s just a talking point that (presumably) rust zealots who don’t actually use python arrive with en mass, but it’s honestly an extremely ineffective way of promoting the product to most python users who don’t have speed of package installation as anything close to a pain point.
It's fast enough that sometimes dependencies can be checked and resolved and installed at program runtime rather than it needing to be a separate step.
You can go from no virtual environment, and just "uv run myfile.py" and it does everything that's needed, nearly instantly.
lol who is using pip so much that .36s of startup time matters to them? This, if presumably uv can do nothing slightly faster, is an absolutely meaningless benefit
In general, whenever you introduce a cache to make software faster (along any dimension), you have to think about cache invalidation and eviction. If your software is fast enough to not need caching, this problem goes away.
Uv is available as a wheel from PyPI, so you can in fact `pip install uv` into an appropriate environment. Since it provides a command-line binary, Pipx will also happily install it into an environment it manages for you. And so on and so forth. (You can even install uv with uv, if you want to, for whatever reason.)
The wheel basically contains a compiled ~53MB (huh, it's grown in recent versions) Rust executable and a few boilerplate files and folders to make that play nice with the Python packaging ecosystem. (It actually does create an importable `uv` module, but this basically just defines a function that tells you the path to the executable.)
The best thing to happen to the Python ecosystem would be something that unites pip and conda. Conda is not going anywhere given how many packages depend on non-python binaries, especially in enterprise settings.
You might be interested in Pixi: https://prefix.dev/
It uses uv under the hood for Python dependencies, while allowing you to also manage Conda dependencies in the same manifest (pixi.toml). The ergonomics are really nice and intuitive imo, and we're on our way to replace our Poetry and Conda usage with only Pixi for Python/C++ astrodynamics projects. The workspace-centric approach along with native lockfiles made most of our package management issues go away. I highly recommend it!
(Not affiliated anyhow, other than contributing with a simple PR for fun)
I'm not sure if you're aware, but there's the Wheel Variants proposal [0] that the WheelNext initiative is working through that was presented at PyCon 2025 [1][2], which hopes to solve some of those problems.
uv has implemented experimental support, which they announced here [3].
The standard approach nowadays is to vendor the binaries, as e.g. Numpy does. This works just fine with pip.
I'm interested if you have any technical documentation about how conda environments are structured. It would be nice to be able to interact with them. But I suspect the main problem is that if you use a non-conda tool to put something into a conda environment, there needs to be a way to make conda properly aware of the change. Fundamentally it's the same issue as with trying to use pip in the system environment on Linux, which will interfere with the system package manager (leading to the PEP 668 protections).
I had this discussion briefly with a buddy who uses python exclusively for his career in austronomy. He was lamenting the pains of colaborting around Conda and seemed convinced it was irreplaceable. Being that I'm not familiar with the exact limitations Conda is providing for, Im curious if you could shed some insight here. Does nix not technically solve the issue? I understand this isn't solely a technical problem and Nix adoption in this space isn't likely, but I'm curious none-the-less!
Mojo stopped saying out loud they are trying to be a Python superset. Maybe they can do it one day but they're keeping that on the DL now because it's a really big ask.
I've been using uv and am pleased that is about as useful as maven was the last time I used it 12 years ago. I'm not really sure why we still need venv.
There have been actually many cases in my experience where venv simply worked but uv failed to install dependencies. uv is really fast but usually you need to install dependencies just once.
Occasionally I have to build Python projects and coming from other languages and package managers, having to deal with a venv is super weird and annoying.
This is it. Later versions of python .11/.12/.13 have significant improvements and differences. Being able to seamlessly test/switch between them is a big QOL improvement.
I don't love that UV is basically tied to a for profit company, Astral. I think such core tooling should be tied to the PSF, but that's a minor point. It's partially the issue I have with Conda too.
> Later versions of python .11/.12/.13 have significant improvements and differences. Being able to seamlessly test/switch between them is a big QOL improvement.
I just... build from source and make virtual environments based off them as necessary. Although I don't really understand why you'd want to keep older patch versions around. (The Windows installers don't even accommodate that, IIRC.) And I can't say I've noticed any of those "significant improvements and differences" between patch versions ever mattering to my own projects.
> I don't love that UV is basically tied to a for profit company, Astral. I think such core tooling should be tied to the PSF, but that's a minor point. It's partially the issue I have with Conda too.
In my book, the less under the PSF's control, the better. The meager funding they do receive now is mostly directed towards making PyCon happen (the main one; others like PyCon Africa get a pittance) and to certain grants, and to a short list of paid staff who are generally speaking board members and other decision makers and not the people actually developing Python. Even without considering "politics" (cf. the latest news turning down a grant for ideological reasons) I consider this gross mismanagement.
Yes. poetry & pyenv was already a big improvement, but now uv wraps everything up, and additionally makes "temporary environments" possible (eg. `uv run --with notebook jupyter-notebook` to run a notebook with my project dependencies)
If that works for you, then that's cool. Personally, I don't want to think about environments, and it's weird that python is the only language that has venvs. Having a tool that handles it completely transparently to me is ideal, to me.
One thing that annoys me about Claude is that it doesn't seem to create a venv by default when it creates a python project. (But who knows, maybe 1/3 of the time it does or something.) But you have to ask each time to be sure.
Honestly though it's a pretty rough indictment of Python that the best thing to happen in a decade is that people started writing Python tools in Rust. Not even a little Rust, uv is 98% Rust. I mean, they just released 3.14 and that was supposed to be a pretty big deal.
I sometimes wonder if many core Python people don’t actually like the language that much. That’s why (a) they’re constantly reinventing it, and (b) they celebrate rewrites from Python into other languages. Long before Rust, it was considered a good thing when a standard library module was rewritten in C.
Compare this to the Go community, who celebrate rewrites from other languages into Go. They rewrote their compiler in Go even though that made it worse (slower) than the original C version, because they enjoy using their own language and recognise the benefits of dogfooding.
No, the "best thing that happened" (in TFA's author's opinion) is that this specific tool exists, with its particular design. Rust is an implementation detail. Most of the benefit that Uv offers over pip, in my analysis, is not a result of being written in Rust.
I don't think Rust is incidental here. First, uv's particular design cargo culted from... well cargo. Which, they should be cause cargo is a great tool, no shade there.
But otherwise, people on this forum and elsewhere are praising uv for: speed, single-file executable, stability, and platform compatibility. That's just a summary of the top reasons to write in Rust!
I agree 3.14 is a big deal as far as Python goes, but it doesn't really move the needle for the language toward being able to author apps like uv.
It's called dogfooding -- writing tools for the language in the language. Not doing so here, where the result is "best thing to happen to the ecosystem in a decade", is a tacit admission that Python isn't up for the task of writing best-in-class Python tooling (the use of Rust wasn't incidental). Having seen uv, people will probably start writing more Python-ecosystem projects in Rust.
But I’m utterly shocked that UV doesn’t support “system dependencies”. It’s not a whole conda replacement. Which is a shame because I bloody hate Conda.
Dependencies like Cuda and random C++ libraries really really ought to be handled by UV. I want a true genuine one stop shop for running Python programs. UV is like 80% of the way there. But the last 20% is still painful.
Ideally UV would obsolete the need for docker. Docker shouldn’t be a requirement to reliable run a program.
> no point in trusting people who are disconnect from reality
My brother in christ, we are all just names without bodies or even faces on this digital ocean of the internet. Letting people know how they should address you isn't "disconnected from reality", it's grounded in the very real reality that we, as people, like talking to each other. We should all be so thankful for their foresight in allowing us the opportunity of avoiding an otherwise unavoidable faux pas of calling everyone in the world "hey you".
Why is this a problem? The ecosystem has developed usable interoperable standards (for example, fundamentally uv manages isolated environments by using the same kind of virtual environment created by the standard library — because that's the only kind that Python cares about; the key component is the `pyvenv.cfg` file, and Python is hard-coded to look for and use that); and you don't have to learn or use more than one.
There are competing options because people have different ideas about what a "package manager" should or shouldn't be responsible for, and about the expectations for those tasks.
It’s definitely an issue for learning the language. Obviously after working with python a bit that doesn’t matter, but fragmentation still makes it more of a hassle to get open source projects up and running if they don’t use something close to your usual package management approach.
All these comments look like advertisement.
"uv is better than python!!",
"8/10 programmers recommend uv",
"I was a terrible programmer before but uv changed my life!!",
"uv is fast!!!"
> All these comments look like advertisement. "uv is better than python!!", "8/10 programmers recommend uv", "I was a terrible programmer before but uv changed my life!!", "uv is fast!!!"
> Does it offer something that standard python tools doesn't?
Other than speed and consolidation, pip, pipx, hatch, virtualenv, and pyenv together roughly do the job (though pyenv itself isn’t a standard python tool.)
> Why uv over, lets say, conda?
Support for Python standard packaging specifications and consequently also easier integration with other tools that leverage them, whether standard or third party.
FWIW I asked the same question last time a uv thread was posted (two weeks ago) - got some legit answers, none that swayed me personally but I can see why people use it. Also lots of inexplicable love for it https://news.ycombinator.com/item?id=45574550
I agree that the speed improvements are inexplicable, as in I can't convince you in writing. "uv is fast!!!" doesn't do it justice. You kinda just have to experience it for yourself.
If you haven't spent 5 minutes trying it out, you don't know what you're missing.
If you're worried about getting addicted like everyone else, I could see that as a valid reason to never try it in the first place.
First time reading one of these threads? It’s a cult, and don’t dare criticize it. I think the same thing used to be true with rust though nobody really talks about it much anymore.
I don’t think people would think twice about the legitimacy (if you want to call it that) of uv except for all the weird fawning over it that happens, as you noticed. It makes it seem more like a religion or something.
A problem remain in that many and still more of the popular repositories don't use uv to manage their dependencies.
So you are back having to use conda and the rest. Now, you have yet another package manager to handle.
I wouldn't be harsh to engineers at astral who developed amazing tooling, but the issue with the python ecosystem isn't lack of tooling, it is the proliferation and fragmentation. To solve dependency management fully would be to incorporate other package descriptors, or convert them.
Rsbuild, another rust library, for the node ecosystem did just that. For building and bundling. They came up with rspack, which has large compatibility with the webpack config.
You find a webpack repo? Just add rsbuild, rspack, and you are pretty much ready to go, without the slow (node native) webpack.
When packages require conda, that has nothing to do with them "not using uv to manage their dependencies".
Conda solves a completely orthogonal set of problems, and is increasingly unnecessary. You can `pip install scipy` for example, and have been able to for a while.
I refered to the interfaces of other packaging tools. I use uv and it's excellent on its own.
You get a repo, it's using playwright, what do you do now ? You install all the dependencies found in the dependency descriptor then sync to create a uv descriptor. or you compose a descriptor that uv understands.
It's repetitive, rather systematic so it could be automated. I should volunteer for a PR but my point is introducing yet another tool to an ecosystem suffering a proliferation of build and deps management tooling expands the issue. It would have been helpful from the get go to support existing and prolific formats.
pnpm understands package.json
It didn't reinvent the wheel be cause we have millions of wheels out there. It created its own pnpm lock file, but that's files a user isn't meant to touch so it goes seamlessly to transition from npm to pnpm. Almost the same when migrating from webpack to rsbuild.
Uv combined with type hints reaching critical mass in the Python ecosystem, and how solid PyLance is in VSCode, feels so good it has made me consider investing in Python as my primary language for everything. But then I remember that Python is dog slow compared to other languages with comparable ergonomics and first-class support for static typing, and...idk it's a tough sell.
I know the performance meta in Python is to...not use python (bind to C, Rust, JVM) - and you can get pretty far with that (see: uv), but I'd rather spend my limited time building expertise in a language that isn't constantly hemorrhaging resources unless your code secretly calls something written in another language :/
There are so many good language options available today that compete. Python has become dominant in certain domains though, so you might not have a choice - which makes me grateful for these big steps forward in improving the tooling and ecosystem.
In two years I bet we’ll be seeing v8 level performance out of CPython.
