Ruby 3.0 and the new FiberScheduler interface
wjwh.eu
wjwh.eu
This is huge. They solved the function-coloring problem by deciding that all functions are now potentially async. It makes it more likely that the ecosystem as a whole actually becomes async-compatible. I wish Python had taken this approach, though I understand why they didn't.
Golang does essentially the same thing and uses channels instead of "futures". However, some APIs will have functions that return a channel that you receive a message on later.. Which is a lot like a future. And if you provide a more simple non-channel API along side(instead of making the consumer wrap your calls in a Go routine or always use a channel), you now have "colored" functions.
I believe this is just a solution to a scheduling problem(pre-emptive vs cooperative).
This is an aside and possibly pedantic, but everyone seems to have forgotten that the whole idea of futures was originally that they weren't some different type. They were the same type, but blocked transparently on demand and had no `get` operation.
The construct that came before futures, called eventual-values, had the different-type thing. The only difference of futures, the big idea, was they said 'maybe we can get rid of that part'.
> These “futures” greatly resemble the “eventual values” of Hibbard’s Algol 68 [39]. The principal difference is that eventual values are declared as a separate data type, distinct from the type of the value they take on, whereas futures cannot be distinguished from the type of the value they take on.
(Halstead, 1985)
But now people write types like `Future[T]`... which completely misses the point! If you start to write `Future[...` then think to yourself... this isn't a future. And that's how you get this colouring problem you describe... it was already fixed and people have forgotten all about it.
Yes - because it means existing code that expects a T cannot accept a Future<T>, which is the colouring problem being discussed.
You end up with half your code accepting futures and the other legacy half not accepting them and there's a divide. If you want to use the legacy code you have to force the future potentially before you really needed it.
> I don’t see the alternative.
The alternative is wherever a function accepts a concrete value T t, also let it accept a future value T t. When t is actually used, then block as if get was called.
Of course because Rust's runtime or lack thereof aims to be zero overhead a Future could never be without the poll function.. right? You need a runtime to make regular functions accept these magic opaque future values.
Yes it requires some kind of special compiler or runtime support.
First of all, this is a programming language specific problem, not a general one. A good language allows you to write:
someFuture.map(someFunction)
Yes, you cannot write someFunction(someFuture), but that is a good thing, since asynchronous execution can very well change the behaviour of the program. Hence, I want the compiler (if there is one) to warn me if a variable that I pass to a function is now a Future instead of a plain value. I then map over the Future and I'm good. Or, maybe I want to wait for completion of the future and I'm good.
In both cases, there is no reason to write a second version of "someFunction".
Therefore...
> If you want to use the legacy code you have to force the future potentially before you really needed it.
... this is only true in incapable programming languages.
You need support in the language to do this. If you're writing a library routine from scratch as Future<... with the compiler knowing nothing about how to handle it specially then you aren't getting a future as they were originally envisioned and you're missing out.
I don't think so - let me try to make it more clear, hold on.
> You need support in the language to do this. If you're writing a library routine from scratch as Future<... with the compiler knowing nothing about how to handle it specially then you aren't getting a future as they were originally envisioned and you're missing out.
You seem to assume that the programming language needs to know of the existence of Futures to help. But that is not the case and probably the reason for the misunderstanding.
It is sufficient that the language supports generics/type-parameters as well as a lightweight syntax (such as in the example I gave) and optimally higher kinded types or a similar mechanism.
If that is given, the language can be totally agnostic of Futures or async programming in general and it is easily possible to define a Future type as a library-type and use it in the way I described. Scala is a good example of such a language.
someFuture.map(someFunction)
This blocks the entire of someFunction until all the values are ready. You haven't adapted someFunction to work with futures at all - you're simply not running it until all the futures are gone and only normal values remain which you then pass into it!
Imagine someFunction comes from a library so cannot be modified trivially, takes three values, and can run successfully with the first two completed values while the third completes. What would you write with map to do that?
So then...
> Imagine someFunction comes from a library takes three values, and can run successfully with the first two completed values while the third completes. What would you write in someFuture.map(...) to do that?
That is a good example actually. I believe that the user (at the callsite) should both _know_ if the library operates in an async way and/or runs code in parallel _and_ the user should be in control of choosing his desired behaviour.
Hence, with a library function that just expects 3 values, this is not possible and I would say the design isn't great.
