The problem is that in order to properly invalidate cached plans we must record the constraint OIDs which the plan depends on. We can do that for PK and UNIQUE constraints, unfortunately, we can't do it for NOT NULL constraints as, at the moment, these are just properties of pg_attribute. For this to be made to work we'd need to store those in pg_constraint too, which is more work that I'm going to do for this
patch.
I was just wondering, since get_primary_key_attnos opens pg_constraint and just skips attributes with other constraint types than primary, what would be the reason we wouldn't just also open pg_attribute and check for the non-nullness of the non-primary key unique attributes? Couldn't we add a function which opens both pg_attribute and pg_constraint to get non-null unique attributes? I suppose that would have a performance impact. And, I did notice the FIXME in the comment before check_functional_grouping which seems to make the argument that there are other use cases for wanting the non-nullness represented in pg_constraint.
* Currently we only check to see if the rel has a primary key that is a * subset of the grouping_columns. We could also use plain unique constraints * if all their columns are known not null, but there's a problem: we need * to be able to represent the not-null-ness as part of the constraints added * to *constraintDeps. FIXME whenever not-null constraints get represented * in pg_constraint. */