It’s wildly optimistic to now expect a 10x speedup in two years, with fewer resources.
I also believe the JIT in v8 and Python are different, the latter relying on copy-and-patch while v8 uses a bunch of different techniques together.
I'd be quite delighted to see, say, 2x Python performance vs. 3.12. The JIT work has potential, but thus far little has come of it, but in fairness it's still the early days for the JIT. The funding is tiny compared to V8. I'm surprised someone at Google, OpenAI et al isn't sending a little more money that way. Talk about shared infrastructure!
But, they don't have the full compatibility with CPython, so nobody really picks them up.
... but then again neither pdm nor uv would have happened without poetry.
I recently had to downgrade one of our projects to 3.12 because of a dependency we needed. With uv, I can be sure that everybody will be running the project on 3.12, it just all happens automatically. Without uv, I'd get the inevitable "but your changes crashed the code, have you even tested them?"
Post like these aptly describe why companies are downsizing in lieu of AI assistants, and they are not wrong for doing so.
Yes, Python is "slow". The thing is, compute is cheap these days and development time is expensive. $1000 per month is considered expensive as hell for an EC2 instance, but no developer would work for $12000 a year.
Furthermore, in modern software dev, most of the bottlenecks is network latency. If your total end to end operation takes 200ms mostly because of network calls, it doesn't matter if you code runs in 10 ms or 5ms as far as compute goes.
When it comes to development, the biggest uses of time are
1. Interfacing with some API or tool, for which you have to write code 2. Making a change, testing a change, fixing bugs.
Python has both covered better than any other language. Just today, it took me literally 10 mins to write code for a menu bar for my Mac using rumps python library so I have most commonly used commands available without typing into a terminal, and that is without using an LLM. Go ahead and try to do the same in Java or Rust or C++ and I promise you that unless you have experience with Mac development, its going to take you way more time. Python has additional things like just putting breakpoint() where you want the debugger, jupyter notebooks for prototyping, and things like lazy imports where you use import inside a function so large modules only get loaded when they run. No compilation step, no complex syntax. Multiprocessing is very easy to use as a replacement for threading, really dunno why people want to get rid of GIL so much. Functionally the only difference is overhead in launching a thread vs launching a process, and shared memory. But with multiprocessing API, you simply spin up a worker pool and send data over Pipes, and its pretty much just as fast as multithreading.
In the end, the things that matter are results. If LLMs can produce code that works, no matter how stringy it is, that code can run in production and start making company money, while they don't have to pay you money for multiple months to write the code yourself. Likewise, if you are able to develop things fast, and a company has to spend a bit more on compute, its a no brainer on using Python.
Meanwhile like strong typing, speed, GIL, and other popular things that get mentioned is all just echos of bullshit education that you learned in CS, and people repeat them without actually having any real world experience. So what if you have weak typing and make mistakes - code fails to run or generate correct results, you go and fix the code, and problem solved. People act like failing code makes your computer explode or something. There is no functional difference between a compilation failure and a code running failure. And as far as production goes, there has never been a case of a strong type language that gets used that gets deployed and doesn't have any bugs, because those bugs are all logic bugs within the actual code. And consequently, with Python, its way easier to fix those bugs.
Youtube, Uber, and a bunch of other well used services all run Python backends for a good reason. And now with skilled LLM usage, a single developer can write services in days that would take a team of engineers to write in weeks.
So TL:DR, if you actually want to stay competitive, use Python. The next set of LLMs are all going to be highly specialized smaller models, and being able to integrate them into services with Pytorch is going to be a very valuable skill, and nobody who is hiring will give a shit how memory safe Rust is.
My hope is that conda goes away completely. I run an ML cluster and we have multi-gigabyte conda directories and researchers who can't reproduce anything because just touching an env breaks the world.
It's still very immature but if you have a mixture of languages (C, C++, Python, Rust, etc.) I highly recommend checking it out.
On the python front, however, I am somehow still an old faithful - poetry works just fine as far as I was every concerned. I do trust the collective wisdom that uv is great, but I just never found a good reason to try it.
It makes building FreeCAD pretty trivial, which is a huge deal considering FreeCAD’s really complex Python and non-python, cross-platform dependencies.
https://docs.metaflow.org/scaling/dependencies https://outerbounds.com/blog/containerize-with-fast-bakery
Want to make sure a software stack works well on a Cray with MPI+cuda+MKL, macOS, and ARM linux, with both C++ and Python libraries? It’s possible with conda-forge.
does that fit the bill?
Conda packaging system and the registry is capable of understanding things like ABI and binary compatibility. It can resolve not only Python dependencies but the binary dependencies too. Think more like dnf, yum, apt but OS-agnostic including Windows.
As far as I know, (apart from blindly bundling wheels), neither PyPI nor Python packaging tools have the knowledge of ABIs or purely C/C++/Rust binary dependencies.
With Conda you can even use it to just have OS-agnostic C compiler toolchains, no Python or anything. I actually use Pixi for shipping an OS-agnostic libprotobuf version for my Rust programs. It is better than containers since you can directly interact with the OS like the Windows GUI and device drivers or Linux compositors. Conda binaries are native binaries.
Until PyPI and setuptools understand the binary intricacies, I don't think it will be able to fully replace Conda. This may mean that they need to have an epoch and API break in their packaging format and the registry.
uv, poetry etc. can be very useful when the binary dependencies are shallow and do not deeply integrate or you are simply happy living behind the Linux kernel and a container and distro binaries are fulfilling your needs.
When you need complex hierarchies of package versions where half of them are not compiled with your current version of the base image and you need to bootstrap half a distro (on all OS kernels too!), Conda is a lifesaver. There is nothing like it.
Conda ecosystem is forced to solve this problem to a point since ML libraries and their binary backends are terrible at keeping their binaries ABI-stable. Moreover different GPUs have different capabilities and support different versions of the GPGPU execution engines like CUDA. There is no easy way out without solving dependency hell.
The curmudgeon in me feels the need to point out that fast, lightweight software has always been possible, it's just becoming easier now with package managers.
Rust is for me similar to C just like you wrote, it is better, bigger but not the overwhelming way like C++ (and Rust has cargo, don't know if C++ has anything).
I stayed for the native functional programming, first class enums, good parts of C++ and the ultimate memory safety.
I also like how you can manage Python versions very easily with it. Everything feels very "batteries-included" and yet local to the project.
I still haven't used it long enough to tell whether it avoids the inevitable bi-yearly "debug a Python environment day" but it's shown enough promise to adopt it as a standard in all my new projects.
You can also prepend the path to the virtual environment's bin/ (or Scripts/ on Windows). Literally all that "activating an environment" does is to manipulate a few environment variables. Generally, it puts the aforementioned directory on the path, sets $VIRTUAL_ENV to the venv root, configures the prompt (on my system that means modifying $PS1) as a reminder, and sets up whatever's necessary to undo the changes (on my system that means defining a "deactivate" function; others may have a separate explicit script for that).
I personally don't like the automatic detection of venvs, or the pressure to put them in a specific place relative to the project root.
> I also like how you can manage Python versions very easily with it.
I still don't understand why people value this so highly, but so it goes.
> the inevitable bi-yearly "debug a Python environment day"
If you're getting this because you have venvs based off the system Python and you upgrade the system Python, then no, uv can't do anything about that. Venvs aren't really designed to be relocated or to have their underlying Python modified. But uv will make it much faster to re-create the environment, and most likely that will be the practical solution for you.
``uv`` accomplishes the same thing, but it is another dependency you need to install. In some envs it's nice that you can do everything with the built-in Python tooling.
At least major and minor, patch is rarely needed for python.
However, I also think many people, even many programmers, basically consider such external state "too confusing" and also don't know how they'd debug such a thing. Which I think is a shame since once you see that it's pretty simple it becomes a tool you can use everywhere. But given that people DON'T want to debug such, I can understand them liking a tool like uv.
I do think automatic compiler/interpreter version management is a pretty killer feature though, that's really annoying otherwise typically afaict, mostly because to get non-system wide installs typically seems to require compiling yourself.
How does the rest of the world manage to survive without venvs? Config files in the directory. Shocking, really :-)))
The problem is, that would require support from the Python runtime itself (so that `sys.path` can be properly configured at startup) and it would have to be done in a way that doesn't degrade the experience for people who aren't using a proper "project" setup.
One of the big selling points of Python is that you can just create a .py file anywhere, willy-nilly, and execute the code with a Python interpreter, just as you would with e.g. a Bash script. And that you can incrementally build up from there, as you start out learning programming, to get a sense of importing files, and then creating meaningful "projects", and then thinking about packaging and distribution.
For .pth files to work, they have to be in a place where the standard library `site` module will look. You can add your own logic to `sitecustomize.py` and/or `usercustomize.py` but then you're really no better off vs. writing the sys.path manipulation logic.
Many years ago, the virtual environment model was considered saner, for whatever reasons. (I've actually heard people cite performance considerations from having an overly long `sys.path`, but I really doubt that matters.) And it's stuck.
source - why are we using an OS level command to activate a programming language's environment
.venv - why is this hidden anyway, doesn't that just make it more confusing for people coming to the language
activate - why is this the most generic name possible as if no other element in a system might need to be called the activate command over something as far down the chain as a python environment
Feels dirty every time I've had to type it out and find it particularly annoying when Python is pushed so much as a good first language and I see people paid at a senior level not understand this command.
Because "activating an environment" means setting environment variables in the parent process (the shell that you use to run the command), which is otherwise impossible on Linux (see for example https://stackoverflow.com/questions/6943208).
> why is this hidden anyway, doesn't that just make it more confusing for people coming to the language
It doesn't have to be. You can call it anything you want, hidden or not, and you can put it anywhere in the filesystem. It so happens that many people adopted this convention because they liked having the venv in that location and hidden; and uv gives such venvs special handling (discovering and using them by default).
> why is this the most generic name possible as if no other element in a system might need to be called the activate command over something as far down the chain as a python environment
Because the entire point is that, when you need to activate the environment, the folder in question is not on the path (the purpose of the script is to put it on the path!).
If activating virtual environments shadows e.g. /usr/bin/activate on your system (because the added path will be earlier in $PATH), you can still access that with a full absolute path; or you can forgo activation and do things like `.venv/bin/python -m foo`, `.venv/bin/my-program-wrapper`, etc.
> Feels dirty every time I've had to type it out
I use this:
Notice that, again, you don't have to put it at .venv . I use a .local folder to store notes that I don't want to publish in my repo nor mention in my project's .gitignore; it in turn has > and I see people paid at a senior level not understand this command.If you know anyone who's hiring....
> which is otherwise impossible on Linux
Node, Rust, etc all manage it.
> Because the entire point is that...
I just mean there is a history of Python using overly generic naming: activate, easy-install. Just feels weird and dirty to me that you'd call such a specific things names like these and I think it's indicative of this ideology that Python is deep in the OS.
Maybe if I'd aliased the activate command a decade ago I wouldn't feel this way or think about it.
uv has increased my usage of python for production purposes because it's maintainable by a larger group of people, and beginners can become competent that much quicker.
If uv makes it invisible it is a step forward.
not that it's great to start with, but it does happen, no?
Either the package manager is invoked with a different PATH (one that contains the desired Node/Java/whatever version as a higher priority item than any other version on the system).
Or the package manager itself has some way to figure that out through its config file.
Or there is a package manager launch tool, just like pyenv or whatever, which does that for you.
In practice it's not that a big of a deal, even for Maven, a tool created 21 years ago. As the average software dev you figure that stuff out a few weeks into using the tool, maybe you get burnt a few times early on for misconfiguring it and then you're on autopilot for the rest of your career.
Wait till you hear about Java's CLASSPATH and the idea of having a SINGLE, UNIFIED package dependency repo on your system, with no need for per-project dependency repos (node_modules), symlinks, or all of that stupidity.
CLASSPATH was introduced by Java in 1996, I think, and popularized for Java dependency management in 2004.
Activating a venv is just setting a few environment variables, including PATH, and storing the old values so that you can put them back to deactivate the environment.
Why else is this discussion getting hundreds of comments?
