Re: Parameterized paths vs index clauses extracted from OR clauses

Robert Haas <robertmhaas(at)gmail(dot)com> writes:
> On Thu, Feb 28, 2013 at 3:03 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> Whichever way we go, the resulting patch is likely to be too large and
>> invasive for me to feel terribly comfortable about back-patching it into
>> 9.2. AFAICT this issue only arises for indexquals extracted out of
>> larger OR conditions, so maybe it's not worth taking such a risk for.

> EnterpriseDB has received a number of complaints from our customers
> resulting from planner behavior changes which were back-patched; so I
> am not sanguine about back-patching unless the situation is pretty
> darn dire and the fix is pretty darn near certain to be an improvement
> in every case.

Well, the point is not so much about whether it's an improvement as that
9.2's current behavior is a regression from 9.1 and earlier. People may
not like changes in minor releases, but they don't like regressions
either.

>> A downside of this approach is that to preserve
>> the same-number-of-rows assumption, we'd end up having to enforce the
>> extracted clauses as filter clauses in parameterized paths, even if
>> they'd not proved to be of any use as index quals.

> I'm not sure I fully grasp why this is a downside. Explain further?

Because we'd be checking redundant clauses. You'd get something like

Nested Loop
Filter: (foo OR (bar AND baz))

... some outer scan here ...

Index Scan:
Filter: (foo OR bar)

If "foo OR bar" is useful as an indexqual condition in the inner scan,
that's one thing. But if it isn't, the cycles expended to check it in
the inner scan are possibly wasted, because we'll still have to check
the full original OR clause later. It's possible that the filter
condition removes enough rows from the inner scan's result to justify
the redundant checks, but it's at least as possible that it doesn't.

> Since there's little point in using a paramaterized path in the first
> place unless it enables you to drastically reduce the number of rows
> being processed, I would anticipate that maybe the consequences aren't
> too bad, but I'm not sure.

Yeah, we could hope that the inner scan is already producing few enough
rows that it doesn't matter much. But I think that we'd end up checking
the added qual even in a non-parameterized scan; there's no mechanism
for pushing quals into the general qual lists and then retracting them
later. (Hm, maybe what we need is a marker for "enforce this clause
only if you feel like it"?)

regards, tom lane