Okay, I think I got your point. Can you please confirm that I fully understand it? So, you’re saying that the biggest problem is the speed of read from hdd, not the intersect operation and the size of the structure with a list of UIDs or network latency potentially needed to execute this query? Basically it means that if I can (or can afford to) map the FS on all Alfa to RAM, even the query will be very fast?
On you other point with testing - yes, I will definitely do that. Right now I’m stuck with the dataset migration here The --upsertPredicate flag is ignored during JSON live upload if you can please help me with that too.