For any random python tool out there, I had about a 60% chance it would work out of the box. uv is the first tool in the python ecosystem that has brought that number basically to 100%. Ironically, it's written in Rust because python does not lend itself well to distributing reliable, fast tools to end users.
But whoever runs this has to install uv first, so not really standalone.
"Lol, no I break into computer systems I am a hacker"
"Geeze hell no I have an axe, I am an OG hacker"
The two main runners I am aware of are uv and pipx. (Any compliant runner can be referenced in the shebang to make a script standalone where shebangs are supported.)
Small price to pay for escaping python dependency hell.
(sadly, uv cannot detect the release date of some packages. I'm looking at you, yaml!)
It will install and use distribution packages, to use PyPA's terminology; the term "module" generally refers to a component of an import package. Which is to say: the names you write here must be the names that you would use in a `uv pip install` command, not the names you `import` in the code, although they may align.
This is an ecosystem standard (https://peps.python.org/pep-0723/) and pipx (https://pipx.pypa.io) also supports it.
linux core utils have supported this since 2018 (coreutils 8.3), amusingly it is the same release that added `cp --reflink`. AFAIK I know you have to opt out by having `POSIX_CORRECT=1` or `POSIX_ME_HARDER=1` or `--pedantic` set in your environment. [1]
freebsd core utils have supported this since 2008
MacOS has basically always supported this.
---
1. Amusingly despite `POSIX_ME_HARDER` not being official a alrge swapt of core utils support it. https://www.gnu.org/prep/standards/html_node/Non_002dGNU-Sta...
This isn't a knock against UV, but more a criticism of dynamic dependency resolution. I'd feel much better about this if UV had a way to whitelist specific dependencies/dependency versions.
uv installing deps is hardly more risky.
Scanning for external dependencies is common but not so much internal private libraries.
I've used Tiger/Saint/Satan/COPS in the distant past. But I think they're somewhat obsoleted by modern packaging and security like apparmor and selinux, not to mention docker and similar isolators.
uv executes http://somemirror.com/some-version
most people like their distro to vet these things. uv et all had a reason when Python2 and 3 were a mess. i think that time is way behind us. pip is mostly to install libraries, and even that is mostly already done by the distros.
It’s the script contents that count, not just dependencies.
Deno-style dependency version pinning doesn’t solve this problem unless you check every hash.
If you don't care about being ecosystem-compliant (and I am sure malware does not), it's only a few lines of Python to download the code and eval it.
curl -LsSf https://astral.sh/uv/install.sh | sh """
Also isn't great. But that's how homebrew is installed, so ... shrug ... ?
Not to bash uv/homebrew, they are better than most _easy_ alternatives.
I will happily copy-paste this from any source I trust, for the same reason I'll happily install their software any other way.
But then I'm a weirdo that takes personal offense at tools hijacking my rc / PATH, and keep things like homebrew at arm's length, explicitly calling shellenv when I need to use it.
The man page tells me:
Without that, the system may try to treat the entirety of "uv run --script" as the program name, and fail to find it. Depending on your env implementation and/or your shell, this may not be needed.See also: https://unix.stackexchange.com/questions/361794
-S causes the string to be split on spaces and so the arguments are passed correctly.
I want to be able to ship a bundle which needs zero network access to run, but will run.
It is still frustratingly difficult to make portable Python programs.
My current hobby language is janet. Creating a statically linked binary from a script in janet is trivial. You can even bring your own C libraries.
As long as you have internet access, and whatever repository it's drawing from is online, and you may get different version of python each time, ...
But, yes, python scripts with in-script dependencies plus uv to run them doesn't change dependency distribution, just streamlines use compared to manual setup of a venv per script.
Morale: follow the rules.
But this is getting a bit off topic, I suppose.
"A scientific truth does not triumph by convincing its opponents and making them see the light, but rather because its opponents eventually die and a new generation grows up that is familiar with it." - Max Planck.
And the game is worse for it :')
Players are incentivized to win due to specific decisions made by the league.
In Bananaball the league says, "practice your choreographed dance number before batting practice." And those same athletes are like, "Wait, which choreographed dance number? The seventh inning stretch, the grand finale, or the one we do in the infield when the guy on stilts is pitching?"
Edit: the grand finale dance number I saw is both teams dancing together. That should be noted.
Baseball has done a terrible job, but at least seems to have turned the corner with the pitch clock. Maybe they'll move the mound back a couple feet, make the ball 5.5oz, reduce the field by a player and then we'll get more entertainment and the players can still try their hardest to win.
Personally, I think it'd be interesting to see how the game plays if you could only have two outfielders (but you could shift however you choose.)
I'd guess MLB The Show video game wouldn't be a bad place to start. They should have a decent simulator built in.
Western Europe in a VERY dense city BTW.
I don't think the implied claim is that there should be specifically a train to every particular address, if that's what you're counting as failure in the game, but rather that with good public transport (including trains) and pedestrian/cyclist-friendly streets it shouldn't be the case that most people need to drive.
Need to move 3 or 4 people? Driving the car may be cheaper.
Don't want to get rained on? Or heatstroke? Or walk through snow? Or carry a bunch of stuff, like a groceries/familyWeek or whatever else? Or go into the countryside/camping? Or move a differently-abled person? Or go somewhere outside public transport hours? Or, or .. or.
Are there many cases where people should take public transport or ride a bike instead of their car? Obviously yes. But once you have a car to cover the exigent circumstances it is easy to use them for personal comfort reasons.
They’re also a joke when it comes to moving large numbers of people. I can’t imagine the chaos if everyone leaving a concert at Wembley Stadium decided to leave by car.
Fort Worth is worse for this!
Strongtowns is definitely worth a listen.
But people claiming that you can live a life without cars don't seem to realise the very many scenarios where cars are often easier and sometimes the only answer.
In the states at least if you're using public transit it's generally as an intentional time / cost tradeoff. That's not a mystery and taking a point-to-point schedule and comparing that against public transit constraints doesn't really prove much.
If you want the freedom to move across vast amounts of open nature, then yeah the private automobile is a good approximation for freedom of mobility. But designing urban areas that necessitate the use of a private vehicle (or even mass transit) for such essentials as groceries or education is enslavement. I don't buy the density argument either. Places that historically had the density to support alternative modes of transportation, densities that are lower than they are today, are only marginally accessible to alternative forms of transportation today. Then there is modern development, where the density is decreased due to infrastructure requirements.
So no, I don't think Europeans who haven't been in America have quite absorbed just how vast America is. It stretches across an entire continent in the E-W direction, and N-S (its shortest border) still takes nearly a full day. (San Diego to Seattle is about 20 hours, and that's not even the full N-S breadth of the country since you can drive another 2.5 hours north of Seattle before reaching the Canadian border). In fact, I can find a route that goes nearly straight N-S the whole way, and takes 25 hours to drive, from McAllen, TX to Pembina, ND: https://maps.app.goo.gl/BpvjrzJvvdjD9vdi9
Train travel is sometimes feasible in America (I am planning Christmas travel with my family, and we are planning to take a train from Illinois to Ohio rather than fly, because the small Illinois town we'll be in has a train station but no airport; counting travel time to get to the airport, the train will be nearly as fast as flying but a lot cheaper). But there are vast stretches of the country where trains just do not make economic sense, and those whose only experience is in Europe usually don't quite realize that until they travel over here. For most people, they might have an intellectual grasp of the vastness of the United States, but it takes experiencing it before you really get it deep down. Hence why the very smart German engineer still misread the map: his instincts weren't quite lined up with the reality of America yet, and so he forgot to check the scale of the map.
There are plenty of city pairs where high speed trains do make economic sense and America still doesn't have them. [1] is a video "56 high speed rail links we should've built already" by CityNerd. And that's aside from providing services for the greater good instead of for profit - subsidizing public transport to make a city center more walkable and more profitable and safer and cleaner can be a worthwhile thing. The US government spends a lot subsidizing air travel.
> So no, I don't think Europeans who haven't been in America have quite absorbed just how vast America is
China had some 26,000 miles of high speed rail two years ago, almost 30,000 miles now connecting 550 cities, and adding another couple of thousand miles by 2030. A hundred plus years ago America had train networks coast to coast. Now all Americans have is excuses why the thing you used to have and tore up is impossible, infeasible, unafordable, unthinkable. You have reusable space rockets that can land on a pillar of fire. If y'all had put as much effort into it as you have into special pleading about why it's impossible, you could have had it years ago.
[1] https://www.youtube.com/watch?v=wE5G1kTndI4
This is, of course, a massively broad generalization, and there will be plenty of voters who don't fit that generalization. But the average American voter, as best I can tell, recoils from the words "high-speed rail" like Dracula would recoil from garlic. And I do believe that California's infamous failure (multiple failures, even) to build the high-speed rail they have been working on for years has a lot to do with that "high-speed rail is a boondoggle and a waste of taxpayer dollars" knee-jerk reaction that so many voters have.
The forests and wilderness of the PNW are much, much, much, much more remote and wild than virtually anywhere you’d go in Europe. Like not even close.
Or have a "car-cabin-without-engine-and-wheels" and treat it like a packet on a network of trains and "skateboard car platforms".
Such "freedom"...
I'm curious how this changes (in your mind) if "trains" can be expanded to "trains, buses, bicycle", or if you consider that to be a separate discussion.
The Atlanta Metro has 6.5 million people across TWENTY THOUSAND square kilometers.
Trains just don't make sense for this. Everything is too spread out. And that's okay. Cites are allowed to have different models of transportation and living.
I like how much road infra we have. That I can visit forests, rivers, mountains, and dense city all within a relatively short amount of time with complete flexibility.
Autonomous driving is going to make this paradise. Cars will be superior to trains when they drive themselves.
Trains lack privacy and personal space.
I live in NYC which has 29,000/sqkm in Manhattan and 11,300/sqkm overall. Public transportation is great here and you don't need a car.
but at 240/sqkm, that's really not much public trans per person!
> Please don't use Hacker News for political or ideological battle. It tramples curiosity.
> Eschew flamebait. Avoid generic tangents. Omit internet tropes.
How did we get here from the post about uv?
I'm so stoked for what uv is doing for the Python ecosystem. requirements.txt and the madness around it has been a hell for over a decade. It's been so pointlessly hard to replicate what the authors of Python projects want the state of your software to be in.
uv has been much needed. It's solving the single biggest pain point for Python.
Public transport is to move people around, not to make money.
Did you forget to support yourself? You're saying Rheinland has three times the population density of Atlanta, with convenient passenger rail, and that demonstrates that low population density isn't an obstacle to passenger rail in Atlanta?
https://www.nytimes.com/interactive/2019/08/14/magazine/traf...
Calgary apparently also does a good job of clearing its bike lanes.
And I do my Costco shopping by bike year-round. I think I've used the car for large purchases at Costco twice in the last year.
I _rarely_ drive my car anywhere in Toronto, and find the streets on bike safer than most of the sidewalks in January -- they get plowed sooner than most homeowners and businesses clear the ice from their sidewalks.
And in Toronto we're rank amateurs at winter biking. Look at Montreal, Oslo, or Helsinki for even better examples. Too bad we've got a addle-brained carhead who doesn't understand public safety or doing his own provincial as our premier.
Personally I've also biked to work (and everywhere, really) in sub-zero degrees many times, because the bicycle lanes are cleared and salted. It's really not too bad. It actually gets a bit too hot even, because you start out by wearing so much.
I used to bike to work in just-above-freezing temperatures. That wasn't so bad.
The one time it started to rain mid-journey, that was bad.
Depending how expensive is gasoline in your country, when using a car people underestimate the cost of a travel by a factor two to five, because they don't count the depreciation of their vehicle's value and the maintenance cost (and sometimes even insurance price) driven by the kilometers ridden during the trip.
I guess Europeans will never find out how great the US is :-)
Actually this idea of just buying things at "the store" is relatively new too. Historically people would make more things themselves, and more food would be purchased directly from farmers who had grown it.
Sure, this is just my experience, but I use Python a lot and use a lot of tools written in Python.
Usually happens to me when I find code for some research paper. Even something that's just three months old can be a real pain to get running
To be fair to the GP comment, this is how I feel about Ruby software. I am not nearly as practiced at installing and upgrading in that ecosystem so if there was a way to install tools in a way that lets me easily and completely blow them away, I would be happier to use them.
God, I hate Python. Why is it so hard to not break code?
This is the entire problem. You gonna put that in a lock file or just tell your colleagues to run the same command?
I guess this is mostly about data science code and maybe people who publish software in those communities are just doing very poor packaging, so this idea of a "lock file" that freezes absolutely everything with zero chance for any kind of variation is useful. Certainly the worst packaged code I've ever seen with very brittle links to certain python versions and all that is typically some ML sort of thing, so yeah.
This is all anathema to those of us who know how to package and publish software.
In 2025, the overall developer experience is much better in (1) Rust compared to C++, and (2) Java/DotNet(C#) compared to Python.
I'm talking about type systems/memory safety, IDEs (incl. debuggers & compilers), package management, etc.
Recently, I came back to Python from Java (for a job). Once you take the drug of a virtual machine (Java/DotNet), it is hard to go back to native binaries.
Last, for anyone unfamiliar with this quote, the original is from Winston Churchill:
I don't really know why this is, at a high level, and I don't care. All I know is that Python is, for me, with the kinds of things I tend to need to build, the absolute fucking worst. I hope uv gets adopted and drives real change.
My last dance with Python was trying to build Ardupilot, which is not written in Python but does have a build that requires a tool written in Python, for whatever reason. I think I was on my Mac, and I couldn't get this tool from Homebrew. Okay, I'll install it with Pip—but now Pip is showing me this error I've never seen before about "externally managed environments", a concept I have no knowledge of. Okay, I'll try a venv—but even with the venv activated, the Ardupilot makefile can't find the tool in its path. Okay, more googling, I'll try Pipx, as recommended broadly by the internet—I don't remember what was wrong with this approach (probably because whatever pipx does is totally incomprehensible to me) but it didn't work either. Okay, what else? I can do the thing everybody is telling me not to do, passing `--break-system-packages` to plain old Pip. Okay, now the fucking version of the tool is wrong. Back it out and install the right version. Now it's working, but at what cost?
This kind of thing always happens, even if I'm on Linux, which is where I more usually build stuff. I see errors nobody has ever posted about before in the entire history of the internet, according to Google. I run into incomprehensible changes to the already incomprehensible constellation of Python tooling, made for incomprehensible reasons, and by incomprehensible I mean I just don't care about any of it, I don't have time to care, and I shouldn't have to care. Because no other language or build system forces me to care as much, and as consistently, as Python does. And then I don't care again for 6 months, a year, 2 years, until I need to do another Python thing, and whatever I remember by then isn't exactly obsolete but it's still somehow totally fucking useless.
The universe has taught me through experience that this is what Python is, uniquely. I would welcome it teaching me otherwise.
UV is making me give python a chance for the first time since 2015s renpy project I did for fun.
One could argue, that this is one difference between npm and such, and what many people use in the Python ecosystem. npm and cargo and so on are automatically creating lock files. Even people, who don't understand why that is important, might commit them to their repositories, while in the Python ecosystem people who don't understand it, think that committing a requirements.txt only (without checksums) is OK.
However, it is wrong, to claim, that in the Python ecosystem we didn't have the tools to do it right. We did have them, and that well before uv. It took a more care though, which is apparently too much for many people already.
C/C++ often had to compile used “make” which I’ll admit to being better at the conda/pip.
I suspect this is because the c/c++ code was developed by people with a more comp Sci background. Configure/make/make install..I remember compiling this one.
https://mafft.cbrc.jp/alignment/software/source.html
If the software made it biogrids life was easier
https://biogrids.org/
But a lot of the languages had their own quirks and challenges (Perl cpan, Java…). Containerization kinda helps.
Honorable mention: Compiling someone else's C code. Come on; C compiles to a binary; don't make the user compile.
I'm assuming a Linux based system here, but consider the case where you have external dependencies. If you don't want to require that the user installs those, then you gotta bundle then or link them statically, which is its own can of worms.
Not to mention that a user with an older glibc may not be able to run your executable, even if they have your dependencies installed. Which you can, for example, solve by building against musl or a similar glibc alternative. But in the case of musl, the cost is a significant overhead if your program does a lot of allocations, due to it lacking many of the optimizations found in glibc's malloc. Mitigating that is yet another can of worms.
There's a reason why tools like Snap, AppImage, Docker, and many more exist, each of which are their own can of worms
$ rustup target add x86_64-unknown-linux-musl
$ cargo build --target x86_64-unknown-linux-musl --release
Similarly for cross-compiling for Windows
Unless you’re on a different architecture, then having the source code is much more useful.
https://hub.docker.com/_/python
pip freeze > requirements.txt
pip install -r requirements.txt
Way before "official" lockfile existed.
Your requirements.txt becomes a lockfile, as long as you accept to not use ranges.
Having this in a single tool etc why not, but I don't understand this hype, when it was basically already there.
With pip you update a dependency, it won't work if it's not compatible, it'll work if they are. Not sure where the issue is?
This is very new behavior in pip. Not so long ago, imagine this:
You `pip install foo` which depends on `bar==1.0`. It installs both of those packages. Now you install `pip install baz` which depends on `bar==2.0`. It installs baz, and updates bar to 2.0. Better hope foo's compatible with the newer version!
I think pip only changed in the last year or two to resolve conflicts, or die noisily explaining why it couldn't be done.
Which makes you part of the people the GP is referring to? Try using it anger for a week, you'll come to understand.
It's like Sisyphus rolling a cube up a hill and being offered a sphere instead: "no thanks, I just push harder when I have to overcome the edges."
It can get complicated. The resolver in uv is part of its magic.
https://docs.astral.sh/uv/reference/internals/resolver/
You include the security patch of whatever your dependencies are into your local vetted pypi repository. You control what you consider liabilities and you don't get shocked by breakages in what should be minor versions.
Of course you have to be able to develop software and not just snap Lego's together to manage a setup like that. Which is why uv is so popular.
Inevitably, these versions are out-of-date. Sometimes, they are very, very out of date. "Sorry, I can only install [version from 5 years ago.]" is always great for productivity.
I ran into this recently with a third-party. You'd think a 5 year old version would trigger alarm bells...
Sensible defaults would completely sidestep this, that's the popularity of uv. Or you can be an ass to people online to feel superior, which I'm sure really helps.
As far as I know, files like requirements.txt, package.json, cargo.toml are intended to be used as a snapshot of the dependencies in your project.
In case you need to update dependency A that also affects dependency B and C, I am not sure how one tool is better than other.
cargo can also update transitive dependencies (you need `--locked` to prevent that).
Ruby's Bundler does not, which is preferred and is the only correct default behaviour. Elixir's mix does not.
I don't know whether uv handles transitive dependencies correctly, but lockfiles should be absolute and strict for reproducible builds. Regardless, uv is an absolute breath of fresh air for this frequent Python tourist.
uv does it a lot faster and generates requirements.txts that are cross-platform, which is a nice improvement.
Pips solver could still cause problems in general on changes.
UV having a better solver is nice. Being fast is also nice. Mainly tho it feeling like it is a tool that is maintained and can be improved upon without ripping one’s hair out is a godsend.
- dev dependencies (or other groups) - distinguishing between direct and indirect dependencies (useful if you want to cut some fat from a project) - dependencies with optional extra dependencies (if you remove the main, it will delete the orphans when relevant)
It's not unachievable with pip and virtualenvs, but verbose and prone to human error.
Like C: if you're careful enough, it can be memory safe. But teams would rather rely on memory safe languages.
But the main reason shouldn't be the "lockfile". I was replying to the parent comment mainly for that particular thing.
That being said, the uv experience is much nicer (also insanely fast).
[1] https://pip.pypa.io/en/stable/user_guide/#constraints-files
Pipenv tried to be what uv is, but it never did seem to work right, and it had too many weird corner cases ("why is it suddenly taking 3 hours to install packages? why it is literally impossible to get it to upgrade one single dependency and not all the others?") to ever be a contender.
Honestly, I can't think of a single good reason not to want to use a venv for Python.
For a long time there were even compatibilities between the RHEL host python version, and the python version the Red Hat Ansible team were shipping.
I would probably use something like this: https://stackoverflow.com/questions/17803829/how-to-customiz...
https://bun.sh/
FWIW I use zsh with auto-auto-completion / auto-completion-as-you-type, so just hitting `p` on an empty command line will remember the most recent command starting with `p` (which was likely `pnpm`), and you can refine with further keystrokes and accept longer prefixes (like I always do that with `git add` to choose between typical ways to complete that statement). IMO people who don't use auto-completion are either people who have a magical ability to hammer text into their keyboards with the speed of light, or people who don't know about anything hence don't know about auto-completion, or terminally obsessive types who believe that only hand-crafting each line is worth while.
I don't know which type of person you are but since typing `pnpm` instead of `npm` bothers you to the degree you refuse to use `pnpm`, I assume you must be of the second type. Did you know you can alias commands? Did you know that no matter your shell it's straightforward to write shell scripts that do nothing but replace obnoxious command invocations with shorter ones? If you're a type 3 person then of course god forbid, no true hacker worth their salt will want to spoil the purity of their artisanal command line incantations with unnatural ersatz-commands, got it.
It even has some (I feel somewhat rudimentary) support for workspaces and isolated installs (what pnpm does)
Maven worked fine without semantic versioning and lock files.
Edit: Changed "semantic versioning" to "version ranging"
No, it actually has the exact same problem. You add a dependency, and that dependency specifies a sub-dependency against, say, version `[1.0,)`. Now you install your dependencies on a new machine and nothing works. Why? Because the sub-dependency released version 2.0 that's incompatible with the dependency you're directly referencing. Nobody likes helping to onboard the new guy when he goes to install dependencies on his laptop and stuff just doesn't work because the versions of sub-dependencies are silently different. Lock files completely avoid this.
Before version ranging, maven dependency resolution was deterministic.
Coming from ruby. However, I think uv has actually now surpassed bundler and the ruby standard toolset for these things. Definitely surpassed npm, which is also not fine. Couldn't speak for cargo.
Why?
It’s almost too easy to add one compared to writing your own functions.
Now compare that to adding a dependency to a c++ project
Some time ago I found out it does work with authentication, but their “counter ascii animation” just covers it… bug has been open for years now…
uv actually works.
Funny how these things get forgotten to history. There's lots of prior art when it comes to replacing pip.
edit: here's an HN thread about pipenv, where many say the same things about it as they are about UV and Poetry before https://news.ycombinator.com/item?id=16302570
However, I have zero reservations about uv. I have not encountered bugs, and when features are present they are ready for complete adoption. Plus there's massive speed improvements. There is zero downside to using uv in any application where it can be used and also there are advantages.
Agree that uv is way way way faster than any of that and really just a joy to use in the simplicity
Also the ability to have a single script with deps using TOML in the headers super eaisly.
Also Also the ability to use a random python tool in effectively seconds with no faffing about.
Yes, though poetry has lock files, and it didn't create the same positive feelings uv does :)
The environment, dependency experience created so much friction compared to everything else. Changed my perspective on Docker for local dev.
Glad to hear it seems to finally be fixed.
I think it's more like Rust devs using Python and thinking what the fuck why isn't this more like rustup+cargo?
Even then though, the core developers made it clear that breaking everyone’s code was the only thing they were willing to do (remember Guido’s big “No 2.8” banner at PyCon?), which left the community with no choice.
And inspired by uv, we now have rv for RoR!
I do prefer uv but it's not like sane python env management hasn't existed
My default feeling towards using python in more ways than I did was default no because the tooling wasn't there for others to handle it, no matter how easy it was for me.
I feel uv will help python go even more mainstream.
You've been able to have the exact same setup forever with pyenv and pyenv-virtualenv except with these nothing ever has to be prefixed. Look, uv is amazing and I would recommend it over everything else but Python devs have had this flow forever.
No, you aren't.
