Polyphony: Fine-Grained Concurrency for Ruby(github.com)
github.com
Polyphony: Fine-Grained Concurrency for Ruby
https://github.com/digital-fabric/polyphony
25 comments
> I am interested in how concurrency can be represented elegantly and efficiently
This was the main driving force behind the development of Polyphony. I've been exploring all the different ways of expressing concurrent operations: callbacks, promises, async/await. Each of these solutions has its drawbacks: callbacks make you split your app logic across different functions; promises are a somewhat better API but do not solve the split logic problem; async/await potentially leads to two types of functions (the "what color is your function" problem). All three kind of break Ruby's exception handling model.
In contrast, Polyphony relies on Ruby fibers, which provide a base for implementing cooperative concurrency, where each fiber does its work, yielding execution whenever it performs a blocking operation, such as reading from a socket, or waiting for a timeout. In addition, fibers are arranged into a tree where each fiber is considered the child of the fiber from which it was spawned, and is guaranteed to stop executing before its immediate parent stops executing. Any uncaught exception raised by a fiber will be propagated up the fiber tree. This implements a pattern known as Structured Concurrency [1]. Fibers can communicate by sending each other messages, in a similar fashion to Erlang/Elixir processes.
There's no direct equivalent in Polyphony to the program you included, which seems to implement two generators running asynchronously. But the following does something similar.
This was the main driving force behind the development of Polyphony. I've been exploring all the different ways of expressing concurrent operations: callbacks, promises, async/await. Each of these solutions has its drawbacks: callbacks make you split your app logic across different functions; promises are a somewhat better API but do not solve the split logic problem; async/await potentially leads to two types of functions (the "what color is your function" problem). All three kind of break Ruby's exception handling model.
In contrast, Polyphony relies on Ruby fibers, which provide a base for implementing cooperative concurrency, where each fiber does its work, yielding execution whenever it performs a blocking operation, such as reading from a socket, or waiting for a timeout. In addition, fibers are arranged into a tree where each fiber is considered the child of the fiber from which it was spawned, and is guaranteed to stop executing before its immediate parent stops executing. Any uncaught exception raised by a fiber will be propagated up the fiber tree. This implements a pattern known as Structured Concurrency [1]. Fibers can communicate by sending each other messages, in a similar fashion to Erlang/Elixir processes.
There's no direct equivalent in Polyphony to the program you included, which seems to implement two generators running asynchronously. But the following does something similar.
def number_generator
n = 0
loop do
client = receive
client << (n += 1)
end
end
task1 = spin { number_generator }
task2 = spin { number_generator }
10.times do
task1 << Fiber.current
puts "next number from task1: #{receive}"
task2 << Fiber.current
puts "next number from task2: #{receive}"
end
So in the above program, two fibers are spun, each running the same number generator, which repeatedly waits for a message which is actually the fiber of the "client", and sends the next number to the client. In Python there's the Trio library [2], which implements structured concurrency, and which was actually one of the main inspirations for developing Polyphony.
Wow, thank you for your really helpful reply, and your thoughts on the matter of concurrency and the various approaches.
I would wonder if pipelines, and Kafka style and functional reactive programming and reactive designs could be adapted to be async.
I would love to know how you would approach the design of a fiber application.
From my limited understanding of fibers, I would start by drawing a diagram, probably a tree or graph of all the components and draw links between the connections between fibers and components.
From what I understand of the code, you can send fibers to other fibers with the "<<" operator.
This reminds me of Golang's channels and send and receive operators on channels.
I notice the use of the word "spin" to "create a fiber" which seems to be an analogy of fibers being literal fibers or strings that are attached to eachother and "pull" and "push".
I asked the following two questions on Stackoverflow.
What's the most efficient way to fork and join control flow multiple times in a multithreaded program?
https://stackoverflow.com/questions/75866966/whats-the-most-...
A while back I was trying to understand Python's asyncio and I asked this question:
What's the canonical approach to creating multiple producers and consumers with coroutines in python asyncio? https://stackoverflow.com/questions/74420108/whats-the-canon...
I feel that wiring up asynchronous system components is kind of not ideal because you have dependencies have to refer to eachother and you don't want to solidify dependencies because you might want to reuse components.
