it's a cost/latency trade-off in production + very use-case dependent