> It doesn't change any of the moving pieces
It literally does, though iyt maintains a mostly-parallel low-level interface, the implementation is replaced with improved (in speed, in dependency solving, and in other areas.) You are using virtual environments (but not venv/virtualenv) and the same sources that pip uses (but not pip).
> You've been able to have the exact same setup forever with pyenv and pyenv-virtualenv except with these nothing ever has to be prefixed.
Yes, you can do a subset of what uv does with those without prefixes, and if you add pipx and hatch (though with hatch you’ll be prefixing for much the same reason as in uv) you’ll get closer to uv’s functionality.
> Look, uv is amazing and I would recommend it over everything else but Python devs have had this flow forever.
If you ignore the parts of the flow built around modern Python packaging standards like pyproject.toml, sure, pieces of the flow have been around and supported by the right constellation of other standard and nonstandard tools for a while.
good god no thank you.
>cargo
more like it.
This is the most insulting take in the ongoing ruination of Python. You used to be able to avoid virtualenvs and install scripts and dependencies directly runnable from any shell. Now you get endlessly chastised for trying to use Python as a general purpose utility. Debian was a bastion of sanity with the split between dist_packages and site_packages but that's ruined now too.
With PEP 723 and confortable tooling (like uv), now you get scripts, that are "actually directly runnable", not just "fake directly runnable oops forgot to apt-get install something sorta runnable", and work reliably even when stuff around you is updated.
This wasn't really the case; in principle anything you installed in the system Python environment, even "at user level", had the potential to pollute that environment and thus interfere with system tools written in Python. And if you did install it at system level, that became files within the environment your system package manager is managing, that it doesn't know how to deal with, because they didn't come from a system package.
But it's worse now because of how many system tools are written in Python — i.e., a mark of Python's success.
Notably, these tools commonly include the system package manager itself. Since you mentioned Debian (actually this is Mint, but ya know):
> Now you get endlessly chastised for trying to use Python as a general purpose utility.No, you don't. Nothing prevents you from running scripts with the system Python that make use of system-provided libraries (including ones that you install later with the system package manager).
If you need something that isn't packaged by your distro, then of course you shouldn't expect your distro to be able to help with it, and of course you should expect to use an environment isolated from the distro's environment. In Python, virtual environments are the method of isolation. All reasonable tooling uses them, including uv.
> Debian was a bastion of sanity with the split between dist_packages and site_packages but that's ruined now too.
It's not "ruined". If you choose to install the system package for pip and to use it with --break-system-packages, the consequences are on you, but you get the legacy behaviour back. And the system packages still put files separately in dist-packages. It's just that... doing this doesn't actually solve all the problems, fundamentally because of how the Python import system works.
Basically the only thing missing from pip install being a smooth experience is something like npx to cleanly run modules/binary files that were installed to that directory. It's still futzing with the PATH variable to run those scripts correctly.
This could still cause problems if you run system tools as that user.
I haven't checked (because I didn't install my distro's system package for pip, and because I use virtual environments properly) but I'm pretty sure that the same marker-file protection would apply to that folder (there's no folder there, on my system).
This ideology is what caused all the problems to begin with, the base python is built as if it's the only thing in the entire operating systems environment when it's entire packaging system is also built in a way that makes that impossible to do without manually having to juggle package conflicts/incompatibilities.
I do agree it is annoying, and what they need to do is just provide an automatic "userspace" virtualenv for anything a user installs themselves... but that is a pandoras box tbh. (Do you do it per user? How does the user become aware of this?)
But that's probably not practical to retrofit given the ecosystem as it is now.
As an outsider to the python ecosystem I've wanted to learn the _how_ behind uv as well, but that hasn't been immediately clear
So far it seems like they have a bunch of these high performance tools. Is this part of an upcoming product suite for python or something? Just curious. I'm not a full-time python developer.
"What I want to do is build software that vertically integrates with our open source tools, and sell that software to companies that are already using Ruff, uv, etc. Alternatives to things that companies already pay for today. An example of what this might look like [...] would be something like an enterprise-focused private package registry."
There's also this interview with Charlie Marsh (Astral founder): https://timclicks.dev/podcast/supercharging-python-tooling-a... (specifically the "Building a commerical company with venture capital " section)
https://astral.sh/pyx
There are apparently 10 million Python developers in the world and pretty soon all of them will be using uv. I doubt it is that hard to monetise.
The "install things that have complex non-Python dependencies using pip" story is much better than several years ago, because of things like pip gaining a new resolver in 2020, but in large part simply because it's now much more likely that the package you want offers a pre-built wheel (and that its dependencies also do). A decade ago, it was common enough that you'd be stuck with source packages even for pure-Python projects, which forced pip to build a wheel locally first (https://pradyunsg.me/blog/2022/12/31/wheels-are-faster-pure-...).
Another important change is that for wheels on PyPI the installer can now obtain separate .metadata files, so it can learn what the transitive dependencies are for a given version of a given project from a small plain-text file rather than having to speculatively download the entire wheel and unpack the METADATA file from it. (This is also possible for source distributions that include PKG-INFO, but they aren't forced to do so, and a source distribution's metadata is allowed to have "dynamic" dependencies that aren't known until the wheel is built (worst case) or a special metadata-only build hook is run (requires additional effort for the build system to support and the developer to implement)).
With uv it just works. With pip, technically you can make it work, and I bet you'll screw something up along the way.
This is different as of Python 3.11. Please see https://peps.python.org/pep-0668/ for details. Nowadays, to install a package globally, you first have to have a global copy of pip (Debian makes you install that separately), then you have to intentionally bypass a security marker using --break-system-packages.
Also, you don't have to activate the venv to use it. You can specify the path to the venv's pip explicitly; or you can use a different copy of pip (e.g. a globally-installed one) passing it the `--python` argument (you have been able to do this for about 3 years now).
(Pedantically, yes, you could use a venv-installed copy of pip to install into the system environment, passing both --python and --break-system-packages. I can't prove that anyone has ever done this, and I can't fathom a reason beyond bragging rights.)
> - really easy to distinguish [dev] and main dependencies
As of 25.1, pip can install from dependency groups described in pyproject.toml, which is the standard way to group your dependencies in metadata.
> distinguish direct dependencies from indirect dependencies, making it easy to find when a package is not needed anymore
As of 25.1, pip can create PEP 751 standard lockfiles.
> easily use different python versions for different projects
If you want something to install Python for you, yes, that was never in pip's purview, by design.
If you want to use an environment based off an existing Python, that's what venv is for.
I'm still mostly on poetry
Wake me up when pip can do any of that.
This is a matter of opinion. Pip exists to install the packages and their dependencies. It does not, by design, exist to manage a project for you.
If anything, pip is a dependency installer, while working with even trivial projects requires a dependency manager. Parent's point was that pip is actually good enough that you don’t even need uv anymore, but as long as pip doesn’t satisfy 80% of the requirements, that’s just plain false.
A majority of HN users might agree with you, but I'd guess that a majority of developers, to paraphrase Don Draper, don't think about it at all.
Some people don't have, or don't care about, the additional requirements you have in mind.
Or by asyncio.
Currently they are a bit pointless. Sure they aid in documentation, but they are effort and cause you pain when making modifications (mind you with halfarse agentic coding its probably less of a problem. )
What would be better is to have a strict mode where instead of duck typing its pre-declared. It would also make a bunch of things faster (along with breaking everything and the spirit of the language)
I still don't get the appeal of UV, but thats possibly because I'm old and have been using pyenv and venv for many many years. This means that anything new is an attack on my very being.
however if it means that conda fucks off and dies, then I'm willing to move to UV.
I've been using it professionally and its been a big improvement for code quality.
It's the python version of fink vs macports vs homebrew. Or apt vs deb. or pkgsrc vs ports.
But I don't think "its just another" gets the value proposition here. It's significantly simpler to deploy in practice for people like me, writing ad hoc scripts and running git downloaded scripts and codelets.
Yes, virtualenv and pip existed. No, they turned out to be a lot more fiddly to run in practice than UV.
That UV is rust is funny, but not in a terrible way. The llvm compiler toolchain is written in C but compiles other languages. Using one language to do things for another language isn't such a terrible outcome.
I hope UV supplants the others. Not to disrespect their authors, but UV is better for end users. If its worse for package maintainers I think the UV authors should be told.
https://peps.python.org/pep-0703/
1. It tries to do too many things. Please just do one thing and do it well. It's simultaneously trying to replace pip, pyenv, virtualenv, and ruff in one command.
2. You end up needing to use `uv pip` so it's not even a full replacement for pip.
3. It does not play well with Docker.
4. It adds more complexity. You end up needing to understand all of these new environmental variables: `UV_TOOL_BIN_DIR`, `UV_SYSTEM_PYTHON`, `UV_LINK_MODE`, etc.
pip and virtualenv also add a ton of complexity and when they break (which happens quite often) debugging it is even harder despite them being "battle tested" tools.
The alternative, of course, is having Python natively support a combined tool. Which you can support while also not liking `uv` for the above reason.
Top that off with first class programming capabilities and modularization and I can share common configuration and packages across systems. And add that those same customized packages can be directly included in a dev shell making all of the amazing software out there available for tooling and support. Really has changed my outlook and I have so much fun now not EVER dealing with tooling issues except when I have explicitly upgrade my shell and nixpkgs version.
I just rebuilt our CI infrastructure with nix and was a able to configure multiple dockerd isolated daemons per host, calculate the subnet spread for all the networks, write scripts configuring the env so you can run docker1 and hit daemon 1. Now we can saturate our CI machines with more parallel work without them fighting over docker system resources like ports. Never would have attempting doing this without nix, being able to generate the entire system config tree and inspect systemd service configs befor even applying to a host reduced my iteration loop to an all time low in the infrastructure land where 10-15mins lead times of building images to find out I misspelling Kafka and kakfa somewhere and now need to rebuild again for 15mins. Now I get almost instant feedback for most of these types of errors.
Yep: Nix
It's the same sort of deal with pyenv--the Python version is itself a dependency of most libraries, so it's a little silly to have a dependency manager that only manages some dependencies.
I started using NodeJS more after lots of Python experience. Packages make so much more sense there. Even imports. You know how hard it is to do the equivalent of "require '../foo.js'" in Python?
`virtualenv` is a heavy-duty third-party library that adds functionality to the standard library venv. Or rather, venv was created as a subset of virtualenv in Python 3.3, and the projects have diverged since.
The standard library `venv` provides "obvious thing that a dependency manager does" functionality, so that every dependency manager has the opportunity to use it, and so that developers can also choose to work at a lower level. And the virtual-environment standard needs to exist so that Python can know about the pool of dependencies thus stored. Otherwise you would be forced to... depend on the dependency manager to start Python and tell it where its dependency pool is.
Fundamentally, the only things a venv needs are the `pyvenv.cfg` config file, the appropriate folder hierarchy, and some symlinks to Python (stub executables on Windows). All it's doing is providing a place for that "pool of dependencies" to exist, and providing configuration info so that Python can understand the dependency path at startup. The venvs created by the standard library module — and by uv — also provide "activation" scripts to manipulate some environment variables for ease of use; but these are completely unnecessary to making the system work.
Fundamentally, tools like uv create the same kind of virtual environment that the standard library does — because there is only one kind. Uv doesn't bootstrap pip into its environments (since that's slow and would be pointless), but you can equally well disable that with the standard library: `python -m venv --without-pip`.
> the Python version is itself a dependency of most libraries
This is a strange way of thinking about it IMO. If you're trying to obtain Python libraries, it's normally because you already have Python, and want to obtain libraries that are compatible with the Python you already have, so that you can write Python code that uses the libraries and works under that Python.
If you're trying to solve the problem of deploying an application to people who don't have Python (or to people who don't understand what Python is), you need another layer of wrapping anyway. You aren't going to get end users to install uv first.
“…I can't see any valid use case for a machine-global pool of dependencies…” - Rhetorical question for OP but how do you run an operating system without having said operating systems dependencies available to everything else?
> how do you run an operating system without having said operating systems dependencies available to everything else?
