My Books

My Slides

rss 아이콘 이미지

Search

'PWJ'에 해당되는 글 1건

  1. 2017.05.24 PostgreSQL 10 - Partition Wise Join (PWJ) 및 조인 Pruning 테스트 (1)

이전 포스트에서 언급한 것과 같이 PostgreSQL은 v9.6까지 파티션 조인 Pruning을 제공하지 않습니다.

 

그런데 EDB 사에서 작성한 문서 (여기 참조)를 보면, 10 버전부터는 선언적인 파티션 기능을 제공할 뿐 아니라 Partition Wise Join (이하 PWJ)도 제공한다고 설명하고 있습니다.

 

매우 반가운 소식이죠.

 

기쁜 마음에, PostgreSQL 베타 10 버전을 설치한 후에 PWJ와 조인 Pruning을 테스트했습니다.

 

음. 예상과 달리 PWJ도 지원되지 않고 조인 Pruning도 여전히 지원되지 않습니다. (정식 버전이 나오면 다시 테스트해봐야겠습니다)

 

테스트 결과를 살펴보겠습니다.

 

 

파티션용 테이블 환경 구성

 

파티션 생성 방법은 PARTITION BY 키워드를 이용합니다. 그리고 각 파티션은 PARTITION OF 키워드를 이용해서 생성합니다.

 

파티션 생성 시에 주의할 점은 "TO" 뒤에 지정된 값은 해당 값을 포함하지 않는다는 점입니다.

 

예를 들어, 2017년 1월 파티션은 "TO" 값을 ('2017-02-01')로 지정해야 합니다.

 

즉, ORACLE로 치면 "VALUES LESS THAN TO_DATE('2017-02-01', 'YYYY-MM-DD')"과 같은 의미입니다.

 

drop table p1 cascade;
drop table p2 cascade;

create table p1 (
    c1        integer,
    logdate   date,
    dummy     char(10)
) partition by range (logdate);

create table p2 (
    c1        integer,
    logdate   date,
	amount    integer,
    dummy     char(10)
) partition by range (logdate);


create table p1_y201701 partition of p1 for values from ('2017-01-01') to ('2017-02-01');
create table p1_y201702 partition of p1 for values from ('2017-02-01') to ('2017-03-01');
create table p1_y201703 partition of p1 for values from ('2017-03-01') to ('2017-04-01');
create table p1_y201704 partition of p1 for values from ('2017-04-01') to ('2017-05-01');
create table p1_y201705 partition of p1 for values from ('2017-05-01') to ('2017-06-01');
create table p1_y201706 partition of p1 for values from ('2017-06-01') to ('2017-07-01');
create table p1_y201707 partition of p1 for values from ('2017-07-01') to ('2017-08-01');
create table p1_y201708 partition of p1 for values from ('2017-08-01') to ('2017-09-01');
create table p1_y201709 partition of p1 for values from ('2017-09-01') to ('2017-10-01');
create table p1_y201710 partition of p1 for values from ('2017-10-01') to ('2017-11-01');
create table p1_y201711 partition of p1 for values from ('2017-11-01') to ('2017-12-01');
create table p1_y201712 partition of p1 for values from ('2017-12-01') to ('2018-01-01');

create table p2_y201701 partition of p2 for values from ('2017-01-01') to ('2017-02-01');
create table p2_y201702 partition of p2 for values from ('2017-02-01') to ('2017-03-01');
create table p2_y201703 partition of p2 for values from ('2017-03-01') to ('2017-04-01');
create table p2_y201704 partition of p2 for values from ('2017-04-01') to ('2017-05-01');
create table p2_y201705 partition of p2 for values from ('2017-05-01') to ('2017-06-01');
create table p2_y201706 partition of p2 for values from ('2017-06-01') to ('2017-07-01');
create table p2_y201707 partition of p2 for values from ('2017-07-01') to ('2017-08-01');
create table p2_y201708 partition of p2 for values from ('2017-08-01') to ('2017-09-01');
create table p2_y201709 partition of p2 for values from ('2017-09-01') to ('2017-10-01');
create table p2_y201710 partition of p2 for values from ('2017-10-01') to ('2017-11-01');
create table p2_y201711 partition of p2 for values from ('2017-11-01') to ('2017-12-01');
create table p2_y201712 partition of p2 for values from ('2017-12-01') to ('2018-01-01');

-- p1 테이블에 일자별로 만건 입력 

do $$
begin
  for i in 1..365 loop 
     for j in 1..10000 loop 
	    insert into p1 values((i-1)*10000+j, to_date('20170101','YYYYMMDD')+i-1, 'dummy');
	 end loop;
  end loop;
end $$;
	
-- p2 테이블에 일자별로 만건 입력

do $$
begin
  for i in 1..365 loop 
     for j in 1..10000 loop 
	    insert into p2 values((i-1)*10000+j, to_date('20170101','YYYYMMDD')+i-1, j, 'dummy');
	 end loop;
  end loop;
end $$;

 

Non-파티션 테이블 환경 구성

 

파티션 테이블의 테스트 결과와 비교하기 위해 Non-파티션 테이블을 생성합니다.

 

drop table np1;
drop table np2;

create table np1 (
    c1        integer,
    logdate   date,
    dummy     char(10)
);

create table np2 (
    c1        integer,
    logdate   date,
	amount    integer,
    dummy     char(10)
);

-- np1 테이블에 일자별로 만건 입력 

do $$	
begin
  for i in 1..365 loop 
     for j in 1..10000 loop 
	    insert into np1 values((i-1)*10000+j, to_date('20170101','YYYYMMDD')+i-1, 'dummy');
	 end loop;
  end loop;
end $$;

-- np2 테이블에 일자별로 만건 입력

do $$
begin
  for i in 1..365 loop 
     for j in 1..10000 loop 
	    insert into np2 values((i-1)*10000+j, to_date('20170101','YYYYMMDD')+i-1, j, 'dummy');
	 end loop;
  end loop;
end $$;

 

통계 정보 생성 및 크기 확인

 

월별 파티션 테이블 크기 및 건수를 확인합니다.

 

analyze np1;
analyze np2;

analyze p1;
analyze p2;

analyze p1_y201701; 
analyze p1_y201702;
analyze p1_y201703;
analyze p1_y201704;
analyze p1_y201705;
analyze p1_y201706;
analyze p1_y201707;
analyze p1_y201708;
analyze p1_y201709;
analyze p1_y201710;
analyze p1_y201711;
analyze p1_y201712;

analyze p2_y201701; 
analyze p2_y201702;
analyze p2_y201703;
analyze p2_y201704;
analyze p2_y201705;
analyze p2_y201706;
analyze p2_y201707;
analyze p2_y201708;
analyze p2_y201709;
analyze p2_y201710;
analyze p2_y201711;
analyze p2_y201712;

select relname, relpages, reltuples::integer 
from   pg_class
where  relname like 'p1%' or
       relname like 'p2%' or 
	   relname like 'np%'
order by 1;
  relname   | relpages | reltuples
------------+----------+-----------
 np1        |    23249 |   3650000
 np2        |    23249 |   3650000
 p1         |        0 |         0
 p1_y201701 |     1975 |    310000
 p1_y201702 |     1784 |    280000
 p1_y201703 |     1975 |    310000
 p1_y201704 |     1911 |    300000
 p1_y201705 |     1975 |    310000
 p1_y201706 |     1911 |    300000
 p1_y201707 |     1975 |    310000
 p1_y201708 |     1975 |    310000
 p1_y201709 |     1911 |    300000
 p1_y201710 |     1975 |    310000
 p1_y201711 |     1911 |    300000
 p1_y201712 |     1975 |    310000
 p2         |        0 |         0
 p2_y201701 |     1975 |    310000
 p2_y201702 |     1784 |    280000
 p2_y201703 |     1975 |    310000
 p2_y201704 |     1911 |    300000
 p2_y201705 |     1975 |    310000
 p2_y201706 |     1911 |    300000
 p2_y201707 |     1975 |    310000
 p2_y201708 |     1975 |    310000
 p2_y201709 |     1911 |    300000
 p2_y201710 |     1975 |    310000
 p2_y201711 |     1911 |    300000
 p2_y201712 |     1975 |    310000

 

PQ Disable

 

1차 테스트 시에는 PQ 수행을 방지하기 위해서 max_parallel_workers_per_gather 파라미터를 0으로 변경합니다. (기본설정값: 2)

 

 

Non-파티션 테이블간에 전체 레코드 조인

 

Non-파티션 테이블 간의 조인 수행 결과, WORK_MEM을 대략 171MB 사용합니다.

