People are increasingly running out of reasons to not use Julia. It really is the future for data science and machine learning. No let me correct that. It really could replace anything R, Python, Perl, Ruby etc is used for today.
Not right now as packages, larger communities etc need to be developed. But long term solutions the potential is obvious. A friendly, powerful and high performance dynamic language ought to have a very broad appeal.
I think Julia really dropped the ball on the execution model. Just-ahead-of-time compilation ends up being the worst of both worlds - you can't compile a small fast binary for deployment, and you can't quickly run a script or REPL for development. It turns out that this really matters for adoption.
And now Julia has competition from Mojo. Mojo makes some compromises for backward compatibility with the Python world, but it's really solving the problems that hurt AI most. And the folks behind Mojo have a lot of real-world experience migrating a community from one language to another.
I think Julia will remain a niche language, confined to science and statistical computing outside of mainstream data science and machine learning.
> it's really solving the problems that hurt AI most
Isn't this a bit premature? Mojo doesn't tangibly exist for most people (we can't run it ourselves), and I am unaware of any ML/ AI applications built with Mojo.
Yeah, definitely early days. But they're saying all the right things: fast code off the bat, easy to vectorize and parallelize, efficient use of memory, easy to create abstractions that execute efficiently on various architectures. The demos are impressive and they have the track record to be credible. I've played with it and the stuff that's been implemented so far works well.
IMO Mojo isn't competing with Julia. It's competing with Rust. Manual memory management, you have to type annotate every variable (if you want good performance), and no overloading doesn't seem to me like a language aimed in the same area as Julia.
If Julia can make good strides in Science/Engineering/Statistics world outside of DS/ML (that too mostly NN) I'm sure the Julia folks would see that as a major win. Basically a unified next generation language taking over R/MATLAB/Fortran space. I think that field is going to grow dramatically (in adoption of tools) over the next decade as opposed to DS/ML which are already overhyped.
If Mojo finds it groove and can convince people that static typing is a price to pay for performance I think Julia will have to adopt some sort of a pure AOT compilation mode and forgo the Dynamism that keeps that Just- portion lying around which does seem like an issue for DS/ML folks.
I'll preface by saying by far my biggest gripe with Julia is the inability to deploy binaries easily. The only hope on easy deployment I suppose is to hope it becomes commonly installed enough that you can rely on it being on systems. But some of your complaints are a bit off (in my own opinion):
>and you can't quickly run a script
What is wrong with the following to run a script?
$ julia myscript.jl
If you have specific needs that demand, after hitting return, the few seconds of delay for the vast majority of scripts is an issue, you can pre-compile it ahead of time or simply use something like https://github.com/dmolina/DaemonMode.jl
Julia has issues as with all languages but "not being able to quickly run a script" is by far one of the easiest to work around.
> and you can't quickly run a script or REPL for development.
REPL- I disagree. Of course you can - that's how many of use Julia.
> And now Julia has competition from Mojo.
...maybe. The code-samples we've seen from Mojo look very similar to Python, obviously. And that is specifically why a lot of poeple love Julia.
The problems people are more and more interested in (machine learning, etc) are at their base mathematical problems. These languages are all tools to translate mathematics into computer instructions. The code should then ideally look as close to that math as possible. Spamming np.linalg, sp.sparse, and so forth over and over again is just ugly, and the entire Python workflow overly encourages object oriented design for concepts that are mathematically functions. And, well, should be functions. If you're working in these or related fields at all and write any code with Julia, it's hard to not fall in love.
Mojo may make Python faster. But it will still be, roughly, Python.
And considering that the Python 2 -> Python 3 transition has still not completed, I think it's premature to conclude Mojo's python upgrades will get anywhere.
> and you can't quickly run a script or REPL for development.
You know that the entire blogpost you're ostensibly commenting under is about how the JIT overhead has been greatly reduced via more static AOT compilation, right?
Python is not compiled and start & load pandas (which is comparable to the libraries loaded in the article) in ~0.4" on my computer, and that's a notoriously slow language.
If I were to use e.g. Rust with polars, load time would be virtually none. And when I have to process ~50k different datasets, I can't afford 0.85" per file, which would translate to ~11 hours of overhead.
> If I were to use e.g. Rust with polars, load time would be virtually none.
Because you're compiling...
And if you need to do the same in Julia, you should also pre-compile or some other method like https://github.com/dmolina/DaemonMode.jl (their demo shows loading a database, with subsequent loads after the first one taking roughly ~0.2% of the first)
No it's not, because I will not architecture my whole pipeline & program around Julia inability to start in maybe a second in a year or 1.7" now, I will just use another language.
