Arc + work stealing scheduler is common. But work stealing schedulers are common (eg libdispatch popularized it). I believe the only alternative is thread-per core but they’re not very common/popular. For what it’s worth zig would look very similar except their novel injectable I/O syntax isn’t compatible with work stealing.

Even then, I’d agree that while Arc is used in lots of places in work stealing runtimes, I disagree that it’s used everywhere or that you can really do anything else if you want to leverage all your cores with minimum effort and not having to build your application specialized to deal with that.

Being possible with minimal effort doesn't really preclude it from it not being the default. The issue I have is huge portions of Tokio's (and other async libs) API have a Send + Sync constraint that destroy the benefit of LocalSet / spawn_local. You can't build and application with the specialized thread-per core or single-threaded runtime thing if you wanted to because of pervasive incidental complexity.

I don't care that they have a good work-stealing event loop, I care that it's the default and their APIs all expect the work-stealing implementation and unnecessarily constrain cases where you don't use that implementation. It's frustrating and I go out of my way to avoid Tokio because of it.

>You can't build and application with the specialized thread-per core or single-threaded runtime thing if you wanted to because of pervasive incidental complexity. [...] It's frustrating and I go out of my way to avoid Tokio because of it.

At $dayjob we have built a large codebase (high-throughput message broker) using the thread-per-core model with tokio (ie one worker thread per CPU, pinned to that CPU, driving a single-threaded tokio Runtime) and have not had any problems. Much of our async code is !Send or !Sync (Rc, RefCell, etc) precisely because we want it to benefit from not needing to run under the default tokio multi-threaded runtime.

We don't use many external libs for async though, which is what seems to be the source of your problems. Mostly just tokio and futures-* crates.

I might be misremembering and the overbearing constraints might be in Axum (which is still a Tokio project). External libs are a huge problem in this area in general, yeah.

Single-threaded runtime doesn't require Send+Sync for spawned futures. AFAIK Tokio doesn't have a thread-per-core backend and as a sibling intimated you could build it yourself (or use something more suited for thread-per-core like Monoio or Glommio).

I mostly wind up using Arc a lot while using async streams. This tends to occur when emulating a Unix-pipeline-like architecture that also supports concurrency. Basically, "pipelines where we can process up to N items in parallel."

But in this case, the data hiding behind the Arc is almost never mutable. It's typically some shared, read-only information that needs to live until all the concurrent workers are done using it. So this is very easy to reason about: Stick a single chunk of read-only data behind the reference count, and let it get reclaimed when the final worker disappears.