I don't see any reason to have a non-poisoning mutex other than a footgun.

I'd be in favor of an interface that panics by default when locking a poisoned mutex, as it is what most people would want anyways.

I didn't know this. It's pretty bad. Non poisoning mutexes means you can easily access inconsostent state.

The main reason I give for picking std mutex over parking_lot is because of poisoning.

No measuring with a scale

I think that while performance at a single-node level is normally not that critical in web development, memory-safety, type-safety and correctness is really important!

The alternative in that domain that seems great is @gleam.run, although I haven't tested it yet, I'm quite excited about it.

Ah good old stack overflow

Enough fairy tales have taught that forests are, indeed, not memory safe.

Makes sense, I'm just not familiar at all with safety-critical software development but I'm really curious about it.

Isn't there a part where you prove that what you are gonna do is correct?

So I think we actually do handle the case of operations failing, and the process invoking the operation not knowing about it by making the operations sequential.

Yeah, that's what I mean, we force consistency by sequentially ordering operations on the same piece of state just like DB transactions. But there's nothing but speed of the I/O that distinguish this from access to state in a distributed system.

You know if it failed or succeeded because you compromise on availability, right? Local operations are normally sequential

But if you view local hardware as distributed you could in theory take parition tolerance in that setting. It's just that we normally work at a rate of failure where it doesn't matter.