I’m not sure if I understand your question, but I’ll answer based on what I think you mean. The OS gets compiled into an artifact, so the dependencies aren’t available to the system itself unless they are explicitly added.
> This is a strange way of thinking about it IMO. If you're trying to obtain Python libraries, it's normally because you already have Python, and want to obtain libraries that are compatible with the Python you already have, so that you can write Python code that uses the libraries and works under that Python.
“normally” is biased by what the tooling supports. If Python tooling supported pinning to an interpreter by default then perhaps it would seem more normal?
I write a lot of Go these days, and the libs pin to a version of Go. When you build a project, the toolchain will resolve and (if necessary) install the necessary Go dependency just like all of the other dependencies. It’s a very natural and pleasant workflow.
I think there are more cases where pip, pyenv, and virtualenv are used together than not. It makes sense to bundle the features of the three into one. uv does not replace ruff.
> 2. You end up needing to use `uv pip` so it's not even a full replacement for pip.
uv pip is there for compatibility and to facilitate migration but once you are full on the uv workflow you rarely need `uv pip` if ever
> 3. It does not play well with Docker.
In what sense?
> 4. It adds more complexity. You end up needing to understand all of these new environmental variables: `UV_TOOL_BIN_DIR`, `UV_SYSTEM_PYTHON`, `UV_LINK_MODE`, etc.
You don't need to touch them at all
uv doesn’t try to replace ruff.
> You end up needing to use `uv pip` so it's not even a full replacement for pip.
"uv pip" doesn't use pip, it provides a low-level pip-compatible interface for uv, so it is, in fact, still uv replacing pip, with the speed and other advantages of uv when using that interface.
Also, while I’ve used uv pip and uv venv as part of familiarizing myself with the tool, I’ve never run into a situation where I need either of those low-level interfaces rather than the normal high-level interface.
> It does not play well with Docker.
How so?
- uv add <package_name>
- uv sync
- uv run <command>
Feels very ergonomic, I don't need to think much, and it's so much faster.
In my experience it generally does all of those well. Are you running into issues with the uv replacements?
> 2. You end up needing to use `uv pip` so it's not even a full replacement for pip.
What do end up needing to use `uv pip` for?
Happened to buy a new machine and decided to jump in the deep end and it's been glorious. I think the difference from your comment (and others in this chain) and my experience is that you're trying to make uv fit how you have done things. Jumping all the way in, I just . . . never needed virtualenvs. Don't really think about them once I sorted out a mistake I was making. uv init and you're pretty much there.
>You end up needing to use `uv pip` so it's not even a full replacement for pip
The only time I've used uv pip is on a project at work that isn't a uv-powered project. uv add should be doing what you need and it really fights you if you're trying to add something to global because it assumes that's an accident, which it probably is (but you can drop back to uv pip for that).
>`UV_TOOL_BIN_DIR`, `UV_SYSTEM_PYTHON`, `UV_LINK_MODE`, etc.
I've been using it for six months and didn't know those existed. I would suggest this is a symptom of trying to make it be what you're used to. I would also gently suggest those of us who have decades of Python experience may have a bit of Stockholm Syndrome around package management, packaging, etc.
I disagree with this principle. Sometimes what I need is a kitset. I don't want to go shopping for things, or browse multiple docs. I just want it taken care of for me. I don't use uv so I don't know if the pieces fit together well but the kitset can work well and so can a la carte.
The uv docs even have a whole page dedicated to Docker; you should definitely check that out if you haven't already: https://docs.astral.sh/uv/guides/integration/docker/
Needing pip and virtualenvs was enough to make me realize uv wasn't what I was looking for. If I still need to manage virtualenvs and call pip I'm just going to do so with both of these directly.
I had been hoping someone would introduce the non-virtualenv package management solution that every single other language has where there's a dependency list and version requirements (including of the language itself) in a manifest file (go.mod, package.json, etc) and everything happens in the context of that directory alone without shell shenanigans.
Isn't that exactly a pyproject.toml via the the uv add/sync/run interface? What is that missing that you need?
Ah ok I was missing this and this does sound like what I was expecting. Thank you!
If you are using uv, you don’t need to do shell shenanigans, you just use uv run. So I'm not sure how uv with pyproject.toml doesn't meet this description (yes, the venv is still there, it is used exactly as you describe.)
Someone, please tell me what's wrong with this. To me, this seems much less complicated that some uv-centric .toml config file, plus some uv-centric commands for more kinds of actions.
I have worked on numerous projects that started with pipenv and it has never "just works" ever. Either there's some trivial dependency conflict that it can't resolve or it's slow as molasses or something or the other. pipenv has been horrible to use. I started switching projects to pip-tools and now I recommend using uv
I'm using uv in two dozen containers with no issues at all. So not sure what you mean that it doesn't play well with Docker.
- resorting to logical fallacies, or
- relying on your unstated assumption that all complexity is bad
No you don't. That's just a set of compatibility approaches for people who can't let go of pip/venv. Move to uv/PEP723, world's your oyster.
> It does not play well with Docker.
Huh? I use uv both during container build and container runtime, and it works just fine?
> You end up needing to understand all of these new environmental variables
Not encountered the need for any of these yet. Your comments on uv are so far out of line of all the uses I've seen, I'd love to hear what you're specifically doing that these become breaking points.
UV is great but I use it as a more convenient pip+venv. Maybe I'm not using it to it's full potential.
uv is probably much more of a game changer for beginner python users who just need to install stuff and don't need to lint. So it's a bigger deal for the broader python ecosystem.
You aren't, but that's fine. Everyone has their own idea about how tooling should work and come together, and I happen to be in your camp (from what I can tell). I actively don't want an all-in-one tool to do "project management".
But where it isn't a matter of opinion is, speed. Never met anyone who given then same interface, would prefer a process taking 10x longer to execute.
pyenv was problematic because you needed the right concoction of system packages to ensure it compiled python with the right features, and we have a mix of MacOS and Linux devs so this was often non-trivial.
uv is much faster than both of these tools, has a more ergonomic CLI, and solves both of the issues I just mentioned.
I'm hoping astral's type checker is suitably good once released, because we're on mypy right now and it's a constant source of frustration (slow and buggy).
> uv is much faster than both of these tools
conda is also (in)famous for being slow at this, although the new mamba solver is much faster. What does uv do in order to resolve dependencies much faster?
- Representing version numbers as single integer for fast comparison.
- Being implemented in rust rather than Python (compared to Poetry)
- Parallel downloads
- Caching individual files rather than zipped wheel, so installation is just hard-linking files, zero copy (on unix at least). Also makes it very storage efficient.
Side rant: yes I get triggered whenever someone tells me "you can just" do this thing that is actually longer and worse than the original.
The `uv run` command is an optional shortcut for avoiding needing to activate the virtual environment. I personally don't like the whole "needing to activate an environment" before I can run commands "natively", so I like `uv run`. (Actually for the last 10 years I've had my `./manage.py` auto-set up the virtual environment for me.)
The `uv add` / `uv lock` / `uv sync` commands are still useful without `uv run`.
There is a new standard mechanism for specifying the same things you would specify when setting up a venv with a python version and dependencies in the header of a single file script, so that tooling can setup up the environment and run the script using only the script file itself as a spec.
uv (and PyPA’s own pipx) support this standard.
> yes I get triggered whenever someone tells me "you can just" do this thing that is actually longer and worse than the original.
"uv run myscript" is neither longer nor worse than separately manually building a venv, activating it, installing dependencies into it, and then running the script.
Apologies for triggering you in advance, but in case you or others find it useful, here’s how to do the equivalent env-activation commands with uv: https://news.ycombinator.com/item?id=44360892
There's also `uv tool install` which will install things in your PATH without infecting your system with Python.
I would love to see them compete with the likes of Conda and try to handle the Python C extension story.
But in the interim, I agree with everyone else who has already commented, Pixi which is partly built atop of UV’s solver is an even bigger deal and I think the longer term winner here.
Having a topologically complete package manager who can speak Conda and PyPi, is amazing.
https://pixi.sh/latest/
In principle, you can ‘activate’ this new virtual environment like any typical virtual environment that you may have seen in other tools, but the most ‘uv-onic’ way to use uv is simply to prepend any command with uv run. This command automatically picks up the correct virtual environment for you and runs your command with it. For instance, to run a script — instead of
you can just doNo; they are plain virtual environments. There is no special kind of virtual environment. Uv simply offers its own command structure for managing those environments. In particular, `uv run` just ensures a venv in a specific location, then uses it.
There is no requirement to activate virtual environments in order to use them (unless you have some other tooling that specifically depends on the environment variables being set). You can, similarly, "just do"
without uv installed.> This command automatically picks up the correct virtual environment for you
Some people dislike such magic, especially since it involves uv having an opinion about where the virtual environment is located.
`uv run` will also sync the environment to be sure it exists and meets the correct specifications.
But yes, it's optional. You can also just do `uv sync` to sync the environment and then activate it like normal.
Or use `uv venv`, `uv pip` commands and just take the speed advantage.
It's just simpler to use, and better overall. It's reduced friction significantly.
I think the Python community should put it as a first preference vehicle, and be respectful to the prior arts, and their developers, but not insist they have primacy.
> I'd get suspicious if a developer is picky about python versions or library versions
Certain library versions only support certain python versions. And they also break API. So moving up/down the python versions also means moving library versions which means stuff no longer works.
Pip can install from dependency groups in a pyproject.toml file, and can write PEP 751 lockfiles, and work is under way to allow it to install from those lockfiles as well.
I don't know what you mean about a "standard dependency dir". When you make a venv yourself, you can call it what you want, and put it where you want. If you want to put it in a "standard" place, you can trivially make a shell alias to do so. (You can also trivially make a shell alias for "activate the venv at a hard-coded relative path", and use that from your project root.)
Yes, pip installation is needlessly slow for a variety of reasons (that mostly do not have to do with being implemented in Python rather than Rust). Resolving dependencies is also slow (and Rust may be more relevant here; I haven't done detailed testing). But your download speed is still going to be primarily limited by your internet connection to PyPI.
> The alternatives are to use higher-level management like uv does,
The question was specifically what's wrong with pip, venv and pyproject toml, i.e. what issues uv is trying to address. Well of course the thing trying to address the problem addresses the problem....
> I don't know what you mean about a "standard dependency dir".
like node's node_modules, or cargo's ~/.cargo/registry. You shouldn't have to manually create and manage that. installing/building should just create it. Which is what uv does and pip doesn't.
> the same as what you get with `python -m venv --without-pip`
The thing that should be automatic. And even if it is not it should at least be less arcane. An important command like that should have been streamlined long ago. One of the many improvements uv brings to the table.
> and work is under way to allow it to install from those lockfiles as well.
Yeah well, the lack up until now is one of those "what is wrong" things.
> But your download speed is still going to be primarily limited by your internet connection to PyPI.
Downloading lots of small packages dependencies serially leaves a lot of performance on the table due to latency and non-instantaneous response from congestion controllers. Downloading and installing concurrently reduces walltime further.
The point is that it is a thing trying to address the "problem", and that not everyone considers it a problem.
> Which is what uv does and pip doesn't.
The point is that you might want to install something not for use in a "project", and that you might want to explicitly hand-craft the full contents of the environment. Pip is fundamentally a lower-level tool than uv.
> The thing that should be automatic.
Bootstrapping pip is the default so that people who have barely learned what Python is don't ask where pip is, or why pip isn't installing into the (right) virtual environment.
Yes, there are lots of flaws in pip. The problem is not virtual environments. Uv uses the same virtual environments. Neither is the problem "being a low-level tool that directly installs packages and their dependencies". I actively want to have that tool, and actively don't want a tool that tries to take over my entire project workflow.
Implementation-wise, there's nothing wrong in my view with venv. Or rather, everything is compelled to use virtual environments, including uv, and venv is just a simple tool for doing so manually. Pip, on the other hand, is slow and bulky due to poor architecture, a problem made worse by the expectation (you can work around it, but it requires additional understanding and setup, and isn't a perfect solution) of re-installing it into each virtual environment.
