Обсуждение: [PERFORM] Cheaper subquery scan not considered unless offset 0
Hello everyone,
Please consider the following three semantically equivalent, but differently written queries:
Query A:
SELECT * FROM items a INNER JOIN ( SELECT item, sum(amount) stock FROM stocktransactions GROUP BY item HAVING
sum(amount)>= 1
) b ON b.item = a. "ID"
Query B:
SELECT * FROM items a INNER JOIN ( SELECT item, sum(amount) stock FROM stocktransactions GROUP BY item
) b ON b.item = a. "ID" WHERE b.stock >= 1
Query C:
SELECT * FROM items a INNER JOIN ( SELECT item, sum(amount) stock FROM stocktransactions b GROUP BY item OFFSET 0
) b ON b.item = a. "ID" WHERE b.stock >= 1
FYI: stocktransactions.item and stocktransactions.amount have not null constraints and stocktransactions.item is a
foreignkey referencing items.ID, the primary key of items.
Queries A + B generate the same plan and execute as follows:
Merge Join (cost=34935.30..51701.59 rows=22285 width=344) (actual time=463.824..659.553 rows=15521 loops=1) Merge
Cond:(a."ID" = b.item) -> Index Scan using "PK_items_ID" on items a (cost=0.42..15592.23 rows=336083 width=332)
(actualtime=0.012..153.899 rows=336064 loops=1) -> Sort (cost=34934.87..34990.59 rows=22285 width=12) (actual
time=463.677..466.146rows=15521 loops=1) Sort Key: b.item Sort Method: quicksort Memory: 1112kB ->
FinalizeHashAggregate (cost=32879.78..33102.62 rows=22285 width=12) (actual time=450.724..458.667 rows=15521 loops=1)
Group Key: b.item Filter: (sum(b.amount) >= '1'::double precision) Rows Removed by
Filter:48277 -> Gather (cost=27865.65..32545.50 rows=44570 width=12) (actual time=343.715..407.243
rows=162152loops=1) Workers Planned: 2 Workers Launched: 2 ->
PartialHashAggregate (cost=26865.65..27088.50 rows=22285 width=12) (actual time=336.416..348.105 rows=54051 loops=3)
Group Key: b.item -> Parallel Seq Scan on stocktransactions b
(cost=0.00..23281.60rows=716810 width=12) (actual time=0.015..170.646 rows=579563 loops=3)
Planning time: 0.277 ms
Execution time: 661.342 ms
Plan C though, thanks to the "offset optimization fence", executes the following, more efficient plan:
Nested Loop (cost=32768.77..41146.56 rows=7428 width=344) (actual time=456.611..525.395 rows=15521 loops=1
total=525.395)-> Subquery Scan on c (cost=32768.35..33269.76 rows=7428 width=12) (actual time=456.591..475.204
rows=15521loops=1 total=475.204) Filter: (c.stock >= '1'::double precision) Rows Removed by Filter: 48277
-> Finalize HashAggregate (cost=32768.35..32991.20 rows=22285 width=12) (actual time=456.582..468.124 rows=63798
loops=1total=468.124) Group Key: b.item -> Gather (cost=27865.65..32545.50 rows=44570
width=12)(actual time=348.479..415.463 rows=162085 loops=1 total=415.463) Workers Planned: 2
Workers Launched: 2 -> Partial HashAggregate (cost=26865.65..27088.50 rows=22285 width=12)
(actualtime=343.952..355.912 rows=54028 loops=3 total=1067.736) Group Key: b.item
-> Parallel Seq Scan on stocktransactions b (cost=0.00..23281.60 rows=716810 width=12) (actual
time=0.015..172.235rows=579563 loops=3 total=516.705) -> Index Scan using "PK_items_ID" on items a (cost=0.42..1.05
rows=1width=332) (actual time=0.003..0.003 rows=1 loops=15521 total=46.563) Index Cond: ("ID" = c.item)
Planning time: 0.223 ms
Execution time: 526.203 ms
I'm wondering, given that Query C's plan has lower overall costs than Query A/B's, why wouldn't the planner choose to
executethat plan for queries A+B as well?
It has lower projected startup cost as well as lower total cost so apparently the optimzer does not consider such a
planwith a subquery scan at all (otherwise it would choose it based on the lower cost estimates, right?) unless one
forcesit to via OFFSET 0.
Though I wouldn't necessarily consider this a bug, it is an issue that one has to explicitly work around with
inadvisableoptimization fences and it would be great if this could be fixed.
