Out of the box, there’s zero security or audit trail. Building that properly isn’t trivial and, even with it in place, many corporate infosec teams would have fits if you suggested that engineers can make arbitrary inspections/modifications to a running production system.
Where it could be appropriate, often you’re running the code in autoscaling containers or something similar. Modifying one instance then is rarely anything but a terrible idea.
Where I have used it is for things like long-running internal batch systems that run a single instance and never touch any sensitive data. Connecting a REPL in those cases is much more flexible and powerful than, say, building a dashboard UI or a control API over http, and you get it for free.
The Lisp REPLcis still superior because it comes with more stuff, like DECOMPILE, INSPECT and so on that can only exist because the language is essentially a compiler even at runtime, which can also be a problem for sensitive domains… but in Java you can do all those things using the IDE so the distance between what is possible in Lisp and a language with good IDE support like Java and Kotlin is now negligible in my opinion.
Unfortunately JRebel has killed their free tier, so I'd now point unwilling-to-pay programmers to something like https://github.com/JetBrains/JetBrainsRuntime which is IntelliJ/Eclipse/whatever-independent. I haven't tried it myself yet though... Given they only address the biggest class reloading concerns, I doubt it's actually comparable to JRebel for business-world Java. JRebel handles among other things dynamic reloading from XML changes and reinitializing autowired Spring beans that other classes use for dependencies.
*Caveat, I've been out of the professional Java grind for a while, I'd be pleasantly surprised if some new version that's come out contradicts me.
Though nowadays the IntelliJ debugger with the OpenJDK is enough for me. I know what works and what doesn’t so I rarely feel frustrated.
Maybe emacs lisp works that way?
(define (displayln msg) (display msg) (newline))
(define (inner x) (+ x 1))
(define (outer x) (inner x))
(displayln (outer 5))
(define (inner x) (+ x 2))
(displayln (outer 5))
Universal hot reload is really a messy beast though. For every "yeah we can just reload this without re-init'ing the structure" there's another "actually reloading causes weird state issues and you have to restart everything anyways" thing.
I've found that hot reloading _specific targetting things_ tends to get you closer to where you want. But even then... sometimes using browser dev tools to experiment on the output will get you where you want faster than trying to hot reload clojurescript but having to "reset" state over and over again or otherwise work around weirdness.
I think this flow works well in Emacs though because you're operating on an editor. So you can change things, press buttons, change things, and have a good mental model. Emacs Lisp methods tend to have very little state to them as well (instead the editor is holding a bunch of exposed state).
Meanwhile React (for example) has _loads_ of hard-to-munge state that means that swapping one component for another inline might be totally fine or might just crash things in a weird way or might not have anything happen. Sometimes just a full page refresh will save you thinking about this
It's a shame that other scripting languages that theoretically have the capabilities to do this don't do this (looking at you, node! Chrome dev tools are fine but way too futzy compared to `import pdb; pdb.set_trace()` and "just" using stdin)
I do also use Emacs, and with Emacs Lisp `trace-function` means you can very quickly get call traces in your running instance without having to pull out a debugger and the like. Not like you can't trace functions with `gdb` of course. But the lowered barrier to entry and the ability to do in-process debugging dynamically means you just have access to richer debugging tools from the outset.
It is obvious why it is not really used or recommended as it really falls flat in a team setting, mostly even when 2 people are involved. But fixing bugs live as they happen and then spitting out a new .exe for clients is still a lot faster than modern alternatives. Far more dangerous too.
Fun fact is that giving AI repls also reduce error rates so much that you can save up to half the tokens/time or more.
I read some Erlang article saying that hot swapping is not actually very useful in production because of some reasons, and instead a blue-green deployment is preferred. Can't find the link atm. This was close: https://learnyousomeerlang.com/relups
Compare to this comment: https://news.ycombinator.com/item?id=42405168 Hot swaps for small patches and bugfixes, and hard restarts for changing data structures and supervisor tree.
But the nicest part is that once you connect to the production application, apart from network lag it's no different than if you were developing and debugging locally on similarly specced hardware to the server, you have all the same tools. Many of the broader activities around "debugging" don't need to happen in a paused thread that was entered with an explicit breakpoint or error, they can happen in a separate thread entirely. You connect, then you can start inspecting (even modifying) any global state, you can define new variables, you can inspect objects, you can define new functions to test hypotheses, redefine existing functions... if you want all requests to pause until you're done, you can make it so. Or if you want to temporarily redirect all requests to some maintenance page, you can make that so instead. A simple thing I like doing sometimes when developing locally (and I could do it on a production binary too) is to define some (namespaced) global variable and redefine a singly-dispatched method to set it to the self object (possibly conditionally), and once I have it I might redefine the method again to have that bit commented out just so I know it won't change underneath me. Alternatively I can (and sometimes do) instead set this where the object is created. Then I have a nice variable independent of any stack frames that I can inspect, pass to other method calls, change properties of, whatever, at my leisure without really impacting the rest of the program's running operation. Another neat trick is being able to dynamically add/remove inherited mixin superclasses to some class, and when you do that it automatically impacts all existing objects of that class as well. Mixin classes are characterized by having aspect-oriented methods associated with them; you can define custom :before, :after, or :around methods independent of the primary method that gets called for some object.