Vik Fearing <vik.fearing@dalibo.com> writes:
> This latest patch is ready for a committer to look at now. The weird
> comments have been changed, superfluous regression tests removed, and
> nothing done about expression pushdown per (brief) discussion.
I started to look at this patch and realized that there's an issue that
isn't covered, which is not too surprising because the existing code fails
to cover it too. Remember that the argument for pushing down being safe
at all is that we expect the pushed-down qual to yield the same result at
all rows of a given partition, so that we either include or exclude the
whole partition and thereby don't change window function results. This
means that not only must the qual expression depend only on partitioning
columns, but *it had better not be volatile*.
In exactly the same way, it isn't safe to push down quals into
subqueries that use DISTINCT unless the quals are non-volatile. This
consideration is missed by the current code, and I think that's a bug.
(Pushing down volatile quals would also be unsafe in subqueries involving
aggregation, except that we put them into HAVING so that they're executed
only once per subquery output row anyway.)
Given the lack of prior complaints, I'm not excited about back-patching a
change to prevent pushing down volatile quals in the presence of DISTINCT;
but I think we probably ought to fix it in 9.5, and maybe 9.4 too.
Thoughts?
regards, tom lane
