Hi Tjark,
I would be suprised if anyone could provide a one-size-fits-all rule of thumb for this. It depends on soo many factors: the size of the datasets, the diversity of the datasets, the amount of joins, the amount of available server memory etc.
It is possbile to reason on this a little bit, see also:
https://docs.mendix.com/refguide/performance-best-practices#mxp007
https://academy.mendix.com/link/module/22/lecture/180/4.5-XPath-Optimizations
But in the end testing out different variations is the proof of the pudding to me.
I hope this helps.