Just a few months back I said I would never use uv. I was already used to venv and pip. No need for another tool I thought.
I now use uv for everything Python. The reason for the switch was a shared server where I did not have root and there were all sorts of broken packages/drivers and I needed pytorch. Nothing was working and pip was taking ages. Each user had 10GB of storage allocated and pip's cache was taking up a ton of space & not letting me change the location properly. Switched to uv and everything just worked
If you're still holding out, really just spend 5 minutes trying it out, you won't regret it.
For me, the big key was: uv is so much easier to explain and especially use - especially for people who sometimes script something in python and don't do this daily.
pip + config file + venv requires you to remember ~2 steps to get the right venv - create one and install stuff into it, and for each test run, script execution and such, you need to remember a weird shebang-format, or to activate the venv. And the error messages don't help. I don't think they could help, as this setup is not standardized or blessed. You just have to beat a connection of "Import Errors" to venvs into your brain.
It's workable, but teaching this to people unfamiliar with it has reminded me how.. squirrely the whole tooling can be, for a better word.
Now, team members need to remember "uv run", "uv add" and "uv sync". It makes the whole thing so much easier and less intimidating to them.
Unfortunately uWSGI (one of the most important libraries) is fundamentally incompatible with uv. had to roll back all my apps that use custom uWSGI for that reason.
Similarly to the sibling I also use both. I let mise manage my uv version (and other tools) and let uv handle Python + PyPI Packages for me. Works great!
Think of uv more as like npm or other thing like that. The new Python pyproject.toml is similar package.json. It defines the project description, list of dependencies, and other hooks. Uv is a package/project tool using pyproject.toml. It is easy to manage dependencies, build and publish to PyPi, add hooks to run tests, linters, or whatever, again much like package.json. It also manages the virtualenv automatically, though you can manage it yourself.
Thanks for mentioning mise. I'm more interested in it for the task running feature. I couldn't figure out how to have shell scripts in Justfile, so I gave up on it.
I use Just inside (and outside) mise, almost exclusively with embedded shell/python scripts. Have used mise tasks a little, but didn't meaningfully add enough for me
Or do you mean justfiles with shebangs?
Either way, curious what problem you hit (and may be able to unblock ya)
From what I remember, the task consisted of several lines of shell and just would just choke at some point. I think it was because I used a new line to visually separate code. I'm not sure, I don't remember exactly as it was some time ago and I just gave up.
Same! I recently set up some dev environments with mise. My old ones are still using poetry, the new ones have uv. uv is incredibly snappy. It's like night and day!
What has been holding me back on uv is my experience with Ruff. Ruff claims "Drop-in parity with Flake8, isort, and Black", but that is simply not true. At least for isort, Ruff only re-implemented what they wanted and then ask you to use Ruff to call out to the old isort tool if there's a feature or setting that's missing in the Ruff re-implementation. So what's the point? Ruff just partially re-implemented many existing different tools and added some new ones. So using Ruff actually increases the amount of tools, yet again, you're having to use because of this and it also not doing everything that Pylint does.
For moving to uv, I haven't heard a good story for what uv provides over Poetry rather than "is fast". The only unique thing that I am currently aware of is that uv can install Python itself, which gets rid of tools like Pyenv. I'm interested because of that, but "is fast" isn't enough of a reason.
My experience is that ruff reimplemented 99% of the most popular features of black, isort, flake8, pylint, etc. and then added 10000% more features on top, that feels like a fair tradeoff to me.
I've converted multiple large Python codebases to ruff, and each time I just configure ruff as close to the previous tools as possible, then reformat the entire codebase with ruff and remove all the previous tools. The speed increase when linting alone is worth the minor formatting changes to me.
If you really insist on keeping isort's sorting then you could at least replace black and pylint, which would reduce the total number of tools by one.
I think you're underestimating how "fast" uv really is. It's night and day different. The sentence that got me to switch and I echo still was: "uv is so fast, I sometimes don't even think it's actually doing anything". The first time I used it, I had to actually double check if uv actually was working and installing the packages that it was saying it was. There's a lot of reasons I've heard not to use it that kind of (not really) make sense (you don't want another package manager, you're happy with just base pip, etc). But if you're already using poetry, you really should just switch. uv is just poetry but without any waiting.
Even your most complicated projects should be able to switch within a day. The only reason I took any time was having to restructure my docker containers to work with uv instead of poetry. And that's mostly with my inexperience with docker, not because uv is complicated.
> Each user had 10GB of storage allocated and pip's cache was taking up a ton of space & not letting me change the location properly. Switched to uv and everything just worked
Is it better about storage use? (And if so, how? Is it just good at sharing what can be shared?)
In those situations, have you had a any luck using UV_LINK_MODE=symlink? I eventually had to resort to `copy` mode because it seemed the folder names (hashes) uv created in the package cache were not fully deterministic. So sometimes my cache volume would change and my Docker build would break. :\
similar story recently with an experimental repo that starts with "its so easy, just $uv a b c" .. under the hood it implies a lot of redundancies? but true enough it worked fine and trouble-free too, on a standard GNU-Debian-Ubuntu host
> Just a few months back I said I would never use uv. I was already used to venv and pip. No need for another tool I thought
Really? :)
requirements.txt is just hell and torture. If you've ever used modern project/dependency management tools like uv, Poetry, PDM, you'll never go back to pip+requirements.txt. It's crazy and a mess.
uv is super fast and a great tool, but still has roughnesses and bugs.
Pip-tools+requirements.txt helped me survive the past few years. I also never thought I needed uv, but after all the talk about it I gave it a spin and never want back. It’s just so blazing fast en convenient.
What for? Support legacy CI/CD pipelines or something like that?
uv.lock already contains locked versions of all dependencies plus a lot of other needed metadata.
I remember using pip and venv back in like 2009. Last time I checked, maybe 5 or 10 years ago, the recommendation of the community was generally to just use Docker instead of all these tools. Did that not catch on?
The advice seems to change every year. For a while it was venv, then pipenv, poetry, docker, and now uv. Maybe the ecosystem will settle on that but who knows.
Docker was always a workaround to Python not having a non-awful dependency manager. uv is that non-awful dependency manager, and I expect in the long term it will reduce the use of Docker.
Docker solves a different problem. Docker is a way to basically ship your whole OS off to another machine. You still have to have a way to install the right version of python and all the python libraries you need inside the Docker container, and uv is great for this.
Secondly Docker only solves a subset of problems. It's fine if you're developing a server that you will be deploying somewhere. It's inconvenient if you're developing an end user application, and it's completely useless if you're developing a library you want people to be able to install.
I love uv. The one gotcha I'll warn people about is: don't touch uvx. I've lost an embarrassing number of hours or days trying to figure out why nothing works properly or makes sense when I tried to run things with uvx. I guess I understand why it's there, but I think it's a built-in foot-gun and not well documented. But if you stay away from it, things work great.
The first time I used `uv`, I was sure that I had made a mistake or typed something wrong because the process finished so much more quickly than anything I had ever experienced with `pip`.
I've had poetry sit there for minutes resolving dependencies at a previous company. I thought something was broken... it probably was... but it did eventually complete.
I've had poetry take so long that I've just deleted the whole cache/poetry list and reinstalled the packages rather than wait for poetry to figure out whatever dependency hell it found itself in. I switched to uv the minute i tried it out
uv and ruff are a great counterexample to all those people who say "never reinvent the wheel". Don't ever do it just for the sake of doing it, but if you have focused goals you can sometimes produce a product that's an order of magnitude better.
I believe most of the time this phrase is said to an inexperienced artisan who has no idea how the current system works, what's the shortcoming of it, and how to improve upon it. Think of an undergraduate student who tries to solve the Goldbach conjecture. Usually what ended up is either he fails to reinvent the wheel, or reinvent the exact same wheel, which has no value. The phrase certainly does not apply to professionals.
Even then, you know what's a good way to learn about how the current system works etc, maybe even the best way? I've got many failed projects behind me, and 0 regrets.
A big part of the "magic" is that there is a team of paid professionals maintaining and improving it. That's more important than it being written in Rust. If uv were forked it would devolve to the level of pip over time.
I'd be curious to know exactly what changed. Python -> Rust won't make network downloads faster nor file I/O faster. My naive guess is that all the speed comes from choosing better algorithms and/or parallelizing things. Not from Python vs Rust (though if it's hard to parallelize in Python and easy in rust that would certainly make a difference)
I've translated code from Ruby to Python, and other code from Rust to Python.
Rust's speed advantages typically come from one of a few places:
1. Fast start-up times, thanks to pre-compiled native binaries.
2. Large amounts of CPU-level concurrency with many fewer bugs. I'm willing to do ridiculous threading tricks in Rust I wouldn't dare try in C++.
3. Much lower levels of malloc/free in Rust compared to some high-level languages, especially if you're willing to work a little for it. Calling malloc in a multithreaded system is basically like watching the Millennium Falcon's hyperdrive fail. Also, Rust encourages abusing the stack to a ridiculous degree, which further reduces allocation. It's hard to "invisibly" call malloc in Rust, even compared to a language like C++.
4. For better or worse, Rust exposes a lot of the machinery behind memory layout and passing references. This means there's a permanent "Rust tax" where you ask yourself "Do I pass this by value or reference? Who owns this, and who just borrows is?" But the payoff for that work is good memory locality.
So if you put in a modest amount of effort, it's fairly easy to make Rust run surprisingly fast. It's not an absolute guarantee, and there are couple of traps for the unwary (like accidentally forgetting to buffer I/O, or benchmarking debug binaries).
NVMe hungers, keeping it fed is hard work. Doing some serial read, decompress, checksum, write loop will leave if starved (QD<1) whenever you're doing anything but the last step.
Disk IO isn't async unless you use io_uring (well ok, writeback caches can be). So threads are almost a must to keep NVMe busy.
Conversely, waiting for blocking IO (e.g. directory enumeration) will keep your CPU starved. Here too the answer is more threads.
packages are in zip files. zip files have their TOC at the end.
So, instead of downloading the entire zip they just get the end
of the file, read the TOC, then from that download just the metadata
part
I've written that code before for my own projects.
2. They cache the results of packages unzipped and then link into your environment
This means there's no files being copied on the 2nd install. Just links.
Both of those are huge time wins that would be possible in any language.
3. They store their metadata as a memory dump
So, on loading there is nothing to parse.
Admittedly this is hard (impossible?) in many languages. Certainly not possible in Python
and JavaScript. You could load binary data but it won't be useful without copying it into
native numbers/strings/ints/floats/doubles etc...
I've done this in game engines to reduce load times in C/C++ and to save memory.
It'd be interesting to write some benchmarks for the first 2. The 3rd is a win but I suspect the first 2 are 95% of the speedup.
From just my observations they basically parallelized the install sequence instead of having it be sequential (among many other optimizations most likely)
It became a bit of a meme, especially in the web development space, that all programs are always waiting on external resources like networks, databases, disks, etc., and so scripting languages being slower than other languages doesn't matter and they'll always be as fast as non-scripting languages.
Even on a single core, this turns out to be simply false. It isn't that hard to either A: be doing enough actual computation that faster languages are in fact perceptibly faster, even, yes, in a web page handler or other such supposedly-blocked computation or B: without realizing it, have stacked up so many expensive abstractions on top of each other in your scripting language that you're multiplying the off-the-top 40x-ish slower with another set of multiplicative penalties that can take you into effectively arbitrarily-slower computations.
If you're never profiled a mature scripting language program, it's worth your time. Especially if nobody on your team has ever profiled it before. It can be an eye-opener.
Then it turns out that for historical path reasons, dynamic scripting languages are also really bad at multithreading and using multiple cores, and if you can write a program that can leverage that you can just blow away the dynamic scripting languages. It's not even hard... it pretty much just happens.
(I say historical path reasons because I don't think an inability to multithread is intrinsic to the dynamic scripting languages. It's just they all came out in an era when they could assume single core, it got ingrained into them for a couple of decades, and the reality is, it's never going to come fully out. I think someone could build a new dynamic language that threaded properly from the beginning, though.)
You really can see big gains just taking a dynamic scripting language program and turning it into a compiled language with no major changes to the algorithms. The 40x-ish penalty off the top is often in practice an underestimate, because that number is generally from highly optimized benchmarks in which the dynamic language implementation is highly tuned to avoid expensive operations; real code that takes advantage of all the conveniences and indirection and such can have even larger gaps.
This is not to say that dynamic scripting languages are bad. Performance is not the only thing that matters. They are quite obviously fast enough for a wide variety of tasks, by the strongest possible proof of that statement. That said, I think it is the case that there are a lot of programmers who have no idea how much performance they are losing in dynamic scripting languages, which can result in suboptimal engineering decisions. It is completely possible to replace a dynamic scripting language program with a compiled one and possibly see 100x+ performance improvements on very realistic code, before adding in multithreading. It is hard for that not to manifest in some sort of user experience improvement. My pitch here is not to give up dynamic scripting languages, but to have a more realistic view of the programming language landscape as a whole.
> Then it turns out that for historical path reasons, dynamic scripting languages are also really bad at multithreading and using multiple cores...
What would a dynamic scripting language look like that wasn't subject to this limitation? Any examples? I don't know of contenders in this design space--- I am not up on it.
The big difference from Python is probably having to use a real tracing GC instead of automatic reference counting. For a single-threaded program, refcounts are beneficial in multiple ways, being fairly cheap, having a smooth performance profile, maintaining low resident set size, and providing deterministic freeing.
But because of the way cache coherency for shared, mutated memory works, parallel refcounting is slow as molasses and will always remain so.
I think Ruby has always used a tracing GC, but it also still has a GIL for some reason?
It would look pretty much the same. It would just have been written to be multithreaded from the beginning, and lack the long list of restrictions and caveats and "but it doesn't work with our C extensions" and such. There wouldn't be a dozen major libraries trying to solve the problem (which, contrary to many people's intuition, is often a sign that a language lacks a good solution). This is part of why I say there's no fundamental reason this can't be done, it's just a historical accident.
There are dynamic languages that were built with concurrency in mind like Clojure. It’s also a surprisingly fast language considering it’s both dynamic and functional.
I'm not trying to suggest that you can't do faster computation in a lower-level language. But, a package manager doesn't do much computation. It mostly downloads, decompresses, and writes files. Yes, it has to solve constraints but that's not a bottleneck given most projects have at most a few 100 dependencies and not millions.
I don't know python but in JavaScript, triggering 1000 downloads in parallel is trivial. Decompressing them, like in python, is calling out to some native function. Decompressing them in parallel in JS would also be trivial (no idea about python). Writing them in parallel is also trivial.
Congratulations! You have proved that it is impossible for uv to be way, way faster than Python-based package managers!