When is 1.10 expected? Just wondering I did a quick google and it seems a lot of the 1.9 improvement were back ported from 1.10? Or is that really 2x on top of current 1.9?
1.10 feature freeze is going to be in the next few weeks. after that it will be a few months depending on how much is required to fix all the bugs that have probably been introduced. 1.9 mostly doesn't shorten loading times (although weak dependencies end up helping a bit). 1.10 has had a bunch of load time optimization which became a lot more obvious once 1.9 got rid of all the stupid stuff. the exact speedups are package dependent, but 2x is a good estimate. some packages get a lot more, some are about the same.
You're not technically wrong, in that Julia has the potential to replace all of those languages. And if it could attract large communities, an amazing interoperating coherent package ecosystem could develop there that's much less of a pain to use than most current systems.
But I don't think either of those are given. Projects with huge potential have unfortunately failed for "trivial" reasons like personpower, funding, approachability, etc. that have nothing to do with underlying technical merit. (edit: but, to be clear, I don't think at this point Julia can exactly "fail", that seems pretty implausible; it will definitely have a strong presence in the MATLAB-ish niche, in the sciences and in engineering. The uncertainty is about more general purpose, popular usage).
I love Julia and would love to see it succeed, but at this point in time its future is still uncertain. (And it's also unclear what if any the effect of LLM code generators is going to be - are they going to reinforce the current popular languages since those are the ones they're good at? Or are they actually going to make less popular languages more approachable, as they grow to learn to generate code in them too?)
> It really could replace anything R, Python, Perl, Ruby etc is used for today.
There is an (admittedly niche) use-case where Julia would be almost incapable to work, as compared to other interpreted languages like for example Lua or Perl. Making a "busybox" style executable that replaces all coreutils (cat, ls, head, tail, wc, sort, uniq, tr, cp, df, echo, printf, ...) with simple Julia implementations. The slow startup-time for each Julia instance would make most shell scripts unbearably sluggish.
I agree that Python would be an equally bad choice for that, but at least it should be slightly faster.
Regarding "serious" uses of Julia, especially in numerical mathematics, I find some basic things still lacking. For example a complete base package for sparse matrices, including kronecker products etc. Octave/Matlab have the "kron" function in the base language. In Julia, should I use things with dubious names like LuxurySparse or what?
> Making a "busybox" style executable that replaces all coreutils
I wonder if DaemonMode (https://github.com/dmolina/DaemonMode.jl) is the right approach for this, having each of those tools' functionality loaded in the server process, and calling out to them from a client when a tool is invoked.
> a complete base package for sparse matrices, including kronecker products etc.
SparseArrays is a standard library package for sparse matrices and arrays. kron is in the LinearAlgebra standard library. So you can just do
using LinearAlgebra
using SparseArrays
and then do any kron product you want, whether sparse matrices with other sparse matrices, or dense ones, or any other array type.
> and then do any kron product you want, whether sparse matrices with other sparse matrices
Thanks! Will try it with the new release. Last time I tried, apparently it converted the sparse matrix to dense before calling kron, and (expectedly) it failed due to memory error. For example, try to compute the adjacency matrix of a grid graph of size 1000x1000. It is the kronecker product of two tridiagonal matrices of that size. It only has four million non-zero entries, which are instantaneous to compute, but if you want to store all the zeros you'll need a few terabites of RAM.
> People are increasingly running out of reasons to not use Julia
I'm sorry, but I can't really let you say that.
I like Julia, it has many nice features, but it still IMHO exhibits a lot of reasons not to use it, all of which I stumbled upon while using it to develop non-trivial pieces of software in bioinformatics.
- parallelism is very confusing, very easy to break, and not very efficient when it works;
- cold-boot time is prohibitive;
- the ecosystem is brittle, and very much underdocumented;
- the dispatch mechanism is both a blessing and a curse;
- the stdlib should include some way of encoding errors, otherwise everyone cook their own stuff in their corner (should I raise an exception? Return -1, or maybe null? Or a Union?);
- not being able to redefine types in REPL is a PITA for REPL-driven development;
- debugging cross-packages bugs with multi-dispatch in time-consuming;
- the stdlib doc website is *excruciatingly* slow;
- memory consumption is hard to control;
- iterators are weird (i.e. eager and hardly composable), and either pipe should be a part of the language or they should be chainable, map(f, filter(g, sort(h, [...]map(t, xs)))) leads to unreadable & uneditable mess;
- performances do not follow what was promised (*C-like order of magnitude*), unless maybe you're (a) a Julia god or (b) only using numeric computations.
