Schemaless databases are good for scenarios where the database isn't a source of truth. If you have a table full of e.g. per-second heartbeats from a bunch of deployed services, there's no fundamental underlying truth anyone's trying to gather from it, and you can't afford to run a full schema migration every time someone adds a new metric.

I recognize some people do try to use schemaless databases in the way you're describing, and I agree that's weird and dangerous.