Here is how I would write it in scala-ish pseudocode:
// library - two separate functions instead of one
def calcX(a, b): X
def calcY(x, c): Y
// One "combined" synchronous function can be created easily if desired:
combinedCalc = calcX.andThen(calcY)
// all given values are (lazy) futures
aFuture = future(...)
aFuture = future(...)
aFuture = future(...)
// callsite - async but _no parallel execution_
xFuture = (aFuture, bFuture).mapN(calcX) // This is actually valid Scala syntax. Returns a future
yFuture = (xFuture, cFuture).mapN(calcY)
// alternative Scala syntax
yFuture = for {
a <- aFuture
b <- bFuture
x <- calcX(a, b)
c <- cFuture
} yield calcY(x, c)
yFuture = for {
a <- aFuture
c <- cFuture
b <- bFuture
x <- calcX(a, b)
} yield calcY(x, c)
How to run it in parallel?
xFuture = (aFuture, bFuture).parMapN(calcX)
yFuture = (xFuture, cFuture).parMapN(calcY)
Sorry the wall of code, but it shows that the code can look rather short even without future or async specific syntax support from the language.
The important part here is that it allows me to have fine-grained control over the execution of my code. There are a lot of situations where I probably don't care and it might be nice to have a way to say "Hey compiler, I don't care, just run that code in the way you think it's best", similar to what an SQL planner does with my queries.
But often I want control. For instance, in your example: what happens if the library decides to always evaluate "c" eagerly even though "a" and "b" often fail and the whole calculation is aborted? That would be a waste of CPU. But worse: the library might change its behaviour any time and keep the same function signature.
While that might be not so bad for mere calculations, what if we start to think about actions? Such as database or file or network IO - or even in-memory transactions that should be atomic. How about _streams_ of these actions, such as running an action every second and then combining this with other streams of actions? Losing control over the fain grained execution will result in terrible hard to debug and fix bugs due to concurrency.
This is just not a true statement - and that was the point of the original thread - the original paper on futures had them implemented transparently and able to be passed down into library code not expecting futures. It's also how Haskell's par works today. So it demonstrably can be implemented.
> I believe that the user (at the callsite) should both _know_
Ok but this is just a 'don't do that' argument. If a user tells us they have a good reason to do it, I'm interested in how we can enable them to do that.
> with a library function that just expects 3 values, this is not possible
But it is in Haskell using par, and other languages with more traditional futures.
> Here is how I would write it
Great... but you've rewritten the library code with 'two separate functions instead of one.' My original point was 'because it means existing code that expects a T cannot accept a Future<T>' and you're proving that back to me so hard that you've accepted you need to completely rewrite the code to do it!
Languages with traditional futures would let you just call the original, unmodified library code, with future values instead of actual values, and it would just work.
If you had library code like this, you could pass in future values of T for a, b, and c, and it would execute the expensive_operation on a and b as soon as it could, while c completes.
def foo(T a, T b, T c)
expensive_operation(a, b) + expensive_operation(b, c)
end
Yes this is a problem with transparent futures, as is the ordering issue you raised. But I wasn't arguing they didn't have their own issues. You're saying it's impossible and that is demonstrably false, because languages do do it.
(I did half of my PhD on these topics, and the other half on Ruby.)
Okay, technically you are right - however it changes the semantics of the program. I admit that I didn't mention this constraint though.
> Great... but you've rewritten the library code with 'two separate functions instead of one.' My original point was 'because it means existing code that expects a T cannot accept a Future<T>' and you're proving that back to me so hard that you've accepted you need to completely rewrite the code to do it!
> Yes this is a problem with transparent futures, as is the ordering issue you raised. But I wasn't arguing they didn't have their own issues.
I don't disagree with any of that. However, what you originally said:
> > Are you saying that generics on futures are bad?
> Yes - because it means existing code that expects a T cannot accept a Future<T>, which is the colouring problem being discussed.
And you imply that transparent futures are better than the ones with generics. So where we disagree is on what is worse. I am convinced that the problems (and there are certainly a lot of unknowns) you introduce with transparent futures is much worse than the performance you might gain in certain cases and the easier syntax in many cases.
I believe it's important to expose the fact that an asynchronous(potentially parallel) task is doing something for you. The runtime semantics of an assignment would appear the same in code by be very different behind the scenes otherwise. Also, a lot of code is written to manage those tasks in use-case specific ways. How would you race or cancel transparent futures?
