I don't understand why people keep talking about multiple dispatch like Julia invented it. You can do that in many other languages, even languages designed for numeric computation. What's cool about Julia is that it has brought 90s compiler technology to scientific computing: a field which still thinks MATLAB is a really good way to develop and communicate scientific ideas.
I think a lot of the reason is that most previous incarnations of multiple dispatch give you slow multiple dispatch, and a different way to define functions that don't have multiple dispatch and are faster to call. As such MD tends not to be commonly used in languages where it exists. Julia isn't the first to have multiple dispatch, but it is the first where everything is multiple dispatch. The result of this is that we put in a ton of work to make multiple dispatch fast, and all the APIs are designed around it, which gives a very different feel to the language.
As I noted around a year ago, there's a little puzzle here: in principle Julia's approach was available in the 90s, when decent JIT technology was becoming available. So why was MD given only toy-like treatment before Julia? (MATLAB did get a JIT in 2002, so it's not really a toy, but the performance was not good enough to prove the concept)
My impression is that the complained-about inflexible division between concrete and abstract types was pretty necessary to achieving good results in practice, but this division is alien to the expressiveness-loving lisp culture.
MATLAB is fairly impressive for development. It works ok for communication. What are the areas for improvement you have in mind, and what are the alternatives I should consider?
The 2 biggest problems with Matlab are price and lack of community. The price is an issue even if you can afford it because it makes it really hard to deploy widely since anyone who wants to run the code also needs to be paying Mathworks. This closely ties in to the community issue. Almost all Matlab libraries are proprietary (and written by MathWorks). If you find a bug, your only option is to file a report and wait 3 years for it to not get fixed. In an open source ecosystem, you can dig into the code and fix the problem yourself if you need to.
I once got on a call with MATLAB compiler engineer and had him fix a bug in an afternoon. I should have written a blog post about it but it is too late, I have very little recollection of the details. It was kinda awesome though. How often can you do just call up someone to fix the compiler. To be fair, I worked for a very large Fortune 50 company and had over 10k licenses of MATLAB.
Also, it is interesting that outside of SV and HN crowd, we thought MATLAB is awesome. We had all the toolboxes, there is a gazillion of them. Even obscure RF related stuff. You just can't find a library for something like a phased array analysis. May be you can, but it won't be as high quality and industry proven as this: https://www.mathworks.com/products/phased-array.html
Another topic that we often understood is that when a jet engine is hoisted up for testing and you need to get something fixed or have a question about in next 2 hours, you need commercial software support to help you out. Gravity of situation cannot be understated. It is stressful. There is OSS software support, but MATLAB is on another level. Absolutely outstanding support. Companies like Apple and SpaceX rely on MATLAB heavily outside of software engineer orgs.
If you are basically using finished code, your own code is mostly a top-level script, and you mainly need polished functionality and professional support, then Matlab toolboxes and support are very good.
But the language itself is primitive compared to OSS alternatives, so when you need to develop your own software on a larger scale, it falls short.
There is of course a sliding scale between 'user' and 'developer' in any language, but I think that the closer you are to the 'user' end of the scale, the better Matlab looks.
I don't find the language primitive at all. It just has a poor deployment story that consists of "autogenerate to hidious unreadable c c++ cuda or vhdl, or writ it again yourself". It is amazing the autogen even works, which it really does, but that code...
But back to the language. I can typically do a line by line syntax change to get pytorch (julia is a but different but not too much). The resulting matlab sometimes runs faster too. If you avoid globals and write everything vectorized it is all really clean. OO code would be hidious, but I try not to use OO or goto in any language unless there is no other option. I like that arguments to a function are pass by reference unless you write to them, in which case it does a smart copy. Julia syntax looks different sometimes because you don't have to vectorize for performance as in matlab or python, in fact sometimes you shouldn't.
All in all, if the deployment story was solved, I probably wouldn't be trying out Julia. In fact I still prototype in Matlab before implementation in pytorch or julia, its just eaier to get that first thing working.
