7 ms·
Re: those infrastructure components written in Erlang not attracting Erlang users in any permanent way—it’s not so much that they have more users than contribut
by derefr 7y ago
Re: those infrastructure components written in Erlang not attracting Erlang users in any permanent way—it’s not so much that they have more users than contributors, but rather just the bare fact of them being infrastructure components. A black-box infrastructure component can’t be a “killer app” for a language.
Rails is a killer-app for Ruby because it’s a framework; a developer who uses Rails in their project is inherently a Ruby developer. People learned Ruby to use Rails.
But you don’t need to write a single line of Erlang to use CouchDB or Riak or Ejabberd. They’re self-contained. (And, in fact, scriptable in not-Erlang languages; CouchDB for example spawns and manages a JavaScript engine to run its “design documents” on.)
Thus, these infra components becoming popular doesn’t cause there to be any more Erlang programmers than there were before.
In that sense, I would say that Erlang has never had a true “killer app” yet. (Outside of a few specialized domains, like telecom and fintech, where the OTP platform itself is Erlang’s killer app.)
And even in the places where Erlang itself is a “killer app”, it’s also not so much a viral thing that the whole industry gets excited about; rather, Erlang is more of a “secret sauce.” Nobody’s learning Erlang in their basement because their friend told them it’d be the next big thing in their industry.
How does anyone learn Erlang, then? Well, usually, engineers are forced to learn Erlang at one job (where the system architect there decided on it); they get to liking it; and then, when they later become a system architect themselves in another job, and choose a stack, they choose Erlang, thus forcing a new wave of engineers to learn it.
- seisvelas 7y agoI don't think this counts, but I imagine a significant portion of people learning today learn it because they are Elixir programmers.
- derefr 7y agoIt actually might. You can write Elixir for years without learning Erlang syntax, but if you’re a perfectionist, you’re gonna notice bugs in your Erlang deps just as much as in your Elixir deps, and you’re gonna want to submit PRs for them, and that leads you to learning Erlang. Also, IMHO, if you want to have a solid foundation when learning Elixir, it’s simpler to learn Erlang/OTP as a cohesive whole (from e.g. ferd’s Learn You Some Erlang book) and then “chase it” with an Elixir book; rather than trying to understand OTP through the lens of how it’s exposed in Elixir. Everyone in the elixir-lang Slack that asks how to get started learning Elixir, I tell them to learn Erlang first. So “being a simpler jumping-off point to Elixir knowledge” is also kind of a killer app. I would guess that this is only true in cases where there are killer apps for a secondary language on a runtime, but no killer apps for the primary language. Like, I wouldn’t say that Clojure is a killer app for Java. But then, is that because Java is already popular? Or is that because Clojure developers don’t necessarily end up knowing Java as well as Elixir developers end up knowing Erlang?
- andrenth 7y agoThere's a lot in the Erlang stdlib that is not available in a wrapped form from Elixir's. As of now I think you need to know at least a bit of Erlang (e.g. to understand the stdlib docs) to be able to use Elixir more productively. I'm not sure if Elixir developers plan on providing complete wrappers for Erlang's stdlib in the future. This is in fact one criticism I have of the language, because even for simple things like serialization you have to drop to Erlang and call :erlang.term_to_binary(), which feels a bit second-class. I really like the language otherwise though.
- klibertp 7y ago> you have to drop to Erlang and call :erlang.term_to_binary(), which feels a bit second-class. Well, to be honest, :erlang module has many functions, many of which have literally nothing in common with each other. That's because that module is imported by default in Erlang - it's like `__builtins__` in Python. Exposing it as `Erlang.term_to_binary/1` module/wrapper wouldn't really fit well with the rest of the Elixir stdlib, which is (commendably) very neatly organized, with each module doing just what its name suggests. On the other hand, wrapping just `erlang:term_to_binary/1` and `erlang:binary_to_term/1` in a `Serializer` module would be probably ok, but then there are tens of such modules you'd be pressed to implement for `:erlang` alone, and it gets worse with a rather extensive Erlang stdlib and OTP apps. One of the Elixir selling points was (and probably is) the simplest possible FFI to (and back from) other BEAM languages (like Erlang or LFE). Calling an Erlang function is exactly the same operation as calling Elixir one - literally, the only difference is that, syntactically, most Elixir functions live in CapitalizedModules. That's it - there's no runtime overhead, no difference in handling arities, arguments, guards, exceptions or return values. The feeling of second-classiness is definitely there - the argument order and naming conventions (for args and functions both) differ, and the Erlang/OTP docs (used to - not sure it's still the case) look less than appealing in comparison with Elixir. But, realistically, Erlang's stdlib is a result of 30 years of accretion of tools and it's not small by any means. Wrapping entirety of it would take many years at least and it would consume a lot of developers' time, possibly hindering the development of the language itself. Elixir is still a young language, without a very large following and secure position in the industry - it's dangerous to divide your forces and pursue to many goals at once in a situation like that. To summarize: in my opinion, the feeling of second-classiness is not a serious enough problem to invest resources in making it go away, at least at this point. While they may look less than pretty, Erlang/OTP docs are very well written and the functionalities tend to be rock solid. The stdlib is... not the prettiest in terms of organization, owing to decades of evolution, but with the FFI as trivial as it is, it's more than usable from Elixir. I think that was the original plan: the stdlib of Elixir was intentionally kept small and very clean, which was supposed to be Elixir's appeal, but it's only possible if reaching for the functionality missing from stdlib is simple. Just a hunch of mine, we'd need to ask Jose to know for sure :)
- brightball 7y agoAgreed. This is actually why I think the killer apps for the BEAM are 2 Elixir based ones. Nerves, which is already being heavily used for embedded systems and only getting stronger thanks to the likes of Scenic. And Phoenix LiveView (not released yet), which I personally think pulls the strengths of Phoenix together in a way that can truly disrupt web development. It’s very difficult to do what LiveView is doing on other tech stacks. The combination of those two will likely fuel things long term.
- woeye 7y agoIsn't Blazor in ASP.NET basically what LiveView is in Phoenix? https://docs.microsoft.com/en-us/aspnet/core/blazor/?view=aspnetcore-3.0 https://docs.microsoft.com/en-us/aspnet/core/blazor/?view=as...
- strmpnk 7y agoNot quite. Elixir code isn't set to the client as Blazor does with .Net IL. LiveView leverages a small runtime loaded onto the client which uses websockets to load deltas. While it sounds expensive at first, the template engine has been extended to support delta tracking and the process model that Elixir inherits fits well for a heavily interactive system with thousands or millions of users. The end result is that it's a very lightweight client setup and outside of using the template system properly, there are no gotchas with learning which code runs where.
- PhilippGille 7y agoThere's two distinct Blazor development models: Client side Blazor uses Mono (running as WebAssembly) to run C# code on the client side, but server side Blazor only updates the DOM via SignalR (like WebSockets), which works like LiveView. I didn't use Blazor yet, but that's what I understand from the documentation.
- strmpnk 7y agoThanks for the correction and clarification. The early parts of the docs didn’t make this distinction clear so I assumed the WASM half was required.