PostgreSQL 對IN,EXISTS,ANY/ALL,JOIN的sql優化方案

測試環境:

postgres=# select version();       
                         version                        
---------------------------------------------------------------------------------------------------------
 PostgreSQL 11.9 on x86_64-pc-linux-gnu, compiled by gcc (GCC) 4.8.5 20150623 (Red Hat 4.8.5-39), 64-bit
(1 row) 
postgres=#

數據準備:

$ pgbench -i -s 10
postgres=# \d
       List of relations
 Schema |    Name    | Type | Owner 
--------+------------------+-------+----------
 public | pgbench_accounts | table | postgres
 public | pgbench_branches | table | postgres
 public | pgbench_history | table | postgres
 public | pgbench_tellers | table | postgres
(4 rows)
 
postgres=# select * from pgbench_accounts limit 1;
 aid | bid | abalance |                    filler                    
-----+-----+----------+--------------------------------------------------------------------------------------
  1 |  1 |    0 |                                          
(1 row)
 
postgres=# select * from pgbench_branches limit 1;
 bid | bbalance | filler
-----+----------+--------
  1 |    0 |
(1 row)
 
postgres=# select * from pgbench_history limit 1;
 tid | bid | aid | delta | mtime | filler
-----+-----+-----+-------+-------+--------
(0 rows)
 
postgres=# select * from pgbench_tellers limit 1;
 tid | bid | tbalance | filler
-----+-----+----------+--------
  1 |  1 |    0 |
(1 row)
 
postgres=# select * from pgbench_branches;
 bid | bbalance | filler
-----+----------+--------
  1 |    0 |
  2 |    0 |
  3 |    0 |
  4 |    0 |
  5 |    0 |
  6 |    0 |
  7 |    0 |
  8 |    0 |
  9 |    0 |
 10 |    0 |
(10 rows)
 
postgres=# update pgbench_branches set bbalance=4500000 where bid in (4,7);
UPDATE 2
postgres=#

IN語句

查詢要求:找出那些餘額(balance)大於0的每個分支(branch)在表在pgbench_accounts中有多少個賬戶

1.使用IN子句 

SELECT
  count( aid ),bid
FROM
  pgbench_accounts
WHERE
  bid IN ( SELECT bid FROM pgbench_branches WHERE bbalance > 0 )
GROUP BY
  bid;
 

2.使用ANY子句

SELECT
  count( aid ),bid
FROM
  pgbench_accounts
WHERE
  bid = ANY ( SELECT bid FROM pgbench_branches WHERE bbalance > 0 )
GROUP BY
  bid;

  

3.使用EXISTS子句

SELECT
  count( aid ),bid
FROM
  pgbench_accounts
WHERE
  EXISTS ( SELECT bid FROM pgbench_branches WHERE bbalance > 0 AND pgbench_accounts.bid = pgbench_branches.bid )
GROUP BY
  bid;

  

4.使用INNER JOIN

SELECT
  count( aid ),a.bid
FROM
  pgbench_accounts a
  JOIN pgbench_branches b ON a.bid = b.bid
WHERE
  b.bbalance > 0
GROUP BY
  a.bid;

在完成這個查詢要求的時候,有人可能會假設exists和inner join性能可能會更好,因為他們可以使用兩表連接的邏輯和優化。而IN和ANY子句需要使用子查詢。

然而,PostgreSQL(10版本之後)已經智能的足以對上面四種寫法產生相同的執行計劃!

所有上面的寫法都會產生相同的執行計劃:

                                      QUERY PLAN                                      
------------------------------------------------------------------------------------------------------------------------------------------------------------------
 Finalize GroupAggregate (cost=23327.73..23330.26 rows=10 width=12) (actual time=97.199..99.014 rows=2 loops=1)
  Group Key: a.bid
  -> Gather Merge (cost=23327.73..23330.06 rows=20 width=12) (actual time=97.191..99.006 rows=6 loops=1)
     Workers Planned: 2
     Workers Launched: 2
     -> Sort (cost=22327.70..22327.73 rows=10 width=12) (actual time=93.762..93.766 rows=2 loops=3)
        Sort Key: a.bid
        Sort Method: quicksort Memory: 25kB
        Worker 0: Sort Method: quicksort Memory: 25kB
        Worker 1: Sort Method: quicksort Memory: 25kB
        -> Partial HashAggregate (cost=22327.44..22327.54 rows=10 width=12) (actual time=93.723..93.727 rows=2 loops=3)
           Group Key: a.bid
           -> Hash Join (cost=1.14..22119.10 rows=41667 width=8) (actual time=24.024..83.263 rows=66667 loops=3)
              Hash Cond: (a.bid = b.bid)
              -> Parallel Seq Scan on pgbench_accounts a (cost=0.00..20560.67 rows=416667 width=8) (actual time=0.023..43.151 rows=333333 loops=3)
              -> Hash (cost=1.12..1.12 rows=1 width=4) (actual time=0.027..0.028 rows=2 loops=3)
                 Buckets: 1024 Batches: 1 Memory Usage: 9kB
                 -> Seq Scan on pgbench_branches b (cost=0.00..1.12 rows=1 width=4) (actual time=0.018..0.020 rows=2 loops=3)
                    Filter: (bbalance > 0)
                    Rows Removed by Filter: 8
 Planning Time: 0.342 ms
 Execution Time: 99.164 ms
(22 rows)

那麼,我們是否可以得出這樣的結論:我們可以隨意地編寫查詢,而PostgreSQL的智能將會處理其餘的問題?!

等等!

如果我們考慮排除情況,事情會變得不同。

排除查詢

查詢要求:找出那些餘額(balance)不大於0的每個分支(branch)在表在pgbench_accounts中有多少個賬戶

1.使用NOT IN

SELECT
  count( aid ),bid
FROM
  pgbench_accounts
WHERE
  bid NOT IN ( SELECT bid FROM pgbench_branches WHERE bbalance > 0 )
GROUP BY
  bid;

執行計劃:

                                    QUERY PLAN                                    
----------------------------------------------------------------------------------------------------------------------------------------------------------
 Finalize GroupAggregate (cost=23645.42..23647.95 rows=10 width=12) (actual time=128.606..130.502 rows=8 loops=1)
  Group Key: pgbench_accounts.bid
  -> Gather Merge (cost=23645.42..23647.75 rows=20 width=12) (actual time=128.598..130.490 rows=24 loops=1)
     Workers Planned: 2
     Workers Launched: 2
     -> Sort (cost=22645.39..22645.42 rows=10 width=12) (actual time=124.960..124.963 rows=8 loops=3)
        Sort Key: pgbench_accounts.bid
        Sort Method: quicksort Memory: 25kB
        Worker 0: Sort Method: quicksort Memory: 25kB
        Worker 1: Sort Method: quicksort Memory: 25kB
        -> Partial HashAggregate (cost=22645.13..22645.23 rows=10 width=12) (actual time=124.917..124.920 rows=8 loops=3)
           Group Key: pgbench_accounts.bid
           -> Parallel Seq Scan on pgbench_accounts (cost=1.13..21603.46 rows=208333 width=8) (actual time=0.078..83.134 rows=266667 loops=3)
              Filter: (NOT (hashed SubPlan 1))
              Rows Removed by Filter: 66667
              SubPlan 1
               -> Seq Scan on pgbench_branches (cost=0.00..1.12 rows=1 width=4) (actual time=0.020..0.021 rows=2 loops=3)
                  Filter: (bbalance > 0)
                  Rows Removed by Filter: 8
 Planning Time: 0.310 ms
 Execution Time: 130.620 ms
(21 rows)
 
postgres=#

2.使用<>ALL

SELECT
  count( aid ),bid
FROM
  pgbench_accounts
WHERE
  bid <> ALL ( SELECT bid FROM pgbench_branches WHERE bbalance > 0 )
GROUP BY
  bid;

執行計劃:

