Julia syntax is so nice and consistent, a welcome relief from the oddball things that are a result of a mishmash in numpy/python/pandas world. Damn shame it hasn't taken off as much. While Multiple dispatch is great, I feel it doesn't scale to large programs. Julia syntax + compile time polymorphism (like C++) + gc + single executable would be a killer language for HPC and Scientific computing.
Julia gets a lot of things right. Going back and forth between Julia and Python, there are things to appreciate about either, but I think it is very clear that Julia has really moved this space forward.
Unfortunately, it has also done so by ignoring a lot of hard-earned wisdom. My personal pet peeve is Concepts/Constraints ala C++. C++ added this after years of pain to make templates more manageable. It's a feature of many different languages, from Haskell to Rust/Mojo to C++. So there seems to be a pretty broad consensus that something like this is good to have. Yet, I don't see Julia getting them any time soon...
The same team for interfaces was first doing the caching of binaries to solve time to first plot for v1.9, and then for v1.12 just finished up the binary trimming. Another big thing landing soon is improvements to escape analysis which should get rid of many small allocations. Interfaces is something next on the list, but the team just choose a few other things as higher priority.
Julia hasn't "taken off" but I venture that it might be "getting there" with recent improvements in being able to generate (relatively) compact binaries.
More generally, I wonder whether the future of software development will be less enamored with new/upstart languages, and more amenable to languages with truly novel offerings which have also had decades to mature. Beyond Julia, OCaml, Erlang/Elixir, and potentially Scheme / Common Lisp are a few of the others I can think of.
CERN is researching into adopting Julia, for the Next Trigger initiative.
Which is quite relevant, 20 years ago when I was there, and Grid Computing was taking off, Python was starting it's baby steps in HPC as scripting language for build tools (CMT), and data analysis.
v1.12 already has merged binary trimming, so we have cases with a full ODE solver that's not 10's of MiB (mostly linear algebra BLAS stuff left), and hello world is 900kB. See https://info.juliahub.com/blog/new-ways-to-compile-julia-blo.... It should release in like 6 months.
Nice, that should help with the nonsense relative-performance-test people run that pans Julia due to the load-times. I suspect it scared away a few potential users evaluating the utility of the language.
I think implicit broadcasting in numpy made it difficult for a lot of people to grasp the concept. When it works, it works like magic. Sometimes it is intuitive. But then it stops working and because you actually never learned the logic behind it it turns into a mess. I had been (mostly successfully) using broadcasting for years before I started playing around with Julia and only then the concept clicked. I think their choice of being explicit here makes a lot of sense.