Обсуждение: No optimization with a partition window in a view
Hello
I tried to avoid implementing window functions inside my ORM by using a=
view,=20
but it seems the optimizer is missing an obvious optimization and thus =
doing a=20
full table scan.
Affected versions : 9.2.4 and 9.3.2 (9.4 not tested yet)
How to reproduce :
=3D> create table test_history (i serial, piece integer not null, date =
timestamp=20
with time zone default now(), location integer not null);
=3D> create table test_history (i serial, piece integer not null, date =
timestamp=20
with time zone default now(), location integer not null);
CREATE TABLE
=3D> insert into test_history (piece, location) select i, 1 from=20
generate_series(1, 1000000) i;
INSERT 0 1000000
=3D> insert into test_history (piece, location) select i, 2 from=20
generate_series(1, 1000000) i;
INSERT 0 1000000
=3D> insert into test_history (piece, location) select i, 3 from=20
generate_series(1, 1000000) i;
^[[AINSERT 0 1000000
=3D> insert into test_history (piece, location) select i, 4 from=20
generate_series(1, 1000000, 2) i;
INSERT 0 500000
=3D> alter table test_history add primary key(i);
ALTER TABLE
=3D> create index on test_history(piece);
CREATE INDEX
=3D> select * from test_history where piece =3D 42;
i | piece | date | location=20
=2D--------+-------+-------------------------------+----------
42 | 42 | 2014-02-15 08:52:50.946586+01 | 1
1000042 | 42 | 2014-02-15 08:52:56.634685+01 | 2
2000042 | 42 | 2014-02-15 08:53:00.762706+01 | 3
(3 rows)
Time: 0.158 ms
=3D> select *, lag(location, 1) over w, lead(location, 1) over w from t=
est_history=20
where piece =3D 42 window w as (partition by piece order by date);
i | piece | date | location | lag | lea=
d=20
=2D--------+-------+-------------------------------+----------+-----+----=
=2D-
42 | 42 | 2014-02-15 08:52:50.946586+01 | 1 | | =
2
1000042 | 42 | 2014-02-15 08:52:56.634685+01 | 2 | 1 | =
3
2000042 | 42 | 2014-02-15 08:53:00.762706+01 | 3 | 2 | =
=20
(3 rows)
Time: 0.203 ms
=3D> create view test_history_lag_lead as select *, lag(location, 1) ov=
er w,=20
lead(location, 1) over w from test_history window w as (partition by pi=
ece order=20
by date);
CREATE VIEW
=3D> select * from test_history_lag_lead where piece =3D 42;
i | piece | date | location | lag | lea=
d=20
=2D--------+-------+-------------------------------+----------+-----+----=
=2D-
42 | 42 | 2014-02-15 08:52:50.946586+01 | 1 | | =
2
1000042 | 42 | 2014-02-15 08:52:56.634685+01 | 2 | 1 | =
3
2000042 | 42 | 2014-02-15 08:53:00.762706+01 | 3 | 2 | =
=20
(3 rows)
Time: 2915.756 ms
=3D> explain analyze select * from test_history_lag_lead where piece =3D=
42;
QUERY P=
LAN =20
=2D----------------------------------------------------------------------=
=2D-------------------------------------------------------------------
Subquery Scan on test_history_lag_lead (cost=3D653058.16..775558.16 r=
ows=3D4=20
width=3D28) (actual time=3D1411.423..3375.794 rows=3D3 loops=3D1)
Filter: (test_history_lag_lead.piece =3D 42)
Rows Removed by Filter: 3499997
-> WindowAgg (cost=3D653058.16..731808.16 rows=3D3500000 width=3D2=
0) (actual=20
time=3D1411.343..3206.959 rows=3D3500000 loops=3D1)
-> Sort (cost=3D653058.16..661808.16 rows=3D3500000 width=3D=
20) (actual=20
time=3D1411.337..1867.895 rows=3D3500000 loops=3D1)
Sort Key: test_history.piece, test_history.date
Sort Method: external merge Disk: 116328kB
-> Seq Scan on test_history (cost=3D0.00..57293.00 row=
s=3D3500000=20
width=3D20) (actual time=3D0.004..310.558 rows=3D3500000 loops=3D1)
Total runtime: 3386.455 ms
(9 rows)
=3D> explain analyze select *, lag(location, 1) over w, lead(location, =
1) over w=20
from=20test_history where piece =3D 42 window w as (partition by piece or=
der by=20
date);
QUE=
RY PLAN =
=20
=2D----------------------------------------------------------------------=
=2D----------------------------------------------------------------------=
=2D----
WindowAgg (cost=3D19.70..19.78 rows=3D4 width=3D20) (actual time=3D0.=
018..0.019 rows=3D3=20
loops=3D1)
-> Sort (cost=3D19.70..19.71 rows=3D4 width=3D20) (actual time=3D0=
.015..0.015=20
rows=3D3 loops=3D1)
Sort Key: date
Sort Method: quicksort Memory: 25kB
-> Index Scan using test_history_piece_idx on test_history =20=
(cost=3D0.43..19.66 rows=3D4 width=3D20) (actual time=3D0.009..0.010 ro=
ws=3D3 loops=3D1)
Index Cond: (piece =3D 42)
Total runtime: 0.037 ms
(7 rows)
As you can see, the optimizer decided to do the WindowAgg first on the =
3,5M=20
lines before filtering, instead of filtering using an Index scan and th=
en doing=20
the aggregation.
