Обсуждение: Bad estimation for "where field not in"

Hello,

We have a table with about 60M records, almost all of which in one of
two statuses ('done', 'failed') and a few of them, usually < 1000, in
different transient statuses. We also have a partial index indexing
the transient items: where status not in ('done', 'failed'). Stats are
about right for the status field:

  stavalues1   |     stanumbers1
---------------+---------------------
 {done,failed} | {0.541767,0.458233}

Trying to access only the transient items leads to very bad records
estimations, and the choice of poor query plans, as the estimate is to
read some 25% of the items table (so joins are often performed against
full scans of other large tables instead of using indexes):

explain analyze select count(*) from items where status not in
('done', 'failed');

 Aggregate  (cost=2879903.86..2879903.87 rows=1 width=0) (actual
time=0.460..0.461 rows=1 loops=1)
   ->  Bitmap Heap Scan on items  (cost=3674.23..2843184.08
rows=14687908 width=0) (actual time=0.393..0.453 rows=20 loops=1)
         Recheck Cond: (((status)::text <> 'done'::text) AND
((status)::text <> 'failed'::text))
         ->  Bitmap Index Scan on i_items_transient_status
(cost=0.00..2.26 rows=14687908 width=0) (actual time=0.381..0.381
rows=38 loops=1)

Looking at these estimate of the rows in the table (59164756) and the
estimate of the filtered rows (14687908), looks like the planner is
calculating the probability of the status being neither done nor
failed as two events independent events:

=# select 59164555 * (1 - 0.541767) * (1 - 0.458233);
 14687927.231665933605

while it is clear (at least in the original query specification,
before splitting the condition in two ANDed parts) that the two events
are disjoint, so the probability should be calculated as 1 - (p(done)
+ p(failed)) instead of (1 - p(done)) * (1 - p(failed)).

Writing the query without the "not", listing the other statuses, leads
to a correct plan instead.

Is this a known planner shortcoming or something unexpected, to be
escalated to -bugs? Server version is 9.0.1.

-- Daniele

On Thu, Mar 1, 2012 at 6:40 PM, Daniele Varrazzo
<daniele.varrazzo@gmail.com> wrote:
> Is this a known planner shortcoming or something unexpected, to be
> escalated to -bugs? Server version is 9.0.1.

The relevant code is in scalararraysel() function. It makes the
assumption that element wise comparisons are completely independent,
while the exact opposite is true. This has been this way since
http://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=290166f93404d8759f4bf60ef1732c8ba9a52785
introduced it to version 8.2.

At least for equality and inequality ops it would be good to rework
the logic to aggregate with
s1 = s1 + s2 and s1 = s1 + s2 - 1 correspondingly.

--
Ants Aasma

Ants Aasma <ants.aasma@eesti.ee> writes:
> On Thu, Mar 1, 2012 at 6:40 PM, Daniele Varrazzo
> <daniele.varrazzo@gmail.com> wrote:
>> Is this a known planner shortcoming or something unexpected, to be
>> escalated to -bugs? Server version is 9.0.1.

> The relevant code is in scalararraysel() function. It makes the
> assumption that element wise comparisons are completely independent,
> while the exact opposite is true. This has been this way since
> http://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=290166f93404d8759f4bf60ef1732c8ba9a52785
> introduced it to version 8.2.

> At least for equality and inequality ops it would be good to rework
> the logic to aggregate with
> s1 = s1 + s2 and s1 = s1 + s2 - 1 correspondingly.

Yeah, I was about to make a similar proposal.  In principle, when
working with a constant array, we could de-dup the array elements
and then arrive at an exact result ... but that seems like expensive
overkill, and in particular it'd be penalizing intelligently-written
queries (which wouldn't have dups in the array to start with) to benefit
badly-written ones.  So it seems like the right thing is for
scalararraysel to (1) check if the operator is equality or inequality,
and if so (2) just assume the array elements are all different and so
the probabilities sum directly.  If the operator is something else
it's probably best to stick with the existing logic.  We could probably
also protect ourselves a bit more by noting if the sum gives an
impossible result (probability > 1 or < 0) and falling back to the
normal calculation in that case.

            regards, tom lane