Jon Gjengset

Yeah, that certainly doesn't look great. Seems like almost all the time is spent on the garbage collection. Which is surprising to me since reads shouldn't generate garbage. This is...

That's crazy, and should definitely not be happening. In a read-only benchmark, we shouldn't be generating _any_ garbage. I'm currently kind of swamped, but maybe @domenicquirl can take a look?

Wow, that would be really bad. This smells like a bug if it allocates even if there is no garbage...

> At least the default handle is re-used all the time Hmm, that doesn't seem right. `LocalHandle` isn't `Send`/`Sync`, so you can't really re-use it except in a thread-local. It's...

> Why are we using a custom collector anyway? I don't see a benefit compared to the global static one. The biggest reasons is for memory reclamation. By using a...

> I was only thinking so far as that the check would cover the case where you try to access with a guard from a wrong local. Yup, the checks...

> Why would we want to have the entire map be local to each thread? Ah, sorry, that wasn't what I meant. What I meant was that you'd store a...

Oh, no, that won't work, since `LocalHandle` isn't `Send`.

Oh, I see, you're proposing changing the implementation of `crossbeam-epoch`! I was thinking of ways to make changes in `flurry` specifically. May be worth raising this on the crossbeam issue...

@domenicquirl Interesting, `flize` doesn't have a way to (unsafely) get a reference when you know no-one else has one? That seems odd.