On Wed, 18 Sept 2024 at 19:28, Peter Eisentraut <pe...@eisentraut.org> wrote: > > On 12.09.24 03:43, David Rowley wrote: > > (Likely it could just look at pg_attribute.attnotnull instead) > > That won't work because you can't record dependencies on that. (This is > one of the reasons for cataloging not-null constraints as real constraints.)
I'm not seeing any need to record constraint dependencies for this optimisation. It would be different for detecting functional dependencies in a view using a unique constraint+not null constraints for ungrouped columns, but that's not what this is. This is just a planner optimisation. The plan can be invalidated by a relcache invalidation, which will happen if someone does ALTER TABLE DROP NOT NULL. For reference, see 5b736e9cf. David