....
Unfortunately, there seems to be a problem here.
When reality and theory conflict, reality wins.
It sounds like you've drunk the same Kool-Aide I was referring to in my post. It's not true. When you're playing with 50x-100x slowdowns, if not more, it's really quite easy to run into user-perceptible slowdowns. A lot of engineers grotesquely underestimate how slow these languages are. I suspect it may be getting worse over time due to evaporative cooling, as engineers who do understand it also tend to have one reason or another to leave the language community at some point, and I believe (though I can not prove) that as a result the dynamic scripting language communities are actually getting worse and worse at realizing how slow their languages are. They're really quite slow.
You seem to be implying rust = fast, the end. I'm implying algorithms and design choices = fast. Those decisions generally (though not always) are far more effective at speed than language choice.
I watched the video linked above on uv. They went over the optimizations. The big wins had nothing to do with rust and everything to do with design/algo choices.
You could have also done without the insults. You have no idea who I am and my experiences. I've shipped several AAA games written in C/C++ and assembly. I know how to optimize. I also know how dynamic languages work. I also know when people are making up bullshit about "it's fast because it's in rust!". No, that is not why it's fast.
I agree there are lot of big wins in uv that tools written in python could take advantage of, and ultimately I think uv is fast because they're obsessed with making it fast, which is why they chose to use rust. I don't see that same level speed obsession with the other tools.
Instead of "It's fast because it's in rust", I'd say: "It's fast because they chose to use rust for their python tool, which means they care a lot about speed."
The package resolution is a big part of it, it's effectively a constraint solver. I.e. if package A requires package B constrained between version 1.0 < X <= 2.X and Package B requires package C between... and so on and so on.
Conda rewrote their package resolver for similar reasons
Python package management was notoriously awful. The problem wasn't that people were trying to do things better, it was that they weren't; every new Python dependency management tool just repeated the mistakes of all the previous Python dependency management tools. uv is the first one to break the cycle (and it's probably not a coincidence that it's the first one to not be written in Python).
Poetry broke the cycle. Unified toolchain, lock file, single configuration file, full dependency graph, dev dependencies. uv is faster which is great but Poetry was a huge step in the right direction and still a good tool.
Worth noting uv-specific implementation has more features than that standard PEP 751 lockfiles, so uv plans to keep using its own implementation by default. https://github.com/astral-sh/uv/issues/12584
Actually not true. One of the main differences with uv is you don't have to think about venvs any more. There's a talk about it from one of the authors at a recent PyCon here: https://www.youtube.com/watch?v=CV8KRvWKYDw (not the same talk I linked elsewhere in the thread).
I said you don't have to think about venvs any more. It's great that we have a standard way to implement them, but this is only necessary in the first place because of the way Python is. Now we have a tool that enforces a workflow that creates virtualenvs without you having to know about them and therefore not screwing them up with ad hoc pip installs etc.
The implementation details don't matter. uv might follow PEP 405 but it could work just as well without doing so. The point is that it doesn't give you the bunch of extra footguns that any other Python package management does.
It matters immensely that it follows PEP 405, it makes uv the implementation detail. You can swap out uv for any other project management tool or build frontend and change nothing needs to change about the development environment.
> You can swap out uv for any other project management tool or build frontend and change nothing needs to change about the development environment.
> This is the entire purpose of the standards.
That seems to amount to saying that the purpose of the standards is to prevent progress and ensure that the mistakes of early Python project management tools are preserved forever. (Which would explain some things about the last ~25 years of Python project management I guess). The parts of uv that follow standards aren't the parts that people are excited about.
The parts of uv that follow standards aren't the parts that people are excited about.
I disagree. Had uv not followed these standards and instead gone off and done their completely own thing, it could not function as a drop in replacement for pip and venv and wouldn't have gotten anywhere near as much traction. I can use uv personally to work on projects that officially have to support pip and venv and have it all be transparent.
Honestly "don't reinvent the wheel" makes absolutely no sense as a saying. We're not still all using wooden discs as wheels, we have invented much better wheels since the neolithic. Why shouldn't we do the same with software?
Quoted keys, strict comma rules, very limited data types, are the main ones. There are a host of others if you view it through the lenses of user-read/write, and a different set of issues if you view it as a machine data interface. Trying to combine the two seems fundamentally misguided.
Lack of comments seems like a big one seeing as it's so widely used for "configuration". It's a big enough downside that VSCode and others have violated it via ad-hoc extensions of the format.
The comma rules introduce diff noise on unrelated lines.
Standards do die off, up to a point. XML is widely used but the last time I really had to edit it in anger working in DevOps/web/Python was a long time ago (10 years ago?).
At this point XML is the backbone of many important technologies that many people won't use or won't use directly anymore.
This wasn't the case circa 2010, when I doubt any dev could have really avoided XML for a bunch of years.
Right, wheels are reinvented every few years. Compare tires of today to the ones 20 years ago and the technology and capability is very different, even though they look identical to a casual eye.
My primary vehicle has off-road capable tires that offer as much grip as a road-only tire would have 20-25 years ago, thanks to technology allowing Michelin to reinvent what a dual-purpose tire can be!
> In the last decade, the spiciest street-legal tires have nearly surpassed the performance of a decade-old racing tire, and computer modeling is a big part of the reason
I always took this saying as meaning that we don’t re-invent the concept of the wheel. For example the Boring company and Tesla hoping to reinvent the concept of the bus/train.. (iirc your car goes underground on some tracks and you get to bypass traffic and not worry about steering)
A metal wheel is still just a wheel. A faster package manager is still just a package manager.
I also like Dan Luu's take (starting with a Joel Spolsky quote)
“Find the dependencies — and eliminate them.” When you're working on a really, really good team with great programmers, everybody else's code, frankly, is bug-infested garbage, and nobody else knows how to ship on time.
We had a similar attitude, although I'd say that we were a bit more humble. We didn't think that everyone else was producing garbage but, we also didn't assume that we couldn't produce something comparable to what we could buy for a tenth of the cost. From talking to folks at some competitors, there was a pretty big cultural difference between how we operated and how they operated. It simply didn't occur to them that they didn't have to buy into the standard American business logic that you should focus on your core competencies, that you can think through whether or not it makes sense to do something in-house on the merits of the particular thing instead of outsourcing your thinking to a pithy saying.[0]
I came here to (wrongly) say that wooden disks were never used as wheels, and that ot all started with spokes.
Some checking showed that, in fact, the oldest known wheels have a lot of solid disks. E.g: https://en.m.wikipedia.org/wiki/Ljubljana_Marshes_Wheel
Hopefully this can disabuse others of similar mistaken memory.
It's meant to signify a step change.
Order of magnitude change = no amount of incremental changes would make up for it.
In common conversation, the multiplier can vary from 2x - 10x. In context of some algorithms, order of magnitudes can be over the delta rather than absolutes. eg: an algorithms sees 1.1x improvement over the previous 10 years. A change that shows a 1.1x improvement by itself, overshadows an an order-of-magnitude more effort.
For salaries, I've used order-of-magnitude to mean 2x. Good way to show a step change in a person's perceived value in the market.
The problem is that 10x appears to be a linear scale. It could mean 9.5x to 10.5x if it's supposed to have 2 significant digits. Or it could be 5x to 15x if it meant to have 1 significant digit.
"10x" has been cheapened / heard enough / de facto, is a more general statement than a literal interpretation would indicate. (i.e. 10x engineer. Don't hear that much around these parts these days)
Order of magnitude faces less of that baggage, until it does :)
Just a warning in case others run into it: on very anemic systems (e.g.: AWS T2.micro running Windows, yes... I know...) uv will try to open too many simultaneous downloads, overloading things, resulting in timeouts.
You can use ent ENV variable UV_CONCURRENT_DOWNLOADS to limit this. In my case it needed to be 1 or 2. Anything else would cause timeouts.
An extreme case, I know, but I think that uv is too aggressive here (a download thread for every module). And should use aggregate speeds from each source server as a way of auto-tuning per-server threading.
Not extreme at all, A lot of people use the cheapest smallest VPS for their hobby work. I know I do (albeit not AWS). Thanks for sharing, hope they improve the automatic detection there.
uv can also run even a beefy linux desktop out of file descriptors for larger projects. And does not have deterministic / reproducible installs. Still needs maturity.
I built a Claude Project with special instructions just teaching it how to do this, which means it can output full scripts for me with inline dependencies based on a single prompt: https://simonwillison.net/2024/Dec/19/one-shot-python-tools/
Claude 4's training cutoff date is March 2025 though, I just checked and it turns out Claude Sonnet 4 can do this without needing any extra instructions:
Python script using uv and inline script dependecies
where I can give it a URL and it scrapes it with httpx
and beautifulsoup and returns a CSV of all links on
the page - their URLs and their link text
Using your system instructions for uv for every LLM now since first seeing your post last year, thanks! It's insanely helpful just asking e.g. Claude to give me a python script for XYZ and just using "uv run". I also added:
If you need to run these scripts, use "uv run script-name.py". It will automatically install the dependencies. Stdlibs don't need to be specified in the dependencies array.
since e.g. Cursor often gets confued because the dependencies are not installed and it doesn't know how to start the script. The last sentence is for when LLMs get confused and want to add "json" for example to the dependency array.
claude sonnet typically forgets about uv script syntax in my experience. I usually find myself having to paste in the docs every time. By default it wants to use uv project syntax.
Started using this recently for personal stuff on my laptop. When you're used to pip, it's just confusingly fast. More than once I thought maybe it didn't work because it returned too quickly..
Tried uv a while ago and I was shocked by how fast and easy it is to use. There's basically no reason to use pip anymore, and if you're using only Python there's basically no reason to use conda either.
Until the moment you will realize that ruff perform only a part of pylint checks and that very obvious mistakes can go through easily like code that can't run because of an obvious error.
It prevents uv from making a virtual environment and does some optimizations like compiling byte code once when your dependencies get installed.
It was well worth the switch. I noticed a ~10x improvement for speed compared to pip (30s to 3s to install all dependencies). Proper lock file support is nice too. Funny enough I wrote about and made a video about switching to uv about a week ago here https://nickjanetakis.com/blog/switching-pip-to-uv-in-a-dock....
Sadly hashbangs are technically limited to:
1) Support only absolute paths, making it necessary to use /usr/bin/env which is in standardized location to look up the uv binary
2) Support only a single argument (everything after the space is passed as a single arg, it's not parsed into multiple args like a shell would), making it necessary to use -S to "S"plit the arguments. It's a feature of env itself, for this very use case.
So there isn't really much to do to make it simpler.
I wasn't really referring to env. I meant change the behavior of uvx. If the first argument passed to uvx is a file path, then execute it exactly the same way as `uv --quiet run --script` does.
Or maybe create a second binary or symlink called `uvs` (short for uv script) that does the same thing.
UV is fast, like FAST. Plus, it removes the need for pyenv (for managing different Python versions) and pip for me. Plus, no need to activate env or anything, `uv run ...` automatically runs your code through the env.
I don't see a way to change current and global versions of python/venvs to run scripts, so that when I type "python" it uses that, without making an alias.
export PATH="$(uv run python -BISc 'import sys; print(sys.base_exec_prefix)')/bin:$PATH"
2. Put this in an executable file called "python" early in your PATH:
#!/bin/sh
exec uv run python $*
Those are basically what pyenv does (via a symlink and PATH entry).
The second option will always pick up the Python local to the project directory you're in, if any. The former (if you put it in your shell profile) will not.
I specifically want to run "python", rather subcommands for some other command, since I often I want to pass in arguments to the python interpreter itself, along with my script.
> I don't see a way to change current and global versions of python
You really shouldn't be doing this. Utilities on your host rely on the system python being stable, unchanged, pinned to a very specific python version; also with exact versions for dependencies.
For example in fedora/rhel, changing the default platform-python will prevent you from using yum/dnf to apply system updates.
It seems like that'd work as long as you restrict yourself entirely to the pip interface. Stray outside of that, and you start accumulating a real dependency on uv itself.
I'm sold. Never going back to pip/twine/requirements.txt again if I don't have to. I have several projects that all share a common wheel hosted on an internal GitLab instance, and I was able to replace like 10 lines of YAML with just "uv build" and "uv publish." Importing is quick and easy, and you can see what your core dependencies are as opposed to everything just being in one mess of a requirements.txt.
This is the most important missing feature to me as well.
I've written a lightweight replacement script to manage named central virtual envs using the same command syntax as virtualenvwrapper. Supports tab completion for zsh and bash: https://github.com/sitic/uv-virtualenvwrapper
A problem I have now, though, is when I I jump to def in my editor it no longer knows which venv to load because it's outside of the project. This somehow used to work with virtualenwrapper but I'm not sure how.
I was surprised to not see earlier laude for the amazing work at pixi.sh too. In our team we use both. UV is great when not dealing with any underlying dependencies. Pixi is our fallback if that fails.
Where I work people work on shared Linux boxes so no one has root access. Conda environments are used to basically install packages you'd normally use a package manager for. Why do you use `uv` if `pixi` is a superset?
For running scripts on my personal computer I really don't care for all the dependency management stuff. I just want a single, globally installed latest version of a library, like what pip does. I've never had the problem of an old script breaking, I guess because I just don't run that much software. These things for writing out explicit versions of everything and reinstalling libraries for every single project just add a lot of overhead, extra typing and an extra layer of complexity that I don't care for. I like just typing "pip install requests" and then being able to do "import requests" in any REPL or script for the next few years, occasionally running a Bash alias to upgrade all dependencies (which is a feature that pip incredibly still doesn't have, 14 years later).
I can see how if you've had issues with dependencies you would rave about systems that let you control down to the commit what an import statement actually means, but I like the system that requires the least amount of typing/thinking and I imagine I'm part of a silent majority.
I've been out of the python world for a while now, but I would think a global install should just be a special case of a local install, one that's shared among scripts, which is basically how npm and yarn work. I'm kind of surprised uv doesn't support something like this already. Maybe it should be up to the distro to base the global python package management on uv then?
but it's more typing. If I type 5 characters per second, making me also type "uv --system" is the same as adding 2 seconds of runtime to the actual command, except even worse because the chance of a typo goes up and typing takes energy and concentration and is annoying.
And when you run the upgrade-all-dependencies script and some dependency has a semver-major/otherwise breaking upgrade, everything depending on it breaks at once.
Hopefully one of those things isn't your backup cronjob.
Word of warning, if you use Linux you can easily break your system like this. Many distros stop you being able to modify the system packages now but that's quite a recent development. You should look into doing user-level installs instead. Don't know about Mac. If you use Windows then you do you. It might break eventually but probably not irrecoverably.