 

set work_mem='200MB';
 
explain (analyze, buffers) 
select count(*)
from   np1 a, np2 b
where  a.logdate = b.logdate 
and    a.c1      = b.c1;
                                                            QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------------------
 Aggregate  (cost=229123.24..229123.25 rows=1 width=8) (actual time=5541.242..5541.242 rows=1 loops=1)
   Buffers: shared hit=46498
   ->  Hash Join  (cost=114499.00..229098.19 rows=10019 width=0) (actual time=1976.580..5091.775 rows=3650000 loops=1)
         Hash Cond: ((a.logdate = b.logdate) AND (a.c1 = b.c1))
         Buffers: shared hit=46498
         ->  Seq Scan on np1 a  (cost=0.00..59749.00 rows=3650000 width=8) (actual time=0.009..508.609 rows=3650000 loops=1)
               Buffers: shared hit=23249
         ->  Hash  (cost=59749.00..59749.00 rows=3650000 width=8) (actual time=1965.800..1965.800 rows=3650000 loops=1)
               Buckets: 4194304  Batches: 1  Memory Usage: 175347kB
               Buffers: shared hit=23249
               ->  Seq Scan on np2 b  (cost=0.00..59749.00 rows=3650000 width=8) (actual time=0.024..657.624 rows=3650000 loops=1)
                     Buffers: shared hit=23249
 Planning time: 0.090 ms
 Execution time: 5551.308 ms

 

 

파티션 테이블간에 전체 레코드 조인

 

그럼 파티션 테이블 간에 조인을 수행해볼까요? 만일 PWJ로 수행된다면 수행 속도도 더 빠르고 WORK_MEM도 더 적게 쓸 것이 분명합니다.
PJW는 파티션 단위로 조인을 수행하므로, 더 적은 WORK_MEM을 가지고 더 빠르게 수행된다는 장점이 있으니까요.

그런데 아래의 수행 결과를 보면 수행 속도도 더 느리고 WORK_MEM도 기존과 동일한 171MB를 사용합니다.

 

set work_mem='200MB';

explain (analyze, buffers) 
select count(*)
from   p1 a, p2 b
where  a.logdate = b.logdate 
and    a.c1      = b.c1;
                                                                  QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------------------------------
 Aggregate  (cost=229131.00..229131.01 rows=1 width=8) (actual time=7544.717..7544.717 rows=1 loops=1)
   Buffers: shared hit=46506
   ->  Hash Join  (cost=114503.00..229106.00 rows=10000 width=0) (actual time=2846.786..7058.230 rows=3650000 loops=1)
         Hash Cond: ((b.logdate = a.logdate) AND (b.c1 = a.c1))
         Buffers: shared hit=46506
         ->  Append  (cost=0.00..59753.00 rows=3650000 width=8) (actual time=0.008..1348.818 rows=3650000 loops=1)
               Buffers: shared hit=23253
               ->  Seq Scan on p2_y201701 b  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.007..54.608 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201702 b_1  (cost=0.00..4584.00 rows=280000 width=8) (actual time=0.044..52.027 rows=280000 loops=1)
                     Buffers: shared hit=1784
               ->  Seq Scan on p2_y201703 b_2  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.015..54.587 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201704 b_3  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.011..53.759 rows=300000 loops=1)
                     Buffers: shared hit=1911
               ->  Seq Scan on p2_y201705 b_4  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.018..60.088 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201706 b_5  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.012..53.712 rows=300000 loops=1)
                     Buffers: shared hit=1911
               ->  Seq Scan on p2_y201707 b_6  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.028..54.309 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201708 b_7  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.011..66.090 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201709 b_8  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.019..53.649 rows=300000 loops=1)
                     Buffers: shared hit=1911
               ->  Seq Scan on p2_y201710 b_9  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.011..57.762 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201711 b_10  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.014..51.902 rows=300000 loops=1)
                     Buffers: shared hit=1911
               ->  Seq Scan on p2_y201712 b_11  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.012..54.026 rows=310000 loops=1)
                     Buffers: shared hit=1975
         ->  Hash  (cost=59753.00..59753.00 rows=3650000 width=8) (actual time=2834.113..2834.113 rows=3650000 loops=1)
               Buckets: 4194304  Batches: 1  Memory Usage: 175347kB
               Buffers: shared hit=23253
               ->  Append  (cost=0.00..59753.00 rows=3650000 width=8) (actual time=0.038..1359.414 rows=3650000 loops=1)
                     Buffers: shared hit=23253
                     ->  Seq Scan on p1_y201701 a  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.021..54.720 rows=310000 loops=1)
                           Buffers: shared hit=1975
                     ->  Seq Scan on p1_y201702 a_1  (cost=0.00..4584.00 rows=280000 width=8) (actual time=0.014..52.013 rows=280000 loops=1)
                           Buffers: shared hit=1784
                     ->  Seq Scan on p1_y201703 a_2  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.010..58.126 rows=310000 loops=1)
                           Buffers: shared hit=1975
                     ->  Seq Scan on p1_y201704 a_3  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.011..54.138 rows=300000 loops=1)
                           Buffers: shared hit=1911
                     ->  Seq Scan on p1_y201705 a_4  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.013..57.886 rows=310000 loops=1)
                           Buffers: shared hit=1975
                     ->  Seq Scan on p1_y201706 a_5  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.010..54.615 rows=300000 loops=1)
                           Buffers: shared hit=1911
                     ->  Seq Scan on p1_y201707 a_6  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.010..56.619 rows=310000 loops=1)
                           Buffers: shared hit=1975
                     ->  Seq Scan on p1_y201708 a_7  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.010..54.834 rows=310000 loops=1)
                           Buffers: shared hit=1975
                     ->  Seq Scan on p1_y201709 a_8  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.009..54.792 rows=300000 loops=1)
                           Buffers: shared hit=1911
                     ->  Seq Scan on p1_y201710 a_9  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.009..57.051 rows=310000 loops=1)
                           Buffers: shared hit=1975
                     ->  Seq Scan on p1_y201711 a_10  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.009..64.463 rows=300000 loops=1)
                           Buffers: shared hit=1911
                     ->  Seq Scan on p1_y201712 a_11  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.009..54.757 rows=310000 loops=1)
                           Buffers: shared hit=1975
 Planning time: 0.434 ms
 Execution time: 7550.042 ms

 

PQ로 수행 - 파티션 테이블간에 전체 레코드 조인

 

설마 PWJ가 PQ인 경우에만 동작하는 걸까요?

 

max_parallel_workers_per_gather 파라미터를 24로 설정한 후에 다시 테스트를 수행했습니다.

 

PQ를 사용해도 PWJ는 동작하지 않습니다.

 

속도도 아주 조금 빨라졌을 뿐이고 WORK_MEM도 기존과 동일하게 171MB를 시용합니다.

 

제가 기대했던 것은 파티션마다 1개의 Parallel Worker가 병렬로 (즉, 12개가 동시에) 해시 조인을 수행하는 것이었으나, 아래의 Explain 결과를 보면 Parallel Worker는 1개만 론칭됩니다.

 

Parallel Worker가 1개만 론칭되는 이유는 파티션 1개의 크기가 15MB이기 때문입니다.

 

즉, Parallel Worker의 개수를 정하는 기준은 쿼리에 사용되는 파티션 개수가 아닌, 개별 파티션의 크기임을 알 수 있습니다. (이점은 개선이 필요할것 같습니다)

 

참고로 Parallel 처리 수행 여부를 결정하는 기준은 테이블 크기이고, 9.6 버전까지는 min_parallel_relation_size 파라미터 (기본설정값: 8MB)였으나, 10 버전부터는 min_parallel_table_scan_size 파라미터 (기본설정값: 1024 * 8Kb)를 사용합니다. 그리고 파라미터값의 3배수마다 Parallel Worker를 1개씩 더 할당하고, 최대 할당 개수는 7개입니다.

 

set work_mem='200MB';

explain (analyze, buffers) 
select count(*)
from   p1 a, p2 b
where  a.logdate = b.logdate 
and    a.c1      = b.c1;
                                                                          QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------------------------
 Finalize Aggregate  (cost=192506.14..192506.15 rows=1 width=8) (actual time=5991.005..5991.006 rows=1 loops=1)
   Buffers: shared hit=70074
   ->  Gather  (cost=192506.03..192506.14 rows=1 width=8) (actual time=5965.123..5990.999 rows=2 loops=1)
         Workers Planned: 1
         Workers Launched: 1
         Buffers: shared hit=70074
         ->  Partial Aggregate  (cost=191506.03..191506.04 rows=1 width=8) (actual time=5962.763..5962.763 rows=1 loops=2)
               Buffers: shared hit=69844
               ->  Hash Join  (cost=114503.00..191491.32 rows=5882 width=0) (actual time=3328.256..5687.283 rows=1825000 loops=2)
                     Hash Cond: ((b.logdate = a.logdate) AND (b.c1 = a.c1))
                     Buffers: shared hit=69844
                     ->  Append  (cost=0.00..44723.59 rows=2147061 width=8) (actual time=0.014..785.853 rows=1825000 loops=2)
                           Buffers: shared hit=23253
                           ->  Parallel Seq Scan on p2_y201701 b  (cost=0.00..3798.53 rows=182353 width=8) (actual time=0.012..35.618 rows=155000 loops=2)
                                 Buffers: shared hit=1975
                           ->  Parallel Seq Scan on p2_y201702 b_1  (cost=0.00..3431.06 rows=164706 width=8) (actual time=0.012..30.147 rows=140000 loops=2)
                                 Buffers: shared hit=1784
                           ->  Parallel Seq Scan on p2_y201703 b_2  (cost=0.00..3798.53 rows=182353 width=8) (actual time=0.009..34.888 rows=155000 loops=2)
                                 Buffers: shared hit=1975
                           ->  Parallel Seq Scan on p2_y201704 b_3  (cost=0.00..3675.71 rows=176471 width=8) (actual time=0.006..33.808 rows=150000 loops=2)
                                 Buffers: shared hit=1911
                           ->  Parallel Seq Scan on p2_y201705 b_4  (cost=0.00..3798.53 rows=182353 width=8) (actual time=0.009..38.101 rows=155000 loops=2)
                                 Buffers: shared hit=1975
                           ->  Parallel Seq Scan on p2_y201706 b_5  (cost=0.00..3675.71 rows=176471 width=8) (actual time=0.008..33.637 rows=150000 loops=2)
                                 Buffers: shared hit=1911
                           ->  Parallel Seq Scan on p2_y201707 b_6  (cost=0.00..3798.53 rows=182353 width=8) (actual time=0.009..35.359 rows=155000 loops=2)
                                 Buffers: shared hit=1975
                           ->  Parallel Seq Scan on p2_y201708 b_7  (cost=0.00..3798.53 rows=182353 width=8) (actual time=0.008..34.228 rows=155000 loops=2)
                                 Buffers: shared hit=1975
                           ->  Parallel Seq Scan on p2_y201709 b_8  (cost=0.00..3675.71 rows=176471 width=8) (actual time=0.009..33.737 rows=150000 loops=2)
                                 Buffers: shared hit=1911
                           ->  Parallel Seq Scan on p2_y201710 b_9  (cost=0.00..3798.53 rows=182353 width=8) (actual time=0.008..34.722 rows=155000 loops=2)
                                 Buffers: shared hit=1975
                           ->  Parallel Seq Scan on p2_y201711 b_10  (cost=0.00..3675.71 rows=176471 width=8) (actual time=0.007..34.295 rows=150000 loops=2)
                                 Buffers: shared hit=1911
                           ->  Parallel Seq Scan on p2_y201712 b_11  (cost=0.00..3798.53 rows=182353 width=8) (actual time=0.009..33.017 rows=155000 loops=2)
                                 Buffers: shared hit=1975
                     ->  Hash  (cost=59753.00..59753.00 rows=3650000 width=8) (actual time=3311.668..3311.668 rows=3650000 loops=2)
                           Buckets: 4194304  Batches: 1  Memory Usage: 175347kB
                           Buffers: shared hit=46506
                           ->  Append  (cost=0.00..59753.00 rows=3650000 width=8) (actual time=0.037..1636.855 rows=3650000 loops=2)
                                 Buffers: shared hit=46506
                                 ->  Seq Scan on p1_y201701 a  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.036..71.783 rows=310000 loops=2)
                                       Buffers: shared hit=3950
                                 ->  Seq Scan on p1_y201702 a_1  (cost=0.00..4584.00 rows=280000 width=8) (actual time=0.032..63.401 rows=280000 loops=2)
                                       Buffers: shared hit=3568
                                 ->  Seq Scan on p1_y201703 a_2  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.019..71.326 rows=310000 loops=2)
                                       Buffers: shared hit=3950
                                 ->  Seq Scan on p1_y201704 a_3  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.020..67.097 rows=300000 loops=2)
                                       Buffers: shared hit=3822
                                 ->  Seq Scan on p1_y201705 a_4  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.029..71.276 rows=310000 loops=2)
                                       Buffers: shared hit=3950
                                 ->  Seq Scan on p1_y201706 a_5  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.021..66.124 rows=300000 loops=2)
                                       Buffers: shared hit=3822
                                 ->  Seq Scan on p1_y201707 a_6  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.038..66.720 rows=310000 loops=2)
                                       Buffers: shared hit=3950
                                 ->  Seq Scan on p1_y201708 a_7  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.022..75.394 rows=310000 loops=2)
                                       Buffers: shared hit=3950
                                 ->  Seq Scan on p1_y201709 a_8  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.021..66.812 rows=300000 loops=2)
                                       Buffers: shared hit=3822
                                 ->  Seq Scan on p1_y201710 a_9  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.022..70.167 rows=310000 loops=2)
                                       Buffers: shared hit=3950
                                 ->  Seq Scan on p1_y201711 a_10  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.022..94.346 rows=300000 loops=2)
                                       Buffers: shared hit=3822
                                 ->  Seq Scan on p1_y201712 a_11  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.029..78.436 rows=310000 loops=2)
                                       Buffers: shared hit=3950
 Planning time: 0.313 ms
 Execution time: 6027.321 ms

 

조인 Pruning 테스트

 

아래의 수행 결과를 보면, 조인 Pruning 또한 동작하지 않는 것을 알 수 있습니다.

 

explain (analyze, buffers)
select count(*)
from   p1 a, p2 b
where  a.logdate = b.logdate 
and    a.c1      = b.c1
and    a.logdate between DATE '2017-05-01'
                     and DATE '2017-05-31';
                                                                 QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------
 Aggregate  (cost=129466.00..129466.01 rows=1 width=8) (actual time=3613.260..3613.260 rows=1 loops=1)
   Buffers: shared hit=25228, temp read=10163 written=10157
   ->  Hash Join  (cost=12486.00..129441.00 rows=10000 width=0) (actual time=1066.850..3569.146 rows=310000 loops=1)
         Hash Cond: ((b.logdate = a.logdate) AND (b.c1 = a.c1))
         Buffers: shared hit=25228, temp read=10163 written=10157
         ->  Append  (cost=0.00..59753.00 rows=3650000 width=8) (actual time=0.017..1411.893 rows=3650000 loops=1)
               Buffers: shared hit=23253
               ->  Seq Scan on p2_y201701 b  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.016..61.122 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201702 b_1  (cost=0.00..4584.00 rows=280000 width=8) (actual time=0.010..55.294 rows=280000 loops=1)
                     Buffers: shared hit=1784
               ->  Seq Scan on p2_y201703 b_2  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.008..57.616 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201704 b_3  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.007..58.504 rows=300000 loops=1)
                     Buffers: shared hit=1911
               ->  Seq Scan on p2_y201705 b_4  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.010..61.685 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201706 b_5  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.014..64.099 rows=300000 loops=1)
                     Buffers: shared hit=1911
               ->  Seq Scan on p2_y201707 b_6  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.007..62.020 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201708 b_7  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.007..59.570 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201709 b_8  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.013..60.802 rows=300000 loops=1)
                     Buffers: shared hit=1911
               ->  Seq Scan on p2_y201710 b_9  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.011..69.859 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201711 b_10  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.007..57.168 rows=300000 loops=1)
                     Buffers: shared hit=1911
               ->  Seq Scan on p2_y201712 b_11  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.010..60.932 rows=310000 loops=1)
                     Buffers: shared hit=1975
         ->  Hash  (cost=6625.00..6625.00 rows=310000 width=8) (actual time=224.806..224.806 rows=310000 loops=1)
               Buckets: 131072  Batches: 4  Memory Usage: 4055kB
               Buffers: shared hit=1975, temp written=793
               ->  Append  (cost=0.00..6625.00 rows=310000 width=8) (actual time=0.015..130.284 rows=310000 loops=1)
                     Buffers: shared hit=1975
                     ->  Seq Scan on p1_y201705 a  (cost=0.00..6625.00 rows=310000 width=8) (actual time=0.014..72.681 rows=310000 loops=1)
                           Filter: ((logdate >= '2017-05-01'::date) AND (logdate <= '2017-05-31'::date))
                           Buffers: shared hit=1975
 Planning time: 0.787 ms
 Execution time: 3613.353 ms

 

Partition WISE Join 테스트 - 파티션 1개

 

PWJ가 지원되면, 아래와 같은 쿼리의 성능도 매우 빨라집니다.

 

즉, P1 테이블은 반드시 전체 스캔을 해야하나, P2 테이블은 a.C1 조건에 해당되는 P1_Y201701 파티션과 매칭되는 P2_Y201701 파티션만 액세스하면 되기 때문인데요.

 

아래 결과를 보면, P2 테이블의 모든 파티션을 액세스합니다. 즉, PWJ가 지원되지 않는다는 것을 알 수 있습니다.

 

explain (analyze, buffers)
select count(*)
from   p1 a, p2 b
where  a.logdate = b.logdate 
and    a.c1      = b.c1
and    a.c1 between 1 and 300000;
                                                                 QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------
 Aggregate  (cost=200620.27..200620.28 rows=1 width=8) (actual time=3494.830..3494.830 rows=1 loops=1)
   Buffers: shared hit=46506, temp read=10139 written=10133
   ->  Hash Join  (cost=83678.26..200595.27 rows=10000 width=0) (actual time=466.617..3458.329 rows=300000 loops=1)
         Hash Cond: ((b.logdate = a.logdate) AND (b.c1 = a.c1))
         Buffers: shared hit=46506, temp read=10139 written=10133
         ->  Append  (cost=0.00..59753.00 rows=3650000 width=8) (actual time=0.007..1321.298 rows=3650000 loops=1)
               Buffers: shared hit=23253
               ->  Seq Scan on p2_y201701 b  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.006..53.575 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201702 b_1  (cost=0.00..4584.00 rows=280000 width=8) (actual time=0.008..57.478 rows=280000 loops=1)
                     Buffers: shared hit=1784
               ->  Seq Scan on p2_y201703 b_2  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.007..56.152 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201704 b_3  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.006..52.546 rows=300000 loops=1)
                     Buffers: shared hit=1911
               ->  Seq Scan on p2_y201705 b_4  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.006..57.502 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201706 b_5  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.007..53.565 rows=300000 loops=1)
                     Buffers: shared hit=1911
               ->  Seq Scan on p2_y201707 b_6  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.007..55.867 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201708 b_7  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.006..56.451 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201709 b_8  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.012..55.934 rows=300000 loops=1)
                     Buffers: shared hit=1911
               ->  Seq Scan on p2_y201710 b_9  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.007..59.008 rows=310000 loops=1)
                     Buffers: shared hit=1975
               ->  Seq Scan on p2_y201711 b_10  (cost=0.00..4911.00 rows=300000 width=8) (actual time=0.010..59.359 rows=300000 loops=1)
                     Buffers: shared hit=1911
               ->  Seq Scan on p2_y201712 b_11  (cost=0.00..5075.00 rows=310000 width=8) (actual time=0.051..64.638 rows=310000 loops=1)
                     Buffers: shared hit=1975
         ->  Hash  (cost=78003.00..78003.00 rows=300151 width=8) (actual time=466.224..466.224 rows=300000 loops=1)
               Buckets: 131072  Batches: 4  Memory Usage: 3947kB
               Buffers: shared hit=23253, temp written=769
               ->  Append  (cost=0.00..78003.00 rows=300151 width=8) (actual time=0.007..383.795 rows=300000 loops=1)
                     Buffers: shared hit=23253
                     ->  Seq Scan on p1_y201701 a  (cost=0.00..6625.00 rows=300140 width=8) (actual time=0.007..65.470 rows=300000 loops=1)
                           Filter: ((c1 >= 1) AND (c1 <= 300000))
                           Rows Removed by Filter: 10000
                           Buffers: shared hit=1975
                     ->  Seq Scan on p1_y201702 a_1  (cost=0.00..5984.00 rows=1 width=8) (actual time=22.058..22.058 rows=0 loops=1)
                           Filter: ((c1 >= 1) AND (c1 <= 300000))
                           Rows Removed by Filter: 280000
                           Buffers: shared hit=1784
                     ->  Seq Scan on p1_y201703 a_2  (cost=0.00..6625.00 rows=1 width=8) (actual time=21.949..21.949 rows=0 loops=1)
                           Filter: ((c1 >= 1) AND (c1 <= 300000))
                           Rows Removed by Filter: 310000
                           Buffers: shared hit=1975
                     ->  Seq Scan on p1_y201704 a_3  (cost=0.00..6411.00 rows=1 width=8) (actual time=22.552..22.552 rows=0 loops=1)
                           Filter: ((c1 >= 1) AND (c1 <= 300000))
                           Rows Removed by Filter: 300000
                           Buffers: shared hit=1911
                     ->  Seq Scan on p1_y201705 a_4  (cost=0.00..6625.00 rows=1 width=8) (actual time=24.022..24.022 rows=0 loops=1)
                           Filter: ((c1 >= 1) AND (c1 <= 300000))
                           Rows Removed by Filter: 310000
                           Buffers: shared hit=1975
                     ->  Seq Scan on p1_y201706 a_5  (cost=0.00..6411.00 rows=1 width=8) (actual time=28.022..28.022 rows=0 loops=1)
                           Filter: ((c1 >= 1) AND (c1 <= 300000))
                           Rows Removed by Filter: 300000
                           Buffers: shared hit=1911
                     ->  Seq Scan on p1_y201707 a_6  (cost=0.00..6625.00 rows=1 width=8) (actual time=23.800..23.800 rows=0 loops=1)
                           Filter: ((c1 >= 1) AND (c1 <= 300000))
                           Rows Removed by Filter: 310000
                           Buffers: shared hit=1975
                     ->  Seq Scan on p1_y201708 a_7  (cost=0.00..6625.00 rows=1 width=8) (actual time=28.858..28.858 rows=0 loops=1)
                           Filter: ((c1 >= 1) AND (c1 <= 300000))
                           Rows Removed by Filter: 310000
                           Buffers: shared hit=1975
                     ->  Seq Scan on p1_y201709 a_8  (cost=0.00..6411.00 rows=1 width=8) (actual time=26.004..26.004 rows=0 loops=1)
                           Filter: ((c1 >= 1) AND (c1 <= 300000))
                           Rows Removed by Filter: 300000
                           Buffers: shared hit=1911
                     ->  Seq Scan on p1_y201710 a_9  (cost=0.00..6625.00 rows=1 width=8) (actual time=24.742..24.742 rows=0 loops=1)
                           Filter: ((c1 >= 1) AND (c1 <= 300000))
                           Rows Removed by Filter: 310000
                           Buffers: shared hit=1975
                     ->  Seq Scan on p1_y201711 a_10  (cost=0.00..6411.00 rows=1 width=8) (actual time=21.484..21.484 rows=0 loops=1)
                           Filter: ((c1 >= 1) AND (c1 <= 300000))
                           Rows Removed by Filter: 300000
                           Buffers: shared hit=1911
                     ->  Seq Scan on p1_y201712 a_11  (cost=0.00..6625.00 rows=1 width=8) (actual time=22.455..22.455 rows=0 loops=1)
                           Filter: ((c1 >= 1) AND (c1 <= 300000))
                           Rows Removed by Filter: 310000
                           Buffers: shared hit=1975
 Planning time: 0.477 ms
 Execution time: 3494.921 ms

 

저작자 표시 비영리 변경 금지
신고
크리에이티브 커먼즈 라이선스
Creative Commons License


 

티스토리 툴바