You also absolutely take the time to think through what should and should not be exposed to end users. Either you creates a namespace for your API with views that expose only the necessary data, or with something like Postgraphile you can specify which functionality, columns, and tables are hidden. Personally I prefer the view approach as it decouples your API from the database implementation.
RLS is difficult to unit test, to configure with flags, not portable to secondary stores, and requires knowledge of arcane database features.
When you put some private data on a cache somewhere and query it, now you need to write your access layer twice since the other one is so tightly coupled to the database.
As an example, consider a query that is fairly complex. How exactly are you going to make resuable security tests for it? How do you know that an error in the query won't expose unexpected data, or that a modification to the query that introduces and error exposes unexpected data? Your security now has to be customized to every query like this.
The alternative case is that you only use super simple queries so that you can write resuable security checks. But now I'm going to ask why you are using a relational database in the first place? Because almost certainly you're doing manual joins in this scenario and incurring huge performance hits.
If you are using a relational database to its full extent, there is nothing as secure as built in row level security. You wouldn't dream of doing integrity checks outside the database, why do security?
This isn’t exactly PostgreSQL’s fault, but mapping application users to database roles is not as easy in pretty much any framework I’ve worked compared to implementing basic post-retrieval filters.
I absolutely think bringing security into the database is good, but it’s also not straightforward to integrate in even assuming you know how to define the necessary database elements.
Row-level security is absolutely not dependent upon DB roles. Table-level security on the other hand is sufficiently coarse-grained that mapping GRANT/REVOKE to applied roles should be feasible.
If you're punting all of this security to the app layer, that's your prerogative, but don't pretend that it's somehow more straightforward or secure. If it seems simpler at the app layer, you may very well be missing something.
The parent's point is that tooling around this is weak, which undermines its value. Not being able to review and easily test authz changes will almost certainly lead to issues.