Обсуждение: Incorrect behavior with CE and ORDER BY
Hello,
We have a problem with CE that I want to verify is either expected
behavior, a bug or something else :).
Yes constraint exclusion is on.
I have tried increasing the default_statistics_target (all the way 1000)
no change in behavior.
Query plan with ORDER BY:
Limit (cost=47110.19..47110.31 rows=50 width=8) (actual
time=6088.013..6088.269 rows=50 loops=1) -> Sort (cost=47110.19..47943.14 rows=333179 width=8) (actual
time=6088.007..6088.104 rows=50 loops=1) Sort Key: public.tbl_profile_search.pse_lastlogin -> Result
(cost=0.00..16547.78rows=333179 width=8) (actual
time=0.020..4339.472 rows=334319 loops=1) -> Append (cost=0.00..16547.78 rows=333179 width=8)
(actual time=0.016..3208.022 rows=334319 loops=1) -> Seq Scan on tbl_profile_search
(cost=0.00..2.27 rows=1 width=8) (actual time=0.012..0.047 rows=2 loops=1) Filter:
(((pse_normalized_text)::text=
'1'::text) AND (pse_interest_type = 10)) -> Index Scan using index_pse_09_on_part_1 on
tbl_profile_search_interest_1 tbl_profile_search (cost=0.00..4.73
rows=1 width=8) (actual time=0.202..0.202 rows=0 loops=1) Index Cond:
((pse_normalized_text)::text=
'1'::text) Filter: (pse_interest_type = 10) -> Bitmap Heap Scan on
tbl_profile_search_interest_10 tbl_profile_search
(cost=3579.12..16540.78 rows=333177 width=8) (actual
time=90.619..2116.224 rows=334317 loops=1) Recheck Cond: ((pse_normalized_text)::text =
'1'::text) Filter: (pse_interest_type = 10) -> Bitmap Index Scan on
index_pse_09_on_part_10 (cost=0.00..3579.12 rows=333177 width=0)
(actual time=89.052..89.052 rows=340964 loops=1) Index Cond:
((pse_normalized_text)::text = '1'::text)Total runtime: 6103.190 ms
Same query, just removed ORDER BY:
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------Limit
(cost=0.00..2.48 rows=50 width=4) (actual time=0.025..57.146
rows=50 loops=1) -> Result (cost=0.00..16549.78 rows=333179 width=4) (actual
time=0.021..56.993 rows=50 loops=1) -> Append (cost=0.00..16549.78 rows=333179 width=4) (actual
time=0.017..56.835 rows=50 loops=1) -> Seq Scan on tbl_profile_search (cost=0.00..2.27
rows=1 width=4) (actual time=0.013..0.050 rows=2 loops=1) Filter: (((pse_normalized_text)::text =
'1'::text)
AND (pse_interest_type = 10)) -> Index Scan using index_pse_09_on_part_1 on
tbl_profile_search_interest_1 tbl_profile_search (cost=0.00..4.73
rows=1 width=4) (actual time=0.051..0.051 rows=0 loops=1) Index Cond: ((pse_normalized_text)::text =
'1'::text) Filter: (pse_interest_type = 10) -> Bitmap Heap Scan on
tbl_profile_search_interest_10
tbl_profile_search (cost=3581.12..16542.78 rows=333177 width=4) (actual
time=56.481..56.573 rows=48 loops=1) Recheck Cond: ((pse_normalized_text)::text = '1'::text)
Filter: (pse_interest_type = 10) -> Bitmap Index Scan on index_pse_09_on_part_10
(cost=0.00..3581.12 rows=333177 width=0) (actual time=54.999..54.999
rows=341233 loops=1) Index Cond: ((pse_normalized_text)::text =
'1'::text)Total runtime: 57.396 ms
--
=== The PostgreSQL Company: Command Prompt, Inc. ===
Sales/Support: +1.503.667.4564 || 24x7/Emergency: +1.800.492.2240 Providing the most comprehensive PostgreSQL
solutionssince 1997 http://www.commandprompt.com/
Donate to the PostgreSQL Project: http://www.postgresql.org/about/donate
"Joshua D. Drake" <jd@commandprompt.com> writes:
> We have a problem with CE that I want to verify is either expected
> behavior, a bug or something else :).
