Oh wow, I never knew Klarna uses BEAM. I always thought BEAM is vastly underrated and it's even hard for me to convince clients of showing how powerful it (Elixir/BEAM) really is, especially when they're already on something like NodeJS. This will be a nice reference to use in conversations with future prospects.
To add on to your statement that the BEAM is vastly underrated, it's the only runtime that I know of that has enabled multiple small sub 50 people teams to scale to millions of daily active users.
I second this. We serve 500 million requests a month ourselves for one of our clients. We are a sub-20 people team. Previously we were on Ruby (which is awesome to get something out there really quickly) and we were on NodeJS / SvelteKit prior to that. I view all of these platforms as tools like hammers and chisels. You should use the right one for the right use case. We still use all 3, but 90% is now in Elixir, 8% about in SvelteKit, maybe 1% in RoR.
There will be winners in this new AI age, and they will be the ones that are technically better, rather than the most popular or convenient for software engineers
I use Elixir (BEAM-based), with LLMs daily, on a fast growing site (current pace is 5 billion HTTP queries per year on a single front node).
Elixir is well handled by LLMs at this point, and we keep the “easy to scale and maintain” aspect (operational characteristics of the runtime are very good).
I can safely say that total cost of ownership is excellent.
Fair enough! That’s how we track growth, it’s easier to communicate to state sponsors.
Roughly 14M per day, but growing 2.5 to 3 on a YoY basis.
My point stands though: much more serenity in ops with that growth, and LLMs do work nicely, especially when most of the code is stateless by language design.
~160 req/s on average, higher peaks. Not trying to impress anyone though, my point was that LLMs work perfectly fine on an Elixir codebase, and that we get future-proof ops on a single node with zero issues. In my experience, it has been easier than with any other stack I've tried.
To be fair, Erlang runtime semantics (cooperative "async threads") never needed changing to tackle C10k problem, they got that right as part of the early designs before C10k was even on the tasble.
Everyone else just moved towards the same kind of cooperative thread runtime semantics (even if the lingo differs, async in C#, callbacks first, then async in NodeJS, virtual threads in Java, async in C++ and so on).
I think an interesting historical tidbit is that Java actually had M:N threading in early versions (1.1, 1.2, 1.3?) that got ripped out because it was deemed complex and error-prone since people back then often did native bindings naively (And perhaps file/socket API's were less complete?) only for them to reintroduce it again in Java 20 I think it was under the virtual threads name?
> Cooperative scheduling is based on the programmer releasing control manually (much like manual memory management) or having yield points baked into the language, often at function calls or IO events. Processes on BEAM can be preempted, they're allotted a number of "reductions" and when those are exhausted, the process is preempted (the process may also yield earlier).
BEAM is actually cooperative. In userspace, function calls and IO (receive and !) are yield points, and reduction count is simply the number of function calls you're allowed. Something something Computer Sciencey reduction basically just means function call.
It just feels preemptive because as a functional language, you have to call functions all the time. Your loop becomes
loop(I) -> loop(I + 1).
And that will yield once it calls itself enough times.
Native code (NIFs, BIFs, and drivers) has to count reductions and/or yield explicitly.
You can easily hog scheduler threads in native code that does significant computation or does i/o without using the runtime facilities; most people who write NIFs will mess it up at least once, I certainly have. It's much harder to do in userspace code, but if you fully unroll a very large loop that does not call functions or anything else that can yield, you can do it; I don't see how this would happen without specific intent, I've never run into it, but it's possible.
A blocked scheduler can be vaguely fine or terrible depending on how long it stays blocked and what the system is up to. Often it won't be terrible until you do something that needs all schedulers to synchronize, such as operations around code loading, but you can probably wedge the whole BEAM if you try.
I'm not 100%, but i don't think think get(...) is charged any reductions, so it won't yield. gc_bif potentially does garbage collection which does get charged reductions and BEAM can yield during gc.
I'm just adding at the end to do something with the values, so the compiler doesn't ignore them. If you wanted to write BEAM assembly, you can probably make a terrible function there without nearly as much work. smallnum math doesn't usually charge reductions from what I was seeing, so you could probably do something almost useful with that.
Right, my point is that C10k solutions basically revolve around M:N threading models where M being OS threads and N userspace threads, so from the view of the OS/Userspace boundary it's cooperative since the OS does nothing for the N userspace threads, only the M.
That the Erlang runtime hides it from the developer, doesn't mean that it isn't cooperative. So yes, for the Erlang programmer it's preemtive but from an OS/C standpoint it's cooperative (Like toast0 mentions in his sibling comment about NIF's exposing the cooperative nature of the runtime to programmers is telling about why it's considered "complicated" and not recommended practice).
My recollection is that M:N ("green") threads were pulled out because in the absence of a runtime that can guarantee a thread is never blocked, you essentially have to reintroduce all the complexity of an OS scheduler without the benefits of OS-wide context.
"Project Loom" spent much of its time (AFAICT) figuring out how to make every
Java system-level API non-blocking without breaking backward compatibility. The only one they didn't get to for the initial release was Java's "interesting" built-in `synchronized` object monitor construct.
bcantrill has strong opinions on M:N and have always been curious what he thinks about Loom.
Exactly, didn't want to go into the details. I think a key point is that back in the green-thread days they were kinda early on trying to do it and "extra" native threads probably felt like makings things worse without benefit since single-core CPU's with little memory every extra "real" thread would cost.
The C# async system iirc has "blocking workers", ie extra threads that takes care of blocking syscalls, so you're limited by a queue of them "waiting" for their respective "async" workload to finish but the rest of the runtime can chug along concurrently, iirc Project Loom also adopted that, so you have M:N for most code but some extra theads to take care of legacy codepaths (that the Java ecosystem has rewritten a lot of things into pure Java helps since native code calling is probably fairly "obscure" these days).
10k requests/sec is not really the same thing as the C10k problem which is 10k concurrent connections. You might serve much more or less than one request per second per connection, so req/s and conncurrent connections are not really related.
I also wouldn't say OTP was built to overcome the C10k problem. It is a very good solution for lots of concurrent connections (with a suitable OS), but OTP was built to manage concurrency needs in the context of telephone switches. The constraints chosen to meet those needs does work (very) well to meet the challenges of large numbers of concurrent IP connections, but I wouldn't say it was built to overcome those challenges.
The important metrics is peak requests per second. 14M / day sounds impressive, but it’s only a sustained 160 per second which literally any language and framework can handle on the lowest end hardware.
Sure, that's why I highlighted peaks as a more important metric than average. Even an order of magnitude higher peaks over the average is still fairly trivial, compared to where BEAM shines. Maintaining hundreds of thousands of concurrent long-lived connections per server is what it excels at, but also not something that's relevant to 99% of web apps out there.
I'd been looking for an excuse to build something in Elixir pretty much since the language was released, and in just the past year and a half I've had the good fortune of getting to use it to build an in house AI agent orchestration system and it's been an absolute pleasure
I've been building with Elixir for the last 5 years, and comparing to Python or NodeJS - it's a breath of fresh air. Such decent developer experience, so easy to understand how your system works because of actors and well-defined primitives. If you haven't - you definitely should give it a try!
In an AI era I feel like the tiny amount of Elixir code in comparison with other languages, but all that code having higher signal to noise ration vastly contributed to how good LLM-written Elixir code is. Especially in a well-maintainted codebase that relies on a decent db schema and Oban.
It judging by many microbenchmarks and my prior experience with RabbitMQ and a few Erlang systems, performance has often been a bottleneck. I’m happy that it’s got some typing. I’ll take it to ChatGPT about that. But I’m wondering about real world experience. Could you share more? I don’t have any meaningful experience with Erlang hands on and have never used Elixir. But I’d be fancy to try.
In real world, we currently serve 20M+ visitors and 500M+ requests with Elixir. Previously the client was on PHP, let's just say they have the lowest cost compared to any of the competitors in the region just by switching to Elixir.
All that gloating aside, a lot of it depends on the architecture itself, which Elixir allows you to do eloquently. Just the fact that there are no object oriented concepts alone should be reason enough to use Elixir.
Appreciate that. What’s the infrastructure that it runs on? Is it about delegating the workloads to BEAM as opposed to something like Node.js scaled on AWS Lambdas or a single large JDK container or VM instance for a Spring project where the JVM would serve the requests?
48 comments
[ 2.4 ms ] story [ 48.5 ms ] threadElixir is well handled by LLMs at this point, and we keep the “easy to scale and maintain” aspect (operational characteristics of the runtime are very good).
