I agree wholeheartedly. If we tell people "Don't write code that panics" then we collectively start writing code that assume panics don't happen. Suddenly panics become a footgun in Rust, and it isn't the safe language we want it to be.
If I'm feeling in a mood, I think we should panic all the time. We should have a "panic monkey" tool that places panics at random parts of the code, to see how panics are handled. Eg to see if they interact poorly with unsafe code.
I can only agree diesel had (maybe still has) a serve bug around connection reuse and panics.
Panics in rust originally where a pretty nice way to have a error kennel pattern and as long as you always kill the whole error kennel in which the panic occurred and don't share mutable state between error kennels this fully save from any inconsistent state.
Sure panics across FFI boundaries can't be done nice at all, but this can still be handled by making sure there is no FFI boundary insider of an error kernel, just in between.
So I would push panic usage for error kennels, and just that. But also the unwind safe traits need an refresh: Enforce them (by linting against inconsistencies) and better defaults for what does and doesn't implement them.
For example a type which is sync but doesn't implement the referential unwind safe trait is broken/bad in 99% of cases. (Either it should implement unwind safe market or not be send). The remaining 1% are edge cases between the usage of catch unwind and thread local storage.
Sure, but having process error kernels isn't at all appropriate for many applications. E.g. you don't start a process for every http request (or actor request) your server handles (even starting a thread for everyone is considered to much overhead).
But not only is the overhead of "starting"/entering a error kernel a problem you also normally do share _immutable_ state (e.g. configs, lookup tables), as well as some _minor_ amount of mutable but well known to be panic safe state (e.g. simple performance counters and some caches). Which is now much harder to get right.
E.g. in a cache you want it potentially to be shared across error-kernels and re-use it after a kernel crashed but only if the panic didn't come from the cache. So the cache needs to be handled in it's own error-kernel which makes it important to make the intercommunication fast.
Sure you _can_ do all this with processes (by using both IPC and mmap together), but it's much harder to get right. Often this leads to either serve performance limitations or braking of the kernels surface (by using mmap in a bad way). It also tend to lead to code which is very dependent on the features of a specific operating system. Problems with managing startup/shutdown (no systemd isn't a solution ), moving resources between kernels etc.
So all in all processes make thinks more complex, less portable with very little to no benefits for many use cases. (Sandboxing of very large error-kernels like done in browsers is one point where it makes sense).
If I'm feeling in a mood, I think we should panic all the time. We should have a "panic monkey" tool that places panics at random parts of the code, to see how panics are handled. Eg to see if they interact poorly with unsafe code.