IMHO, that optimization would be a special case possible only with wind=
ows on a=20
partition that is contained in the filtering fields, but I aint no expe=
rt here=20
:)
Thanks
Pierre Ducroquet
Hello, this seems to be a matter of subquery pushdown, query
transform involving subqueries or how to deal with views in
planner, rather than a bug.
> I tried to avoid implementing window functions inside my ORM by using a view,
> but it seems the optimizer is missing an obvious optimization and thus doing a
> full table scan.
>
> Affected versions : 9.2.4 and 9.3.2 (9.4 not tested yet)
9.4dev is also "affected".
> How to reproduce :
<snip>
OK_A> => select * from test_history where piece = 42;
OK_B> => select *, lag(location, 1) over w, lead(location, 1) over w from test_history
where piece = 42 window w as (partition by piece order by date);
> => create view test_history_lag_lead as select *, lag(location, 1) over w,
> lead(location, 1) over w from test_history window w as (partition by piece order
> by date);
NG_C> => select * from test_history_lag_lead where piece = 42;
The equivalent for NG_C is not OK_B but this,
select * from
(select *, lag(location, 1) over w, lead(location,1) over w
from test_history window w as (partition by piece order
by date)) test_history_lag_lead
where piece = 42;
You will see this also falls down to SeqScan. Since views are
treated as monolithic substances. The planner simplly replaces a
view with a subquery and the subquery is planned individually -
separately from the upper part so that result comes. No amendment
is seen in my poor sight so far..
If your objective is simplification of queries or concealing
details and you are allowed to do these in different way,
functions could be usable.
CREATE OR REPLACE FUNCTION test_history_lag_lead(integer)
RETURNS TABLE(i integer, piece integer, date timestamp with time zone,
location integer, lag integer, lead integer) AS $$
SELECT *, lag(location, 1) OVER w, lead(location, 1) OVER w
FROM test_history WHERE piece = $1
WINDOW w AS (PARTITION BY piece ORDER BY date);
$$ LANGUAGE SQL;
select * from test_history_lag_lead(42);
...
Total runtime: 0.594 ms
(Seems slow but seqscan takes 13 sec on my rig..(altough without
compile optimizations))
regards,
--
Kyotaro Horiguchi
NTT Open Source Software Center
Kyotaro HORIGUCHI <horiguchi.kyotaro@lab.ntt.co.jp> writes:
> Hello, this seems to be a matter of subquery pushdown, query
> transform involving subqueries or how to deal with views in
> planner, rather than a bug.
In general, pushing a WHERE clause down through a window function call
*would* be a bug. I think it should be safe in this case because the
WHERE clause matches the window functions' partition clauses, so that
applying the WHERE removes either all or none of the rows of any
particular partition. We've not gotten around to that type of refinement
in window function planning, yet. It would take some infrastructure that
doesn't exist now --- I don't recall that we have any code that tries to
make that particular kind of proof.
regards, tom lane
Hello, At Mon, 17 Feb 2014 23:10:53 -0500, Tom Lane wrote > Kyotaro HORIGUCHI <horiguchi.kyotaro@lab.ntt.co.jp> writes: > > Hello, this seems to be a matter of subquery pushdown, query > > transform involving subqueries or how to deal with views in > > planner, rather than a bug. > > In general, pushing a WHERE clause down through a window function call > *would* be a bug. I presumed without thought the first two queries are equivalent, but http://www.postgresql.org/docs/9.2/static/tutorial-window.html | The rows considered by a window function are those of the | "virtual table" produced by the query's FROM clause as filtered | by its WHERE, GROUP BY, and HAVING clauses if any. For example, | a row removed because it does not meet the WHERE condition is | not seen by any window function. Pushing down WHERE clause relating to window functions is certainly a bug for window functions in most cases. > I think it should be safe in this case because the WHERE clause > matches the window functions' partition clauses, so that > applying the WHERE removes either all or none of the rows of > any particular partition. I understood that after examining data. How stupid! > We've not gotten around to that type of refinement in window > function planning, yet. It would take some infrastructure that > doesn't exist now --- I don't recall that we have any code that > tries to make that particular kind of proof. # Althgough I dont't see the landscape of what kind of these # optimizations could there be, Subquery comes to me to seem to be a firm obstacle for many things :-( But it makes everything simpler, too. The infrastructure is, in simple imagination(?), that for examining condition, pushing down, repeated exec time estimation and selection involving upper nodes of the subquery, but this is too naive design as it takes too long time to make exec plan as you mentioned before , besides of ugliness. (But multiple plannings for one subquery seems not needed for this case.) Larger scale transformations which can dissolve subqueries would be better. Feeling overwhelmed.. regards, -- Kyotaro Horiguchi NTT Open Source Software Center