They’re still figuring it out and I believe that they’re also still pre-revenue. With that being said, if you’re able to capture a large part of the python ecosystem, I imagine business ideas won’t be too hard to come by!
Worth using just for 'uv pip' as a replacement for pip on the grounds of being much much faster. I've completely switched to uv and I can't imagine ever going back to pip for this and lots of other outstanding features.
This has the same issue as so many package managers for Python, namely, it doesn't provide a way for --no-binary to remain sticky.
There are times when you do NOT want the wheel version to be installed (which is what --no-binary implements in pip), but so many package managers including uv don't provide that core, basic functionality. At least for those that do use pip behind the scenes, like pipenv, one can still use the PIP_NO_BINARY environment variable to ensure this.
It helps, that's for sure. But this sort of knowledge should not exist in the environment in any case. It should be part of the canonical package list, and not hidden away elsewhere. The whole idea of a dependency manager should be a centralized and consistent way to install everything you need and not be dependent on what values may or may not exist as a env-var.
Can you elaborate on the reasons why a package would need to declare that its dependencies must be installed from source rather than from pre-built binaries? I'm having trouble imagining a scenario where that capability would be used as anything other than a workaround for a deeper problem with how your dependencies are packaged.
Let's say you're developing a python package that relies on a native shared library. If you distribute it as a pre-compiled wheel, you're locking your users into the exact version and compilation flags, options, etc that you choose. So your question effectively boils down to "why do people need to compile their own software?".
Many reasons: you need more control, specialized hardware, testing newer versions of the library, maintaining an internal fork of a library, security, performance, the dev team maintains both the native library and python package and needs to keep them independent, or simply preference for dynamic linking against system libraries to avoid duplication.
> So your question effectively boils down to "why do people need to compile their own software?".
No, the question is why a package would need to decide for its users that the package and its dependencies must be installed the gentoo way. That's quite obviously different from why an end user would decide to install from source despite the availability of binary packages.
There are some wheels, for example, 'lxml' that bundle in their binary possibly incompatible external libraries, or older libraries than what you would like. This can cause library conflicts.
I'm in ML-land. I thought we were all hopelessly tied to conda. But I moved all my own projects to uv effortlesly and have never looked back. Now first thing I do when pulling down another reseacher's code is add a pyproject toml (if they don't have one), `uv add -r`, and `uv run` off into the sunset. I especially like how good uv is with non-pypi-published dependencies: GitHub, dumb folders, internal repos, etc.
uv is almost perfect. my only pet peeve is updating dependencies. sometimes I just want to go "uv, bump all my dependencies to the as latest version as possible while respecting their constraints". I still haven't found an elegant way to do this, but I have written a script that parses pyproject.toml, removes the deps, and invokes `uv add --upgrade` with them.
other than that, it's invaluable to me, with the best features being uvx and PEP 723
Unfortunately, no. Only `uv.lock` gets updated, but the dependencies in `pyproject.toml` are frozen at their original constraints.
What I want is, if my project depends on `package1==0.4.0` and there are new versions of package1, for uv to try install the newer version. and to do that for a) all the deps, simultaneously, b) without me explicitly stating the dependencies in the command line since they're already written in the pyproject.toml. an `uv refresh` of sorts
If you specify your constraints in pyproject.toml like this: `package1==0.4.0`; then that is the latest (and only) version satisfying your constraints. Not upgrading is expected behavior, because upgrading would violate constraints.
pyproject.toml’s dependency list specifies compatibility: we expect the program to run with versions that satisfy constraints.
If you want to specify an exact version as a validated configuration for a reproducible build with guaranteed functionality, well, that’s what the lock file is for.
In serious projects, I usually write that dependency section by hand so that I can specify the constraints that match my needs (e.g., what is the earliest version receiving security patches or the earliest version with the functionality I need?). In unserious projects, I’ll leave the constraints off entirely until a breakage is discovered in practice.
If `uv` is adding things with `==` constraints, that’s why upgrades are not occurring, but the solution is to relax the constraints to indicate where you are okay with upgrades happening.
> ... the solution is to relax the constraints to indicate where you are okay with upgrades happening.
Yeah, that's pretty much what I've been doing with my workaround script. And btw most of my projects are deeply unserious, and I do understand why one should not do that in any other scenario.
Why not depend on package1>=0.4.0 rather than specifying an explicit version? Then uv will upgrade it to the latest version.
pyproject.toml is meant to encode the actual constraints for when your app will function correctly, not hardcode exact versions, which is what the lockfile is for.
Because then you don't get to use the new features in 0.5.0.
Though I do think with Python in particular it's probably better to manually upgrade when needed, rather than opportunistically require the latest, because Python can't handle two versions of the same package in one venv.
Actually, it won't work. I tried it and running `uv run script.py` just reinstalls the deps back... which is, I admit, the behaviour I expect and want as a user.
Astral's tools are so fast in general, when you first try them out, you double check what went wrong because you are sure nothing happened.
Same with uv. They are doing very nice tricks, like sending Range requests to only download the metadata part from the ZIP file from PyPI, resolve them in memory and only after that downloading the packages. No other package manager does this kind of crazy optimization.
There is nothing Rust specific as far as I’m aware, other than the developers are comfortable with the language. They have a very fast dependency resolver/algo, but presumably you could implement that in C or Zig or whatever. It is much faster than if it was implemented in Python or another interpreted language though. Ruff is way, way faster than the equivalent Python tools. I like Rust for the type safety and tooling so that may slow them to build faster?
I've started on a CMake integration to take multiple pyproject.toml files (in a monorepo or such), turn them into an editable workspace venv for development (with dependencies), and then on `make install` package the workspace up into a venv (at some point I'll add a package target...). The lock file ends up checked in and updated whenever you build w/ changes to any of the pyproject files.
This is very very useful for robotics or other places where you might end up with a big list of pip installed packages in some Dockerfile, want them pinned, with just one version of each library across the whole space, but don't want to necessarily do the song and dance of a full python package. Because uv is fast it just kind of works and checking the projects doesn't take any time at all compared to the other build work being done.
I've been happily using it for a long time, and rye before that.
Just today I set it up on 20 PCs in a computer lab that doesn't have internet, along with vs code and some main packages. Just downloaded the files, made a powershell script and it's all working great with Jupyter etc... Now to get kids to be interested in it...
Its sad that pyrhon only tooling is aparently so imcapable that you have to write it in a compiled language.
After that many years of optimization pure python seems still to be wishfull thinking. It's ai/mk success is also only as a shim language around library calls.
What was super unclear was how I develop locally with uv. Figuring out I needed `aider sync --extra` then `aider run --projrct /opt/aider aider` to run was a lot of bumbling in the dark. I still struggle to find good references for everyday running projects use with uv.
It was amazing though. There were so many pyproject and other concerns that it just knew how to do. I kept assuming I was going to have to do a lot more steps.
Never thought to see if Jane Street had a YouTube or if they had programming related stuff but here we are thanks to you! Makes sense they do, and that it’s python related.
I continue to be puzzled why sometime running uvx (uv tool run) will redownload all dependencies even though it just downloaded them for another tool. Downloading torch 15 times per day gets old even on 500 mbits
I haven't been able to find any kind of rhyme or rhythm to it, so I don't know how to explain when it happens or how to better debug it for a bug report.
Initially, I used `uv tool run <name>`, but later discovered `uv tool install <name>`. The `install` command downloads dependencies once and caches them for future use. It is similar to how `pipx install <name>` works.
It's not just the programs that are fast, but the company's response. I reported an issue on Discord and the founder fixed it right away, collaborating with me on the PR I created.
Switched all my python stuff to poetry several years ago. Still kinda tired and shagged out from that endeavour. Yeah yeah, ok ok, if y'all keep raving THIS much about uv, I guess it must be pretty good ... still, sigh, yet another One True (tm) python package management solution. It's on my todo list, I'll get around to it, I reckon within the next 12 months, currently I have no pressing need to move away from poetry.
You don't know what you're missing if you haven't spent 5 minutes trying out uv.
If you're intentionally not trying it simply because you don't want to get addicted like everyone else clearly is, I could see that as a valid reason to never try it in the first place.
I usually avoid jumping on bandwagons, so I've always stuck with vanilla pip/venv, but at this point it really is clear to me that uv really is the "One True (tm) python package management solution", and probably will be for the next 10 years.
uv is indeed fast and easy. But I'm also finding that the maintainers' efforts to make it work like Cargo mean it can be more difficult to use in more complex project structures. As sensible as Rust's project management ethos is, you're never going to escape Python's underlying design in a Python project, and friction between the two philosophies may not be avoidable.
One possible alternative is Pants. It's also written in Rust for performance, but has more flexibility baked into the design.
Pants and uv are two very different tools, with very different use-cases.
uv is basically souped-up pip.
Pants is an entire build/tooling system, analogous to something like Bazel. It can handle multiple dependency trees, multiple types of source code, building and packaging, even running tests.
uv is so much more than just a souped-up pip. It can also, for example, do building and packaging.
It also has workspaces and subprojects features that are sneaking toward something like the full-fledged multi-artifact project support you get in Pants (or Poetry, with plugins). Except that their decision that the entire project must have one and only one global dependency solution means that there's no escape hatch if you ever end up in a dependency hell situation. Which is fairly common in Python for a variety of reasons. And, even when it does work, subprojects all sharing a single virtualenvironment means it's really easy to accidentally create an undeclared dependency and never notice at development time because it's already in the venv due to a sister project already having declared it. So if you have a project that builds multiple packages, you're on your own for implementing a solution to ensure packages with sane dependency specs get published to the archive.
That's a major reason why Python development culture decided to go with many project-specific virtualenvironments instead of a single global one like what uv is trying to drive toward. And it's true that it still allows you to do independent projects that reference each other at development time using path dependencies. But, unlike some of its more mature alternatives, it doesn't give you any help with replacing those path dependencies with named, version dependencies at build time. So if you have a project that builds multiple packages, you're on your own for implementing a solution to ensure packages with sane dependency specs get published to the archive.
I have no understanding of python really but very glad I settled on uv. It’s such an improvement user wise over poetry or pip and it’s extremely fast too.
I have one complaint though, I want ./src to be the root of my python packages such that
> from models.attention import Attention
Works if I have a directory called models with a file called attention.py in it (and __init__.py) etc. The only way this seems to work correctly is if I set PYTHONPATH=./src
Surely the environment manager could set this up for me? Am I just doing it wrong?
src layout _should_ work. The pyproject.toml needs to live next to src, not inside of it. You might need to `uv run python` in order to pick up the right PYTHONPATH?
ChatGPT says two things to this “uv is a package manager not a file runner” um uv run seems to be one of the best parts of uv… and that “You’re doing things correctly you are just hitting one of Python’s most notorious pain points: managing import paths cleanly”.
I have read a few tickets saying uv won’t support this so everyone running my project will have to read the README first to get anything to run. Terrible UX.
I love uv. I am a convert, I use it for everything. One area I find it incredible for: deployment. If I have to launch a docker container and install a lot of python packages it saves so much time and compute.
I also appreciate that it handles most package conflicts and it constantly maintains the list of packages as you move. I have gotten myself into a hole or two now with packages and dependencies, I can usually solve it by deleting venv an just using uv to reinstall.
I’ve been writing Python for over a decade and was still using the Pyenv Virtualenv (dated I know). Lately I had been trying to switch over to Poetry. Somehow it never works smoothly for me.
Then I looked at uv and oh boy is it better. Everything just works and is blazing fast. Kudos to the team.
Started out with conda mainly because it played nicely with jupyter, but after switching to fish and vscode on mac, managing conda environments for Jupyter kernels turned into a pain. Honestly, half my personal projects procrastination comes down to fighting with conda setup. Been hearing a lot of buzz about UV, and this post finally convinced me to give it a shot on my next project.
Here’s my #1 complaint about uv: I’m new to the python ecosystem. I don’t know anything about pip or the existing tools. I would love for uv to at least have documentation and a user guide that doesn’t assume knowledge of the old bad tools that uv replaces.
Perhaps uv will continue its ascendancy and get there naturally. But I’d like to see uv be a little more aggressive with “uv native” workflows. If that makes sense.
That exists! [0] But IMHO the guides should be linked in big text as the first thing on the homepage. Right now you have to read through a bunch of meaningless comparative bullet points and reassurances then click a few times to get to the guides. If it weren’t for everyone telling me I need to switch, I might not have had the patience to find them.
I switched to uv in an open source project I maintain as our builds were taking dozens of minutes to resolve dependencies using Poetry. The migration itself was pretty easy as both tools use pyproject.toml, so it was a matter of changing headers and minor adjustments.
After the switch, the same dependency resolution was done in seconds. This tool single-handedly made iteration possible again.
I'm setting up a new dev workstation and I'm trying NixOS and for Python work I'm going to try living with "uv" for the bulk of my Python work. An experiment I've just started, but I've been using uv for ~6 months and it's been spectacular and I think it's going to be able to live up to my expectations.
I don't use uv for my (django) projects because my pip/venv workflow is good enough and had never failed me in 15 years.
However I really like installing uv globally on my Windows systems and then using uvx to run stuff without caring about venvs and putting stuff to path.
I’ve been writing an equivalent of uv for the R language at work, and it’s quite daunting / unwieldy. I was feeling bad about it, but then I reminded myself that uv has hundreds of contributors and my project only has one.
The biggest speed-up in uv comes from the way it uses caching and hard links. When you install a package into a virtual environment uv uses a hard link to a previously cached version rather than copying files.
Using Rust is responsible for a lot of speed gains too, but I believe it's the hard linking trick (which could be implemented in any language) that's the biggest win.
There's also the assumptions around wheels having consistent metadata (which isn't required in older versions of the specs), so as long as you don't need to support the long tail of packages, then those assumptions are fine (it's worth noting uv weren't the first to make these assumptions, poetry did this too, but pip does not).
I don't want to charge people money to use our tools, and I don't want to create an incentive structure whereby our open source offerings are competing with any commercial offerings (which is what you see with a lost of hosted-open-source-SaaS business models).
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 (we may not do this, but it's helpful to have a concrete example of the strategy) would be something like an enterprise-focused private package registry. A lot of big companies use uv. We spend time talking to them. They all spend money on private package registries, and have issues with them. We could build a private registry that integrates well with uv, and sell it to those companies. [...]
But the core of what I want to do is this: build great tools, hopefully people like them, hopefully they grow, hopefully companies adopt them; then sell software to those companies that represents the natural next thing they need when building with Python. Hopefully we can build something better than the alternatives by playing well with our OSS, and hopefully we are the natural choice if they're already using our OSS.