Looking at the API for Fiber.. It looks like "resume" returns the yield(generator) or fiber return value? So Fibers can be treated as a sort of Promise? As soon as you start returning fibers from methods you have the coloring problem..
Not really, they've just made it super implicit. Any FFI calls are now implicitly coloured, same with anything CPU-heavy.
Like their approach to type annotations, I think this will be a mistake in hindsight.
That said, you're right in that one may have to make some changes to code to get around some new problems. The problem with FFI/CPU-heavy functions is that they prevent _other_ fibers from being scheduled while they run.
If it becomes possible to implement work-stealing, then that would mitigate the problem. It would also be solvable by sprinkling "yield to scheduler" calls throughout such functions. Annoying and not always possible, but, since in these cases the signature would not change, technically not coloring.
[1]: https://glyph.twistedmatrix.com/2014/02/unyielding.html
I kinda agree but yielding only on I/O is not enough IMO. Sooner or later we'll want to do CPU intensive work in fibers as well, which in the current implementation will block all other fibers on the same thread. Like I mentioned in the article, I'd like to see work stealing between different threads as that would allow fibers to migrate away from threads stuck in CPU intensive work. An alternative way could be to adopt a model similar to Haskells lightweight threads, where the runtime forces a `yield` after N milliseconds (configurable). That would make sure that CPU intensive work would not block other fibers "too" much.
Then they added a guarantee that if your loop called a function, the scheduler would be able to make the goroutine yield. https://golang.org/doc/go1.2#preemption
This mostly worked although you could still have a CPU-bound thread not making any function calls. I also personally ran into a pathological issue where the scheduler was being invoked, but a heuristic kept the current goroutine running so others were still starved.
Finally they added true pre-emption (not yielding) in 1.14 https://golang.org/doc/go1.14#runtime it looks like it just sends signals and saves state.
Once nice thing is that if I understand the go runtime correctly, work stealing by scheduling goroutines on different threads has been a thing for a long time.
This limitation is present in every modern asynchronous programming system I've ever seen. They simply can't deal with long function calls. There's a loop under the hood which constantly executes queued functions. Any long-running function will stall the execution pipeline and increase latency. It's just like cooperative multitasking...
And for just semantically long running functions, it is not difficult to write a scheduler that tries to be "fair", however you want to interpret that. But of course it requires to fully embrace asynchronous programming, something that will probably difficult in Ruby.
I would just suggest that sometimes performance isn't worth sacrificing clarity. Obviously sometimes you have to (Big O can be an unforgiving beast) but not always.
This new scheduler interface just gives us a nicer and lighter weight abstraction for handling it rather than, for example, using OS thread.
That's blocking IO with multi-threading. The executing thread stops and waits.
If you can only do parallel IO in a thread why not use a light weight fibre (green thread) instead of an OS thread? Hence this new scheduler and fibres.
The GVL isn't automatically released when a thread blocks on IO. Each bit of native code performing IO has to explicitly release it like in the pg gem here: https://github.com/ged/ruby-pg/blob/fb465855ce1dd12cf7eb69c9...
If you don't need high concurrency then using threads is simpler.
Right, but it still blocks the entire calling thread, no? Been a few years since I've been deep into Rails.
The GVL will release when the IO call blocks allowing another OS thread, including the parent, to continue.
That allows you to make blocking calls in parallel but is heavier weight than using the new scheduler and fibres for those parallel IO calls.
In the case of ActiveRecord I think it’s down to the database driver implementation, anyone who knows care to comment?
And, maybe it becomes more useful for using ruby in smaller services outside of the Rails context, especially when the go-to solution for a bunch of problems is to run a separate copy of your server in 'worker mode'. Even with Rails, not relying on Delayed Job or Sidekiq by default would be nice. Of course, I'm thinking about Ractor there too.
So in practice we glady throw more memory and processes at the problem.
I have used ruby since 2013 and started on rails but quit using rails in 2015. I still use ruby all the time though, mostly for scripts/automation that violate my rule on bash v. other lang for scripts (basically do I need arrays, maps, or to parse json beyond simple extractions that jq is great at). I do have a couple sinatra services now though that I maintain. Sinatra is wonderful with simple needs like mine.