Everything is a matrix, that's horrific (especially the Nx1 vs 1xN ambiguity which pops up all the time). No default argument values (and until recently, no support for keyword args, though they have a bad version now), meaning half your code lines are input parsing. Forced vectorization does the opposite of making code clean, it makes complicated code a lovecraftian mess (unlike Julia, which lets you vectorize efficiently at the top level.) Every function must live in an m-file(#$@&%*!) Everything in your path is in scope :(
The varargin/varargout/nargout mess, with outputs specified in the signature line, instead of proper return statements.
Also, for a language that requires vectorization for performance, there should really be a proper `map`, instead of the mess that is `arrayfun`/`cellfun`/`structfun`. Their arrays are super-limited, no mixed-element arrays (use cell-arrays!), so `[3, [4,5,6], 7]` is just concatenated, while `[3, [4;5;6], 7]` errors (and check out what `[3, "hello", 7]` does(!) or `[3, 'hello', 7]`)
Poor support for integers, 2 is a double(!), and `int8(3)/int8(2) == int8(2)` (yikes.)
Their OOP is actually not that bad, though it's slow. And their graphics system is pretty ok.
Also annoying: now you have two types of strings, old-fashioned 'abc', and new-fangled "abc", which are very different, and sort-of, half-way work together. Though I think moving to the new strings is actually a good move, but painful now.
I want Nx1 and 1xN to act differently. It is a form of type safety. I would be fine with defaulting to Nx1x1...x1 as well and have everything be a tensor where you specify which side all the singleton dimensions proceed. Disagree, but they could do this better, matrix just happens to cover most cases.
I consder default arguments code smell in every language. Or really more foot-guns than code smell. I discourage them whenever anyone will listen. Disagree on this one.
Not super fond of kwargs either but they certainly have their place. If they are going to support it, they should do so well. Agreed
I actually prefer to read vectorized notation, I wish it was more consistently performant in Julia, sometimes the for loops run faster, but they take longer for me to read and understand. The exception is if there is an einsum in there somewhere, or the equivalent auto expansions in Matlab, that takes me a few. Personal preference I guess?
I've never come across a use for mixed arrays, but everytime I come across an api that returns them I begin cursing. Agreed
Typecast rounds instead of floors? That is kinda odd, but not wrong I guess? I haven't ever run across this because I use floor or round explicitly.
In 20 years ivrmever noticed the string thing. I'll jave to read about that, thanks!
Yes. Of course you do. The problem is, you never know what you are going to get. What you actually want is a vector, but Matlab has no such thing, so then you have to write your code in a way that anticipates either Nx1 or 1xN, and handle both. Sounds, simple enough, but I have a lot of code lines dedicated to checking and handling row/column orientation. If only there were real 1D vectors!
> I consder default arguments code smell in every language. Or really more foot-guns than code smell.
I don't really understand what you mean here, but the problem is that in Matlab you handle 'default arg values' with `if nargin < 5, par5 = default_value` etc. It's just worse, and it's perfectly idiomatic matlab.
> I actually prefer to read vectorized notation
My argument is actually that vectorization works much more cleanly in Julia. If in Matlab you have `foo`, which calls `bar` which calls `baz`, etc, then you must make sure that each of these explicitly can handle array inputs. You have to think about arrays on every level, including what happens with axis broadcasting (and probably checking 1xN vs Nx1 orientations on multiple levels). In Julia, on the other hand, you can write your functions `foo`, `bar` and `baz` to handle scalar arguments, and then you vectorize the whole thing with `foo.(args)`.
So vectorized code in Julia is much simpler to write, and also simpler to read. (Just to be clear: writing vectorized Matlab code and vectorized Julia code are things I do all day, every day, so I have a decent basis for comparison.)
> sometimes the for loops run faster, but they take longer for me to read and understand.
Loops vs broadcasting should be basically the same for performance, but sometimes you can get extra performance from a loop, by exploiting algorithmic advantages. But a simple vectorization is just a dot away.
> mixed arrays
It's something one generally tries to avoid, but they are often necessary for passing along arguments to inner functions etc. Tuples are great for this, but it doesn't exist in Matlab.
> Typecast rounds instead of floors?
It's odd, but I don't mind. Mainly, I am annoyed that integers are not well supported in Matlab.
