venice

Complexity is one downside. Because it is a stateful client, that adds some operational constraints. We cannot expand the fleet of clients near-instantly to deal with a traffic surge elastically. New (empty) client instances need to bootstrap their state which makes their startup take longer.

Cost is another downside. It takes local RAM or local SSD to host the state.

Feasibility can also be a limiting factor. If the dataset is larger than what can reasonably fit in one node, then Da Vinci will not be a good fit. Whereas data can be partitioned in the Venice servers and scale horizontally, and the stateless thin client can query the right partition and get the data back.

That makes a lot of sense! Thank you for expanding on that