Hacker News new | past | comments | ask | show | jobs | submit login

EEx does macros well. Note that EEx.function_from_x makes you declare the name of the function it instantiates, so it's searchable.

GenServer is kind of a "bad example" of a macro. It creates a totally hidden function. Though I guess to be fair 99.9% of the time you really shouldn't be writing genservers in elixir.




>It creates a totally hidden function.

Are you talking about child_spec/1? `use GenServer` honestly provides some pretty good QoL enhancements over `@behaviour GenServer`. I don't know when or if there was a change to recommending `use GenServer`, but I was pleasantly surprised when I started using it. Especially because I'm not writing new GenServers all the time, having guard rails is nice to avoid being slowed down by remembering all the boilerplate: init/1, handle_x/y, child_spec/1

It's even nice enough to give additional information about why its crashing so you don't get confused why e.g. the handle_call function is working. I make this mistake almost every time I write a new GenServer:

    def handle_call(:ping, state) do
      {:reply, :pong, state}
    end
Before, you would just get this meaningless gen_server stacktrace that basically just says "it crashed lol". Now you get the additional hint of "hey you didn't define handle_call/3" when it crashes.


Last I checked you need to do init/1 by hand. Honestly being more transparent about creating child_spec like `GenServer.create_default_child_spec` would be great. The rest of use GenServer is pretty good, IMO (this is why I wrote 'kind of a bad' macro). I normally hate default implementation of functions but defaulting with an error saying "you forgot this" is reasonable.

Edit: just checked, it warns if you don't... I have compiler warnings as error on all my personal projects so that's why I thought you had to do it manually


> 99.9% of the time you really shouldn't be writing genservers in elixir

Woah! Why is that?

I've written at least one GenServer in nearly all the Elixir projects I've worked on. They seem like one of the base building-blocks to me. Also, if you squint a bit, you'll see that many libraries you're working with are essentially exposing a GenServer interface, with a few extra features.


The only thing I can think of is when developers overuse genserver not realizing it's involving another process and then serializing your requests (behind a queue even) into its mailbox. That may be desirable for some but in-process handling can sometimes be the better choice.


I've never written a genserver in any professional project I've worked on.... Well, except maybe for a trivial 5 line wrapper enables a global ets.

> Also, if you squint a bit, you'll see that many libraries you're working with are essentially exposing a GenServer interface, with a few extra features.

I never said you shouldn't be using GenServers. I said you shouldn't write GenServers.


> I've never written a genserver in any professional project I've worked on

Wow, that's quite surprising! We use GenServers quite extensively at work. The most interesting usage is probably this application, which uses GenServers to model the state of real-world hardware (in this case, transit stop countdown clocks): https://github.com/mbta/realtime_signs/blob/main/lib/signs/s...

Curious what you use instead of GenServers. Plain processes? Agents? Something else? GenServers are probably overused but I still reach for them when I need a concurrent bundle of state with a specific interface and behavior.

I think we're generally trying to write more GenStages than GenServers at the MBTA these days, but that's mainly because we're moving towards being event-driven. GenServers still have their place.


I use other people's GenServers. Never agents. Tasks for concurrency (but almost never those either, just Oban). Ets for stateful datastore.

Pedantically I guess gen_statem isn't a genServer, but I'm generally referring to the pattern, which gen_statem is.

Mostly webdev, so, that stuff is all taken care of for me.

MBTA is modeling reactive systems that exist in the real world, so yeah, that's a place where i would be surprised if you didn't use genservers, though I suppose you could just use a database-backed state machine. Surprised you use GenServers instead of gen_statem (though you might have been, as I was, less than pedantic)


Ah I see. Yeah, Phoenix definitely gives you a lot out of the box. I think for webdev it's probably enough. This is a pretty straightforward Phoenix app for example, only uses a couple GenServers: https://github.com/mbta/arrow

The question of when to introduce databases is a really interesting one and also one I'm still coming to terms with a few years into writing Elixir full-time. Many of our apps don't use them at all, but I think at least a few could benefit from more structure and durability in their datastores.


I see after I posted you edited to mention Agents, Tasks, Oban, and gen_statem.

What do you use gen_statem for? I'll admit its intended use-case has always seemed pretty narrow to me, but maybe I just don't understand it very well.


I don't use gen_statem, its clunky as hell, does a weird thing with response aliasing, and it's even harder to read than Genserver. Plus I don't need it.

I understand why the callbacks in gen_statem are so messed up though, they want you to be able to organize your code by state. Since Erlang and (weakly) elixir requires you to group handle_x callbacks linearly in the module, you can't group by state unless you have a `handle_anything`. It's all so messed up :(


Are you actively trying not to use GenServers in your application code or is it just that you have a set of 3rd party dependencies that remove the need?


Yes


There's no reason not to write GenServers if you need them. Just don't use them as a class replacement :) But if you need to model some simple runtime state or concurrency, that's what they're there for.


use Tasks for simple concurrency, Jose has built an awesome module for us.


Oh I agree! I have a couple of gen servers acting as very simple message queues—the kind that don't matter if they are nuked during a deploy.


> Also, if you squint a bit, you'll see that many libraries you're working with are essentially exposing a GenServer interface, with a few extra features.

Like LiveViews :)




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: