(And btw when I say Erlang I also mean Elixir, they both share the same VM so most things apply to both).
But Erlang is a the secret sauce (so to speak) for a high concurrency fault tolerant backend. "Fault tolerant" should probably go first. The reason is the observation that the higher the concurrency or complexity in the system, the higher the need for fault tolerance. If you only serve 10 connections and the backend segfaults, 10 clients have lost connectivity, you get 10 angry phone calls. If you serve 1M connections and your backend segfaults, you get 1M phone calls. Exaggerating here of course to get the idea across. But this is not just a marketing gimmick, this translates to money in the pocket in operational costs. Some subsystem crashes and restarts? Fine, let it do that if it is 4am, no need to wake people up, will fix in the morning. I've heard of stories of subsystems crashing and restarting for years with the main service staying up.
The lack of centralized state might seem minor but unless you have been debugging a shared memory systems with locks, threads and manual memory allocations, with classes that have pages of attributes defined, and trying to understand why it crashes on Wednesdays only at customer A's site, it is easy to miss the benefit. This comes through using small lightweight processes with an isolated heap and also using functional constructs.
Then in general, the ability to reason about concurrency using OTP constructs (Erlang's standard library) and processes is like going from Assembly to C in terms of distributed and concurrent systems. Can express ideas at a higher level. This means having less code to look through and maintain.
There are other niceties like good garbage collection performance, awesome tracing and hot code loading capability (used this a few times, so started to appreciate it more now).
Now, individually all of those features can probably be found in other systems and frameworks, but they are just not integrated or not quite there -- Java has code loading but it is not the same. Can always spawn an OS process to get fault tolerance, but can only spawn so many before system falls down, Go has goroutines but they also share memory, so fault tolerance is not there. Other languages have garbage collection, but most still have to stop the world sometimes and so on.
. What kind of web applications are you building with it? I'm asking what kind of web apps or scenarios do you think Elixir is particularly well suited for?
I saw a thread on Elixir a couple of days ago and it piqued my interest and I saw a couple of videos that were posted there, one from some Ruby Conf that claimed that Elixir was giving better results (in request time) than rails. He never explained how that was posible or what would have been the results if he would have been using a cache (it's always faster to hit an in memory cache than hitting a db that has to touch disks, so if he speeded things up without a cache he would speed things even more with a cache). Then I watched another videos from some conference in Oslo or something, and from what I could understand he was doing away with the db completely.
. So I have another question, how do you architect your application in Elixir to keep application state though multiple requests (sessions) on multiple boxes without using something like memcached or redis (or a network disk)?
. Even if it's running on only one Box? Where does data reside if you are not using a db?
I have a basic understanding of Erlang processes (what's explained here: http://stackoverflow.com/questions/2708033/technically-why-a...) and how it's particularly well suited for concurrency. My questions are about Web Apps and Elixir and scaling.
select * from visits, plus conversion to JSON and delivery to a client on local loop. About 5000 records.
* Phoenix 140 ms
* Rails 248 ms
* Ruby without AR 219 ms
* PostgreSQL 2.97 ms, with no JSON generation and no delivery
select started_at, duration from visits -- JSON and delivery
* Phoenix 74 ms
* Rails 116 ms
* Ruby without AR 88 ms
* PostgreSQL 3.47 ms, no JSON no delivery
Single process, so maybe Phonix could get a larger advantage as the number of processes/requests increase. For the typical none to low traffic site there is little difference, the tool the programmer is more familiar with wins.
Edit: improved formatting.
>Ruby Conf that claimed that Elixir was giving better results (in request time) than rails
Ruby Conf that claimed that Elixir was giving better results (in request time) without a cache than rails with a cache
2) db is rarely your problem. Interpretation can be, and tons of other things. Also erlang schedules things far better and use all the core of the machine, so it will block far less, etc It also means that the thourgput and latency will stay stable despite the amount of users growing. (up to a limit, but it will far higher)
3) Store it in memory. And link other node to yours. Tadaaa
4) Where data live depends of your use case but why use a db if you do not need it anyway?
The thing in general is that web framework tend to use db to hide concurrency.
Also why no cache : because cache is complex, cost you another dependency and another program, make debugging harder, make your app less deterministic, etc.
If you need cache use it. But caching is a really hard thing.
Sure, np. I am mainly an Erlang developer but the same properties apply to Elixir, because both share the same battle tested VM -- BEAM.
> I'm asking what kind of web apps or scenarios do you think Elixir is particularly well suited for?
I am not currently working on building web app directly. But in general Elixir/Erlang would be good for a scenario where they'd be multiple connected clients at the same time. Think maybe something like a car sharing service where map updates in realtime as cars move through the city. Lots of users chatting together, or playing a game. Maybe they are bidding on ads, or tagging posts with "likes" and so on.