Thanks to the developer community for delivering this great product, I hope this helps in enhancing it.
Cheers,
Benjamin
--
Bejamin Coutu
ben.coutu@zeyos.com
ZeyOS, Inc.
http://www.zeyos.com
--
Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-performance
On 30 October 2017 at 00:24, Benjamin Coutu <ben.coutu@zeyos.com> wrote: > -> Index Scan using "PK_items_ID" on items a (cost=0.42..1.05 rows=1 width=332) (actual time=0.003..0.003 rows=1 loops=15521total=46.563) I've never seen EXPLAIN output like that before. Is this some modified version of PostgreSQL? -- David Rowley http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services -- Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-performance
It's not a modified postgres version. It's simply for my convenience that my tooling calculats "total" as "actual time"
multipliedby "loops". Looks like I didn't properly strip that away when copy-pasting.
Here are the queries and original plans again, sorry for the confusion.
Query A:
SELECT * FROM items a INNER JOIN ( SELECT item, sum(amount) stock FROM stocktransactions b GROUP BY item HAVING
sum(amount)>= 1
) c ON c.item = a."ID"
Query B:
SELECT * FROM items a INNER JOIN ( SELECT item, sum(amount) stock FROM stocktransactions b GROUP BY item
) c ON c.item = a."ID" WHERE c.stock >= 1
Query C:
SELECT * FROM items a INNER JOIN ( SELECT item, sum(amount) stock FROM stocktransactions b GROUP BY item OFFSET 0
) c ON c.item = a."ID" WHERE c.stock >= 1
Queries A + B generate the same plan and execute as follows:
Merge Join (cost=34935.30..51701.59 rows=22285 width=344) (actual time=463.824..659.553 rows=15521 loops=1) Merge
Cond:(a."ID" = b.item) -> Index Scan using "PK_items_ID" on items a (cost=0.42..15592.23 rows=336083 width=332)
(actualtime=0.012..153.899 rows=336064 loops=1) -> Sort (cost=34934.87..34990.59 rows=22285 width=12) (actual
time=463.677..466.146rows=15521 loops=1) Sort Key: b.item Sort Method: quicksort Memory: 1112kB ->
FinalizeHashAggregate (cost=32879.78..33102.62 rows=22285 width=12) (actual time=450.724..458.667 rows=15521 loops=1)
Group Key: b.item Filter: (sum(b.amount) >= '1'::double precision) Rows Removed by
Filter:48277 -> Gather (cost=27865.65..32545.50 rows=44570 width=12) (actual time=343.715..407.243
rows=162152loops=1) Workers Planned: 2 Workers Launched: 2 ->
PartialHashAggregate (cost=26865.65..27088.50 rows=22285 width=12) (actual time=336.416..348.105 rows=54051 loops=3)
Group Key: b.item -> Parallel Seq Scan on stocktransactions b
(cost=0.00..23281.60rows=716810 width=12) (actual time=0.015..170.646 rows=579563 loops=3)
Planning time: 0.277 ms
Execution time: 661.342 ms
Plan C though, thanks to the "offset optimization fence", executes the following, more efficient plan:
Nested Loop (cost=32768.77..41146.56 rows=7428 width=344) (actual time=456.611..525.395 rows=15521 loops=1) ->
SubqueryScan on c (cost=32768.35..33269.76 rows=7428 width=12) (actual time=456.591..475.204 rows=15521 loops=1)
Filter:(c.stock >= '1'::double precision) Rows Removed by Filter: 48277 -> Finalize HashAggregate
(cost=32768.35..32991.20rows=22285 width=12) (actual time=456.582..468.124 rows=63798 loops=1) Group Key:
b.item -> Gather (cost=27865.65..32545.50 rows=44570 width=12) (actual time=348.479..415.463 rows=162085
loops=1) Workers Planned: 2 Workers Launched: 2 -> Partial
HashAggregate (cost=26865.65..27088.50 rows=22285 width=12) (actual time=343.952..355.912 rows=54028 loops=3)
Group Key: b.item -> Parallel Seq Scan on stocktransactions b
(cost=0.00..23281.60rows=716810 width=12) (actual time=0.015..172.235 rows=579563 loops=3) -> Index Scan using
"PK_items_ID"on items a (cost=0.42..1.05 rows=1 width=332) (actual time=0.003..0.003 rows=1 loops=15521) Index
Cond:("ID" = c.item)
Planning time: 0.223 ms
Execution time: 526.203 ms
========== Original ==========
From: David Rowley <david.rowley@2ndquadrant.com>
To: Benjamin Coutu <ben.coutu@zeyos.com>
Date: Sun, 29 Oct 2017 12:46:42 +0100
Subject: Re: [PERFORM] Cheaper subquery scan not considered unless offset 0
>
>
> On 30 October 2017 at 00:24, Benjamin Coutu <ben.coutu@zeyos.com> wrote:
> > -> Index Scan using "PK_items_ID" on items a (cost=0.42..1.05 rows=1 width=332) (actual time=0.003..0.003
rows=1loops=15521 total=46.563)
>
> I've never seen EXPLAIN output like that before.
