Is your feature request related to a problem? Please describe.
this feature request is a follow up of #1836. ideally multiple preprocessing threads would 1. read from a shared cache (storing non random transform outcomes) and 2. have independent transform instances for each of them. 2 is not currently implemented as a result the transform instance properties might be modified by multiple threads simultaneously and causing issues #1030 #1836.

the caching logic could be improved so that it is easier to make customised transforms for the cache based datasets.