Now none of these flaws are intrinsic, and I very much hope that all these edges will be polished in the coming times. But as a ‶grunt on the field″, it is exasperating to see the core Julia community just telling you that you're holding it wrong and that all is best in the best of worlds.
This kind of thinking is very old school. Its 2023 now.
1.GPT-4 and co are really, really good at python, and will write 90% of boilerplate for you. That makes a python dev way more productive than the Julia one. There is way, way less data on Julia, so the ceiling for AI assisted Julia is way lower.
2.Data science is not just running fancy statistical functions. Its also API calls to like 10 services. You can setup a PDF->OCR->GPT processed text->Send to vector database with a few API calls in python. There's a python package for everything, and everything put into one script.
3.Did you know everyone using stable diffusion (millions), has to install pytorch and like 50 packages in python? Now 95% will just use auto-installers. But the remainder 5%, will be exposed to python, and forced to learn it if they want to use AI art professionally (scripting) or do further development in it. When you see hordes of nerds desperately learning python to generate waifu pictures on google collab notebooks, you know a language will be dominant for decades to come.
Until there is a Julia package for everything it is not a replacement for python. I find it hard to make a case for Julia if performance is not really a consideration.
It's jarring initially, but becomes natural very quickly. Writing loops like "for i in 1:length(arr) ... end" is pretty neat compared to C++ or even python. Plus in math sequences typically start at index 1.
I would go as far as to say that it is now largely an archaic idiom to write «for i in 1:length(arr) ... end», and there is no reason to write such a code to process a collection of elements (an array, a list etc) in its entirety. Yet, people keep on writing it and then spend countless man-hours chasing subtle bugs that blow out later in production.
Most modern languages have a «foreach» operator or its functional equivalent as a collection item generator. «foreach elem in my_collection {}» also has a clearly defined semantics: «process everything». The code is cleaner, concise and a bit shorter and reduces the cognitive overload. And since many languages now support lambdas, the code can quickly and error free be converted into a staged data processing pipeline without the use of temporary variables that serve no purpose, e.g. «foreach elem in my_collection.filter (_ -> _.my_selection_criteria ("42"); ) { do_something (elem); }». Whether the collection is indexed from 0 or from 1 becomes irrelevant.
You should never write loops that way, at least in code you're going to share (assuming that you’re going to have some arr[i] in the loop body—if not, you would just do "for e in arr").
Assuming that arrays start at 1 is a source of occasional bugs in public packages. The existence of OffsetArrays means that arrays can start at any index (so for people who get nauseated by 1-based arrays, you can change it).
Instead, write "for i in eachindex(arr)".
In fact, Julia’s array syntax means you can loop over and manipulate arrays without using indexes much of the time, so don’t even need to know how they’re based.
> Plus in math sequences typically start at index 1.
I'm not quite sure that this is the case enough to say "typically". In terms of undergraduate exposure to math, I think more people have taken Calculus (or Analysis) than Linear Algebra, and I think Calculus textbooks tend to index from 0 while Linear Algebra textbooks tend to index from 1.
That is such a weird argument to me. I can count in the fingers of one hand the number of times in my life that I've had to compose ranges, it seems a strange thing to optimize for.
You seem to have a weird fetish for composable ranges seeing your history of comments on exactly this topic (even this thread!). The number of times I have had to think about composability as you mention is close to zero. In fact, the off-by-one errors with C class of languages due to the open ended nature of the interval have bit me about 3-4 orders of magnitude more. There is an entire class of security bugs because of this. I find this obsession very strange, YMMV obviously.
I used Julia for a project back in 2016, and had the same adverse reaction to 1-based indexing. It's the same reason I have a hard time with Lua. Why? I suppose it's all rather superficial, but it's one of those things that just grates on me and I'll avoid it if I can. I'm not sure there's a logical explanation, but I've found the sentiment to be rather common among developers.
Same, but probably it's more of a familiarity bias rather than a logical one. I think there is a strong argument to index from 1 as is done in MATLAB, R etc. since it matches the way most people refer and think about lists in reality.
A brave decision from the authors though, I imagine they would have got less flak matching the more popular 0-index languages.
Ha, I was thinking of replying to your parent comment, "Half-open ranges are such an unpleasant thing as to be unforgivable to me. Different subjective preferences."
I've always found 0-indexing mildly unpleasant too, even though C was the language I learnt programming with, and felt like I found home when I came across 1-indexed languages.
