investigate converting graph hashmap into an MRU/LRU

[JimFuller-RedHat]

Which will naturally partition data along the memory and/or temporal axis.

Specifically we want to investigate (maybe an ADR) on using something like this https://crates.io/crates/lru on analysis graph hashmap.

[chirino]

did a little searching, and found https://github.com/marmeladema/clru-rs which is a weighted lru. Does anyone have better options?

[JimFuller-RedHat]

comment: this recent fix #1211 fixed graph hashmap to actually persist a graph eg. in some previous work broke it and were actually (lazily) continually reloading graphs into memory.

[chirino]

Took a peek at replacing the Map in the GraphMap with CLruCache. It was mostly straightforward until you got to the point where we implemented the GraphMap::get :

• since the LRU cache needs to be updated on get to track that it was recently used, the get method needs be changed to take a &mut self ref.
• this has the ripple effect of needing a write lock instead of read lock in query_graph.

I don't think we want to do that. You don't want your read path to serialize on a writelock. Is there a better way to restructure this?

Or should we be looking at more complex cache implementations like https://github.com/moka-rs/moka ?