(The standard library venv defaults to such installation; you can disable this, but then you have to have a global pip set up, and you have to direct it to install into the necessary environment. One sneaky way to do this is to install Pipx, and then set up some script wrappers that use Pipx's vendored copy of pip. I describe my techniques for this in https://zahlman.github.io/posts/2025/01/07/python-packaging-....)
Edit: by "design" above I meant the broad strokes of how you use pip, installing single packages with their transitive dependencies etc. There's a lot I would change about the CLI syntax, and other design issues like that.
How many commands are required to build up a locally consistent workspace?
Modern package managers do that for you.
Pip also generates PEP 751 lockfiles, and installing from those is on the roadmap still (https://github.com/pypa/pip/issues/13334).
venv is lower-level tooling. Literally all it does is create a virtual environment — the same kind that uv creates and manages. There's nothing to "integrate".
But you don't have to. Brew and other package managers hold uv in their registries.
It’s really excellent stuff
virtualenv, venv, pyenv, pipenv... I think at one point the recommended option changed because it was integrated into Python, but I can't even remember which is which anymore.
Such a pleasure to finally have just one, for maybe... ~99% of my needs.
Definitely lightyears faster than mypy though.
UV means getting more strings attached with VC funded companies and leaning on their infrastructure. This is a high risk for any FOSS community and history tells us how this ends….
Speaking of history, I was very sympathetic to the "we are open-source volunteers, give us a break" kind of stuff for the first N years.. but pypa has a pattern of creating problems, ignoring them, ignoring criticism, ignoring people who are trying to help, and pushing talent+interest elsewhere. This has fragmented the packaging ecosystem in a way that confuses newcomers, forces constant maintenance and training burden on experts, and damages the credibility of the language and its users. Hatch is frankly too little too late, and even if it becomes a wonderful standard, it would just force more maintenance, more confusion for a "temporary" period that lasts many, many years. Confidence is too far gone.
As mentioned elsewhere in the thread, there are tons of conflicting tools in the space already, and due to the fragmentation, poetry etc could never get critical mass. That's partly because pypa stuff felt most "official" and a safer long term bet than anything else, but partly because 33% better was never good enough to encourage widespread adoption until it was closer to 200% better. But uv actually IS that much better. Just let it win.
And let pypa be a case-study in how to NOT do FOSS. Fragmentation is fine up to a point, but you know what? If it wasn't for KDE / Gnome reinventing the wheel for every single kind of individual GUI then we'd have already seen the glorious "year of the linux desktop" by now.
yep, I've been saying this for years, and astral have proved it in the best way: with brilliant, working software
python was a dying project 10 years ago, after the python 3000 debacle
the talent left/lost interest
then the machine learning thing kicked off (for some reason using python), and now python is everywhere and suddenly massively important
and the supporting bureaucracies, still in their death throes, are unable to handle a project of its importance
uv is MIT licensed so if they rug pull, you can fork.
Maybe if you trust the software, then trusting the install script isn't that big of a stretch?
Also, many of the "distribution" tools like brew, scoop, winget, and more are just "PR a YAML file with your zip file URL, name of your EXE to add to a PATH, and a checksum hash of the zip to this git repository". We're about at a minimum effort needed to generate a "distribution" point in software history, so seems interesting shell scripts to install things seem to have picked up instead.
There have also been PoCs on serving malicious content only when piped to sh rather than saved to file.
If you want to execute shell code from the internet, at the very least store it in a file first and store that file somewhere persistent before executing it. It will make forensics easier
Versioning OTOH is often more problematic with distro package managers that can't support multiple versions of the same package.
Also inability to do user install is a big problem with distro managers.
You can `pip install uv` or manually download and extract the right uv-*.tar.gz file from github: https://github.com/astral-sh/uv/releases
Also, most reasonable developers should already be running with the ExecutionPolicy RemoteSigned, it would be nice if code signing these install script was a little more common, too. (There was even a proposal for icm [Invoke-Command] to take signed script URLs directly for a much safer alternative code-golfed version of iwr|iex. Maybe that proposal should be picked back up.)
/just guessing, haven't tried it
no. thats how you get malware. Make a package. Add it to a distro. then we will talk.
Maven has always been a very good solution. I think Bazel is too, but haven't had much experience with it.
Python for me is great when things can remain as simple to wrap your head around as possible.
It's moving pretty quick.
> Do they have good influence on what python's main ecosystem is moving to?
Yes, they're an early adaptor/implementer of the recent pyproject.toml standards.
It’s hard to demonstrate the speed difference in a pitch deck.
Hopeful that a lot of this will be even more resolved next time I'm looking to make decisions.
1. https://pixi.sh/latest/
Now with uv everything just works and I can play around easily with all the great Python projects that exist.
It has always been enough to place installations in separate directories, and use the same bash scripts for environment variables configuration for all these years.
> uv is an incredibly powerful simplification for us that we use across our entire tech stack. As developers, we can all work with identical Python installations, which is especially important given a number of semi-experimental dependencies that we use that have breaking changes with every version. On GitHub Actions, we’re planning to use uv to quickly build a Python environment and run our unit tests. In production, uv already manages Python for all of our servers.
> It’s just so nice to always know that Python and package installation will always be handled consistently and correctly across all of our machines. That’s why uv is the best thing to happen to the Python ecosystem in a decade.
I can only conclude, that the author of the article, and perhaps even the organization they work in, is unaware of other tools that did the job long before uv. If they really value reproducibility that much, how come they didn't look into the matter before? Things much have been really hastily stitched together, if no one ever looked at existing tooling before, and only now they make things reproducible.
I guess reproducibility is still very much a huge problem, especially in jobs, where it should be one of the most important things to take care of: Research. ("Astronomer & Science Communicator" it says on the website). My recommendation is: Get an actual software developer (at least mid-level) to support your research team. A capable and responsibly acting developer would have sorted this problem out right from the beginning.
I am glad they improved their project setups to the level they should be at, if they want to call it research.
Yes, Poetry has had lock files for years, and pyenv has been able to manage installations, but uv is "an incredibly powerful simplification" that makes it easy to do everything really well with just one tool.
There’s a bigger conversation about open source maintenance there, but if I have to get my job done it’s increasingly tempting to take the simplifications and speed.
To me, Python's best feature is the ability to quickly experiment without a second thought. Conda is nice since it keeps everything installed globally so I can just run `python` or iPython/Jupyter anywhere and know I won't have to reinstall everything every single time.
One thing I did recently was create a one-off script with functions to exercise a piece of equipment connected to the PC via USB, and pass that to my coworkers. I created a `main.py` and uv add'ed the library. Then when I wanted to use the script in the REPL, I just did `uv run python -i main.py`.
This let me just call functions I defined in there, like `set_led_on_equipment(led='green', on=True)` directly in the REPL, rather than having to modify the script body and re-run it every time.
Edit: another idea that I just had is to use just[0] and modify your justfile accordingly, e.g. `just pything` and in your justfile, `pything` target is actually `uv run --with x,y,z ipython`
Edit edit: I guess the above doesn't even require just, it could be a command alias or something, I probably am overengineering that lol.
[0]: https://github.com/casey/just
Over the years, I've tried venv, conda, pipenv, petry, plain pip with requirements.txt. I've played with uv on some recent projects and it's a definite step up. I like it.
Uv actually fixes most of the issues with what came before and actually builds on existing things. Which is not a small compliment because the state of the art before uv was pretty bad. Venv, pip, etc. are fine. They are just not enough by themselves. Uv embraces both. Without that, all we had was just a lot of puzzle pieces that barely worked together and didn't really fit together that well. I tried making conda + pipenv work at some point. Pipenv shell just makes using your shell state-full just adds a lot of complexity. None of the IDEs I tried figured that out properly. I had high hopes for poetry but it ended up a bit underwhelming and still left a lot of stuff to solve. Uv succeeds in providing a bit more of an end to end solution. Everything from having project specific python installation, venv by default without hassle, dependency management, etc.
My basic needs are simple. I don't want to pollute my system python with random crap I need for some project. So, like uv, I need to have whatever solution deal with installing the right python version. Besides, the system python is usually out of date and behind the current stable version of python which is what I would use for new projects.
But why is it the Windows installation is to execute a script off the Internet with bypassed security isolations?
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
How do I install it globally on a system? Debian doesn't let me install packages via pip outside of a venv or similar.
You may have a library that's been globally installed, and you have multiple projects that rely on it. One day you may need to upgrade the library for use in one project, but there are backward incompatibile changes in the upgrade, so now all of your other projects break when you upgrade the global library.
In general, when projects are used by multiple people across multiple computers, it's best to have the specific dependencies and versions specified in the project itself so that everyone using that project is using the exact same version of each dependency.
For recreational projects it's not as big of a deal. It's just harder to do a recreation of your environment.
Because it being available in the system environment could cause problems for system tools, which are expecting to find something else with the same name.
And because those tools could include your system's package manager (like Apt).
> So there is a massive possibility I am simply wrong and pip-installing something globally is a huge risk. I'm just not understanding it.
I assume you're referring to the new protections created by the EXTERNALLY-MANAGED marker file, which will throw up a large boilerplate warning if you try to use pip to install packages in the system environment (even with --user, where they can still cause problems when you run the system tools without sudo).
You should read one or more of:
* the PEP where this protection was introduced (https://peps.python.org/pep-0668/);
* the Python forum discussion explaining the need for the PEP (https://discuss.python.org/t/_/10302);
* my blog post (https://zahlman.github.io/posts/2024/12/24/python-packaging-...) where I describe in a bit more detail (along with explaining a few other common grumblings about how Python packaging works);
* my Q&A on Codidact (https://software.codidact.com/posts/291839/) where I explain more comprehensively;
* the original motivating Stack Overflow Q&A (https://stackoverflow.com/questions/75608323/);
* the Python forum discussion (https://discuss.python.org/t/_/56900) where it was originally noticed that the Stack Overflow Q&A was advising people to circumvent the protection without understanding it, and a coordinated attempt was made to remedy that problem.
Or you can watch Brodie Robertson's video about the implementation of the PEP in Arch: https://www.youtube.com/watch?v=35PQrzG0rG4.
The Machine-Learning world, especially "Google Brain" research team figured out that NumPy was an awesome piece of software for dealing with large arrays of numbers and matrix multiplication. They built "TensorFlow" on top of it around 2015 which became very popular. Facebook followed suit and released PyTorch in 2016.
IPython/Jupiter notebooks (for Julia, Python and R) from 2015 were another factor, also adopted by the AI/ML community.
The alternative data-science languages at the time were Mathematica, MATLAB, SAS, Fortran, Julia, R, etc, but Python probably won because it was general purpose and open source.
I suspect Python would not have survived the 2/3 split very well if it wasn't for AI/ML adopting Python as its main language.
> when the tooling was so inferior
Since 2012, Conda/Anaconda has been the go-to installer in the SciPy/NumPy world which also solves a lot of problems that uv solves.
No need to clone/manually install packages first. E.g. `uvx --from "git+https://github.com/richstokes/meshtastic_terminal.git" meshtastic-tui`
[0]: https://blog.toolkami.com/mcp-server-in-a-file/
The home page should be a simplified version of this page buried way down in the docs: https://docs.astral.sh/uv/guides/projects/
The fact that it's a binary, not written in python, also simplifies bootstrapping. So you don't need python+dependencies installed in order to install your python+dependencies.
Some foundations have moved into the stdlib. This means that newer tools are much more compatible with each other and mainly just differ in implementation rather than doing different things altogether. The new stuff is working on a much more standard base and can leave behind many dark crufty corners.
Unravelling the legacy stuff and putting the standards in place seems to have taken 15+ years?
Standards are developed to allow existing tools to inter-operate; this entails allowing new tools to appear (and inter-operate), too.
This system was in some regards deliberate, specifically to support competition in "build backends". The background here is that many popular Python projects must interface to non-Python code provided with the project; in many cases this is code in compiled languages (typically C, Fortran or Rust) and it's not always possible to pre-build for the user's system. This can get really, really complicated, and people need to connect to heavyweight build systems in some cases. The Python ecosystem standards are designed with the idea that installers can automatically obtain and use those systems when necessary.
