They work well when designed well - and for me that's more often designed organically. If they're designed badly, most bugs eventually reduce to trying to do distributed transactions across microservices.

That's my yardstick - if I see poor man's distributed transaction attempts happening across service boundaries I'll start to assume over-engineering / cargo culting. If I see services acting a nodes in an pipeline or a flowchart, preferably with multiple incoming lines, it smells better to me.