David Rowley <dgrowleyml@gmail.com> writes:
> For just planner smarts such as LEFT JOIN removal, Unique Joins, and
> all this Unique Key stuff, we really don't need to record the
> dependency as if the index or constraint is dropped, then that'll
> cause a relcache invalidation and we'll see the invalidation when we
> attempt to execute the cached plan. That will cause the statement to
> be re-planned and we'll not see the unique index when we do that.
You need to make sure that the thing you're concerned about will actually
cause a relcache invalidation of a table in the query. But yeah, if it
will then there's not a need to have any other invalidation mechanism.
(It occurs to me BTW that we've been overly conservative about using
NOT NULL constraints in planning, because of failing to consider that.
Addition or drop of NOT NULL has to cause a change in
pg_attribute.attnotnull, which will definitely cause a relcache inval
on its table, cf rules in CacheInvalidateHeapTuple(). So we *don't*
need to have a pg_constraint entry corresponding to the NOT NULL, as
we've mistakenly supposed in some past discussions.)
regards, tom lane