They are hijacking the entire python's ecosystem in a very smart way, that's all. At some point we, probably, will find us vendor locked-in, just because the initial offer was so appealing. Take a closer look at it: package manager, formatter/linter, types, lsp. What's left before it will poke cpython one way or another? Maybe cloud-based IDE, some interesting WASM relationship (but RustPython is not there yet, they just don't have enough money). Otherwise, Astral is on a pretty straightforward way to `touchdown` in a few years. It's both, the blessing, and the curse.
Let's be honest, all tries to bring a cpython alternative failed (niche boosters like PyPy is a separate story, but it's not up-to-date, and not entirely exact). For some reason, people think that 1:1 compatibility is not critical and too costly to pursue (hello, all LLVM-based compilers). I think, it's doable and there's a solid way to solve it. What if Astral thinks so too?
Honestly... I don't really care. If in 5 years they turn around and try to charge for uv we'll still be in a much better place than if we'd all stuck with the catastrofuck that is pip.
Anaconda makes on the order of 100M a year “solving” data science package management. I would argue it has a significantly worse product, attacking a much smaller part of the ecosystem.
It seems easy to imagine Astral following a similar path and making a significant amount of money in the process.
Anaconda isn't free. I don't want to pay per-seat fees for slightly improved versions of open source tools which is why I'm very skeptical of Astral and uv.
One day they're going to tell me I have to pay $10/month per user and add a bunch of features I really don't need just because nobody wants to prioritize the speed of pip.
And most of that fee isn't going to go towards engineers maintaining "pip but faster", it's going to fund a bunch of engineers building new things I probably don't want to use, but once you have a company and paying subscribers, you have to have developers actively doing things to justify the cost.
I don't think Anaconda is targeting a smaller part of the ecosystem, rather a different but overlapping ecosystem (remember Anaconda does more than just Python, and there's more to data science than Python). Astral I suspect won't try to move outside of Python, which limits what its tools can be used for.
In theory, Anaconda solves the next higher level of the Python package management nightmare, namely knowing what versions are compatible with each other. But that could presumably be done on top of uv.
Apparently VC(1) so far, I'd assume there will be LTS support contracts and tailored enterprise features down the line; for the moment I'd assume it's just a bunch of talented devs fixing problems they've been tired off / see as long term existencial threats to the python ecosystems.
The Astral projects are all great. I hope the company finds a revenue stream in the future with some hosted something or the other because these tools are so useful I don't want them to become pay-for etc.
I still mostly use pip (largely due to habit) but uv has been incredible for speeding up ci workflows! We shaved off a full minute of CI for a python package I maintain when we started using uv.
I love uv, not just for local development, but it also makes it WAY easier to manage python environments you setup for running python workers / user code in the cloud.
I moved from pip to poetry a while back, but was introduced to uv recently while working on MCP stuff. I now use uv for everything and haven’t looked back.
Moved to this and have no need for anything else, especially since uv pip install whatever works with uv and is faster than pip (though I usually use uv add).
uv is great! One surprise I recently ran into, though, was that `uv run path/to/foo.py` doesn't look up the nearest pyproject.toml / .venv. Similarly, a script with a shebang `#!/usr/bin/env -S uv run` only executes correctly when executed from within the folder containing the .venv. Am I the only one finding this weird?
I think the thought is to be obviously deterministic. It’s not exactly trivial to make those decisions for users since uv is not locked to its .venv - you may want the env, the system python, a pyenv, a brew python, or whatever variation can co-exist seamlessly.
I’d be surprised if there wasn’t an env var for it though.
uv is great and we have largely switched over to using it. however, since it doesn't handle compiled binaries, we have to continue using conda for environments where those are needed (i.e., mkl).
Seems a lot of people like this and are happy about it, but I for one am tired of the proliferation of python package management tools.
Many languages have many package management tools but most languages there are one or two really popular ones.
For python you just have to memorize this basically:
- Does the project have a setup.py? if so, first run several other commands before you can run it. python -m venv .venv && source .venv/bin/activate && pip install -e .
- else does it have a requirements.txt? if so python -m venv .venv && source .venv/bin/activate && pip install -r requirements.txt
- else does it have a pyproject.toml? if so poetry install and then prefix all commands with poetry run
- else does it have a pipfile? pipenv install and then prefix all commands with pipenv run
- else does it have an environment.yml? if so conda env create -f environment.yml and then look inside the file and conda activate <environment_name>
- else I have not had to learn the rules for uv yet
Thank goodness these days I just open up a cursor tab and say "get this project running"
> - else does it have a pyproject.toml? if so poetry install and then prefix all commands with poetry run
That's not even correct. Not all projects with pyproject.toml use poetry (but poetry will handle everything with a pyproject.toml)
Just try uv first. `uv pip install .` should work in a large majority of cases.
pipenv is on the way out. bare `setup.py` is on the way out. `pyproject.toml` is the present and future, and the nice thing about it is it is self-describing in the tooling used to package.
I wish setup.py was actually on the way out, but sadly, it's the only straightforward way to handle packages that use cython or interop. In these cases, libraries use setup.py to compile the dll/so/dylib at install time. Naturally this is a bit of nightmare fuel since installing gets arbitrary code execution privileges but there's no real standard for privileges in python package installs.
I use the two together with no issues. Micromamba works great for conda dependencies that’s not on PyPi (for many reasons), and everything else I go for uv.
I can't reproduce this, possibly it's another environment error or the problem has been fixed in the versions I am using. (uv 0.7.11, zsh 5.9, python 3.13.5, MacOS 15.3.1)
Just a few months back I said I would never use uv. I was already used to venv and pip. No need for another tool I thought.
I now use uv for everything Python. The reason for the switch was a shared server where I did not have root and there were all sorts of broken packages/drivers and I needed pytorch. Nothing was working and pip was taking ages. Each user had 10GB of storage allocated and pip's cache was taking up a ton of space & not letting me change the location properly. Switched to uv and everything just worked
If you're still holding out, really just spend 5 minutes trying it out, you won't regret it.
For me, the big key was: uv is so much easier to explain and especially use - especially for people who sometimes script something in python and don't do this daily.
pip + config file + venv requires you to remember ~2 steps to get the right venv - create one and install stuff into it, and for each test run, script execution and such, you need to remember a weird shebang-format, or to activate the venv. And the error messages don't help. I don't think they could help, as this setup is not standardized or blessed. You just have to beat a connection of "Import Errors" to venvs into your brain.
It's workable, but teaching this to people unfamiliar with it has reminded me how.. squirrely the whole tooling can be, for a better word.
Now, team members need to remember "uv run", "uv add" and "uv sync". It makes the whole thing so much easier and less intimidating to them.
This is similar to things like poetry (poetry run ..., poetry add, poetry install), but yeah. Does look nice.
The absolute killer feature for me of uv is that it's still compatible with all of my old venv-based workflows. Just run `uv venv`.
Unfortunately uWSGI (one of the most important libraries) is fundamentally incompatible with uv. had to roll back all my apps that use custom uWSGI for that reason.
Could you explain how? Does it do something funky in the venv that uwsgi doesn't understand?
I wonder how it compares with something more green generalist like "mise", to which I migrated after using "ASDF" for some time.
Similarly to the sibling I also use both. I let mise manage my uv version (and other tools) and let uv handle Python + PyPI Packages for me. Works great!
There's also some additional integration which I haven't tried yet: https://mise.jdx.dev/mise-cookbook/python.html#mise-uv
Think of uv more as like npm or other thing like that. The new Python pyproject.toml is similar package.json. It defines the project description, list of dependencies, and other hooks. Uv is a package/project tool using pyproject.toml. It is easy to manage dependencies, build and publish to PyPi, add hooks to run tests, linters, or whatever, again much like package.json. It also manages the virtualenv automatically, though you can manage it yourself.
Thanks for mentioning mise. I'm more interested in it for the task running feature. I couldn't figure out how to have shell scripts in Justfile, so I gave up on it.
Out of interest - What wasn't working for you?
I use Just inside (and outside) mise, almost exclusively with embedded shell/python scripts. Have used mise tasks a little, but didn't meaningfully add enough for me
Or do you mean justfiles with shebangs?
Either way, curious what problem you hit (and may be able to unblock ya)
From what I remember, the task consisted of several lines of shell and just would just choke at some point. I think it was because I used a new line to visually separate code. I'm not sure, I don't remember exactly as it was some time ago and I just gave up.
Ah fair enough, thanks!
possibly patched now but definitely going to try breaking it now before moving any more stuff into it
I use both! uv installed globally with mise, and uv tools can then be managed via “miss use -g pipx:foo”.
Same! I recently set up some dev environments with mise. My old ones are still using poetry, the new ones have uv. uv is incredibly snappy. It's like night and day!
What has been holding me back on uv is my experience with Ruff. Ruff claims "Drop-in parity with Flake8, isort, and Black", but that is simply not true. At least for isort, Ruff only re-implemented what they wanted and then ask you to use Ruff to call out to the old isort tool if there's a feature or setting that's missing in the Ruff re-implementation. So what's the point? Ruff just partially re-implemented many existing different tools and added some new ones. So using Ruff actually increases the amount of tools, yet again, you're having to use because of this and it also not doing everything that Pylint does.
For moving to uv, I haven't heard a good story for what uv provides over Poetry rather than "is fast". The only unique thing that I am currently aware of is that uv can install Python itself, which gets rid of tools like Pyenv. I'm interested because of that, but "is fast" isn't enough of a reason.
My experience is that ruff reimplemented 99% of the most popular features of black, isort, flake8, pylint, etc. and then added 10000% more features on top, that feels like a fair tradeoff to me.
I've converted multiple large Python codebases to ruff, and each time I just configure ruff as close to the previous tools as possible, then reformat the entire codebase with ruff and remove all the previous tools. The speed increase when linting alone is worth the minor formatting changes to me.
If you really insist on keeping isort's sorting then you could at least replace black and pylint, which would reduce the total number of tools by one.
I think you're underestimating how "fast" uv really is. It's night and day different. The sentence that got me to switch and I echo still was: "uv is so fast, I sometimes don't even think it's actually doing anything". The first time I used it, I had to actually double check if uv actually was working and installing the packages that it was saying it was. There's a lot of reasons I've heard not to use it that kind of (not really) make sense (you don't want another package manager, you're happy with just base pip, etc). But if you're already using poetry, you really should just switch. uv is just poetry but without any waiting.
Even your most complicated projects should be able to switch within a day. The only reason I took any time was having to restructure my docker containers to work with uv instead of poetry. And that's mostly with my inexperience with docker, not because uv is complicated.
> Each user had 10GB of storage allocated and pip's cache was taking up a ton of space & not letting me change the location properly. Switched to uv and everything just worked
Is it better about storage use? (And if so, how? Is it just good at sharing what can be shared?)
uv hardlinks identical packages, so adding virtual envs takes up very little space.
Unless you cross mount points, which uv will helpfully warn about.
In those situations, have you had a any luck using UV_LINK_MODE=symlink? I eventually had to resort to `copy` mode because it seemed the folder names (hashes) uv created in the package cache were not fully deterministic. So sometimes my cache volume would change and my Docker build would break. :\
Both pip and uv cache packages to ~/.cache. Uv lets you change it to /tmp and symlink instead of copying
Note: /tmp will take awhile to get rid of, but it's definitely on the chopping block.
I'd avoid workflows that lean on it, if anything else for security's sake.
There is a global cache for all installed packages in the user home cache dir.
similar story recently with an experimental repo that starts with "its so easy, just $uv a b c" .. under the hood it implies a lot of redundancies? but true enough it worked fine and trouble-free too, on a standard GNU-Debian-Ubuntu host
Just made the switch myself. Was far easier and smoother than I expected.
I came here to comment that I don’t see any reason to bother - that’s for the comment, I will try it now!
> Just a few months back I said I would never use uv. I was already used to venv and pip. No need for another tool I thought
Really? :)
requirements.txt is just hell and torture. If you've ever used modern project/dependency management tools like uv, Poetry, PDM, you'll never go back to pip+requirements.txt. It's crazy and a mess.
uv is super fast and a great tool, but still has roughnesses and bugs.
Pip-tools+requirements.txt helped me survive the past few years. I also never thought I needed uv, but after all the talk about it I gave it a spin and never want back. It’s just so blazing fast en convenient.
We use uv to compile requirements.txt from pyproject.toml to get the locked versions.
What for? Support legacy CI/CD pipelines or something like that? uv.lock already contains locked versions of all dependencies plus a lot of other needed metadata.
> What for? Support legacy CI/CD pipelines
Yes. Azure, for instance, looks for requirements.txt if you deploy a web app to Azure App Service.
If you’re doing a code-based deployment, it works really well. Push to GitHub, it deploys.
You can of course do a container-based deployment to Azure App Service and I’d assume that will work with uv.
"legacy CI/CD pipelines"
Damn I'm getting old
pip also works with pyproject.toml. Sticking with requirements.txt is a self-imposed constraint.
[dead]
I remember using pip and venv back in like 2009. Last time I checked, maybe 5 or 10 years ago, the recommendation of the community was generally to just use Docker instead of all these tools. Did that not catch on?
The advice seems to change every year. For a while it was venv, then pipenv, poetry, docker, and now uv. Maybe the ecosystem will settle on that but who knows.
I mean docker is orthogonal to package manager. It makes it easier to deploy but none of the other thing also have managers do are relevant.
Docker was always a workaround to Python not having a non-awful dependency manager. uv is that non-awful dependency manager, and I expect in the long term it will reduce the use of Docker.
Docker solves a different problem. Docker is a way to basically ship your whole OS off to another machine. You still have to have a way to install the right version of python and all the python libraries you need inside the Docker container, and uv is great for this.
Secondly Docker only solves a subset of problems. It's fine if you're developing a server that you will be deploying somewhere. It's inconvenient if you're developing an end user application, and it's completely useless if you're developing a library you want people to be able to install.
I love uv. The one gotcha I'll warn people about is: don't touch uvx. I've lost an embarrassing number of hours or days trying to figure out why nothing works properly or makes sense when I tried to run things with uvx. I guess I understand why it's there, but I think it's a built-in foot-gun and not well documented. But if you stay away from it, things work great.
I suppose you didn't accompany the command `uvx` with the necessary `--with=` arguments[1] for each dependency.
[1] https://docs.astral.sh/uv/guides/tools/#commands-with-plugin...
What issues are you having with uvx? It replaces tools like pipx that set up implicit venvs to run specific tools. Works great for me.
uvx is fine. I use it to run executable packages all the time. What is your problem with it?
The first time I used `uv`, I was sure that I had made a mistake or typed something wrong because the process finished so much more quickly than anything I had ever experienced with `pip`.
I've sometimes had uv take up to 200ms to install packages, so you could feel a slight delay between pressing enter and the next shell prompt
You don't have that problem with Poetry. You go make a cup of coffee for a couple minutes, and it's usually done when you come back.
