Hacker News Reader

Comment by serguzest | Hacker News Reader

I have the conclusion opt-in async/await is a mistake. I think high level language/runtime like c# should handle it automatically like it is a sync call but a way to opt-out should be still possible. In my career, 99% of the time my code awaited async calls. I remember one case we tried to run tasks parallel, but the client for DB(ravendb) started throwing exceptions, possible race condition.

Also, they did terrible work when it was first introduced, low level details were leaked to app surface and it was impossible to comprehend for a regular programmer what the hell is going on? Do you remember ConfigureAwait(false)?

Years later such guidance compiled: https://github.com/davidfowl/AspNetCoreDiagnosticScenarios/b...

View on HN

Nope, that's a fallacy. Sync and Async code have inherently different semantics and trying to unify them will always fail, no matter what technique you try to apply. It has been tried more than often enough.

Completely agree with this. Having ported a lot of old non async c# to async, even mixing the notions for a moment is a disaster. This was a world of pain in the late .net classic lifecycle (4.5-4.8) where Microsoft kept fucking it up as well as us.

If you have legacy or UI code you are best served using Microsoft.VisualStudio.Threading to ease the pain. Not using it will land you in a world of random deadlocks and disappearing data from variables.

vs-threading should be part of the .NET standard library but somehow isn't.

Hello Task.Run() scattered all over the place.

I cried a little when I read this :(

Go did that, and it's working out very well.

Agree with the grandparent. Async/await is unnecessary complexity in all but the most performance-oriented languages (like Rust).

Very curious to see how Java Loom will pan out in practice.

Java has a better approach, because you can still decide which kind of thread to use, which makes FFI performance still much better than cgo.

Once you develop in a non single threaded ui framework it is not only about performance.

Could you please expand on that?

You need to keep the UI responsive, so you hand off all resource or cpu heavy work off the UI thread, give back control to rendering and responding to user input ... and later need to collect the results in the UI thread to change the UI state.

It is about doing things in parallel. There are many ways on how to achieve that, I am just nit picking on performance.

Then I think we misunderstood each-other.

Go is asynchronous-first, it just doesn't use the (flawed) async-await semantic to model asynchronicity.

Not really, look at doing FFI using Go. Its very slow compared to C# specially because of how Goroutines work.

Trading this for simplifying the remaining 99% of my code is a tradeoff I'm willing to make.

The vast majority of projects don't use CGo, and of those that do, the majority doesn't call CGo in a tight loop, making the performance impact mostly irrelevant.

So it's really coming back to the performance-orientedness of a language.

The practical end result is an insular ecosystem that tends to rewrite code "in pure Go" rather than reusing existing libraries. To the point where this attitude resulted in the standard library of Go breaking on occasions because they tried to go around libc on macOS and use syscalls directly (which are not guaranteed to be stable).

.NET is the opposite - it was designed from the get-go from cross-language and cross-runtime interop. Back when Java folk still had to struggle with JNI to use anything native, .NET 1.0 had P/Invoke and COM Interop. And now that it has async/await, it's all mapped down to a C-compatible ABI for interop purposes, so you can easily write async code in C# that calls into async code written in C++ that calls back into C# etc.

I think that’s one of the reasons why Java went with virtual threads instead, as the Async/await pattern is hard to do? As that does give you asynchronous code without any of the coding hustle and hide it from the programmer and you can keep your APIs the same.

> trying to unify them will always fail

Wrong. Languages that definitely got it right:

Erlang/Elixir (high level), especially elixir's Task.async

Zig (low level)

Very close runner-up: Go

Erm, just look at the name: "Task.async". So there is the distinct concept auf async, otherwise why would you name it like that. Mind that I'm not talking about async/await syntax here.

No there isn’t. Elixir, and by extension Erlang and the BEAM VM, doesn’t have the concept of stackless coroutines.

Task.async/1 (or Task.async/3) at a high level just spawns another BEAM process links it to the currently running process, sets up a process monitor, and then runs. Task.await/2 just pulls from the mailbox of the current process for a returned message.

But that is exactly what I mean.

There is a semantic difference between just running a line of code or running ith with Task.async right? So then there is a distinct concept of sync vs async.

The BEAM VM doesn't have faculties to accommodate any other form of concurrency other than synchronous, pre-empted green threads. Put simply, there is nothing other than synchronous in Elixir (edit: or Erlang or any other BEAM VM language). The Task module is basically just convenience functions (edit: convenience functions around existing BEAM VM functions for spawning BEAM processes or Erlang threads) meant to facilitate short term execution of, well, tasks.

Task.async/3 and Task.await/2 are, from the thousand foot view, the equivalent of:

  parent = self()
  spawn(fn -> send(parent, do_something()) end)
  receive do
    result -> IO.puts result
  end

In Elixir/Erlang parlance, I have just spawned a thread where the BEAM VM handles both the internal data structures to facilitate concurrent-safe communications between threads and the scheduling in userspace.

Put more generally, the mentality of concurrency in Elixir/Erlang is the equivalent as that of traditional pthreads in C.

I think you are misunderstanding me, so let me ask a simple question:

Is it (except for performance) always 100% equivalent whether a calculation is done "locally" or whether it is ran on some other, potentially just spawned, thread and then retrieved from this thread?

If your answer is "no it is not always 100% equivalent" then we agree (and I think we do).

No, they're equivalent. It makes no difference whether I calculate in the current process, spawned a process, started a process via the Supervisor, Task.async, or what not. They are all equivalent in the sense that they use BEAM processes.

An Elixir, Erlang, or any other developer using the BEAM VM (or at the very least myself since I don't want to generalize) doesn't make a semantic distinction between spawn and Task.async/3 because there is none. The significant differentiation is whether or not the caller is expecting a returned value, if not then Task.start_link can be used instead.

I don't need to worry about function colouring because there is none. I also don't need to worry about async vs Task.run like in C# because in Elixir there is none. To be asynchronous in Elixir/Erlang/etc... is to be synchronous. To equate it to modern backends, we basically have a microservice but every service can directly call each other.

I think I also might have been talking around you a bit, the BEAM VM sets itself apart from C#'s async/await or the TPL due to its implementation of the actor model. This video by Sasa Juric can explain better than I can the general overview of the BEAM: [1].

But a rough tl;dw, a BEAM process is to the BEAM, what an OS process is to the OS. We don't define a main function as the process supervisor in BEAM spawns processes we define instead and those are our "entrypoints."

[1] https://www.youtube.com/watch?v=JvBT4XBdoUE

> They are all equivalent in the sense that they use BEAM processes.

Sure, but that was not my question. 100% equivalent means that there cannot be any observable change in the behaviour of the whole program/application (except for performance).

Or in other words: is the sole purpose of the existance of `Thread.async` (or spawn, pick whichever you want) to change performance characteristics? If not, what is the purpose of its existence then?

> Sure, but that was not my question. 100% equivalent means that there cannot be any observable change in the behaviour of the whole program/application (except for performance).

In a vacuum (i.e. via spawn/1 or spawn/3) there is no observable change in behaviour throughout the whole system.

If they're linked via spawn_link then a child process crashing means the parent process dies with it unless handled in some way.

So ultimately, no there isn't observable changes. My logging process doesn't care or even know that my WebSocket process crashed.

> Or in other words: is the sole purpose of the existance of `Thread.async` (or spawn, pick whichever you want) to change performance characteristics? If not, what is the purpose of its existence then?

As stated before Task.async/3 is just a convenience wrapper around low-level BEAM primitives, there isn't anything special about Task.async/3 that you couldn't do via spawn.

The reason being that the BEAM VM schedulers prioritizes overall latency of the system over throughput by enforcing reduction calls (parlance for a function call, but specifically for tracking looping in the BEAM) at around 4000 calls, then it moves the current process back into the queue; repeat ad infinitum.

So you don't get massive speedups just spawning a new BEAM process but your independent units of execution still enjoy the same overall latency as before. If you want performance increases in the BEAM you would need to jump to dirty schedulers and NIFs which carry their own dangers.