But there might be a significant improvemtent even in the plain old request-goes-to-database scenario, simply because the VM (BEAM) is better equipped at handling multiple connected sockets, all streaming data in and out. For example, it knows how to take advantage of multiple CPU cores, it handles GC better and so on. Not sure how deeply you want this explanation to go, so will stop here.
But, yes, if it all goes to a single MySQL instance running in another datacenter, that might be the bottleneck so there might not be a speedup seen by using something else. So you have to measure. In that case maybe a cache in front of it will work just as well to speed things up.
> he was doing away with the db completely.
Haven't seen the video. But with Erlang can think of those processes as in-memory storage of state. Can spawn hundreds of thousands of them, and they can live as long as the node stays up. Then can also connect multiple BEAM VM instances on different machines, create a cluster and so can refer to these processes as if they are local, in a rather transparrent manner. Or maybe they mean they used Erlang's built-in database (Mnesia), that had some limitations, but recently in a new release (19.0) it has the ability to scale much better as it can handle a pluggable storage such as LevelDB for example. The advantage there is the database is integrated into the langauge. So queries are not in a different query language, via a driver, so some other process, but queries looks like list comprehensions and transactions are just function closures. That can simplify things immensly in some cases.
> how do you architect your application in Elixir to keep application state though multiple requests (sessions) on multiple boxes without using something like memcached or redis (or a network disk)?
Can create a cluster so can keep state on another node that doesn't go down? (as mentioned), or can still use a database (say Postgres), so it really depends. Hard to give a general advice here.
Part of the problem I had with the videos had to do with the examples they were using (a shopping cart that stores data on process memory rather than an external db seems a little risky to me).
I will investigate Erlang Clusters. Again, thank you for your answer.
I use it in Java/Kotlin almost every time I do something involving more than one thread.
Add a blocking-queue, define incoming messages, define outgoing messages and forget about it.
But "I switched" means you switched your main language, right? I wouldn't switch my main tool to something with a specific usecase. Your main language should be something general purpose, because most of your life's problems have a wide range of usecases to consider.
Therefore I would have expected an argument for why Erlang may be a beter general purpose language or a headline like "Why <project/company/service x> switched from Python to Erlang".
Second for performance as much as possible from the SQL building should be outside the function. Also if the query is that static, you should probably use the plain SQL anyway ;) Or maybe even build the SQL query string by using SQL alchemy outside the function, and then using the SQL generated by SQLAlchemy inside the function (if you don't like pure SQL, like for example compatibility).
- It looks like the SQL queries produced by pure_python() and simple_sql() are structurally different, which would make any comparison worthless.
- How does the author solve the long running tasks (> 500 ms in the article) problem with Erlang? In Python, it looks like he was using Celery with Redis then RabbitMQ. What does he use in Erlang? (Erlang message queues have no built-in persistence.)
There is also RabbitMQ, which itself is written in Erlang.
A lot of the problems are more or less solved problems with well known workarounds but it's still a pain and probably why there's so much buzz around finding the "next (insert your favorite language/framework)".
Erlang generated a lot of buzz several years ago but aside from the recent success of WhatsApp it never quite stuck. I'm curious to know how this author thinks differently.
Erlang is quietly running a lot of stuff out there. It's almost certainly running some bits of infrastructure that you use daily, or in a supporting role for it. People who are not old Erlang hackers are picking it for new projects.
Erlang hasn't been and probably will never be quite mainstream, but to say it "hasn't stuck" might be missing the point: it's a mature and performant language that has some very nice standard features in OTP that are lacking or less mature in other languages. If your problem domain is smack in the middle of Erlang's strong points then it's worth considering. It's not going away any time soon.
This is all of course compared to the current buzz surrounding Elixir/Phoenix. I didn't intend to dismiss Erlang's amazing track record and applications to the domains where it continues to shine.
This is roughly the time span I was referring to when I said "buzz":
All these languages have their advantages and followers, but the thing they all have in common is first-class concurrency, which everyone now realizes they need (and old Erlang devs will tell you "duh").
Honestly, while I write Go almost exclusively anymore, I think Elixr is the most exciting. The more I learn about Beam the more magical it becomes in my eyes, considering just how old and stable it is. Phoenix is a game changer in many ways.
Alright -- I'll bite.
Rarely anyone "needs to scale." Plenty of $1B businesses can succeed on a fistful of c4.large's running Ruby on Rails. The same Ruby on Rails that ran their MVP. Scaling via language change isn't the path to victory, unless you picked a language that was poorly suited to your domain to begin with.
Too many companies end up following this cargo-cult advice, and spend more time grappling with their tools than innovating. Because "concurrency." Someday.