It's funny when the exact same thing was probably said about pipenv and poetry
I've had poetry sit there for minutes resolving dependencies at a previous company. I thought something was broken... it probably was... but it did eventually complete.
I've had poetry take so long that I've just deleted the whole cache/poetry list and reinstalled the packages rather than wait for poetry to figure out whatever dependency hell it found itself in. I switched to uv the minute i tried it out
lol. Yes, I've done the same thing. haha. I couldn't do that at a previous company cause they were pretty anal about mass updates to poetry.lock
Same here lol! The experience is so smooth it doesn't feel like python
That makes sense, because it's Rust. :)
I just had this same experience last week, and was certain it wasn’t working correctly as well. I’m a convert.
Likewise. I was skeptical, then I tried it and won’t go back.
uv and ruff are a great counterexample to all those people who say "never reinvent the wheel". Don't ever do it just for the sake of doing it, but if you have focused goals you can sometimes produce a product that's an order of magnitude better.
I believe most of the time this phrase is said to an inexperienced artisan who has no idea how the current system works, what's the shortcoming of it, and how to improve upon it. Think of an undergraduate student who tries to solve the Goldbach conjecture. Usually what ended up is either he fails to reinvent the wheel, or reinvent the exact same wheel, which has no value. The phrase certainly does not apply to professionals.
Even then, you know what's a good way to learn about how the current system works etc, maybe even the best way? I've got many failed projects behind me, and 0 regrets.
They didn't reinvent the wheel, "just" replaced all the wood with more durable materials to make it handle rotation at 10 times the speed
A big part of the "magic" is that there is a team of paid professionals maintaining and improving it. That's more important than it being written in Rust. If uv were forked it would devolve to the level of pip over time.
I'd be curious to know exactly what changed. Python -> Rust won't make network downloads faster nor file I/O faster. My naive guess is that all the speed comes from choosing better algorithms and/or parallelizing things. Not from Python vs Rust (though if it's hard to parallelize in Python and easy in rust that would certainly make a difference)
I've translated code from Ruby to Python, and other code from Rust to Python.
Rust's speed advantages typically come from one of a few places:
1. Fast start-up times, thanks to pre-compiled native binaries.
2. Large amounts of CPU-level concurrency with many fewer bugs. I'm willing to do ridiculous threading tricks in Rust I wouldn't dare try in C++.
3. Much lower levels of malloc/free in Rust compared to some high-level languages, especially if you're willing to work a little for it. Calling malloc in a multithreaded system is basically like watching the Millennium Falcon's hyperdrive fail. Also, Rust encourages abusing the stack to a ridiculous degree, which further reduces allocation. It's hard to "invisibly" call malloc in Rust, even compared to a language like C++.
4. For better or worse, Rust exposes a lot of the machinery behind memory layout and passing references. This means there's a permanent "Rust tax" where you ask yourself "Do I pass this by value or reference? Who owns this, and who just borrows is?" But the payoff for that work is good memory locality.
So if you put in a modest amount of effort, it's fairly easy to make Rust run surprisingly fast. It's not an absolute guarantee, and there are couple of traps for the unwary (like accidentally forgetting to buffer I/O, or benchmarking debug binaries).
NVMe hungers, keeping it fed is hard work. Doing some serial read, decompress, checksum, write loop will leave if starved (QD<1) whenever you're doing anything but the last step. Disk IO isn't async unless you use io_uring (well ok, writeback caches can be). So threads are almost a must to keep NVMe busy. Conversely, waiting for blocking IO (e.g. directory enumeration) will keep your CPU starved. Here too the answer is more threads.
There is a talk about it from one of the authors here: https://www.youtube.com/watch?v=gSKTfG1GXYQ
tl;dw Rust, a fast SAT solver, micro-optimisation of key components, caching, and hardlinks/CoW.
Extremely interesting presentation from Charlie Marsh about all the optimizations https://youtu.be/gSKTfG1GXYQ?si=CTc2EwQptMmKxBwG
Thanks. So from the video the biggest wins were
1. they way get the metadata for a package.
packages are in zip files. zip files have their TOC at the end. So, instead of downloading the entire zip they just get the end of the file, read the TOC, then from that download just the metadata part
I've written that code before for my own projects.
2. They cache the results of packages unzipped and then link into your environment
This means there's no files being copied on the 2nd install. Just links.
Both of those are huge time wins that would be possible in any language.
3. They store their metadata as a memory dump
So, on loading there is nothing to parse.
Admittedly this is hard (impossible?) in many languages. Certainly not possible in Python and JavaScript. You could load binary data but it won't be useful without copying it into native numbers/strings/ints/floats/doubles etc...
I've done this in game engines to reduce load times in C/C++ and to save memory.
It'd be interesting to write some benchmarks for the first 2. The 3rd is a win but I suspect the first 2 are 95% of the speedup.
From just my observations they basically parallelized the install sequence instead of having it be sequential (among many other optimizations most likely)
It became a bit of a meme, especially in the web development space, that all programs are always waiting on external resources like networks, databases, disks, etc., and so scripting languages being slower than other languages doesn't matter and they'll always be as fast as non-scripting languages.
Even on a single core, this turns out to be simply false. It isn't that hard to either A: be doing enough actual computation that faster languages are in fact perceptibly faster, even, yes, in a web page handler or other such supposedly-blocked computation or B: without realizing it, have stacked up so many expensive abstractions on top of each other in your scripting language that you're multiplying the off-the-top 40x-ish slower with another set of multiplicative penalties that can take you into effectively arbitrarily-slower computations.
If you're never profiled a mature scripting language program, it's worth your time. Especially if nobody on your team has ever profiled it before. It can be an eye-opener.
Then it turns out that for historical path reasons, dynamic scripting languages are also really bad at multithreading and using multiple cores, and if you can write a program that can leverage that you can just blow away the dynamic scripting languages. It's not even hard... it pretty much just happens.
(I say historical path reasons because I don't think an inability to multithread is intrinsic to the dynamic scripting languages. It's just they all came out in an era when they could assume single core, it got ingrained into them for a couple of decades, and the reality is, it's never going to come fully out. I think someone could build a new dynamic language that threaded properly from the beginning, though.)
You really can see big gains just taking a dynamic scripting language program and turning it into a compiled language with no major changes to the algorithms. The 40x-ish penalty off the top is often in practice an underestimate, because that number is generally from highly optimized benchmarks in which the dynamic language implementation is highly tuned to avoid expensive operations; real code that takes advantage of all the conveniences and indirection and such can have even larger gaps.
This is not to say that dynamic scripting languages are bad. Performance is not the only thing that matters. They are quite obviously fast enough for a wide variety of tasks, by the strongest possible proof of that statement. That said, I think it is the case that there are a lot of programmers who have no idea how much performance they are losing in dynamic scripting languages, which can result in suboptimal engineering decisions. It is completely possible to replace a dynamic scripting language program with a compiled one and possibly see 100x+ performance improvements on very realistic code, before adding in multithreading. It is hard for that not to manifest in some sort of user experience improvement. My pitch here is not to give up dynamic scripting languages, but to have a more realistic view of the programming language landscape as a whole.
> Then it turns out that for historical path reasons, dynamic scripting languages are also really bad at multithreading and using multiple cores...
What would a dynamic scripting language look like that wasn't subject to this limitation? Any examples? I don't know of contenders in this design space--- I am not up on it.
The big difference from Python is probably having to use a real tracing GC instead of automatic reference counting. For a single-threaded program, refcounts are beneficial in multiple ways, being fairly cheap, having a smooth performance profile, maintaining low resident set size, and providing deterministic freeing.
But because of the way cache coherency for shared, mutated memory works, parallel refcounting is slow as molasses and will always remain so.
I think Ruby has always used a tracing GC, but it also still has a GIL for some reason?
It would look pretty much the same. It would just have been written to be multithreaded from the beginning, and lack the long list of restrictions and caveats and "but it doesn't work with our C extensions" and such. There wouldn't be a dozen major libraries trying to solve the problem (which, contrary to many people's intuition, is often a sign that a language lacks a good solution). This is part of why I say there's no fundamental reason this can't be done, it's just a historical accident.
There are dynamic languages that were built with concurrency in mind like Clojure. It’s also a surprisingly fast language considering it’s both dynamic and functional.
I'm not trying to suggest that you can't do faster computation in a lower-level language. But, a package manager doesn't do much computation. It mostly downloads, decompresses, and writes files. Yes, it has to solve constraints but that's not a bottleneck given most projects have at most a few 100 dependencies and not millions.
I don't know python but in JavaScript, triggering 1000 downloads in parallel is trivial. Decompressing them, like in python, is calling out to some native function. Decompressing them in parallel in JS would also be trivial (no idea about python). Writing them in parallel is also trivial.
Congratulations! You have proved that it is impossible for uv to be way, way faster than Python-based package managers!
....
Unfortunately, there seems to be a problem here.
When reality and theory conflict, reality wins.
It sounds like you've drunk the same Kool-Aide I was referring to in my post. It's not true. When you're playing with 50x-100x slowdowns, if not more, it's really quite easy to run into user-perceptible slowdowns. A lot of engineers grotesquely underestimate how slow these languages are. I suspect it may be getting worse over time due to evaporative cooling, as engineers who do understand it also tend to have one reason or another to leave the language community at some point, and I believe (though I can not prove) that as a result the dynamic scripting language communities are actually getting worse and worse at realizing how slow their languages are. They're really quite slow.
You seem to be implying rust = fast, the end. I'm implying algorithms and design choices = fast. Those decisions generally (though not always) are far more effective at speed than language choice.
I watched the video linked above on uv. They went over the optimizations. The big wins had nothing to do with rust and everything to do with design/algo choices.
You could have also done without the insults. You have no idea who I am and my experiences. I've shipped several AAA games written in C/C++ and assembly. I know how to optimize. I also know how dynamic languages work. I also know when people are making up bullshit about "it's fast because it's in rust!". No, that is not why it's fast.
I agree there are lot of big wins in uv that tools written in python could take advantage of, and ultimately I think uv is fast because they're obsessed with making it fast, which is why they chose to use rust. I don't see that same level speed obsession with the other tools.
Instead of "It's fast because it's in rust", I'd say: "It's fast because they chose to use rust for their python tool, which means they care a lot about speed."
The package resolution is a big part of it, it's effectively a constraint solver. I.e. if package A requires package B constrained between version 1.0 < X <= 2.X and Package B requires package C between... and so on and so on.
Conda rewrote their package resolver for similar reasons
The history of Python package management is clear that everyone thinks they can do a better job than the status quo.
In this case, they were right.
I would say in many cases they were right; the history of Python package management is littered with winners as well as losers.
Python package management was notoriously awful. The problem wasn't that people were trying to do things better, it was that they weren't; every new Python dependency management tool just repeated the mistakes of all the previous Python dependency management tools. uv is the first one to break the cycle (and it's probably not a coincidence that it's the first one to not be written in Python).
Poetry broke the cycle. Unified toolchain, lock file, single configuration file, full dependency graph, dev dependencies. uv is faster which is great but Poetry was a huge step in the right direction and still a good tool.
uv is purely a performance improvement, it changes nothing about the mechanics of Python environment management or packaging.
The improvements came from lots of work from the entire python build system ecosystem and consensus building.
Disagree in that uv makes switching out the underlying interpreter so straightforward. Becomes trivial to swap from say 3.11 to 3.12. The pybi idea.
Sure, other tools could handle the situation, but being baked into the tooling makes it much easier to bootstrap different configurations.
Yes, it's faster and better than pyenv, but the mechanism it's using (virtual environments) is not a uv invention.
uv does the Python ecosystem better than any other tool, but it's still the standard Python ecosystem as defined in the relevant PEPs.
Are the lock files standardised, or a uv-specific thing?
uv has both a uv-specific implementation, and support for standard PEP 751 lockfiles
Worth noting uv-specific implementation has more features than that standard PEP 751 lockfiles, so uv plans to keep using its own implementation by default. https://github.com/astral-sh/uv/issues/12584
Actually not true. One of the main differences with uv is you don't have to think about venvs any more. There's a talk about it from one of the authors at a recent PyCon here: https://www.youtube.com/watch?v=CV8KRvWKYDw (not the same talk I linked elsewhere in the thread).
How do you think uv works?
It creates a venv. Note were talking about the concept of a virtual environment here, PEP 405, not the Python module "venv".
I said you don't have to think about venvs any more. It's great that we have a standard way to implement them, but this is only necessary in the first place because of the way Python is. Now we have a tool that enforces a workflow that creates virtualenvs without you having to know about them and therefore not screwing them up with ad hoc pip installs etc.
The implementation details don't matter. uv might follow PEP 405 but it could work just as well without doing so. The point is that it doesn't give you the bunch of extra footguns that any other Python package management does.
It matters immensely that it follows PEP 405, it makes uv the implementation detail. You can swap out uv for any other project management tool or build frontend and change nothing needs to change about the development environment.
This is the entire purpose of the standards.
> You can swap out uv for any other project management tool or build frontend and change nothing needs to change about the development environment.
> This is the entire purpose of the standards.
That seems to amount to saying that the purpose of the standards is to prevent progress and ensure that the mistakes of early Python project management tools are preserved forever. (Which would explain some things about the last ~25 years of Python project management I guess). The parts of uv that follow standards aren't the parts that people are excited about.
The parts of uv that follow standards aren't the parts that people are excited about.
I disagree. Had uv not followed these standards and instead gone off and done their completely own thing, it could not function as a drop in replacement for pip and venv and wouldn't have gotten anywhere near as much traction. I can use uv personally to work on projects that officially have to support pip and venv and have it all be transparent.
There are no parts of uv that don't follow standards.
The standards have nothing to do with the last 25 years of Python project management, the most import ones (PEP 517/518) are less than 10 years old.
uv only exists because of those standards and therefore can make assumptions that earlier tools could not.
> How do you think uv works?
Dont know, dont care. It thinks about these things not me.
True, but then all software is developed for this reason.
Of course they do, this tends to happen when the history is it being hot flaming garbage.
Honestly "don't reinvent the wheel" makes absolutely no sense as a saying. We're not still all using wooden discs as wheels, we have invented much better wheels since the neolithic. Why shouldn't we do the same with software?
When asked why he had invented JSON when XML already existed, Douglas Crockford said:
The good thing about reinventing the wheel is that you can get a round one.
https://scripting.wordpress.com/2006/12/20/scripting-news-fo...
You can get a round one. Or you can make yet another wonky shaped one to add to the collection, as ended up being the case with JSON.
What makes JSON wonky?
Personally the only thing I miss from it is support for binary data - you end up having to base64 binary content which is a little messy.