Uh, what's your problem exactly? The example only seems to demonstrate
that if you don't ask for a sort, you don't get one.
regards, tom lane
Tom Lane wrote:
> "Joshua D. Drake" <jd@commandprompt.com> writes:
>> We have a problem with CE that I want to verify is either expected
>> behavior, a bug or something else :).
>
> Uh, what's your problem exactly? The example only seems to demonstrate
> that if you don't ask for a sort, you don't get one.
Sorry. The problem is, if I ask for an ORDER BY it scans all partitions
versus only scanning the partition that has the data in it.
Sincerely,
Joshua D. Drake
>
> regards, tom lane
>
--
=== The PostgreSQL Company: Command Prompt, Inc. ===
Sales/Support: +1.503.667.4564 || 24x7/Emergency: +1.800.492.2240
Providing the most comprehensive PostgreSQL solutions since 1997 http://www.commandprompt.com/
Donate to the PostgreSQL Project: http://www.postgresql.org/about/donate
Joshua D. Drake wrote: > Tom Lane wrote: > > "Joshua D. Drake" <jd@commandprompt.com> writes: > >> We have a problem with CE that I want to verify is either expected > >> behavior, a bug or something else :). > > > > Uh, what's your problem exactly? The example only seems to demonstrate > > that if you don't ask for a sort, you don't get one. > > Sorry. The problem is, if I ask for an ORDER BY it scans all partitions > versus only scanning the partition that has the data in it. Huh, but that's not what the EXPLAIN ANALYZE you posted says ... -- Alvaro Herrera http://www.CommandPrompt.com/ PostgreSQL Replication, Consulting, Custom Development, 24x7 support
Alvaro Herrera wrote:
> Joshua D. Drake wrote:
>> Tom Lane wrote:
>>> "Joshua D. Drake" <jd@commandprompt.com> writes:
>>>> We have a problem with CE that I want to verify is either expected
>>>> behavior, a bug or something else :).
>>> Uh, what's your problem exactly? The example only seems to demonstrate
>>> that if you don't ask for a sort, you don't get one.
>> Sorry. The problem is, if I ask for an ORDER BY it scans all partitions
>> versus only scanning the partition that has the data in it.
>
> Huh, but that's not what the EXPLAIN ANALYZE you posted says ...
>
Sorry I realize the error of my ways. It isn't that it is scanning all
partitions, it is that it is scanning all of a single partition (subject
to the WHERE clause). That is correct behavior.
Sincerely,
Joshua D. Drake
--
=== The PostgreSQL Company: Command Prompt, Inc. ===
Sales/Support: +1.503.667.4564 || 24x7/Emergency: +1.800.492.2240
Providing the most comprehensive PostgreSQL solutions since 1997 http://www.commandprompt.com/
Donate to the PostgreSQL Project: http://www.postgresql.org/about/donate
I followed up with Joshua on Jabber. This is the query: SELECT pse_userid FROM tbl_profile_search WHERE pse_normalized_text='1' and pse_interest_type = 10 order by pse_lastlogin DESC limit 50 offset 0 I suggested adding an index on (pse_normalized_text, pse_lastlogin), on the assumption that the planner would get the sorted output from there and be able to push the LIMIT clause, just below the indexscan, thus saving the big heap scan (and a sort across a large result set). But it turns out the index is already there. So it seems to me to be a planner shortcoming. Is this correct? My idea of the plan would be (tabs=8 spaces) Limit (50) Sort (key: pse_lastlogin) Result Append SeqScan tbl_profile_search Indexscan tbl_profile_search_interest_1 Limit (50) IndexScan on the index mentioned above Is this possible? It would be very fast. Maybe it should be like this instead: Limit (50) Sort (key: pse_lastlogin) Result Append Limit (50) SeqScan tbl_profile_search Limit (50) Indexscan tbl_profile_search_interest_1 Limit (50) IndexScan on the index mentioned above This is the actual plan: > Limit (cost=47110.19..47110.31 rows=50 width=8) (actual > time=6088.013..6088.269 rows=50 loops=1) > -> Sort (cost=47110.19..47943.14 rows=333179 width=8) (actual > time=6088.007..6088.104 rows=50 loops=1) > Sort Key: public.tbl_profile_search.pse_lastlogin > -> Result (cost=0.00..16547.78 rows=333179 width=8) (actual > time=0.020..4339.472 rows=334319 loops=1) > -> Append (cost=0.00..16547.78 rows=333179 width=8) > (actual