Optimizing your choice of language for anything other than a linear relationship between Real Complexity and Implementation Time is a fool's errand and a fiduciary travesty.
In all fairness, Elixir/Phoenix could become as well-learned as Ruby on Rails. We'll end up in a best-of-both-worlds scenario. And at that point, I'll happily eat my words.
But in the meantime, solve your "scaling" problems by measuring and optimizing, instead of re-writing your app in the flavor-of-the-week.
Then there are plenty of 7 figure companies that have also had scale issues. Game companies come to mind first and foremost but they're hardly the only ones.
You want to run your marketing website on Rails? Yeah of course. But billion dollar companies on Rails?
Most of Github's stack is Ruby on Rails, with specialized components/sub-systems written in C. This is a common theme amongst companies that use Rails or Python at scale.
There's a reason why people keep those languages around. Its the same reason why they tend to be used for MVPs: the tool gets out of your way so you can focus on solving the hard problems of your domain. The longer you can preserve those ergonomics, the better.
Agree, and I say it as an Erlang full-time dev. Chances are you don't need 9 9s reliability and super scalability.
I've heard and seen cases of companies rewriting stuff unnecessarily due to some incompetent people following latest fads. "Oh I read Python has the GIL so we need to switch to Go. Saw on HN everyone is switching" -- cue years wasted in rewriting and destabilizing a solid usable code base.
How so? What makes Erlang more functional? And how is Erlang simpler for you?
A big part of the renaissance in programming languages come from developers leveraging rock solid VMs such as BEAM and JVM. It enables developers to have the expressiveness of dynamic languages and without sacrificing performance or scalability later down the road.
In the case of Elixir, many feel that it adds some innovative language features like the pipe ("|>") operator and metaprogramming facilities that ease many of the pain points with Erlang. I'm not an expert in Erlang but there are quite a few notable Erlang developers that enthusiastically support Elixir (including some nice words from Joe Armstrong).
I'm not arguing with you but curious about your thoughts. Looking forward to the next installments in your series.
Well both points are valid. Elixir does add some great concepts like the pipe, usable macros and others. But with it comes some complexity -- it is a trade-off. I don't think one is a superset of another. Both will appeal to some people and that's fine. I prefer Erlang, but I hear a lot of people put off by Erlang's syntax, so well maybe Elixir will look better for them.
One thing I am impressed about Elixir is the community -- Jose and the team, did a fantastic job building not just the language but making a friendly and welcoming community, that is hard overstate.
To be clear, are you suggesting that's reasonable to do a rewrite (or at least for certain parts) from Python/Ruby when you need to scale? Is that going to be reasonably accomplishable in many cases? Honest question. Just wondering what you have in mind.
In this case, you probably first start by switching your message queuing to Erlang (RabbitMQ), since you had a problem with that anyway; then you can expand outwards from there if it makes sense or makes you happy: job running in Erlang, webservices in Erlang, putting the data in mnesia instead of postgres, etc. You can do the same with any language, of course.
Among the languages you cited, C# has async/await, Clojure has core.async, Elixir has processes and messages, Go has goroutines and channels, but the other languages (Python, Ruby, Java, Rust) have nothing built-in specifically related to concurrency, do they?
Java, Rust & Ruby have built-in support for concurrency, but I haven't used those languages that much.
> Python 3 has async coroutines built-in
Yes, the asyncio library was introduced in Python 3.4.
> there are a number of libraries for Python 2 (e.g. twisted, tornado)
Yes, but they are not part of the language per se.
> Python 2 & 3 also have threading and multiprocessing for concurrency.
Threading and multiprocessing are really not a good fit when the number of concurrent processes increases, because each thread allocates its own stack, which can be quite expensive in terms of memory. This is the reason why Go (and maybe Erlang too but I'm not sure about the implementation) map "green" threads to OS threads, and implement a resizable stack. Libraries like gevent offer something similar for Python, with some low-level trickery.
> Java, Rust & Ruby have built-in support for concurrency
In the language and/or the standard library?
In Java, there is Quasar and Akka, for example, but they are not part of the language/standard library.
> In the language and/or the standard library?
In Ruby, there's Thread and Fiber. MRI's GIL hurts it here, but other implementations don't have this problem.
Same question for Ruby :-)
I don't know much about the details, as I haven't used them myself, I just know they exist.
My experience with Go has been that it is almost as easy to get it up and running just as fast as with Python.
Edit: to be fair the variable proj_id sounds like it's not taken from untrusted input, but nevertheless a bad idea.
I sincerely hope nobody is making changes to a working production stack based on posts like this.
Edit: ok, close enough --
"Please explain me Byshtrtv"
Edit2: replier and op were much more helpful than google, thank you!
"Well, you are trying to add Python code some Erlang features."
and it wasn't obvious to me what you meant.