Quoted keys, strict comma rules, very limited data types, are the main ones. There are a host of others if you view it through the lenses of user-read/write, and a different set of issues if you view it as a machine data interface. Trying to combine the two seems fundamentally misguided.
I consider JSON's very limited data types to be part of what makes it so good.
Lack of comments seems like a big one seeing as it's so widely used for "configuration". It's a big enough downside that VSCode and others have violated it via ad-hoc extensions of the format.
The comma rules introduce diff noise on unrelated lines.
Insert the xkcd about 15 competing standards . . .
Standards do die off, up to a point. XML is widely used but the last time I really had to edit it in anger working in DevOps/web/Python was a long time ago (10 years ago?).
At this point XML is the backbone of many important technologies that many people won't use or won't use directly anymore.
This wasn't the case circa 2010, when I doubt any dev could have really avoided XML for a bunch of years.
I do like XML, though.
Right, wheels are reinvented every few years. Compare tires of today to the ones 20 years ago and the technology and capability is very different, even though they look identical to a casual eye.
My primary vehicle has off-road capable tires that offer as much grip as a road-only tire would have 20-25 years ago, thanks to technology allowing Michelin to reinvent what a dual-purpose tire can be!
> Compare tires of today to the ones 20 years ago and the technology and capability is very different, even though they look identical to a casual eye
Can you share more about this? What has changed between tires of 2005 and 2025?
In short: Better materials and better computational models.
https://www.caranddriver.com/features/a15078050/we-drive-the...
> In the last decade, the spiciest street-legal tires have nearly surpassed the performance of a decade-old racing tire, and computer modeling is a big part of the reason
(written about 8 years ago)
I always took this saying as meaning that we don’t re-invent the concept of the wheel. For example the Boring company and Tesla hoping to reinvent the concept of the bus/train.. (iirc your car goes underground on some tracks and you get to bypass traffic and not worry about steering)
A metal wheel is still just a wheel. A faster package manager is still just a package manager.
That's not how I've ever seen it used in practice. People use it to mean "don't build a replacement for anything functional."
I also like Dan Luu's take (starting with a Joel Spolsky quote)
“Find the dependencies — and eliminate them.” When you're working on a really, really good team with great programmers, everybody else's code, frankly, is bug-infested garbage, and nobody else knows how to ship on time.
We had a similar attitude, although I'd say that we were a bit more humble. We didn't think that everyone else was producing garbage but, we also didn't assume that we couldn't produce something comparable to what we could buy for a tenth of the cost. From talking to folks at some competitors, there was a pretty big cultural difference between how we operated and how they operated. It simply didn't occur to them that they didn't have to buy into the standard American business logic that you should focus on your core competencies, that you can think through whether or not it makes sense to do something in-house on the merits of the particular thing instead of outsourcing your thinking to a pithy saying.[0]
[0] https://danluu.com/nothing-works/
I came here to (wrongly) say that wooden disks were never used as wheels, and that ot all started with spokes. Some checking showed that, in fact, the oldest known wheels have a lot of solid disks. E.g: https://en.m.wikipedia.org/wiki/Ljubljana_Marshes_Wheel
Hopefully this can disabuse others of similar mistaken memory.
Ruff is actually a good example of the danger of rewrites. They rewrote tools but not all of the parts of the tools.
> an order of magnitude better
off topic, but i wonder why that phrase gets used rather than 10x which is much shorter.
Short answer: Because the base may not be 10.
Long answer: Because if you put a number, people expect it to be accurate. If it was 6x faster, and you said 10x, people may call you out on it.
It's meant to signify a step change. Order of magnitude change = no amount of incremental changes would make up for it.
In common conversation, the multiplier can vary from 2x - 10x. In context of some algorithms, order of magnitudes can be over the delta rather than absolutes. eg: an algorithms sees 1.1x improvement over the previous 10 years. A change that shows a 1.1x improvement by itself, overshadows an an order-of-magnitude more effort.
For salaries, I've used order-of-magnitude to mean 2x. Good way to show a step change in a person's perceived value in the market.
I think of "an order of magnitude" as a log scale. It means somewhere between 3.16X and 31.6X.
yeah that's what i meant with 10x, like it's +1 on the exponent, if base is 10. but i'm guessing what others are thinking, hence the question.
The problem is that 10x appears to be a linear scale. It could mean 9.5x to 10.5x if it's supposed to have 2 significant digits. Or it could be 5x to 15x if it meant to have 1 significant digit.
good point
"10x" has been cheapened / heard enough / de facto, is a more general statement than a literal interpretation would indicate. (i.e. 10x engineer. Don't hear that much around these parts these days)
Order of magnitude faces less of that baggage, until it does :)
Would you say it faces . . . orders of magnitude less baggage?
- sounds cooler
- 10x is a meme
- what if it's 12x better
10x is too precise.
Because it's not 10x?
Because "magnitude" has cool gravitas, something in how it's pronounced. And it's not meant to be precise, it just means "a whole lot more".
5x faster is an order of magnitude bc of rounding
ruff does not support custom plugins so is useless to me
Just a warning in case others run into it: on very anemic systems (e.g.: AWS T2.micro running Windows, yes... I know...) uv will try to open too many simultaneous downloads, overloading things, resulting in timeouts.
You can use ent ENV variable UV_CONCURRENT_DOWNLOADS to limit this. In my case it needed to be 1 or 2. Anything else would cause timeouts.
An extreme case, I know, but I think that uv is too aggressive here (a download thread for every module). And should use aggregate speeds from each source server as a way of auto-tuning per-server threading.
Not extreme at all, A lot of people use the cheapest smallest VPS for their hobby work. I know I do (albeit not AWS). Thanks for sharing, hope they improve the automatic detection there.
uv can also run even a beefy linux desktop out of file descriptors for larger projects. And does not have deterministic / reproducible installs. Still needs maturity.
I adore the
And then shoving on top so I can run Python scripts easily. It's great!I built a Claude Project with special instructions just teaching it how to do this, which means it can output full scripts for me with inline dependencies based on a single prompt: https://simonwillison.net/2024/Dec/19/one-shot-python-tools/
Claude 4's training cutoff date is March 2025 though, I just checked and it turns out Claude Sonnet 4 can do this without needing any extra instructions:
Here's the output, it did the right thing with regards to those dependencies: https://claude.ai/share/57d5c886-d5d3-4a9b-901f-27a3667a8581Using your system instructions for uv for every LLM now since first seeing your post last year, thanks! It's insanely helpful just asking e.g. Claude to give me a python script for XYZ and just using "uv run". I also added:
since e.g. Cursor often gets confued because the dependencies are not installed and it doesn't know how to start the script. The last sentence is for when LLMs get confused and want to add "json" for example to the dependency array.claude sonnet typically forgets about uv script syntax in my experience. I usually find myself having to paste in the docs every time. By default it wants to use uv project syntax.
Bro I gotchu
https://github.com/shakefu/cursor-rules/blob/f4dffbbf04ce99f...
Using this trick with Marimo.io notebooks in app-mode.
Instant reactive reproducible app that can be sent to others with minimal prerequisites (only uv needs to be installed).
Such a hot combo.
It's so cool. I now habitually vibe-code little scripts that I can immediately run. So much nicer than having to manage environments and dependencies:
- https://everything.intellectronica.net/p/the-little-scripter
- https://www.youtube.com/watch?v=8LB7e2tKWoI
- https://github.com/intellectronica/ez-mcp
e: I misread your example, disregard below irrelevant pattern matching of 'uv add --script' to 'uv add' in the project context!
~~That mutates the project/env in your cwd. They have a lot in their docs, but I think you’d like run --with or uv’s PEP723 support a lot more~~
https://docs.astral.sh/uv/guides/scripts/
PEP723 support is exactly what the poster is using?
Ach, missed the --script, thanks.
What's "mydependencies"? A config file?
My apologies: "mydependencies" are packages from PyPI (https://pypi.org/), as if you did "pip install ...".
Started using this recently for personal stuff on my laptop. When you're used to pip, it's just confusingly fast. More than once I thought maybe it didn't work because it returned too quickly..
Tried uv a while ago and I was shocked by how fast and easy it is to use. There's basically no reason to use pip anymore, and if you're using only Python there's basically no reason to use conda either.
It seems to make pyenv and poetry droppable too.
and pipx.
Love UV!
Also love Ruff from the Astral team. We just cut our linting + formatting across from pylint + Black to Ruff.
Saw lint times drop from 90 seconds to < 1.5 seconds. crazy stuff.
Until the moment you will realize that ruff perform only a part of pylint checks and that very obvious mistakes can go through easily like code that can't run because of an obvious error.
Do you have an example of this?
It's in the documentation:
https://docs.astral.sh/ruff/faq/#how-does-ruffs-linter-compa...
https://github.com/astral-sh/ruff/issues/970
Which checks is it missing?
Many, but it also has many that pylint does not. I use them both, ruff first, followed by pylint.
https://docs.astral.sh/ruff/faq/#how-does-ruffs-linter-compa...
https://github.com/astral-sh/ruff/issues/970
Seems though mypy/pyright/pyre/pyrefly or the future ty could bridge the gap, if one is writing typed python?
The big question will be to know if tools will still be fast once they will have added the tests that are missing.
If anyone is looking for examples on using uv with Docker, I have Flask and Django example apps at https://github.com/nickjj/docker-flask-example and https://github.com/nickjj/docker-django-example.
It prevents uv from making a virtual environment and does some optimizations like compiling byte code once when your dependencies get installed.
It was well worth the switch. I noticed a ~10x improvement for speed compared to pip (30s to 3s to install all dependencies). Proper lock file support is nice too. Funny enough I wrote about and made a video about switching to uv about a week ago here https://nickjanetakis.com/blog/switching-pip-to-uv-in-a-dock....
this is my new fav for running small executable scripts:
I really wish that hashbang line was something way WAY easier to remember like `#!/usr/bin/env uvx`. I have to look this up every single time I do it.
Sadly hashbangs are technically limited to: 1) Support only absolute paths, making it necessary to use /usr/bin/env which is in standardized location to look up the uv binary 2) Support only a single argument (everything after the space is passed as a single arg, it's not parsed into multiple args like a shell would), making it necessary to use -S to "S"plit the arguments. It's a feature of env itself, for this very use case.
So there isn't really much to do to make it simpler.
I wasn't really referring to env. I meant change the behavior of uvx. If the first argument passed to uvx is a file path, then execute it exactly the same way as `uv --quiet run --script` does.
Or maybe create a second binary or symlink called `uvs` (short for uv script) that does the same thing.
UV is fast, like FAST. Plus, it removes the need for pyenv (for managing different Python versions) and pip for me. Plus, no need to activate env or anything, `uv run ...` automatically runs your code through the env.
It's a nice software.
> Plus, it removes the need for pyenv
I don't see a way to change current and global versions of python/venvs to run scripts, so that when I type "python" it uses that, without making an alias.
Two options other than aliases:
1. Modify your PATH:
2. Put this in an executable file called "python" early in your PATH: Those are basically what pyenv does (via a symlink and PATH entry).The second option will always pick up the Python local to the project directory you're in, if any. The former (if you put it in your shell profile) will not.
If they're your scripts (i.e. your writing/editing them) then you can declare dependencies following the PEP723 format and uv will respect that.
https://docs.astral.sh/uv/guides/scripts/#declaring-script-d...
> uv run example.py
I specifically want to run "python", rather subcommands for some other command, since I often I want to pass in arguments to the python interpreter itself, along with my script.
You can use "uv run python script.py" to the same effect.
> I don't see a way to change current and global versions of python
You really shouldn't be doing this. Utilities on your host rely on the system python being stable, unchanged, pinned to a very specific python version; also with exact versions for dependencies.
For example in fedora/rhel, changing the default platform-python will prevent you from using yum/dnf to apply system updates.
Utilities on my host that require system python won't use /usr/bin/env python.
It is a venture capital startup. If I start use uv, what's our protection against the company going rogue?
Community will fork it and move on. See the following examples:
* Redis -> redict, valkey
* elastic search -> opensearch
* terraform -> opentofu
(Probably a few more but those are the ones that come to mind when they "go rogue")
Does the current license allow it?
Yes, uv has a standard permissive open source license (Apache-2.0 OR MIT): https://github.com/astral-sh/uv/?tab=readme-ov-file#license
Apache & MIT - Yes
I brought this up when leaders at my place were discussing implementing uv. I am verging on paranoid about this kind of thing.
The answer is that we replaced all our of 'pip' with 'uv pip'. Shit goes south and we simply change it back at the cost of speed.
Why wouldn't you be able to switch back to using pip ?
It seems like that'd work as long as you restrict yourself entirely to the pip interface. Stray outside of that, and you start accumulating a real dependency on uv itself.
It's trivial to switch from uv to other projects like pdm or poetry. uv follows the standard PEPs really closely, so most things are interoperable.
If uv disappeared tomorrow, five projects would spring up to make compatible implementations of its functionality.
Because after a few years of using a non-awful tool you lose your tolerance for awfulness?
I'm sold. Never going back to pip/twine/requirements.txt again if I don't have to. I have several projects that all share a common wheel hosted on an internal GitLab instance, and I was able to replace like 10 lines of YAML with just "uv build" and "uv publish." Importing is quick and easy, and you can see what your core dependencies are as opposed to everything just being in one mess of a requirements.txt.
Still no option to store virtual envs outside projects after a year
https://github.com/astral-sh/uv/issues/1495
This is the most important missing feature to me as well.
I've written a lightweight replacement script to manage named central virtual envs using the same command syntax as virtualenvwrapper. Supports tab completion for zsh and bash: https://github.com/sitic/uv-virtualenvwrapper
Inspired by this I made a similar thing for direnv: https://github.com/georgek/dotfiles/blob/01fb53619adc7269824...
A problem I have now, though, is when I I jump to def in my editor it no longer knows which venv to load because it's outside of the project. This somehow used to work with virtualenwrapper but I'm not sure how.
The UV_PROJECT_ENVIRONMENT variable can be used as a workaround: https://docs.astral.sh/uv/concepts/projects/config/#project-...
`UV_PROJECT_ENVIRONMENT` works fine for that. That's what I use for my venv-managing Fish functions: https://gist.github.com/akx/c357bbda9facbe41f08747a2622795c4
I think http://pixi.sh is much cooler because it supports conda environments so you can install non-python packages as well (e.g, gcc).
I was surprised to not see earlier laude for the amazing work at pixi.sh too. In our team we use both. UV is great when not dealing with any underlying dependencies. Pixi is our fallback if that fails.
Where I work people work on shared Linux boxes so no one has root access. Conda environments are used to basically install packages you'd normally use a package manager for. Why do you use `uv` if `pixi` is a superset?