> In a vacuum (i.e. via spawn/1 or spawn/3) there is no observable change in behaviour throughout the whole system.

Okay, let me follow up on this one. So could we take arbitrary Erlang/Elixir programs and automatically/mechanically rewrite them to push previously inlined calculations to be run on a different process (using spawn) or the otherway around - and that without causing any observable difference in behaviour in any situation except for difference in performance?

> As stated before Task.async/3 is just a convenience wrapper around low-level BEAM primitives, there isn't anything special about Task.async/3 that you couldn't do via spawn.

You can likewise answer my question for the BEAM primitives (like processes/threads): is the sole purpose of the existance of those primitives to change performance characteristics? If not, what is the purpose of its existence then?

Also, just so that we understand each other: I'm a big fan of the BEAM. This is not a criticism or anything; I just want to explain why I believe that function coloring (or however you call it) can never be conceptionally be removed from a language if sync/async is to be supported.

> Also, just so that we understand each other: I'm a big fan of the BEAM.

We probably should've established this way earlier, sorry if I came across as pretentious or off putting.

I'll answer in a bit of reverse order.

> I just want to explain why I believe that function coloring (or however you call it) can never be conceptionally be removed from a language if sync/async is to be supported.

To respond in more of a roundabout way, why doesn't C have colouring issues if it too supports asynchronicity (I'm pretty sure I've made up a word but hope the point gets accross fine) via pthreads?

The colouring issue is, at least under my understanding of C#'s async/await system is to accommodate for the fact that Rosyln transforms the C# source into .NET IL with a state machine. It's, to me at least, just a consequence from what layer of the stack we talk at. Erlang doesn't need to worry about since the runtime was designed to run with this specific model of concurrency in mind; but if we look at C#, .NET only supports using OS threads (though David Fowler is investigating green threads in the, iirc, labs repository on the DotNet Github) the async/await is, for lack of a more better word, bolted on. It's to my understanding that IronPython (or maybe it was PythonNet) has difficulties calling into C# async for that reason; though if I'm wrong please correct me.

> is the sole purpose of the existance of those primitives to change performance characteristics? If not, what is the purpose of its existence then?

Fault tolerance, latency, scalability, perhaps minor throughput increases (yeah this is a bit of a contradiction to what I said earlier, but I've revised my opinion that you wouldn't see massive performance gains unless resorting to NIFs).

The Phoenix Framework's 2 million websocket connection challenge I think is a best demonstrator for the use case of BEAM processes. A specific websocket connection dies? No problem, the Supervisor respawns it. Two million connections receiving a Wikipedia length article on a topic without the entire system coming to a crawl? No problem. There isn't a need to worry about kqueue, poll/epoll, or IOCP here, just fire off a task for the purpose and let it do it's thing.

But overall you wouldn't spawn a BEAM process just so you could compute the matrices behind RAID6/RAIDZ2, you'd delegate those to NIFs. The biggest gains in overall performance came from BeamAsm instead.

> Okay, let me follow up on this one. So could we take arbitrary Erlang/Elixir programs and automatically/mechanically rewrite them to push previously inlined calculations to be run on a different process (using spawn) or the otherway around - and that without causing any observable difference in behaviour in any situation except for difference in performance?

Yes. Using the case of the 2 million websocket connection challenge earlier, it doesn't really matter if each of those connections spawned another process as part of its routines, the other processes don't know and don't care. Taking a more generic case of Elixir's GenServers (or gen_server in Erlang), when I perform a call (as in the GenServer behaviour) it doesn't matter to the calling process what happens behind the scenes, it just blocks and waits for a response. The GenServer could fire off any number of processes it wants but the calling process doesn't care about that.

> why doesn't C have colouring issues if it too supports asynchronicity

But it does have those issues. In fact it is worse than in most other languages - just because there is no distinct support of async in a language does not mean it doesn't have async capabilities. C does have those, you just don't have any language support whatsoever, you have to take care yourself. Delegating the execution to the OS doesn't change that.

For example, if you want to transform elements in an array concurrently, you now have to use pthreads. You cannot do that while using the same code as before (which is would removing function coloring is all about).