I don't know how much of an anomaly I am though, interested in hearing from others as well.
Edit: not looking to debate bash vs other langs here. Of course you're free to do such below, but I won't be engaging (got to focus on work and have had the debate many times and I don't think any of us will ever convince the other. It's become religion at this point).
Also I use Elixir/Phoenix these days for use cases that used to be rails
What held me back so far is that a lot of simple B2B, low volume stuff doesn't seem to benefit from the high parallelism that BEAM brings.
But I'm wondering if it brings enough other advantages that I shouldn't view it that simplistically.
So to start with, Phoenix LiveView really is a game changer, you should have a look at Elixir just for that. This is the talk which really bluffed me when I first saw it: https://www.youtube.com/watch?v=MZvmYaFkNJI.
For the other upsides, I like a little bit better the way everything is architectured in Phoenix, there's much less magic and it's easier to follow the data flow and what is going on.
I would recommend trying it and seeing how you like it. I've basically dropped Python which was my daily language in favor of Elixir. I find I get a higher top bound to what I can do, high-level expressive but less magical code and a bunch of capabilities that aren't typically feasible with other runtimes (state handling, resiliency and stuff). The parallelism I get for free with Phoenix whether I try or not.
I'd watch Sasa Jurics talk on the heart of elixir and erlang to get more of the technical advantages laid out quite well.
I've run both Rubinius and jruby (with rails), both gave me significant performance gains.
Shrug Having tried like every possible method of async I/O with Ruby to eek out moar good perfz back in the day... including various actor implementations (sup, JRuby and Akka)... nothing compares to just using BEAM if you want the actor paradigm. The cooperative scheduling problem is a really enormous pain in the ass most of the time.
Do you really want an actor system?
Actors are non-determinsitic, extremely prone to difficult-to-debug race conditions, and inherently stateful. They're a classic concurrency foot-gun from the dark old ways of doing things.
Don't we want to be moving away from these models that we know trip people up? Can't we do better than this for Ruby?
Sure you're not thinking of threads? Actor's have only really been done in Erlang, and more recenetly in Akka and Orlean's. AFAIK, there aren't any other paradigms other than CSP or threads for concurrency.
Actor's can have race conditions, same as any other concurrent system but I've personally never actually had any given the design patterns in OTP and Elixir. It's really helpful (to me) that each actor is always deterministic in itself and doesn't share data, only messages. It also maps nicely onto multicore.
No I'm thinking of actors - stateful, non-determinstic, racey actors. A minefield of classic concurrency bugs!
> Actor's can have race conditions, same as any other concurrent system
Deterministic systems like fork-join don't ever have race-conditions.
> is always deterministic in itself and doesn't share data
But the global state of all the actors is shared. If an actor you send a message to responds differently due to its state then let's be honest you're implicitly sharing that state.
> It also maps nicely onto multicore.
It maps directly to multicore, right. Don't we want something higher level and safer than directly mapping to our hardware?
https://hexdocs.pm/elixir/Task.html#module-async-and-await
[0] determinism can go out the window once there's an error. Can you safely say you know what happens in ruby if you are running async and one of your concurrently running async functions throws?
The nice thing in elixir there is zero ambiguity as to what happens if you have (N) async functions running and one of them crashes -- it brings down all of the other threads and triggers their respective resource cleanups in exactly the way that they do as actors, since Task is a fairly trivial wrapping of erlang's actors. (Or you have choices!!, you can async_nolink if you want your async function to plow forward even if its parent dies due to the crash of one of its siblings)
I think there is still some confusion since you suggest fork/join but that's only for threads and won't work over a network.
Map-reduce is fork-join and runs just fine over a network.
Well yes systems of actors contain race conditions. However, so does Rust's or Go's concurrency primitives. The only way to completely avoid race conditions I know of would be formal methods via TLA+ or similar. From my brief looks into TLA+ it'd appear to match actor implementation models pretty well (messages to actors yield new states).
> But the global state of all the actors is shared. If an actor you send a message to responds differently due to its state then let's be honest you're implicitly sharing that state.
It's not really whether a program has state, but how that state is accessed and modified. In the actor model, only an actor itself can modify it's own data which makes it easier to reason about many classes of concurrency problems. Especially actions within an actor itself are deterministic (generally speaking). This makes a lot of problems become very straightforward to model without needing locks everywhere. The concurrency problem is shifted to modeling the behavior of a group of actors, which helps enforce separation of concurrency concerns from the lower level implementation details (with a good framework at least).