This article highlights that problem of wiring up coroutines/fibers in the right order and which side refers to which. https://www.chiark.greenend.org.uk/~sgtatham/coroutines.html
I wonder if there could be a pattern where every fiber receives a sequence of fibers that "configure" the fiber, like a calling convention of a method, this way runtime configuration can be separate from the definition of asynchronous an pipeline.
In my original code sample, my code refers to the children tasks. In your code sample, the parent refers to the children tasks. This seems to be a property of structured concurrency.
I would wonder if pipelines, and Kafka style and functional reactive programming and reactive designs could be adapted to be async.
I would love to know how you would approach the design of a fiber application.
From my limited understanding of fibers, I would start by drawing a diagram, probably a tree or graph of all the components and draw links between the connections between fibers and components.
From what I understand of the code, you can send fibers to other fibers with the "<<" operator.
This reminds me of Golang's channels and send and receive operators on channels.
I notice the use of the word "spin" to "create a fiber" which seems to be an analogy of fibers being literal fibers or strings that are attached to eachother and "pull" and "push".
I asked the following two questions on Stackoverflow.
What's the most efficient way to fork and join control flow multiple times in a multithreaded program?
https://stackoverflow.com/questions/75866966/whats-the-most-...
A while back I was trying to understand Python's asyncio and I asked this question:
What's the canonical approach to creating multiple producers and consumers with coroutines in python asyncio? https://stackoverflow.com/questions/74420108/whats-the-canon...
I feel that wiring up asynchronous system components is kind of not ideal because you have dependencies have to refer to eachother and you don't want to solidify dependencies because you might want to reuse components.
This article highlights that problem of wiring up coroutines/fibers in the right order and which side refers to which. https://www.chiark.greenend.org.uk/~sgtatham/coroutines.html
I wonder if there could be a pattern where every fiber receives a sequence of fibers that "configure" the fiber, like a calling convention of a method, this way runtime configuration can be separate from the definition of asynchronous an pipeline.
In my original code sample, my code refers to the children tasks. In your code sample, the parent refers to the children tasks. This seems to be a property of structured concurrency.
> From what I understand of the code, you can send fibers to other fibers with the "<<" operator.
You can pass any data between fibers using "receive" and "<<", including fiber refs.
> In my original code sample, my code refers to the children tasks. In your code sample, the parent refers to the children tasks. This seems to be a property of structured concurrency.
In general, yes, but in Polyphony at least you can easily walk the fiber tree using "Fiber#parent" and "Fiber#children".
You can pass any data between fibers using "receive" and "<<", including fiber refs.
> In my original code sample, my code refers to the children tasks. In your code sample, the parent refers to the children tasks. This seems to be a property of structured concurrency.
In general, yes, but in Polyphony at least you can easily walk the fiber tree using "Fiber#parent" and "Fiber#children".
I'd love a performance comparison against concurrent-ruby on the Readme.
That's assuming Polyphony is aiming to replace most or all of it.
That's assuming Polyphony is aiming to replace most or all of it.
> That's assuming Polyphony is aiming to replace most or all of it.
Not really. Concurrent-Ruby implements a lot of different concurrency patterns. Polyphony has a much smaller scope, though I imagine you could reimplement a lot of concurrent-ruby on top of Polyphony. I think there's a place for the two in the Ruby ecosystem.
Not really. Concurrent-Ruby implements a lot of different concurrency patterns. Polyphony has a much smaller scope, though I imagine you could reimplement a lot of concurrent-ruby on top of Polyphony. I think there's a place for the two in the Ruby ecosystem.
I'm the author of Polyphony. I'd be happy to answer any questions you have.
Does this use the Fiber Scheduler API to figure out which fibers are blocking? If so, why does this need libev - I'm guessing it's to do with performance?
No, Polyphony has a design that diverges from the model proposed by the FibeScheduler API. Note that the FiberScheduler API included with Ruby is only an interface, not an implementation. Polyphony uses io_uring on recent Linux kernels, or libev otherwise to perform I/O operations and implement timeouts.
does it use anything specific on darwin and/or freebsd or does it fall back to libev?
On non-Linux systems it just uses libev to for I/O readiness and for timers, and uses normal syscalls for the actual I/O.