Maybe we have a misunderstanding here: I'm talking abouyt sync/async not the async/await syntax that many languages have. The latter is to help deal with colored functions and the former is what enables the existance of concurrency. Any language that is - in one way or the other - capable of running code concurrently is suffering from that problem. It can be avoided by enforcing synchronous execution only, which some languages do (especially some DSLs), but most general purpose languages support concurrency.

And once they do, there's two choices: the programmer has to deal with it (using Thread.async, spawn, pthreads, go channels, promises, ...) or they don't. If they don't then sync and async code must look exactly the same (or be able to be converted forth and back automatically). And this goal (either making it look the same or converting it automatically) is just not solvable in general. It has been tried for decades.

For instance, check the paper "A Critique of the Remote Procedure Call Paradigm" from 1988. I think it was from Tanenbaum. It lists some of the problems that just cannot be solved in general. Some programs will always behave different after the automatic conversion, it's impossible to prevent that.

> > Okay, let me follow up on this one. So could we take arbitrary Erlang/Elixir programs and automatically/mechanically rewrite them to push previously inlined calculations to be run on a different process (using spawn) or the otherway around - and that without causing any observable difference in behaviour in any situation except for difference in performance?

> Yes.

What I'm trying to explain is that you are mistaken. It just is not possible. Maybe in specific cases, sure, but not in a safe way for every usecase that can be thought of.

And neither Erlang nor Elixir (nor the latest BEAM language, I forgot the name), luckily, try to do that. They give this controller to the developer and let them deal with the decision of how and when to use concurrency. They just try to make that as easy as possible, but they don't try to make sync and async code look and be the same.

They are not equivalent. Running a function in another process incurs the overhead of copying the data to the other process (potentially over the network). In both directions. Task.async introduces extra necessities like a linkage between the two processes, a timeout receiver, etc, and the result receiver.

Moreover, there is a context switch and it's more likely the code will run on a different core if you run it async than if you run it in the same process (which is very likely but not guaranteed) to run on the same core.

The way in which they are equivalent, is that the code that you write is identical, and the bytecode that gets run, and the existence of implicit yields is identical between the async and "sync" code.

> They are not equivalent. Running a function in another process incurs the overhead of copying the data to the other process (potentially over the network). In both directions.

> Moreover, there is a context switch and it's more likely the code will run on a different core if you run it async than if you run it in the same process (which is very likely but not guaranteed) to run on the same core.

Yes I agree, but the question as presented by valenterry asks us whether or not there is some existing semantic difference and not under the confines of performance (both gains or losses). Regardless what you have stated are all true.

But as an aside, and not directed towards you, Task.async/3 doesn't do anything that the developer cannot already do. Even in the tutorial for the learning Elixir, the fledgling developer is exposed to the different mechanisms that power Task.async/3, the source code [1] reflects this, although supervisors are covered much later down the line. The documentation for

> The way in which they are equivalent, is that the code that you write is identical, and the bytecode that gets run, and the existence of implicit yields is identical between the async and "sync" code.

And just to add on for those not familiar with the BIFs, receive, which is what Task.await/2 and Task.yield/2 use under the hood, yields execution. NIFs are another one.

[1] https://github.com/elixir-lang/elixir/blob/fb729784e5504f499...

Everything in the beam is "low-level" async with implicit yield points at function boundaries. Task.async just gives you a convenient future so you can get the value back, instead of the beam throwing it away (goroutines also throw away the result). You could also write that boilerplate by hand. There are also other nice things that Task.async gives you but that's not relevant to this discussion.

IOW, async/await in the BEAM is a high level thing (like goroutines) which is one of two reasonable choices, the other reasonable choice being "what zig did".

Yet David Fowler has been researching how to integrate Go and Java's approach into .NET.

