How to correctly design multi-thread cache that's transparent to user iteration?

Why task_local_storage is not ideal

task local storage is no good because if you have 100 ev in a cluster, and the user spawned a few tasks within that cluster, you end up decompressing that cluster more than 1 time.

they may hit OOM if they used @sync for ... @spawn pattern, because then you decompress every cluster and the same time, with high multiplicity