Have you looked into whether that 1-2M threshold still applies if the tables are partitioned? Such as a multi-tenant application where a search might be constrained to a single user_id. I wonder if that would provide, at least temporarily, a way to get past that constraint.