>
> Is this some modified version of PostgreSQL?
>
--
Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-performance
There is actually another separate issue here apart from that the planner obviously choosing the wrong plan as
originallydescribed in my last message, a plan it knows to be more expensive based on cost estimates.
Take a look at the way the filter condition is treated differently when estimating the number of returned rows when
appliedin different nodes.
Queries A/B:
-> Finalize HashAggregate (cost=32879.78..33102.62 rows=22285 width=12) (actual time=450.724..458.667 rows=15521
loops=1) Group Key: b.item Filter: (sum(b.amount) >= '1'::double precision) Rows
Removedby Filter: 48277 -> Gather ...
Query C:
-> Subquery Scan on c (cost=32768.35..33269.76 rows=7428 width=12) (actual time=456.591..475.204 rows=15521 loops=1)
Filter: (c.stock >= '1'::double precision) Rows Removed by Filter: 48277 -> Finalize HashAggregate
(cost=32768.35..32991.20rows=22285 width=12) (actual time=456.582..468.124 rows=63798 loops=1) Group Key:
b.item -> Gather ...
Interestingly enough the subquery scan with query C correctly accounts for the filter when estimating rows=7428, while
A/Bdoesn't seem to account for the filter in the HasAggregate node (estimated rows=22285). This looks like a bug.
--
Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-performance
Benjamin Coutu <ben.coutu@zeyos.com> writes:
> Please consider the following three semantically equivalent, but differently written queries:
> ...
> Queries A + B generate the same plan and execute as follows:
> -> Finalize HashAggregate (cost=32879.78..33102.62 rows=22285 width=12) (actual time=450.724..458.667
rows=15521loops=1)
> Group Key: b.item
> Filter: (sum(b.amount) >= '1'::double precision)
> Rows Removed by Filter: 48277
> Plan C though, thanks to the "offset optimization fence", executes the following, more efficient plan:
> -> Subquery Scan on c (cost=32768.35..33269.76 rows=7428 width=12) (actual time=456.591..475.204 rows=15521
loops=1total=475.204)
> Filter: (c.stock >= '1'::double precision)
> Rows Removed by Filter: 48277
> -> Finalize HashAggregate (cost=32768.35..32991.20 rows=22285 width=12) (actual time=456.582..468.124
rows=63798loops=1 total=468.124)
> Group Key: b.item
Huh. So we can see that the grouping step produces 63798 rows in reality,
of which 15521 pass the >= filter condition. In Plan C, the planner
estimates the total number of group rows at 22285; then, having no
information about the statistics of c.stock, it uses DEFAULT_INEQ_SEL
(0.333) as the filter selectivity estimate, arriving at 7428 as the
estimated number of result rows for the subquery.
In Plan A+B, the planner presumably estimated the number of group rows at
22285 as well, but then it comes up with 22285 as the overall result.
Uh, what about the HAVING?
Evidently, the difference between 7428 and 22285 estimated rows out of
the subquery is enough to prompt a change in join plan for this query.
Since the true number is in between, it's just luck that Plan C is faster.
I don't put any great amount of stock in one join plan or the other
having been chosen for this case based on those estimates.
But ... what about the HAVING? I took a quick look around and couldn't
find anyplace where the selectivity of an aggregate's filter condition
gets accounted for, which explains this observed behavior. That seems
like a big oversight :-(
Now, it's true that we're basically never gonna be able to do better than
default selectivity estimates for post-aggregation filter conditions.
Maybe, at some point in the dim past, somebody intentionally decided that
applying the standard selectivity estimation logic to HAVING clauses was a
loser. But I don't see any comments to that effect, and anyway taking the
selectivity as 1.0 all the time doesn't seem very bright either.
Changing this in back branches might be too much of a behavioral change,
but it seems like we oughta change HEAD to apply standard selectivity
estimation to the HAVING clause.
regards, tom lane
--
Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-performance