Since I was asking and thinking about Ruby getting (preemptive) actors initially, this frame of thought seems odd to me. The entire foundation of Ruby is built on objects that contain implicit state and on those objects behaving differently based on their state even to the point of checking for unhandled "messages" via `#method_missing`.
Th fork/join model can be emulated rather well as @dnautics described in a sibling comment. It's close to the model used in Elixir's Phoenix web server where each request just forks an Erlang process to handle it.
At the end of the day, issues with races in the actor paradigm are little different than communicating with remote programs as either fork'ed processes or network programs. Just fork'ing a program may avoid some concurrency issues by not doing any useful work by itself. If you do use fork/join to service http requests via sockets you're now relying on implicit state from the OS to do the heavy lifting.
If anything Linux's new `io_uring` works by treating a process like an actor where the process & kernel pass messages via a "messsage queue" which means your forks will just be acting as pseudo-actors anyway. Side note, on Linux doing this saves a lot of context switches by not preempting the OS process' call stack via a direct syscall every time it has something to communicate to the kernel. That gives a big speed bump to IO heavy programs. There can be significant performance benefits to making programming models work more like the hardware. I expect that actor models will grow to be a better performance match as core counts continue rising.
> Don't we want something higher level and safer than directly mapping to our hardware?
Certainly! ... if it makes acceptable tradeoff's in other areas. Though Actor's are really more of a base system to build higher level constructs that are easier/safer to use. Most programming shouldn't be done using your own concurrency primitives, but building on well thought out and tested libraries.
Mainly though, the idea of a Ruby with preemptive actors seems fun to me. Perhaps something you could map to lego robot's for kids.
The fork-join model does allow you to completely avoid race conditions, because tasks are always created in a deterministic order, each can only work with the immutable input they get, and tasks join also in a deterministic order.
(I think when I say fork-join you're thinking processes fork and join - I mean the abstract concept of fork-join. Map-reduce is an example of a fork-join model and is deterministic and race-free by construction with no proof system needed.)
That way you don't need to make a complex fiber scheme just for resource management. Spin up as many as you need an the runtime will do the task pool for you.
https://ruby-concurrency.github.io/concurrent-ruby/1.1.4/Con...
require "open-uri"
require "json"
results = []
(0..100).each_slice(10) do |slice|
slice.map do |i|
Thread.new do
results << JSON.parse(URI.open("https://httpbin.org/get?i=#{i}").read)["args"]
end
end.each(&:join)
end
p resultsSo I guess as with most things in software engineering, if there is something wrong with threads or not depends on your use case.
I'd wager that the great majority of Web apps (i.e., the not-terribly-busy ones) would scale reasonably well on thread-per-request.
Of course, you could implement a more robust algorithm for scheduling threads, but it sounds like an implementation of FiberScheduler would make the best use of a single thread's resources (and you could scale that to multiple threads as needed using something like a shared queue). Since fibers are lighter-weight, it's presumably more efficient to run parallel IO operations in a single thread, and only use more threads if your work requires it.
I'm not sure about Ruby's specific implementation, but the threads need not wait around. When the thread entry returns pthread_exit is called. I believe the OS thread would go away regardless of how you are waiting on them.
But the promise of Fibers + a fiber scheduler is that you won’t have to think about the performance impact of blocking IO as much as you might normally.
For example, in a rails app (or any ruby web server) I am very scared to make slow external http requests because that might tie up a thread for a while, and request threads are a pretty limited resource in typical configurations. (16 threads/process) * (1 process/cpu core) might be a reasonable default, giving you at most 128 concurrent requests in flight on an 8 core server.
Suppose you make a really slow 3rd party api call that takes 10 seconds to service a request. Your server can handle no more than 128reqs/10s = 12.8 of these requests per second, even though it will spend most of its time doing nothing but waiting.
If we bring fibers in to the mix and do one fiber per request, then they can yield to another fiber while waiting on the api response - you can then easily have thousands of requests in flight instead of 128.
Granted, if you are entirely IO bound then you can probably get away with a much larger thread pool, or do thread-per-request. But if you have some other endpoints that are more compute intensive then bumping up your thread count might be dangerous... then what do you do? Fibers might be the answer.