uv: An extremely fast Python package and project manager, written in Rust

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)

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.

From just my observations they basically parallelized the install sequence instead of having it be sequential (among many other optimizations most likely)

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).

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

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

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.

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.

> 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?

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.

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.

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.

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."