This is a good suggestion because of the flexibility and depending on usage, RAM, and strategy choices seems likely to me to be much faster than file based stuff. If it’s a few hundred smallish JSON “views,” it could all live in memory and only need IO on cache misses/expirations. If it’s mostly needing IO then it would be slower since it’s just one more layer of housekeeping on top of it all. As you say, the CHI backends make trying it out in different ways quite easy.