The failure-mode matrix: what throws, what silently returns zero
Concept 53 handed you a table and asked you to memorize it. Fair warning: memorizing a table and being able to diagnose a live incident from it are different skills. In production you never get to see the table -- you get exactly one signal, a row count or an error code, and from that signal alone you have to say which clause fired and why. This concept is where you stop reciting the matrix and start reading it off a real, running database, including the two corners that memorizing prose never prepares you for: a write that throws after it already happened, and a relationship that leaks through a door RLS was never watching.
try to break it
you drive.
BOTH run, at different moments. USING asks 'may you touch this row?' before the change — fail it and zero rows are affected, silently. WITH CHECK asks 'may it become that?' after — fail it, by giving your row away, and it throws 42501. Omit WITH CHECK and Postgres reuses USING.
USING clause silently filters — zero rows, no error. A failing WITH CHECK clause throws 42501.So it is not true that "only
INSERT throws": an UPDATE that tries to move a row out of your ownership fails WITH CHECK and raises too. And because three of the four commands can fail in total silence, a broken policy and a working one look identical from the client. Test by asserting on the emptiness.