> In 20 years ivrmever noticed the string thing
The "strings" are new, a couple of versions back. Of course, you will notice that "string" is actually a 1x1 matrix of strings, so length("hello") equals 1. And indexing into strings is actually indexing into the array of strings. So `str = "hello"` then `str(1)` returns the string itself, and `str(2)` errors.
I don't want 1d vectors to exist, and ideally, I'd want everything to be infinite rank tensors with som clean notation for singletons x N or N x singletons or NxMx singletons etc. The dimension mismatch has caught more errors for me than I could ever possibly count. My functions don't check to see if something is scalar or columnar, I let the error tell the user their mistake. And using a column vector when a row vector is expected is a math error. And the same function usually works for scalar or vector or matrix for the most part.
I do like the dot notation and Julia in general, but I just don't have many complaints about Matlab. Auto expansion maybe. They do allow helper functions in the same file now, finally.
Well, in many cases, dimension mismatch causes errors, not just lets you find them. And in fact, with the recently-ish introduced broadcasting behavior, you don't even get an error, just surprising dimensional-increasing behavior.
As for "math error", it often is nothing to do with that, you just want a list of values, what should such a list be? Column or row? And often, I'm the user, calling into other people's code that sometimes wants columns, sometimes rows, and behaves surprisingly different for each (sometimes with errors, sometimes not.)
There should be a convention in Matlab about whether fundamentally 1d structures should be columns or rows, but there isn't. (Just for example: reductions work down columns, except when the columns are length-1, while iteration only occurs along rows...)
This is a huge, huge problem in my daily work. I can give many examples, but am on my phone now.
Let me put it like this: if I could change only one single thing about Matlab, I would introduce a proper vector.
> Also, it is interesting that outside of SV and HN crowd, we thought MATLAB is awesome. We had all the toolboxes, there is a gazillion of them. Even obscure RF related stuff. You just can't find a library for something like a phased array analysis. May be you can, but it won't be as high quality and industry proven as this: https://www.mathworks.com/products/phased-array.html
As a matter of fact I find most comments about Matlab here supportive. I actually do find Matlab to be a polished product, specially in the IDE and documentation part. In the end you use whatever saves you time, that's why most people that use Python have chosen Python: because of the libraries. I have personally found Matlab toolboxes quite simplistic for my own purposes. Most people, myself included, actually trust more that SW is used and deployed in the millions than any kind of support.
>How often can you do just call up someone to fix the compiler. To be fair, I worked for a very large Fortune 50 company and had over 10k licenses of MATLAB.
Very often if you're paying upwards of $10,000,000 in licenses a month.
The community issue also means that code sharing is very limited compared to other languages. If it isn't in one of the Mathworks toolboxes (that you can afford), you'll probably have to end up at the Mathworks "File Exchange" site, where you take a look at the existing code that "solves your problem" (in a non-generic, buggy way), cry your eyes out, and end up implementing it yourself.
Both fair points. And the alternative? I like Julia, but i still haven't found a workflow I like. I still typically prototype in Matlab and deploy in c++/cuda.
I'm trying to work entirely inside the repl. Its not bad but for large stuff the lack of go to definition hurts, and there isn't an easy way to copy commands that worked into my module. I'll try vscode next. That is what I use for c++/c. Is there anything else I should try?
I use VSCode+repl which works well for me (although the language server for Julia isn't as good as a C/C++ one). The one other workflow to look at is Pluto notebooks. They're similar to jupyter notebooks, but they track cell dependencies and automatically dependent cells, to guarantee that you maintain consistent state.
I worked in pluto for a bit. I should give that another shot. If there is an easy way that someone simply runs the script, and it doesn't try to fire off plots, that would be the ultimate in self documenting code.
If you read the small print on that license it almost certain says something like "for hobby use only. Not for academic or commercial use". A MATLAB license that allows commercial use starts at around 2000$/user.
Which Lisp? What applications do you think it is slow at? Hint: It can be faster than C due to compiler macros, and infact, its regular expression engine is much faster than Perl's, which is written in C, just to give a concrete example.
My experience is that even writing "idiomatic", CLOS heavy code tends to be faster out of the box than most dynlangs out there. Writing close to C++ code requires a lot of manual work and is probably not even worth it unless for optimizing hot-paths.
