Path unpacking takes up a considerable amount of the query time for MLD (and CH actually). A simple technique to mitigate this is to create an unpacking cache. Generally this cache would store an EdgeID or source, target pair to unpacked path in a LRU manner.

There are a lot of knobs one can tune to make this perform best some of the basic considerations:

1. What do we store in the cache: Just EdgeIDs that we need to turn into PathData again to full lists of PathData
2. What is a good cache size and how do we determine it
3. What kind of edges should be cached: For MLD we have a lot of choices here (e.g. only store clique arcs from certain levels), for CH there is no good mechanism to decide when a shortcut should be cached and when not.
4. Should we have a thread-local cache or a shared cache

This would hook into SearchEngineData (since it needs read-write and thread-safe access) and would need to be cleared on every data update.

/cc @oxidase