Fiber::Scheduler is only an interface. You still need an IO multiplexer underneath it.
This is not what I don’t understand with the Ruby Team, why did they not ship Fiber Scheduler with an implementation and just an interface?
Dis they expect everyone to implement the scheduler om their own?
I am aware of the existing fiber scheduler implementations you can get on Ruby Gems but there should be official implementation.
Dis they expect everyone to implement the scheduler om their own?
I am aware of the existing fiber scheduler implementations you can get on Ruby Gems but there should be official implementation.
Excellent explanation and very well thought out. Can we have something similar for C++ and Rust please?
Without support from Rails, all of these libraries unfortunately can't be used by 99% of Ruby code out there and using Ruby in any sort of fan out architecture remains a pain (ie almost impossible in practice).
I haven't really had time to explore the usage of Polyphony in Rails apps. Personally, I don't use Rails in my work, and instead develop custom Ruby apps. I believe there's a place for Ruby development work without any connection to Rails.
There is a lot more to Ruby than Rails.
I can see it being useful for many projects, web or non-web.
I can see it being useful for many projects, web or non-web.
I don't see why you couldn't use Polyphony inside a Rails app.
This is where the "in practice" part comes in. Ruby culture thinks it's a great idea to use global mutable state in lieu of request local state all over the place because the code looks so beautiful if you don't have all of these parameters and just grab stuff from the ether. So unless all of your dependency tree (three quarters of which probably had its last release in 2017) fixes their usage of global variables, you won't have much fun even if Rails itself got rid of their global variables (to be fair, they've done a great job over the last year or so).
> Ruby culture thinks it's a great idea to use global mutable state in lieu of request local state all over the place because the code looks so beautiful if you don't have all of these parameters and just grab stuff from the ether
That is utterly incorrect. Make it s/Ruby/Rails and you'd be somewhat more correct. Cue Rack, Sinatra, Hanami, Roda, and that's just web "frameworks".
> all of your dependency tree (three quarters of which probably had its last release in 2017)
I don't see how that could make sense. Some libraries can be mature and thus not warrant much, if any, update.
Also, the aforementioned pattern is only present on a small subset of Ruby gems, usually ones that are close in mindset to Rails.
Nothing prevents anyone to code in Ruby in a functional-like style, nor pick dependencies that are written with such sensibility. But if you wilfully pick Rails you can't really expect to veer too far from the philosophy that comes with it.
That is utterly incorrect. Make it s/Ruby/Rails and you'd be somewhat more correct. Cue Rack, Sinatra, Hanami, Roda, and that's just web "frameworks".
> all of your dependency tree (three quarters of which probably had its last release in 2017)
I don't see how that could make sense. Some libraries can be mature and thus not warrant much, if any, update.
Also, the aforementioned pattern is only present on a small subset of Ruby gems, usually ones that are close in mindset to Rails.
Nothing prevents anyone to code in Ruby in a functional-like style, nor pick dependencies that are written with such sensibility. But if you wilfully pick Rails you can't really expect to veer too far from the philosophy that comes with it.
Because eventually a dependency will make it non concurrent.
Rails is the driving force for adoption in the ruby ecosystem. If it isn't a rails requirement to be concurrent friendly, it will be broadly ignored.
I am interested in how concurrency can be represented elegantly and efficiently, so I am interested in how libraries can simplify async and make it easier to reason about and write
What kind of code do you want to be capable of writing? Any imaginary syntax you can think of?
The libev and ioring support is great for IO scalability (https://github.com/enki/libev not sure if this is the official repo)
In Python I use the "select" module and use epoll on Linux.
I am currently thinking of designing an API that allows the registration of epoll-like listeners to arbitrary objects, including business objects, so you can efficiently register a listener on multiple behaviours of multiple arbitrary objects.
I wrote an async/await simulation in Java and my scheduler is really simple, it's just a for loop that checks to see if there are any tasks that can progress. I notice the switch_fiber in polyphony must do something similar. This is similar to a yield in a coroutine.
My async/await simulation takes the following program:
I had someone on HN explain to me how async works in Rust, but I've not written enough Rust to truly understand async in rust. https://news.ycombinator.com/item?id=33592334