And by doing all of this, Python core developers get to focus on Python itself.
Another important concern is that some bad choices were made initially with Setuptools, and we have been seeing a very long transition because of a very careful attitude towards backwards compatibility (even if it doesn't seem that way!) which in turn is motivated by the battle scars of the 2->3 transition. In particular, it used to be normal and expected that your project would use arbitrary Python code (in `setup.py` at the project root) simply to specify metadata. Further, `setup.py` generally expects to `import setuptools`, and might require a specific version of Setuptools; but it can't express its build-time Setuptools version requirement until the file is already running - a chicken-and-egg scenario.
Modern projects use a declarative TOML file for "abstract" metadata instead (which is the source for concrete metadata included in the actual build artifacts), but the whole ecosystem still has to support a lot of really outdated ways of doing things, because in part of how much abandonware is out there.
[0]: Wheels are zip-compressed, and Python can run code from a zip file, with some restrictions. The pip project is designed to make sure that this will work. The standard library provides a module "ensurepip" which locates this wheel and runs a bootstrap script from that wheel, which will then install into the current environment. Further, the standard library "venv", used to create virtual environments, defaults to using this bootstrap in the newly created environment.
With python over the years i can think of pip, pipx, setuptools, easy_install, distutils, venv, conda, wheel, .egg, wheel (formats) , now uv.
PHP stabilized with composer, perl with cpan , go with `go mod` and `go get` (builtin).
Java and Swift had some competition with Gradle/maven and swiftPM / cocoapods, but nothing as egregious.
file tree, dep tree, task DAG. how many ways can they be written?
Almost literally: https://wheelnext.dev/
> how many ways can they be written?
It's not just a matter of how they're written. For Python specifically, build orchestration is a big deal. But also, you know, there are all the architecture ideas that make uv faster than pip. Smarter (and more generous) caching; hard-linking files where possible rather than copying them; parallel downloads (I tend to write this off but it probably does help a bit, even though the downloading process is intermingled with resolution); using multiple cores for precompiling bytecode (the one real CPU-intensive task for a large pure-Python installation).
(Transparently, I'm posting this before I've completed the article.)
uv's biggest advantage is speed. It claims a 10-100x performance speedup over pip and Conda [1]. uv can also manage python versions and supports using Python scripts as executables via inline dependencies [2].
But Conda is better for non-Python usage and is more mature, especially for data science related uses.
[1]: https://github.com/astral-sh/uv/blob/main/BENCHMARKS.md [2]: https://docs.astral.sh/uv/#scripts
Please see https://news.ycombinator.com/item?id=45753142.
I had to update some messy python code and I was looking for a tool that could handle python versions, package updates, etc. with the least amount of documentation needing be read and troubleshooting.
Rye was that for me! Next time I write python I'm definitely going to use uv.
What strikes me about uv is that it seems to understand that not everyone launching a Python-based project has a CS degree. That accessibility matters—especially in the era where more non-engineers are building products.
Curious: for those who've switched to uv, did you notice any friction when collaborating with team members who were still on traditional setups? I'm thinking about adoption challenges when you're not a solo builder.
I'm very happy the python community has better tooling.
I'm teaching (strongly recommending/forcing using) uv in all my courses now.
I've always wondered why Linux OSes that rely on python scripts don't make their own default venv and instead clobber the user's default python environment...
Since I am mostly avoiding non-reproducible use-cases, like for example stating dependencies inside the python scripts themselves, without checksums, only with versions, and stuff like that, I am not really benefiting that much. I guess, I am just not writing enough throwaway code, to benefit from those use-cases.
Some people here act, like uv is the first tool ever to install dependencies like npm and cargo and so on. Well, I guess they didn't use poetry before, which did just that.
Pip is also not conda, but uv is way faster than pip.
> Reminds me of that competing standards xkcd.
Yes, for years I've sat on the sidelines avoiding the fragmented Poetry, ppyenv, pipenv, pipx, pip-tools/pip-compile, rye, etc, but uv does now finally seem to be the all-in-one solution that seems to be succeeding where other tools have failed.
In general, you can use your preferred package management tool with their code. The developers are just showing you their own workflow, typically.
not a python developer, so not sure it's equivalent as the npm registry is shared between all.
uv is a clear improvement over pip and venv, for sure.
But I do everything in dev containers these days. Very few things get to install on my laptop itself outside a container. I've gotten so used to this that tools that uninstall/install packages on my box on the fly give me the heebie-jeebies.
Yes, it was the NPM supply chain issues that really forced this one me. Now I install, fetch, build in an interactive Docker container
This whole discussion has the same vibes like digital photography 15 years ago. Back then some people spent more time on discussing the tech spec their cameras than takin photos. Now some people spend more time on discussing the pros and cons of different Python environment management solutions than building real things.
The last time I had to touch one of my dockerized environments was when Miniconda and Miniforge were merged. I said the agent "fix the dockerfile", and the third attempt worked. Another time, one dependency was updated and I had to switch to Poetry. Once again, I said the agent "refactor the repository to Poetry" and it worked. Maybe because all my Python package versions are frozen and I only update them when they break or when I need the functionality of the new version.
Whenever this topic pops up in real life, I always ask back what was the longest time they managed the same Python service in the cloud. In the most cases, the answer is never. The last time someone said one year. After a while this service was turned into two .py files.
I don't know. Maybe I'm just too far away from FAANG level sorcery. Everything is a hammer if all you have to deal with are nails.
You can go from no virtual environment, and just "uv run myfile.py" and it does everything that's needed, nearly instantly.
https://danluu.com/productivity-velocity
https://danluu.com/input-lag/
The wheel basically contains a compiled ~53MB (huh, it's grown in recent versions) Rust executable and a few boilerplate files and folders to make that play nice with the Python packaging ecosystem. (It actually does create an importable `uv` module, but this basically just defines a function that tells you the path to the executable.)
If you want it in your system environment, you may be out of luck, but check your full set of options at https://docs.astral.sh/uv/getting-started/installation/ .
The install script does a ton of system introspection. It seems to be structured quite similarly to the Julia installer, actually.
For example, installing on an air gapped system, where uv barely has support.
uv has implemented experimental support, which they announced here [3].
[0] https://wheelnext.dev/proposals/pepxxx_wheel_variant_support...
[1] https://us.pycon.org/2025/schedule/presentation/100/
[2] https://www.youtube.com/watch?v=1Oki8vAWb1Q
[3] https://astral.sh/blog/wheel-variants
I'm interested if you have any technical documentation about how conda environments are structured. It would be nice to be able to interact with them. But I suspect the main problem is that if you use a non-conda tool to put something into a conda environment, there needs to be a way to make conda properly aware of the change. Fundamentally it's the same issue as with trying to use pip in the system environment on Linux, which will interfere with the system package manager (leading to the PEP 668 protections).
I don’t really get that uv solves all these problems ve never encountered. Just make a venv and use it seems to work fine.
For me package installation is way, way faster with uv, and I appreciate not needing to activate the virtual environment.
I don't love that UV is basically tied to a for profit company, Astral. I think such core tooling should be tied to the PSF, but that's a minor point. It's partially the issue I have with Conda too.
I just... build from source and make virtual environments based off them as necessary. Although I don't really understand why you'd want to keep older patch versions around. (The Windows installers don't even accommodate that, IIRC.) And I can't say I've noticed any of those "significant improvements and differences" between patch versions ever mattering to my own projects.
> I don't love that UV is basically tied to a for profit company, Astral. I think such core tooling should be tied to the PSF, but that's a minor point. It's partially the issue I have with Conda too.
In my book, the less under the PSF's control, the better. The meager funding they do receive now is mostly directed towards making PyCon happen (the main one; others like PyCon Africa get a pittance) and to certain grants, and to a short list of paid staff who are generally speaking board members and other decision makers and not the people actually developing Python. Even without considering "politics" (cf. the latest news turning down a grant for ideological reasons) I consider this gross mismanagement.
The PSF is busy with social issues and doesn't concern itself with trivia like this.
Edit: or was it ruff? Either way. I thought they created the tools first, then the company.
Wonderful project
Compare this to the Go community, who celebrate rewrites from other languages into Go. They rewrote their compiler in Go even though that made it worse (slower) than the original C version, because they enjoy using their own language and recognise the benefits of dogfooding.
3.14 is a big deal.
But otherwise, people on this forum and elsewhere are praising uv for: speed, single-file executable, stability, and platform compatibility. That's just a summary of the top reasons to write in Rust!
I agree 3.14 is a big deal as far as Python goes, but it doesn't really move the needle for the language toward being able to author apps like uv.
Which is fine, Python is not for everything.
But I’m utterly shocked that UV doesn’t support “system dependencies”. It’s not a whole conda replacement. Which is a shame because I bloody hate Conda.
Dependencies like Cuda and random C++ libraries really really ought to be handled by UV. I want a true genuine one stop shop for running Python programs. UV is like 80% of the way there. But the last 20% is still painful.
Ideally UV would obsolete the need for docker. Docker shouldn’t be a requirement to reliable run a program.
EDIT: Looks like I fell hook, line, and sinker for the troll. Shame on me.
My brother in christ, we are all just names without bodies or even faces on this digital ocean of the internet. Letting people know how they should address you isn't "disconnected from reality", it's grounded in the very real reality that we, as people, like talking to each other. We should all be so thankful for their foresight in allowing us the opportunity of avoiding an otherwise unavoidable faux pas of calling everyone in the world "hey you".
No, the same uv that people have been regularly (https://hn.algolia.com/?q=uv) posting about on HN since its first public releases in February of 2024 (see e.g. https://news.ycombinator.com/item?id=39387641).
> How many are there now?
Why is this a problem? The ecosystem has developed usable interoperable standards (for example, fundamentally uv manages isolated environments by using the same kind of virtual environment created by the standard library — because that's the only kind that Python cares about; the key component is the `pyvenv.cfg` file, and Python is hard-coded to look for and use that); and you don't have to learn or use more than one.
There are competing options because people have different ideas about what a "package manager" should or shouldn't be responsible for, and about the expectations for those tasks.
Have you tried uv?
Other than speed and consolidation, pip, pipx, hatch, virtualenv, and pyenv together roughly do the job (though pyenv itself isn’t a standard python tool.)
> Why uv over, lets say, conda?
Support for Python standard packaging specifications and consequently also easier integration with other tools that leverage them, whether standard or third party.
If you haven't spent 5 minutes trying it out, you don't know what you're missing.
If you're worried about getting addicted like everyone else, I could see that as a valid reason to never try it in the first place.
I don’t think people would think twice about the legitimacy (if you want to call it that) of uv except for all the weird fawning over it that happens, as you noticed. It makes it seem more like a religion or something.
So you are back having to use conda and the rest. Now, you have yet another package manager to handle.
I wouldn't be harsh to engineers at astral who developed amazing tooling, but the issue with the python ecosystem isn't lack of tooling, it is the proliferation and fragmentation. To solve dependency management fully would be to incorporate other package descriptors, or convert them.
Rsbuild, another rust library, for the node ecosystem did just that. For building and bundling. They came up with rspack, which has large compatibility with the webpack config.
You find a webpack repo? Just add rsbuild, rspack, and you are pretty much ready to go, without the slow (node native) webpack.
Conda solves a completely orthogonal set of problems, and is increasingly unnecessary. You can `pip install scipy` for example, and have been able to for a while.
I refered to the interfaces of other packaging tools. I use uv and it's excellent on its own.
You get a repo, it's using playwright, what do you do now ? You install all the dependencies found in the dependency descriptor then sync to create a uv descriptor. or you compose a descriptor that uv understands.
It's repetitive, rather systematic so it could be automated. I should volunteer for a PR but my point is introducing yet another tool to an ecosystem suffering a proliferation of build and deps management tooling expands the issue. It would have been helpful from the get go to support existing and prolific formats.
pnpm understands package.json It didn't reinvent the wheel be cause we have millions of wheels out there. It created its own pnpm lock file, but that's files a user isn't meant to touch so it goes seamlessly to transition from npm to pnpm. Almost the same when migrating from webpack to rsbuild.