time=0.016..3208.022 rows=334319 loops=1) > -> Seq Scan on tbl_profile_search > (cost=0.00..2.27 rows=1 width=8) (actual time=0.012..0.047 rows=2 loops=1) > Filter: (((pse_normalized_text)::text = > '1'::text) AND (pse_interest_type = 10)) > -> Index Scan using index_pse_09_on_part_1 on > tbl_profile_search_interest_1 tbl_profile_search (cost=0.00..4.73 > rows=1 width=8) (actual time=0.202..0.202 rows=0 loops=1) > Index Cond: ((pse_normalized_text)::text = > '1'::text) > Filter: (pse_interest_type = 10) > -> Bitmap Heap Scan on > tbl_profile_search_interest_10 tbl_profile_search > (cost=3579.12..16540.78 rows=333177 width=8) (actual > time=90.619..2116.224 rows=334317 loops=1) > Recheck Cond: ((pse_normalized_text)::text = > '1'::text) > Filter: (pse_interest_type = 10) > -> Bitmap Index Scan on > index_pse_09_on_part_10 (cost=0.00..3579.12 rows=333177 width=0) > (actual time=89.052..89.052 rows=340964 loops=1) > Index Cond: > ((pse_normalized_text)::text = '1'::text) > Total runtime: 6103.190 ms -- Alvaro Herrera http://www.CommandPrompt.com/ The PostgreSQL Company - Command Prompt, Inc.
Alvaro Herrera <alvherre@commandprompt.com> writes:
> Is this possible? It would be very fast.
It's possible but not exactly simple. As an example, your proposed
plan:
> Limit (50)
> Sort (key: pse_lastlogin)
> Result
> Append
> Limit (50)
> SeqScan tbl_profile_search
> Limit (50)
> Indexscan tbl_profile_search_interest_1
> Limit (50)
> IndexScan on the index mentioned above
is wrong because there's no guarantee that the first 50 elements of a
seqscan will be anything special. You could imagine dealing with that
by sorting the seqscan results and limiting to 50, or by not
sorting/limiting that data at all but letting the upper sort see all the
seqscan entries. Offhand I think either of those could win depending on
how many elements the seqscan will yield. Also, it might be interesting
to consider inventing a "merge" plan node type that takes N
already-sorted inputs and produces a sorted output stream. Then we'd
need to trade off this approach versus doing the top-level sort, which
could cope with some of its inputs not being pre-sorted.
This seems to have some aspects in common with the recent discussion
about how to optimize min/max aggregates across an appendrel set.
regards, tom lane
Tom Lane ha scritto: > Alvaro Herrera <alvherre@commandprompt.com> writes: >> Limit (50) >> Sort (key: pse_lastlogin) >> Result >> Append >> Limit (50) >> SeqScan tbl_profile_search >> Limit (50) >> Indexscan tbl_profile_search_interest_1 >> Limit (50) >> IndexScan on the index mentioned above > > is wrong because there's no guarantee that the first 50 elements of a > seqscan will be anything special. You could imagine dealing with that > by sorting the seqscan results and limiting to 50, or by not > sorting/limiting that data at all but letting the upper sort see all the > seqscan entries. Offhand I think either of those could win depending on > how many elements the seqscan will yield. Also, it might be interesting > to consider inventing a "merge" plan node type that takes N > already-sorted inputs and produces a sorted output stream. Then we'd > need to trade off this approach versus doing the top-level sort, which > could cope with some of its inputs not being pre-sorted. > > This seems to have some aspects in common with the recent discussion > about how to optimize min/max aggregates across an appendrel set. The plan proposed by Alvaro reminds me of: http://archives.postgresql.org/pgsql-performance/2005-09/msg00047.php My proposal was in fact (Alvaro's plan + first Tom's suggested change): Limit (50) Sort (key: pse_lastlogin) Result Append Limit (50) Sort (key: pse_lastlogin) SeqScantbl_profile_search Limit (50) Indexscan tbl_profile_search_interest_1 Limit (50) IndexScan on the indexmentioned above The plan was generated rewriting the query to use explicit subselect and forcing the planner to order by and limit for each subquery. I've tried a few times to write a patch to handle it, but I wasn't able to do it because of my lack of internals knowledge and spare time. Best regards -- Matteo Beccati http://phpadsnew.com http://phppgads.com