Can you try and express what the semantic differences are exactly? Most approaches for expressing semantics yield equivalent results or nearly so. For example, using a Hoare triple approach, what conditions in P and Q would yield different assertion results between,

    assert P;
    foo(); // blocking call
    assert Q;

and

    assert P;
    await foo(); // await on an async call
    assert Q;

There may be some (e.g. thread identity is preserved in the former but possibly not in the latter, depending on implementation, and that could be expressed with proper assertions), but they only highlight how the differences are merely technical and certainly not "inherent".

In your examples you are explicitly handling the async nature of foo(). So you acknowedlge that there is a difference and you even use specific syntax to deal with it.

The question is rather: is it possible to remove the await and just automatically rewrite every program to use await where necessary?

And the answer is no, because with only sync functions you have the guarantee that if your thread crashes, every code immediately stops executing. Not so with async functions. Once they are run there is a chance that your own thread crashes but some other thread keeps executing more code, at least for some time, potentially executing side effects.

It cannot be automatically decided if this is okay or not. Or to turn it around: the behaviour of a program that automatically awaits async functions (which formerly where sync) can be different.

I think we're talking past each other. There is a semantic difference between `asyncFoo()` and `blockingFoo()`, but not between `await asyncFoo()` and `blockingFoo()`, and not between `asyncFoo()` and `spawn blockingFoo()`.

There's obviously a difference between launching an operation to run concurrently with the spawning code and launching an operation and awaiting its completion, but that difference does not require the introduction of an async/await feature, and is perfectly expressible with nothing but threads and synchronous code. In particular, there is no need for two separate syntactic domains between async and sync subroutines. Indeed, in Erlang, Go, and Java, all the same behaviour of async/await is achieved with no such need, and no need for different kinds of APIs.

> but not between `await asyncFoo()` and `blockingFoo()`

Of course there is. The former constraints the type (or return-type) of the function it is used in, the latter does not.

> but that difference does not require the introduction of an async/await feature

I don't remember saying anything about async/await in my original post though.

> Of course there is. The former constraints the type (or return-type) of the function it is used in, the latter does not.

That's not a semantic difference, but a syntactic one. Syntactic differences are expressed by which programs are accepted or rejected by the language; semantic differences are expressed by programs exhibiting different behaviours.

To show a semantic difference you need to find P and Q such that when the value of P is the same in the two programs, the value of Q could be different; that's what expresses the semantics of the code between P and Q. Again, there are some (e.g. P: `(t = currentThread(), true)`, Q: `currentThread() == t`), but they're purely technical.

> I don't remember saying anything about async/await in my original post though.

You said that the lack of need of a syntactic kind for asynchronous routines is a fallacy. It is not. Everything that you can describe with sync and async routines your can describe, with the same ease and semantics, with just sync routines and threads. There is no need for a special syntactic category of asynchronous subroutines.

> Everything that you can describe with sync and async routines your can describe, with the same ease and semantics, with just sync routines and threads. There is no need for a special syntactic category of asynchronous subroutines.

I agree with that statement.

> semantic differences are expressed by programs exhibiting different behaviours.

It can cause different behaviour whether code runs on the same thread or on a different one (see one of my other posts here). That doesn't have to happen in every program and it doesn't have to happen on each execution, but that still makes it a different behaviour - at least in my terminology.

> You said that the lack of need of a syntactic kind for asynchronous routines is a fallacy. It is not.

That's not what I said (or at least not what I meant). I meant that it is doomed to fail to try to generally unify sync and async code on a language (or runtime) level.

When people say a language "unifies sync and async code" all they mean is that there is no distinct syntactic category for "async routines". Rather, asynchronous operations are expressed as regular (i.e. synchronous, blocking) methods spawned on another thread. Either you agree that there is no loss in what you can express or believe it is doomed to fail. I don't understand how you can do both.

You should elaborate. I just want to read a file, program flow should stop until it is done. sync or awaited async call for this semantically sounds exactly same to me.

I knew Go can do it, so it doesn't "always fail"

> You should elaborate. I just want to read a file, program flow should stop until it is done. sync or awaited async call for this semantically sounds exactly same to me.

What if you are reading the file and somewhere else in the code there is some logging happening. Should the logging flow stop until reading the file is done? Should it not? If not, what should happen if the "thing" (e.g. thread) that executes the code to read the file crashes?

