dashmap VS stretto

The problem is, every single one of these half-dozen crates has at least one known major issue (including UAF), exactly like C++ implementations (which isn't surprising since it's the kind of things where the ownership isn't clear and then the borrow checker can't help us).

Deadlocks and leaks are easy as other languages.

Have you considered https://github.com/xacrimon/dashmap ?

Dashmap made some api changes compared to the stdlibs hashmap, which leads to some oddities, as highlighted here: https://github.com/xacrimon/dashmap/issues/175

Some additional notes are in this slide deck and the implementation javadoc. You'd probably want to use something like DashMap for the hash table.

You can look at existing concurrent maps like Dashmap https://github.com/xacrimon/dashmap or Cashmap https://gitlab.redox-os.org/redox-os/chashmap

Also dashmap

The only one I've used is Dashmap, it's a concurrent interior-mutability hashmap. Very convenient crate in the case you need that.

Hi, I think this link is a good explanation https://github.com/al8n/stretto/pull/7

Ya, I saw concache but I looked into it and it doesn't implement what is needed. Each bucket has its own linked-list backing (hence "lock-free linked list buckets"). An LRU needs each value in each bucket to be part of one linked list I believe. After posting this I realized my line of research was failing because it was state of the art five years ago. Caffeine replaced `concurrentlinkedhashmap` in the java world (by the same author). A rust version of that is Moka. These are much more complicated than a concurrent LRU but faster (aka more state of the art). Another rust crate is Stretto which is a port of dgraph's Ristretto (in go). The question becomes is it worth it to essentially port `concurrentlinkedhashmap` to have a great concurrent LRU when there are more state of the art caches out there.

For the case in the benches folder(a very roughly bench case), stretto is around 20 - 30 ms(sync version is around 30 - 40 ms) faster than moka, for 120, 000+ operations. I set stretto to collect metrics when benching, collecting metrics will make around 10% overhead. Moka seems not to provide a configuration to collect the metrics, so the hit-ratio is not compared.