David Rowley <dgrowleyml@gmail.com> writes:
> There are quite a lot of things which would have to be restricted,
> much more than you might think. Any conditions which could cause an
> error would have to be evaluated last in a WHERE clause and that might
> result in being unable to use indexes because some other (possibly
> unindexed) expression would need to be evaluated first.
That particular aspect might not be too awful, because there's already
a good deal of pressure for index opclass members to not fail ---
certainly I'd expect that all standard btree comparison functions
could be marked non-failing. (Let us slide quietly past the question
of exactly how strong the guarantee should be; for example any function
that takes toastable argument types is potentially at risk of OOM,
but do you really want to exclude numeric_eq and texteq from the
set of safe operations? See also past discussions on how strict
we should be about the related concept of leakproofness.)
In any case I agree with your larger point that ordering things such
that potentially-failing tests are always done last would be a
performance disaster, even if it's possible at all.
regards, tom lane