Go does not treat sync and async the same. It specifically has syntax to deal with async, e.g. go routines.

> Do you remember ConfigureAwait(false)?

I remember it and still use it, and you should, too. It can be elided in ASP.NET Core, which is the web application framework. Outside of that, there’s no escaping that you have to know what you’re doing.

It's this type of stuff that causes me to prefer the Sync implementations over the Async wherever I can. Did I get it right? Is it going to deadlock in production? How can I be sure? Who knows? Straight threads are much easier for me to understand than async in C#.

For desktop apps, you have to design the UI completely differently if it async or blocking. If it is async then suddenly the user can push other buttons while the previous action is executing which makes it a magnitude more complex.

You never want to block your UI thread for any amount of time a user would notice say more than 20ms. And if you use async or another thread you can always update the UI to disabled with spinner state. This is all less convenient for the programmer but much better for the UX. Blocking the UI blocks it all, even the close button!

> I think high level language/runtime should handle it automatically like it is a sync call

I understand that such a design is called "green threads" https://en.wikipedia.org/wiki/Green_thread

As far as I know, there have been recent prototypes in .NET, but nothing shippable.

https://stackoverflow.com/questions/37285366/green-threads-i...

https://www.reddit.com/r/csharp/comments/vapgxn/net_experime...

Oh, thanks for the reminder. ConfigureAwait still lingers in a few of my libraries.

It took me quite a bit of coffee to fully appreciate the simple phrase "it's async/await all the way down"

> I have the conclusion opt-in async/await is a mistake. I think high level language/runtime like c# should handle it automatically...

This is not a good idea unless you like having to make all your code thread-safe (well, async-safe). Explicit declaration of "hey this method might not run all at once or be interleaved with other methods" is a good idea; explicit opt-in is the same in every other implementation of async/await (ex: js).

I don't understand. Async methods can call sync methods already, how it is not mixed already?

Nope, it is not same in every other implementation. e.g Golang.

> ConfigureAwait(false)

It’s literally the configuration, to be used according to your needs

> Do you remember ConfigureAwait(false)?

What makes you think that this no longer applies? It still does to my knowledge. The page you linked to has a section for it that just says "TBD".

I actually have no idea if it still applies, but I haven't seen code that uses it for a long time. Maybe it was important for desktop apps?

It's important to have ConfigureAwait(true) (the default) any time you need the continuation to be on the same thread. This is used/needed for many UI frameworks (WPF), which is why it is the default. Otherwise you should use (false) for performance reasons.

Yes, this is very important.

I experienced deadlocks in ASP.NET Core 2.1 when I connected a legacy application to the web with it and the developers of Microsoft.AspNetCore.Mvc.Formatters.Json forgot to add .ConfigureAwait(false) to a async call into Newtonsoft.Json.

This was easily fixed by writing a small warper around the JsonInputFormatter but it shows how important ConfigureAwait is. When writing a library for anything netstandard, always put ConfigureAwait(false) after each async call where it is not mandatory to return to the same thread again.

> It's important to have ConfigureAwait(true) ... Otherwise you should use (false)

This is not universal. In modern ASP.NET core (i.e. web services), ConfigureAwait does nothing and is not recommended. I can confidently say that I don't want it in any of the code that I work on. We can make blanket statements, but clearly we do different things using different patterns and one size does not fit all.

https://stackoverflow.com/questions/42053135/configureawaitf...

ConfigureAwait (mostly) no longer applies. You might want it in .NET 4.x apps, and in libraries that might be used in them. Or in apps with a "UI thread" - i.e. desktop GUIs.

In modern ASP .NET, it does nothing, so we don't use it.

https://stackoverflow.com/questions/42053135/configureawaitf...

Your advice is correct for ASP.NET Core (the web application framework), but not for all of .NET (Core or suffixless).

Thanks, I have edited. Even inside .NET different people work on very different things, and my niche might not be your niche.

Not even (fully) for ASP.NET Core, as Blazor Server does use a SynchronizationContext.

Reply on news.ycombinator.com