With half open ranges, the length of the range is the end minus the beginning, which is nice. It's also much more simpler to write an empty half inclusive range. For instance if left=right then left..right would be the empty range. Whereas in Julia I'd need... left:(right-1)? But is 1:0 the empty range or the right-to-left inclusive range [1,0]? Very confusing and hard to work with all around.
> With half open ranges, the length of the range is the end minus the beginning, which is nice.
But I already know from real life that if there's maintanence in blocks 5 to 15, that's eleven blocks under maintanence. With half-open ranges, it once again introduces confusion and makes things unintuitive.
We can each find countless examples where each style comes out better, for eg `1:N` most often expresses the intent better than `range(N+1)`. I find people who prefer half-open ranges to be weird and incomprehensible, you probably find my preferences the same, I just don't like the categorical statement often made in this regard that one is inherently and universally superior to the other.
Yes, range(n+1) is bad. I hate python as well. You want something like Rust’s m..n and m..=n to have the choice. Anyway, if you have half open ranges, it's trivial to get a closed range (add one to the right endpoint), but going the other direction is not so simple.
Idk typing +1 seems easy enough... 1:N == union(1:(N÷2), (N÷2+1):N).
But really, these discussions are funny to me - each side pretending their convention is how God intended indexing to be done. You see it's naturally composable because N/2 shows up twice, which is really the perfect amount of times to show up, and as you recall I just defined composable in that way (not to mention it matches the fact that N/2 occurs twice in [0, N)!)
I'm not a math or data scientist so personally when I touch Python or Ruby it's for a system scripting, small CLI tool use case. Would you say Julia fills that corner as well?
you can do it but it's not great. there's no batteries-included standard argparse library and even with the new improvements, there is still some noticeable overhead of a few seconds for some scripts i've written.
I'm curious about this too. The latest 1.9 changes make it so that package precompilations are stored and reused, so the actual workflow will have to be to generate a package (which Julia has tools for), and have the script be just a tiny driver that calls into the package. With that workflow, what kind of latency can we expect? And how consistent would that be?
Same. I went down this road with GPT and was kinda disappointed TBH, since IIRC it recommended the language for the purpose.
I use ABS, Perl, PHP or Nim for most of that stuff but am always looking for new ones to try. (Preferably single binary these days since I have to migrate systems from time to time)
Julia is a language that you write more or less like python, but in cases where python would chase a million pointers to figure out what the type of a variable is, get an instance variable, call a method, etc, Julia (if written correctly) does just in time compilation. The first time it has to call a function on a combination of types it hasn't seen before, it compiles the function for those specific types (into LLVM IR, and then native code) and stores the compiled code so that subsequent dispatches of that function with those types are instantaneous, calling directly into the compiled code. You do not need to add type annotations for this to happen; at runtime, Julia checks the types of variables (which may have even been determined when the function was compiled, eliding the runtime check, if it was used in a context where the types were statically known; see below) to determine which version of a function to call.
This is a strength, as the only thing you have to do to get the performance of fully compiled code (modulo things like GC and bounds checking) is make sure your functions are “type stable”, i.e., their type information can be determined statically. (In fact, in Julia you get better performance by writing more, smaller functions because you'll have more regions within which the types are statically inferrable.) But it's also a weakness because the first time you start up a REPL and call a bunch of functions, each of those functions has to be compiled from scratch, which takes a long time (google “Julia time to first plot”).
Julia has other niceties such as very flexible math (promotion between every pair of numeric types) and a lisp-like macro system with homoiconic code, which make writing numerical code and scientific algorithms highly ergonomic.
> But it's also a weakness because the first time you start up a REPL and call a bunch of functions, each of those functions has to be compiled from scratch, which takes a long time (google “Julia time to first plot”).
But also to further clarify on this, the context of the article is that Julia 1.9 makes it so that for code in packages, the package authors can set up a "call a bunch of functions" section in the package itself, and the compiled results of those calls will be stored as native code. So anything from those parts will have much lower "time to first X".
Matlab that compiles to LLVM. It's fast, pretty easy to write, and surprisingly expressive.
It's also 1-indexed.
I just think it's neat, but it's never solved a meaningful problem for me. My data gets filtered and well-structured upstream of my analysis, so python+pandas is perfect. Visualization tools aren't better in Julia than JS or Python. Really crunchy simulations tend to be written in C or C++.
Of course, Fortran and Matlab had heard of each other. How could they not
have, given so many mutual friends? Many of those friends thought that they
should meet up. Some even thought they would hit it off. But the common view
was that they would dislike each other. Elderly Fortran, who thought deeply
before even considering to act on something, didn't see much to gain in meeting
Matlab. And Matlab, long the darling of the just-go-for-it crowd, was by now
middle-aged, and saw little benefit in dating at all.
