It's a leaky abstraction (https://www.joelonsoftware.com/2002/11/11/the-law-of-leaky-a...), but maybe there is no helping it and for some reason it is a necessary tradeoff for performance?
I suppose that a wrapper like "aprint" (a convenience function labelled async, like with an "a" prefix), would be a bit better than having people continually try using print, not await it, and not getting the expected output in stdout (or whatever stream it's sent to), while they are in the middle of trying to test something or otherwise get something working because I'm of the opinion that common things should be easy. Maybe "people would generally expect a print function to just work and not return a promise or something" is an abstraction? "aprint" might actually be the wrong name I'm not sure I've really thought about it right.
Either all side effects should be marked or none should. Ret-connecting await annotations as an useful feature instead of a necessary evil is baffling.
Memory allocation by comparison are extremely quick, and generally very reliable. Your system’s memory subsystem isn’t a smorgasbord of different memory drivers and controllers. It one memory system, taking to one memory controller, via an API that been standardised for decades, and where every implementation of that API is basically tested to the extreme every time a computer turns on. That’s assuming your language even bother asking the OS for memory on every allocation, which it probably doesn’t. Most language runtimes request large blocks of memory from the OS, then allocate out of those block on demand. So most “allocating functions” never result in syscall at all.
The fact the most allocations are fulfilled via internal pools is immaterial, at some point the allocator needs to ask the OS for more memory. This parallels the way that most I/O doesn't actually performs syscalls because of buffering.
Also allocations might end up performing arbitrary I/O indirectly if the OS needs to flush dirty pages to disk to free up memory.
I write an in memory kv cache. It’s in memory so no async needed. Now I create a trait and implement a second version with file backing. Now the children are crying because async needs to be retroactively added and also why, makes no sense etc.
Perhaps anything involving syscalls should be exposed and contractual. I doubt it, but maybe it’s important for some obscure ownership-of-resources reason. But then why the inconsistency between traditional and pooled syscalls? The only difference is whether the runtime sits in the kernel or in user space. The only one who should care is the runtime folks.
My take has been for years that this is throwing complexity over the fence and shaming users for not getting it. And even when they do get it, they Arc<Mutex> everything anyways in which case you are throwing the baby out with the bathwater (RAII, single ownership, static borrowing).
Because the kernel doesn't expose that contract, so they don't have that behaviour.
> The only difference is whether the runtime sits in the kernel or in user space.
In other words, what contracts you have control over and are allowed to provide.
> My take has been for years that this is throwing complexity over the fence and shaming users for not getting it.
I'm sure how we got here would seem baffling if you're going to just ignore the history of the C10K problem that led us to this point.
You can of course paper over any platform-specific quirks and provide a uniform interface if you like, at the cost of some runtime overhead, but eliminating as much of this kind of implicit runtime overhead as possible seems like one of Rust's goals. Other languages, like Go, have a different set of goals and so can provide that uniform interface.
It's probably also possible to have some of that uniform interface via a crate, if some were so inclined, but that doesn't mean it should be in the core which has a broader goal.
I am not unaware of pooled syscalls. I worked on the internals of an async Rust runtime, although that should not matter for critiquing language features.
The archeological dig into why things are the way they can come up with a perfectly reasonable story, yet at the same time lead to a suboptimal state for a given goal - which is where the opinion space lies - the space where I’m expressing my own.
> but eliminating as much of this kind of implicit runtime overhead as possible seems like one of Rust's goals
Yes, certainly. And this is where the perplexity manifests from my pov. Async is a higher level feature, with important contractual ecosystem-wide implications. My thesis is that async in rust is not a good solution to the higher level problems, because it interacts poorly with other core features of Rust, and because it modularizes poorly. Once you take the event loop(s) and lift it up into a runtime, the entire point (afaik - I don’t see any other?) to abstract away tedious lower level event and buffer maintenance. If you just want performance and total control, it’s already right there with the much simpler event loop primitives.
In short, I fail to see how arguments for async can stand on performance merits alone. Some people disagree about the ergonomics issues, which I am always happy to argue in good faith.
>Because the kernel doesn't expose that contract, so they don't have that behaviour
Which OS are we talking about? Linux doesn't really have mutices as primitives. You can build async mutexes on top of eventfd and soon even on top of futexes with io_uring.
Implementations of the standard library mutexes are here https://github.com/rust-lang/rust/tree/master/library/std/sr...
And of course it’s pthreads on Linux.
As an aside, it is interesting that rust uses pthread_mutex for its standard library mutex. GCC/libstdc++ regrets that decision as its std::mutex is now way larger than it needs to be but it is now permanently baked in the ABI. I guess rust still doesn't guarantee ABI stability so the decision could be reversed in the future.