when do you actually need to shard a database, considering all the hassle involved

we're hitting some serious scaling issues with our primary relational database. read and write loads are getting pretty high, and we're starting to feel the impact on performance. we've talked about sharding, but the complexity of managing a sharded setup, handling cross-shard queries, and ensuring data consistency seems like a huge undertaking. are there specific metrics or thresholds you look for that signal sharding is the right move, or is it more of a 'when all else fails' kind of situation? what have been your experiences, good or bad?