But on day, as the joke goes, they both walked into the same bar. Fortran
ambled in slowly carrying a musty book and exuding an air of superiority.
Matlab arrived a bit later and, busy chatting on the phone, did not even notice
Fortran sitting alone in the shadows. But after a while, Matlab did notice
a flickering over in that dark corner of the bar. It was a candle. Intrigued,
Matlab strode towards the light in the darkness.
Fortran, as always, was deep in thought. Buried in the book, sure, but more
than that. Fortran had been thinking about this thing for several weeks. And
not just thinking with a single mind, like others in the bar, but at this point
with ten thousand minds. The idea was only now just starting to take shape,
with these many minds and these many hours.
Now, Matlab was no fool. Matlab could recognize brilliance when it was sitting
right there, with the candle, the book and the furrowed brow. The rest
of the bar faded away. The music, gone. The conversations, the laughter, the
flirting, the clinking of glasses -- all gone. For Matlab had realized something: Fortran was sexy as hell.
So Matlab sat down. And Fortran, to the astonishment of anyone who had been
around for the past half-century, looked up. Wow. In this light, in this
place, at this moment Matlab seemed like an angel descended from heaven.
They said very few words. And the words didn't really even matter much. It
was the tone of voice that mattered. The pauses. The crinkling around the
eyes. After a while, Matlab's hand reached for Fortran's. And that was the
start of something new. Something neither expected. Something their friends
could not have imagined, let alone hoped for.
An attempt at an easy-to-use dynamic language which tries to compile things early as much as possible via aggressive type inference (helped by the type system's design and multiple dispatch), so that you don't leave performance on the table where it's possible to have it, while also having the flexibility of dynamism where you need it.
That's my interpretation, I'd be curious what Julia folks who actually know what they're talking about think of it.
No classes, weak typing despite having multiple dispatch, 1-based indexing, no real DL libraries with comparable baseline to pytorch/weight portability
There are no reasons to not use it if you like gesturing from a distance
No classes - thank the lord. This was a 30 year experiment that failed. Parametric polymorphism is the way and the light.
>weak typing despite having multiple dispatch
I'm struggling to understand, do you mean that it doesn't have type safety? Could you explain a bit?
1-based indexing - I don't care about...
the "no DL libraries" and the community effects around having weight portability is significant I think. I need to spend more time with Transformers.jl.
It isn't weakly typed, it has strong, dynamic typing. GP is either confusing weak typing with dynamic typing or making some other kind of error. To demonstrate, in contrast to say perl, try `1+"1"` in Julia. It'll happily error out for you instead of coercing that string into a number to perform arithmetic with.
I thought more or less the same. Sometimes people think strong typing is type safety. Weak/strong typing is a very confusing concept now, I think dynamic / static is a much clearer division.
Personally I have had a lot of value using the Julia type system for runtime inference in programs, I love it!
Julia:Python::RISC:CISC imo. The first example that comes to mind is that in Python, you'd write:
`A = [ x for x in range(10) if not isprime(x) ]`
In Julia, you'd instead write:
`A = [ x for x in 0:10 if !isprime(x) ]`
It feels like as if having an infix operator for not is redundant - why add a separate & distinct symbol when it means the exact same thing?
There's also the experience of writing iterables/generators, where in Julia you have to subtype the very specific Base.iterate function, but in Python you can just append the magical `yield` to your function.
Overall in Julia I feel like there is less 'magic', but with the added benefit of having a very clear picture of what's happening.
> groupby by two columns is not very common, so DataFrames.jl decided to leave it out from precompilation. For this reason when you run groupby(flights, [:origin, :dest]) native code for such a scenario is not cached. This is indeed a hard design decision for package maintainers. You could add more and more precompilation statements to improve the coverage of cached native code, but it also costs as it would impact: package installation time and package load time.
How feasible is it to let the users provide their own precompilation code? If I know that two-column groupbys are important to me (or some other operation that takes tens of seconds), it would be nice to be able to pay the one time precompilation cost for them to have much better TTFX later.
Thank you. Tim Holy linked to the same in a previous thread (https://news.ycombinator.com/item?id=35885133), but I couldn't understand the context in which I'd want to use it when I looked at it then. Now I understand the idea of Startup packages much better.
These aren't runtime microbenchmarks. They're benchmarks showing how much the JIT compilation overhead has been reduced, leading to faster startups and lower latency.
Not right now as packages, larger communities etc need to be developed. But long term solutions the potential is obvious. A friendly, powerful and high performance dynamic language ought to have a very broad appeal.