I can safely say that total cost of ownership is excellent.
Roughly 14M per day, but growing 2.5 to 3 on a YoY basis.
My point stands though: much more serenity in ops with that growth, and LLMs do work nicely, especially when most of the code is stateless by language design.
Everyone else just moved towards the same kind of cooperative thread runtime semantics (even if the lingo differs, async in C#, callbacks first, then async in NodeJS, virtual threads in Java, async in C++ and so on).
I think an interesting historical tidbit is that Java actually had M:N threading in early versions (1.1, 1.2, 1.3?) that got ripped out because it was deemed complex and error-prone since people back then often did native bindings naively (And perhaps file/socket API's were less complete?) only for them to reintroduce it again in Java 20 I think it was under the virtual threads name?
Yeah, they called that “green threads”. https://en.wikipedia.org/wiki/Green_thread
BEAM is actually cooperative. In userspace, function calls and IO (receive and !) are yield points, and reduction count is simply the number of function calls you're allowed. Something something Computer Sciencey reduction basically just means function call.
It just feels preemptive because as a functional language, you have to call functions all the time. Your loop becomes
And that will yield once it calls itself enough times.Native code (NIFs, BIFs, and drivers) has to count reductions and/or yield explicitly.
You can easily hog scheduler threads in native code that does significant computation or does i/o without using the runtime facilities; most people who write NIFs will mess it up at least once, I certainly have. It's much harder to do in userspace code, but if you fully unroll a very large loop that does not call functions or anything else that can yield, you can do it; I don't see how this would happen without specific intent, I've never run into it, but it's possible.
A blocked scheduler can be vaguely fine or terrible depending on how long it stays blocked and what the system is up to. Often it won't be terrible until you do something that needs all schedulers to synchronize, such as operations around code loading, but you can probably wedge the whole BEAM if you try.
I'm just adding at the end to do something with the values, so the compiler doesn't ignore them. If you wanted to write BEAM assembly, you can probably make a terrible function there without nearly as much work. smallnum math doesn't usually charge reductions from what I was seeing, so you could probably do something almost useful with that.
That the Erlang runtime hides it from the developer, doesn't mean that it isn't cooperative. So yes, for the Erlang programmer it's preemtive but from an OS/C standpoint it's cooperative (Like toast0 mentions in his sibling comment about NIF's exposing the cooperative nature of the runtime to programmers is telling about why it's considered "complicated" and not recommended practice).
"Project Loom" spent much of its time (AFAICT) figuring out how to make every Java system-level API non-blocking without breaking backward compatibility. The only one they didn't get to for the initial release was Java's "interesting" built-in `synchronized` object monitor construct.
bcantrill has strong opinions on M:N and have always been curious what he thinks about Loom.
The C# async system iirc has "blocking workers", ie extra threads that takes care of blocking syscalls, so you're limited by a queue of them "waiting" for their respective "async" workload to finish but the rest of the runtime can chug along concurrently, iirc Project Loom also adopted that, so you have M:N for most code but some extra theads to take care of legacy codepaths (that the Java ecosystem has rewritten a lot of things into pure Java helps since native code calling is probably fairly "obscure" these days).
I also wouldn't say OTP was built to overcome the C10k problem. It is a very good solution for lots of concurrent connections (with a suitable OS), but OTP was built to manage concurrency needs in the context of telephone switches. The constraints chosen to meet those needs does work (very) well to meet the challenges of large numbers of concurrent IP connections, but I wouldn't say it was built to overcome those challenges.
I have done several experiments where, the only difference are a couple of paragraphs describing the technology stack for the AI tools.
Hardly any different than changing the target CPU architecture on a compiler output.
Quite sad, as now the fun of why language X instead of Y is gone.
Progress!
I'd been looking for an excuse to build something in Elixir pretty much since the language was released, and in just the past year and a half I've had the good fortune of getting to use it to build an in house AI agent orchestration system and it's been an absolute pleasure
In an AI era I feel like the tiny amount of Elixir code in comparison with other languages, but all that code having higher signal to noise ration vastly contributed to how good LLM-written Elixir code is. Especially in a well-maintainted codebase that relies on a decent db schema and Oban.
All that gloating aside, a lot of it depends on the architecture itself, which Elixir allows you to do eloquently. Just the fact that there are no object oriented concepts alone should be reason enough to use Elixir.