For running scripts on my personal computer I really don't care for all the dependency management stuff. I just want a single, globally installed latest version of a library, like what pip does. I've never had the problem of an old script breaking, I guess because I just don't run that much software. These things for writing out explicit versions of everything and reinstalling libraries for every single project just add a lot of overhead, extra typing and an extra layer of complexity that I don't care for. I like just typing "pip install requests" and then being able to do "import requests" in any REPL or script for the next few years, occasionally running a Bash alias to upgrade all dependencies (which is a feature that pip incredibly still doesn't have, 14 years later).
I can see how if you've had issues with dependencies you would rave about systems that let you control down to the commit what an import statement actually means, but I like the system that requires the least amount of typing/thinking and I imagine I'm part of a silent majority.
Python dependency management is a tragedy and it causes widespread pain, ergo the many solutions.
You are probably part of the silent majority because yes, most people have relatively simple needs for every popular tool.
I've been out of the python world for a while now, but I would think a global install should just be a special case of a local install, one that's shared among scripts, which is basically how npm and yarn work. I'm kind of surprised uv doesn't support something like this already. Maybe it should be up to the distro to base the global python package management on uv then?
I think it does support it, like this
but it's more typing. If I type 5 characters per second, making me also type "uv --system" is the same as adding 2 seconds of runtime to the actual command, except even worse because the chance of a typo goes up and typing takes energy and concentration and is annoying.If only there were a way to alias commands, eh?
And when you run the upgrade-all-dependencies script and some dependency has a semver-major/otherwise breaking upgrade, everything depending on it breaks at once.
Hopefully one of those things isn't your backup cronjob.
Word of warning, if you use Linux you can easily break your system like this. Many distros stop you being able to modify the system packages now but that's quite a recent development. You should look into doing user-level installs instead. Don't know about Mac. If you use Windows then you do you. It might break eventually but probably not irrecoverably.
What’s their motivation to build tools like uv and ruff? What’s their business model?
They’re still figuring it out and I believe that they’re also still pre-revenue. With that being said, if you’re able to capture a large part of the python ecosystem, I imagine business ideas won’t be too hard to come by!
See also this comment which has more information from an interview with their founder.
https://news.ycombinator.com/item?id=44358482
A private package registry is a pretty exciting idea! I honestly wouldn't have expected that idea from them.
If they do pursue that idea, I wonder if they'd also have a public component to it that could act as a better PyPI?
Worth using just for 'uv pip' as a replacement for pip on the grounds of being much much faster. I've completely switched to uv and I can't imagine ever going back to pip for this and lots of other outstanding features.
UV solved any issue I had getting python to run on a machine.
Just `git clone someproject`, `uv run somescript.py`, then mic drop and walk away.
`uvx` directly from the repo also works nicely
Why even `git clone`? With a properly packaged project, you can just `uvx --from=git+https://github.com/akx/bindl.git bindl`
This has the same issue as so many package managers for Python, namely, it doesn't provide a way for --no-binary to remain sticky.
There are times when you do NOT want the wheel version to be installed (which is what --no-binary implements in pip), but so many package managers including uv don't provide that core, basic functionality. At least for those that do use pip behind the scenes, like pipenv, one can still use the PIP_NO_BINARY environment variable to ensure this.
So I'll not be migrating any time soon.
Maybe not exactly what you need (sticky) but you can set UV_NO_BINARY=1
See https://docs.astral.sh/uv/reference/environment/#uv_no_binar...
It helps, that's for sure. But this sort of knowledge should not exist in the environment in any case. It should be part of the canonical package list, and not hidden away elsewhere. The whole idea of a dependency manager should be a centralized and consistent way to install everything you need and not be dependent on what values may or may not exist as a env-var.
Can you elaborate on the reasons why a package would need to declare that its dependencies must be installed from source rather than from pre-built binaries? I'm having trouble imagining a scenario where that capability would be used as anything other than a workaround for a deeper problem with how your dependencies are packaged.
Let's say you're developing a python package that relies on a native shared library. If you distribute it as a pre-compiled wheel, you're locking your users into the exact version and compilation flags, options, etc that you choose. So your question effectively boils down to "why do people need to compile their own software?".
Many reasons: you need more control, specialized hardware, testing newer versions of the library, maintaining an internal fork of a library, security, performance, the dev team maintains both the native library and python package and needs to keep them independent, or simply preference for dynamic linking against system libraries to avoid duplication.
> So your question effectively boils down to "why do people need to compile their own software?".
No, the question is why a package would need to decide for its users that the package and its dependencies must be installed the gentoo way. That's quite obviously different from why an end user would decide to install from source despite the availability of binary packages.
Yes, I can see how it would make sense to be able to set this in pyproject.toml (typically for private package)
uv is still quite new though. Perhaps you can open an issue and ask for that?
It looks like there's already a pyproject.toml no-binary option?
Or for a specific package: https://docs.astral.sh/uv/reference/settings/#no-binary> There are times when you do NOT want the wheel version to be installed
When, why? Should I be doing this?
There are some wheels, for example, 'lxml' that bundle in their binary possibly incompatible external libraries, or older libraries than what you would like. This can cause library conflicts.
I'm in ML-land. I thought we were all hopelessly tied to conda. But I moved all my own projects to uv effortlesly and have never looked back. Now first thing I do when pulling down another reseacher's code is add a pyproject toml (if they don't have one), `uv add -r`, and `uv run` off into the sunset. I especially like how good uv is with non-pypi-published dependencies: GitHub, dumb folders, internal repos, etc.
uv is almost perfect. my only pet peeve is updating dependencies. sometimes I just want to go "uv, bump all my dependencies to the as latest version as possible while respecting their constraints". I still haven't found an elegant way to do this, but I have written a script that parses pyproject.toml, removes the deps, and invokes `uv add --upgrade` with them.
other than that, it's invaluable to me, with the best features being uvx and PEP 723
This worked pretty well for me:
Agree that something like this should be built in.https://pypi.org/project/uv-bump/
Does `uv lock —upgrade` not do what you want?
Unfortunately, no. Only `uv.lock` gets updated, but the dependencies in `pyproject.toml` are frozen at their original constraints.
What I want is, if my project depends on `package1==0.4.0` and there are new versions of package1, for uv to try install the newer version. and to do that for a) all the deps, simultaneously, b) without me explicitly stating the dependencies in the command line since they're already written in the pyproject.toml. an `uv refresh` of sorts
If you specify your constraints in pyproject.toml like this: `package1==0.4.0`; then that is the latest (and only) version satisfying your constraints. Not upgrading is expected behavior, because upgrading would violate constraints.
pyproject.toml’s dependency list specifies compatibility: we expect the program to run with versions that satisfy constraints.
If you want to specify an exact version as a validated configuration for a reproducible build with guaranteed functionality, well, that’s what the lock file is for.
In serious projects, I usually write that dependency section by hand so that I can specify the constraints that match my needs (e.g., what is the earliest version receiving security patches or the earliest version with the functionality I need?). In unserious projects, I’ll leave the constraints off entirely until a breakage is discovered in practice.
If `uv` is adding things with `==` constraints, that’s why upgrades are not occurring, but the solution is to relax the constraints to indicate where you are okay with upgrades happening.
> ... the solution is to relax the constraints to indicate where you are okay with upgrades happening.
Yeah, that's pretty much what I've been doing with my workaround script. And btw most of my projects are deeply unserious, and I do understand why one should not do that in any other scenario.
Still, I dream of `uv refresh` :D
There's an open issue for "Upgrade dependencies in pyproject.toml (uv upgrade)":
https://github.com/astral-sh/uv/issues/6794
Why not depend on package1>=0.4.0 rather than specifying an explicit version? Then uv will upgrade it to the latest version.
pyproject.toml is meant to encode the actual constraints for when your app will function correctly, not hardcode exact versions, which is what the lockfile is for.
Because then you don't get to use the new features in 0.5.0.
Though I do think with Python in particular it's probably better to manually upgrade when needed, rather than opportunistically require the latest, because Python can't handle two versions of the same package in one venv.
Then make your dependency package1>=0.4
https://packaging.python.org/en/latest/specifications/depend...
> then you don't get to use the new features in 0.5.0.
Yes you do
package1>=0.4.0 means 0.4.0, 0.4.1, 0.4.100, 0.4.100.1 and so on
package1>=0.4 includes the above plus 0.5.0, 0.5.1, 0.6.0, 0.100.0 and so on
> What I want is, if my project depends on `package1==0.4.0` and there are new version of package1, for uv to try install the newer version.
I think you're just specifying your dependency constraints wrong. What you're asking for is not what the `==` operator is for; you probably want `~=`.
You are writing your project file incorrectly. It's not a lock file
I never, ever, write my project file[1]. uv {add,remove} is all I ever use.
[1]: I do sometimes write the title or the description. But never the deps themselves
Even using `uv add`, you don't have to limit yourself to declaring exact versions when your intention is to allow newer versions.
Whichever method you use, you need to produce a different file that better reflects your intentions
You can specify bounds when using uv add:
uv add example>=0.4.0
Then it will update as you are thinking.
I think what you want is `uv sync --upgrade`
You could either delete the .venv and recreate it or run `uv pip install --upgrade .`
Much prefer not thinking about venvs.
Actually, it won't work. I tried it and running `uv run script.py` just reinstalls the deps back... which is, I admit, the behaviour I expect and want as a user.
Astral's tools are so fast in general, when you first try them out, you double check what went wrong because you are sure nothing happened.
Same with uv. They are doing very nice tricks, like sending Range requests to only download the metadata part from the ZIP file from PyPI, resolve them in memory and only after that downloading the packages. No other package manager does this kind of crazy optimization.
Can anyone point to the top 3 Rust specific things that UV leverages to make it run so much faster?
Yes I know Rust is not a GC language… go a level deeper
There is nothing Rust specific as far as I’m aware, other than the developers are comfortable with the language. They have a very fast dependency resolver/algo, but presumably you could implement that in C or Zig or whatever. It is much faster than if it was implemented in Python or another interpreted language though. Ruff is way, way faster than the equivalent Python tools. I like Rust for the type safety and tooling so that may slow them to build faster?
1. Zero-Copy Deserialization with rkyv and Unsafe Code 2. Platform-Specific High-Performance Memory Allocators 3. High-Performance Hash Maps with FxHasher and SwissTable
https://deepwiki.com/search/point-to-the-top-3-rust-specif_4...
I've started on a CMake integration to take multiple pyproject.toml files (in a monorepo or such), turn them into an editable workspace venv for development (with dependencies), and then on `make install` package the workspace up into a venv (at some point I'll add a package target...). The lock file ends up checked in and updated whenever you build w/ changes to any of the pyproject files.
This is very very useful for robotics or other places where you might end up with a big list of pip installed packages in some Dockerfile, want them pinned, with just one version of each library across the whole space, but don't want to necessarily do the song and dance of a full python package. Because uv is fast it just kind of works and checking the projects doesn't take any time at all compared to the other build work being done.
I've been happily using it for a long time, and rye before that.
Just today I set it up on 20 PCs in a computer lab that doesn't have internet, along with vs code and some main packages. Just downloaded the files, made a powershell script and it's all working great with Jupyter etc... Now to get kids to be interested in it...
Its sad that pyrhon only tooling is aparently so imcapable that you have to write it in a compiled language.
After that many years of optimization pure python seems still to be wishfull thinking. It's ai/mk success is also only as a shim language around library calls.
So? It's the best language there is for shimming around library calls. Use the right tool for the job. There's no single language that can do it all.
I ran into this via Aider. Who centers their user experience on using uv as an installer. https://aider.chat/2025/01/15/uv.html
What was super unclear was how I develop locally with uv. Figuring out I needed `aider sync --extra` then `aider run --projrct /opt/aider aider` to run was a lot of bumbling in the dark. I still struggle to find good references for everyday running projects use with uv.
It was amazing though. There were so many pyproject and other concerns that it just knew how to do. I kept assuming I was going to have to do a lot more steps.
Very relevant talk if anyone is interested in learning a bit more about how the project is so fast. https://www.youtube.com/watch?v=gSKTfG1GXYQ
Never thought to see if Jane Street had a YouTube or if they had programming related stuff but here we are thanks to you! Makes sense they do, and that it’s python related.
uv has converted people in the same way Google search, Chrome, Git and SSDs did.
Fast is a massive factor.
I haven't used it much, but being so fast, I didn't even stop to think "is it perfect at dependency management?" "does it lack any features?".
Ease of use was the key for me
I continue to be puzzled why sometime running uvx (uv tool run) will redownload all dependencies even though it just downloaded them for another tool. Downloading torch 15 times per day gets old even on 500 mbits
Might be worth opening an issue, the project states it aims to aggressively cache dependencies: https://docs.astral.sh/uv/concepts/cache
Maybe that functionality isnt implemented the same way for uvx.
You could try this equivalent command that is under "uv run" to see if it behaves differently: https://docs.astral.sh/uv/concepts/tools/#relationship-to-uv...
I haven't been able to find any kind of rhyme or rhythm to it, so I don't know how to explain when it happens or how to better debug it for a bug report.
Initially, I used `uv tool run <name>`, but later discovered `uv tool install <name>`. The `install` command downloads dependencies once and caches them for future use. It is similar to how `pipx install <name>` works.
e.g.
$ uv tool install asciinema
$ asciinema play example.cast
It's not just the programs that are fast, but the company's response. I reported an issue on Discord and the founder fixed it right away, collaborating with me on the PR I created.
Switched all my python stuff to poetry several years ago. Still kinda tired and shagged out from that endeavour. Yeah yeah, ok ok, if y'all keep raving THIS much about uv, I guess it must be pretty good ... still, sigh, yet another One True (tm) python package management solution. It's on my todo list, I'll get around to it, I reckon within the next 12 months, currently I have no pressing need to move away from poetry.
You don't know what you're missing if you haven't spent 5 minutes trying out uv.
If you're intentionally not trying it simply because you don't want to get addicted like everyone else clearly is, I could see that as a valid reason to never try it in the first place.
I usually avoid jumping on bandwagons, so I've always stuck with vanilla pip/venv, but at this point it really is clear to me that uv really is the "One True (tm) python package management solution", and probably will be for the next 10 years.
uv is indeed fast and easy. But I'm also finding that the maintainers' efforts to make it work like Cargo mean it can be more difficult to use in more complex project structures. As sensible as Rust's project management ethos is, you're never going to escape Python's underlying design in a Python project, and friction between the two philosophies may not be avoidable.
One possible alternative is Pants. It's also written in Rust for performance, but has more flexibility baked into the design.
Pants and uv are two very different tools, with very different use-cases.
uv is basically souped-up pip.
Pants is an entire build/tooling system, analogous to something like Bazel. It can handle multiple dependency trees, multiple types of source code, building and packaging, even running tests.
uv is so much more than just a souped-up pip. It can also, for example, do building and packaging.
It also has workspaces and subprojects features that are sneaking toward something like the full-fledged multi-artifact project support you get in Pants (or Poetry, with plugins). Except that their decision that the entire project must have one and only one global dependency solution means that there's no escape hatch if you ever end up in a dependency hell situation. Which is fairly common in Python for a variety of reasons. And, even when it does work, subprojects all sharing a single virtualenvironment means it's really easy to accidentally create an undeclared dependency and never notice at development time because it's already in the venv due to a sister project already having declared it. So if you have a project that builds multiple packages, you're on your own for implementing a solution to ensure packages with sane dependency specs get published to the archive.
That's a major reason why Python development culture decided to go with many project-specific virtualenvironments instead of a single global one like what uv is trying to drive toward. And it's true that it still allows you to do independent projects that reference each other at development time using path dependencies. But, unlike some of its more mature alternatives, it doesn't give you any help with replacing those path dependencies with named, version dependencies at build time. So if you have a project that builds multiple packages, you're on your own for implementing a solution to ensure packages with sane dependency specs get published to the archive.
I have no understanding of python really but very glad I settled on uv. It’s such an improvement user wise over poetry or pip and it’s extremely fast too.
I have one complaint though, I want ./src to be the root of my python packages such that
> from models.attention import Attention
Works if I have a directory called models with a file called attention.py in it (and __init__.py) etc. The only way this seems to work correctly is if I set PYTHONPATH=./src
Surely the environment manager could set this up for me? Am I just doing it wrong?
src layout _should_ work. The pyproject.toml needs to live next to src, not inside of it. You might need to `uv run python` in order to pick up the right PYTHONPATH?
https://docs.astral.sh/uv/concepts/projects/init/#packaged-a...
ChatGPT says two things to this “uv is a package manager not a file runner” um uv run seems to be one of the best parts of uv… and that “You’re doing things correctly you are just hitting one of Python’s most notorious pain points: managing import paths cleanly”.
I have read a few tickets saying uv won’t support this so everyone running my project will have to read the README first to get anything to run. Terrible UX.
What is it that uv won’t support?
Setting PYTHONPATH (or for that matter any other Env vars) by default. Seems useful to me.
I love uv. I am a convert, I use it for everything. One area I find it incredible for: deployment. If I have to launch a docker container and install a lot of python packages it saves so much time and compute.
I also appreciate that it handles most package conflicts and it constantly maintains the list of packages as you move. I have gotten myself into a hole or two now with packages and dependencies, I can usually solve it by deleting venv an just using uv to reinstall.
I’ve been writing Python for over a decade and was still using the Pyenv Virtualenv (dated I know). Lately I had been trying to switch over to Poetry. Somehow it never works smoothly for me.
Then I looked at uv and oh boy is it better. Everything just works and is blazing fast. Kudos to the team.
Started out with conda mainly because it played nicely with jupyter, but after switching to fish and vscode on mac, managing conda environments for Jupyter kernels turned into a pain. Honestly, half my personal projects procrastination comes down to fighting with conda setup. Been hearing a lot of buzz about UV, and this post finally convinced me to give it a shot on my next project.
Sources
Biggest dev QoL improvement for me since TypeScript has been going all in on uv.
I switched everything over and haven’t looked back.
It’s everything I hoped poetry would be, but 10x less flakey.
Here’s my #1 complaint about uv: I’m new to the python ecosystem. I don’t know anything about pip or the existing tools. I would love for uv to at least have documentation and a user guide that doesn’t assume knowledge of the old bad tools that uv replaces.
Perhaps uv will continue its ascendancy and get there naturally. But I’d like to see uv be a little more aggressive with “uv native” workflows. If that makes sense.
That exists! [0] But IMHO the guides should be linked in big text as the first thing on the homepage. Right now you have to read through a bunch of meaningless comparative bullet points and reassurances then click a few times to get to the guides. If it weren’t for everyone telling me I need to switch, I might not have had the patience to find them.
[0] https://docs.astral.sh/uv/guides/
I'm looking forward to adopting uv but I'm blocked until Snyk supports it. https://support.snyk.io/s/question/0D5PU00000u1G4n0AE/suppor...
Export a requirements.txt from UV and analyze that?
I switched to uv in an open source project I maintain as our builds were taking dozens of minutes to resolve dependencies using Poetry. The migration itself was pretty easy as both tools use pyproject.toml, so it was a matter of changing headers and minor adjustments.
After the switch, the same dependency resolution was done in seconds. This tool single-handedly made iteration possible again.
I'm setting up a new dev workstation and I'm trying NixOS and for Python work I'm going to try living with "uv" for the bulk of my Python work. An experiment I've just started, but I've been using uv for ~6 months and it's been spectacular and I think it's going to be able to live up to my expectations.
I don't use uv for my (django) projects because my pip/venv workflow is good enough and had never failed me in 15 years.
However I really like installing uv globally on my Windows systems and then using uvx to run stuff without caring about venvs and putting stuff to path.
I’ve been writing an equivalent of uv for the R language at work, and it’s quite daunting / unwieldy. I was feeling bad about it, but then I reminded myself that uv has hundreds of contributors and my project only has one.
Is there something inheret about rust that means its faster at dep resolution than Python? Like where is the speed up coming from?
Or would it be possible to go this fast in python if you cared enough about speed?
Is it a specific thing that rust has an amazing library for? Like Network or SerDe or something?
The biggest speed-up in uv comes from the way it uses caching and hard links. When you install a package into a virtual environment uv uses a hard link to a previously cached version rather than copying files.
Using Rust is responsible for a lot of speed gains too, but I believe it's the hard linking trick (which could be implemented in any language) that's the biggest win.
There's also the assumptions around wheels having consistent metadata (which isn't required in older versions of the specs), so as long as you don't need to support the long tail of packages, then those assumptions are fine (it's worth noting uv weren't the first to make these assumptions, poetry did this too, but pip does not).
They came up with a faster resolution algorithm, in addition to implementing it in a faster language.
pip could be made faster based on this, but maybe not quite as fast.
The resolution algorithm is the pubgrub algorithm from Darts package manager pub, implemented in Rust for Cargo https://youtu.be/LGXx5Bqcgq8
So how does Astral plan to make money?
Here's a loose answer to that question from uv founder Charlie Marsh last September: https://hachyderm.io/@charliermarsh/113103564055291456
I don't want to charge people money to use our tools, and I don't want to create an incentive structure whereby our open source offerings are competing with any commercial offerings (which is what you see with a lost of hosted-open-source-SaaS business models).
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 (we may not do this, but it's helpful to have a concrete example of the strategy) would be something like an enterprise-focused private package registry. A lot of big companies use uv. We spend time talking to them. They all spend money on private package registries, and have issues with them. We could build a private registry that integrates well with uv, and sell it to those companies. [...]
But the core of what I want to do is this: build great tools, hopefully people like them, hopefully they grow, hopefully companies adopt them; then sell software to those companies that represents the natural next thing they need when building with Python. Hopefully we can build something better than the alternatives by playing well with our OSS, and hopefully we are the natural choice if they're already using our OSS.
They are hijacking the entire python's ecosystem in a very smart way, that's all. At some point we, probably, will find us vendor locked-in, just because the initial offer was so appealing. Take a closer look at it: package manager, formatter/linter, types, lsp. What's left before it will poke cpython one way or another? Maybe cloud-based IDE, some interesting WASM relationship (but RustPython is not there yet, they just don't have enough money). Otherwise, Astral is on a pretty straightforward way to `touchdown` in a few years. It's both, the blessing, and the curse.
Let's be honest, all tries to bring a cpython alternative failed (niche boosters like PyPy is a separate story, but it's not up-to-date, and not entirely exact). For some reason, people think that 1:1 compatibility is not critical and too costly to pursue (hello, all LLVM-based compilers). I think, it's doable and there's a solid way to solve it. What if Astral thinks so too?
Honestly... I don't really care. If in 5 years they turn around and try to charge for uv we'll still be in a much better place than if we'd all stuck with the catastrofuck that is pip.
Anaconda makes on the order of 100M a year “solving” data science package management. I would argue it has a significantly worse product, attacking a much smaller part of the ecosystem.
It seems easy to imagine Astral following a similar path and making a significant amount of money in the process.
Anaconda isn't free. I don't want to pay per-seat fees for slightly improved versions of open source tools which is why I'm very skeptical of Astral and uv.
One day they're going to tell me I have to pay $10/month per user and add a bunch of features I really don't need just because nobody wants to prioritize the speed of pip.
And most of that fee isn't going to go towards engineers maintaining "pip but faster", it's going to fund a bunch of engineers building new things I probably don't want to use, but once you have a company and paying subscribers, you have to have developers actively doing things to justify the cost.
Enterprises don't care about faster, but they do care an enormous amount about security. Astral is very well positioned here.
I don't think Anaconda is targeting a smaller part of the ecosystem, rather a different but overlapping ecosystem (remember Anaconda does more than just Python, and there's more to data science than Python). Astral I suspect won't try to move outside of Python, which limits what its tools can be used for.
In theory, Anaconda solves the next higher level of the Python package management nightmare, namely knowing what versions are compatible with each other. But that could presumably be done on top of uv.
Apparently VC(1) so far, I'd assume there will be LTS support contracts and tailored enterprise features down the line; for the moment I'd assume it's just a bunch of talented devs fixing problems they've been tired off / see as long term existencial threats to the python ecosystems.
1: https://old.reddit.com/r/Python/comments/12rk41t/astral_next...
Relevant: https://sgoel.dev/posts/building-cython-or-c-extensions-usin...
The Astral projects are all great. I hope the company finds a revenue stream in the future with some hosted something or the other because these tools are so useful I don't want them to become pay-for etc.
I still mostly use pip (largely due to habit) but uv has been incredible for speeding up ci workflows! We shaved off a full minute of CI for a python package I maintain when we started using uv.
https://github.com/skanga/JarGet is a slightly similar tool for Java
Bragging that your program is faster than anything written in Python is a low bar, lol.
Also, it seems like a sign that even Python tooling needs to not be written in Python now to get reasonable performance.
Yes `uv` is great but using it to manage dependencies in an AWS Lambda Docker container is really not as simple as using pip or poetry.
I love uv, not just for local development, but it also makes it WAY easier to manage python environments you setup for running python workers / user code in the cloud.
I moved from pip to poetry a while back, but was introduced to uv recently while working on MCP stuff. I now use uv for everything and haven’t looked back.
What made you switch from poetry?
Moved to this and have no need for anything else, especially since uv pip install whatever works with uv and is faster than pip (though I usually use uv add).
uv is great! One surprise I recently ran into, though, was that `uv run path/to/foo.py` doesn't look up the nearest pyproject.toml / .venv. Similarly, a script with a shebang `#!/usr/bin/env -S uv run` only executes correctly when executed from within the folder containing the .venv. Am I the only one finding this weird?
I think the thought is to be obviously deterministic. It’s not exactly trivial to make those decisions for users since uv is not locked to its .venv - you may want the env, the system python, a pyenv, a brew python, or whatever variation can co-exist seamlessly.
I’d be surprised if there wasn’t an env var for it though.
uv is so good, I would recommend it to anyone.
The suffix "written in Rust" is getting cringy though.
I was just thinking the dearth of package managers for Python is such a missed opportunity
One of may favorite party tricks with uv:
echo 'import antigravity' | uv run -
Thank you astral!
uv is great and we have largely switched over to using it. however, since it doesn't handle compiled binaries, we have to continue using conda for environments where those are needed (i.e., mkl).
For MacOS the install doesnt mention if on homebrew?
You can `brew install uv`
Seems a lot of people like this and are happy about it, but I for one am tired of the proliferation of python package management tools.
Many languages have many package management tools but most languages there are one or two really popular ones.
For python you just have to memorize this basically:
- Does the project have a setup.py? if so, first run several other commands before you can run it. python -m venv .venv && source .venv/bin/activate && pip install -e .
- else does it have a requirements.txt? if so python -m venv .venv && source .venv/bin/activate && pip install -r requirements.txt
- else does it have a pyproject.toml? if so poetry install and then prefix all commands with poetry run
- else does it have a pipfile? pipenv install and then prefix all commands with pipenv run
- else does it have an environment.yml? if so conda env create -f environment.yml and then look inside the file and conda activate <environment_name>
- else I have not had to learn the rules for uv yet
Thank goodness these days I just open up a cursor tab and say "get this project running"
uv handles most, if not all, of those cases.
> - else does it have a pyproject.toml? if so poetry install and then prefix all commands with poetry run
That's not even correct. Not all projects with pyproject.toml use poetry (but poetry will handle everything with a pyproject.toml)
Just try uv first. `uv pip install .` should work in a large majority of cases.
pipenv is on the way out. bare `setup.py` is on the way out. `pyproject.toml` is the present and future, and the nice thing about it is it is self-describing in the tooling used to package.
I wish setup.py was actually on the way out, but sadly, it's the only straightforward way to handle packages that use cython or interop. In these cases, libraries use setup.py to compile the dll/so/dylib at install time. Naturally this is a bit of nightmare fuel since installing gets arbitrary code execution privileges but there's no real standard for privileges in python package installs.
> That's not even correct. Not all projects with pyproject.toml use poetry
I didn't say "all projects with pyproject.toml use poetry"
Or just use uv and then you don't have to think about any of that.
Uv will quickly become the dominant player in the space. It’s that much better. There is no reason to not use it.
uv has completely changed the experience of Python for me.
Why the "written in Rust" in the title?
It's +5hp, like these two white stripes on a car hood.
uv has become essential for me. conda and virtualenv never worked smoothly for me, but uv was easy and "just worked" from day 1.
The only thing keeping Python usable.
Why use this over micromamba?
I use the two together with no issues. Micromamba works great for conda dependencies that’s not on PyPi (for many reasons), and everything else I go for uv.
was forced to use it for work and was converted!
People always mention the speed, but that's rarely a factor for me.
Rather, pip was broken intentionally two years ago and they are still not interested in fixing it:
https://github.com/pypa/packaging/issues/774
I tried uv and it just worked.
uvx is the killer app imo
unfairly fast
Now.. can they make Python itself as fast as the package manager??
wake me when venv works on mac
It does? What problems have you had?
source .venv/bin/activate .venv/bin/activate:37: parse error near `deactivate'
Automatically generated, not compatible with zsh
I can't reproduce this, possibly it's another environment error or the problem has been fixed in the versions I am using. (uv 0.7.11, zsh 5.9, python 3.13.5, MacOS 15.3.1)
[dead]
[dead]