                                     QUERY PLAN                                    
------------------------------------------------------------------------------------------------------------------------------------------------------------
 Finalize GroupAggregate (cost=259581.79..259584.32 rows=10 width=12) (actual time=418.220..419.913 rows=8 loops=1)
  Group Key: pgbench_accounts.bid
  -> Gather Merge (cost=259581.79..259584.12 rows=20 width=12) (actual time=418.212..419.902 rows=24 loops=1)
     Workers Planned: 2
     Workers Launched: 2
     -> Sort (cost=258581.76..258581.79 rows=10 width=12) (actual time=413.906..413.909 rows=8 loops=3)
        Sort Key: pgbench_accounts.bid
        Sort Method: quicksort Memory: 25kB
        Worker 0: Sort Method: quicksort Memory: 25kB
        Worker 1: Sort Method: quicksort Memory: 25kB
        -> Partial HashAggregate (cost=258581.50..258581.60 rows=10 width=12) (actual time=413.872..413.875 rows=8 loops=3)
           Group Key: pgbench_accounts.bid
           -> Parallel Seq Scan on pgbench_accounts (cost=0.00..257539.83 rows=208333 width=8) (actual time=0.054..367.244 rows=266667 loops=3)
              Filter: (SubPlan 1)
              Rows Removed by Filter: 66667
              SubPlan 1
               -> Materialize (cost=0.00..1.13 rows=1 width=4) (actual time=0.000..0.001 rows=2 loops=1000000)
                  -> Seq Scan on pgbench_branches (cost=0.00..1.12 rows=1 width=4) (actual time=0.001..0.001 rows=2 loops=337880)
                     Filter: (bbalance > 0)
                     Rows Removed by Filter: 8
 Planning Time: 0.218 ms
 Execution Time: 420.035 ms
(22 rows) 
postgres=#

3.使用NOT EXISTS

SELECT
  count( aid ),bid
FROM
  pgbench_accounts
WHERE
  NOT EXISTS ( SELECT bid FROM pgbench_branches WHERE bbalance > 0 AND pgbench_accounts.bid = pgbench_branches.bid )
GROUP BY
  bid;

執行計劃:

                                      QUERY PLAN                                     
----------------------------------------------------------------------------------------------------------------------------------------------------------------
 Finalize GroupAggregate (cost=28327.72..28330.25 rows=10 width=12) (actual time=152.024..153.931 rows=8 loops=1)
  Group Key: pgbench_accounts.bid
  -> Gather Merge (cost=28327.72..28330.05 rows=20 width=12) (actual time=152.014..153.917 rows=24 loops=1)
     Workers Planned: 2
     Workers Launched: 2
     -> Sort (cost=27327.70..27327.72 rows=10 width=12) (actual time=147.782..147.786 rows=8 loops=3)
        Sort Key: pgbench_accounts.bid
        Sort Method: quicksort Memory: 25kB
        Worker 0: Sort Method: quicksort Memory: 25kB
        Worker 1: Sort Method: quicksort Memory: 25kB
        -> Partial HashAggregate (cost=27327.43..27327.53 rows=10 width=12) (actual time=147.732..147.737 rows=8 loops=3)
           Group Key: pgbench_accounts.bid
           -> Hash Anti Join (cost=1.14..25452.43 rows=375000 width=8) (actual time=0.134..101.884 rows=266667 loops=3)
              Hash Cond: (pgbench_accounts.bid = pgbench_branches.bid)
              -> Parallel Seq Scan on pgbench_accounts (cost=0.00..20560.67 rows=416667 width=8) (actual time=0.032..45.174 rows=333333 loops=3)
              -> Hash (cost=1.12..1.12 rows=1 width=4) (actual time=0.036..0.037 rows=2 loops=3)
                 Buckets: 1024 Batches: 1 Memory Usage: 9kB
                 -> Seq Scan on pgbench_branches (cost=0.00..1.12 rows=1 width=4) (actual time=0.025..0.027 rows=2 loops=3)
                    Filter: (bbalance > 0)
                    Rows Removed by Filter: 8
 Planning Time: 0.322 ms
 Execution Time: 154.040 ms
(22 rows) 
postgres=#

4.使用LEFT JOIN和IS NULL

SELECT
  count( aid ),a.bid
FROM
  pgbench_accounts a
  LEFT JOIN pgbench_branches b ON a.bid = b.bid AND b.bbalance > 0
WHERE
  b.bid IS NULL
GROUP BY
  a.bid;

執行計劃:

                                      QUERY PLAN                                      
------------------------------------------------------------------------------------------------------------------------------------------------------------------
 Finalize GroupAggregate (cost=28327.72..28330.25 rows=10 width=12) (actual time=145.298..147.096 rows=8 loops=1)
  Group Key: a.bid
  -> Gather Merge (cost=28327.72..28330.05 rows=20 width=12) (actual time=145.288..147.083 rows=24 loops=1)
     Workers Planned: 2
     Workers Launched: 2
     -> Sort (cost=27327.70..27327.72 rows=10 width=12) (actual time=141.883..141.887 rows=8 loops=3)
        Sort Key: a.bid
        Sort Method: quicksort Memory: 25kB
        Worker 0: Sort Method: quicksort Memory: 25kB
        Worker 1: Sort Method: quicksort Memory: 25kB
        -> Partial HashAggregate (cost=27327.43..27327.53 rows=10 width=12) (actual time=141.842..141.847 rows=8 loops=3)
           Group Key: a.bid
           -> Hash Anti Join (cost=1.14..25452.43 rows=375000 width=8) (actual time=0.087..99.535 rows=266667 loops=3)
              Hash Cond: (a.bid = b.bid)
              -> Parallel Seq Scan on pgbench_accounts a (cost=0.00..20560.67 rows=416667 width=8) (actual time=0.025..44.337 rows=333333 loops=3)
              -> Hash (cost=1.12..1.12 rows=1 width=4) (actual time=0.026..0.027 rows=2 loops=3)
                 Buckets: 1024 Batches: 1 Memory Usage: 9kB
                 -> Seq Scan on pgbench_branches b (cost=0.00..1.12 rows=1 width=4) (actual time=0.019..0.020 rows=2 loops=3)
                    Filter: (bbalance > 0)
                    Rows Removed by Filter: 8
 Planning Time: 0.231 ms
 Execution Time: 147.180 ms
(22 rows) 
postgres=#

NOT IN 和 <> ALL生成執行計劃都包含瞭一個子查詢。他們是各自獨立的。

而NOT EXISTS和LEFT JOIN生成瞭相同的執行計劃。

這些hash連接(或hash anti join)是完成查詢要求的最靈活的方式。這也是推薦exists或join的原因。因此,推薦使用exists或join的經驗法則是有效的。

但是,我們繼續往下看! 即使有瞭子查詢執行計劃,NOT IN子句的執行時間也會更好?

是的。PostgreSQL做瞭出色的優化,PostgreSQL將子查詢計劃進行瞭hash處理。因此PostgreSQL對如何處理IN子句有瞭更好的理解,這是一種邏輯思維方式,因為很多人傾向於使用IN子句。子查詢返回的行很少,但即使子查詢返回幾百行,也會發生同樣的情況。

但是,如果子查詢返回大量行(幾十萬行)怎麼辦?讓我們嘗試一個簡單的測試:

CREATE TABLE t1 AS
SELECT * FROM generate_series(0, 500000) id;
 
CREATE TABLE t2 AS
SELECT (random() * 4000000)::integer id
FROM generate_series(0, 4000000);
 
ANALYZE t1;
ANALYZE t2;
 
EXPLAIN SELECT id
FROM t1
WHERE id NOT IN (SELECT id FROM t2);

執行計劃:

    QUERY PLAN                 
--------------------------------------------------------------------------------
 Gather (cost=1000.00..15195064853.01 rows=250000 width=4)
  Workers Planned: 1
  -> Parallel Seq Scan on t1 (cost=0.00..15195038853.01 rows=147059 width=4)
     Filter: (NOT (SubPlan 1))
     SubPlan 1
      -> Materialize (cost=0.00..93326.01 rows=4000001 width=4)
         -> Seq Scan on t2 (cost=0.00..57700.01 rows=4000001 width=4)
(7 rows)
 
postgres=#

這裡,執行計劃將子查詢進行瞭物化。代價評估變成瞭15195038853.01。(PostgreSQL的默認設置,如果t2表的行低於100k,會將子查詢進行hash)。這樣就會嚴重影響性能。因此,對於那種子查詢返回的行數很少的場景,IN子句可以起到很好的作用。

其它註意點

有的!在我們用不同的方式寫查詢的時候,可能有數據類型的轉換。

比如,語句:

EXPLAIN ANALYZE SELECT * FROM emp WHERE gen = ANY(ARRAY['M', 'F']);

就會發生隱式的類型轉換:

Seq Scan on emp (cost=0.00..1.04 rows=2 width=43) (actual time=0.023..0.026 rows=3 loops=1)
 Filter: ((gen)::text = ANY ('{M,F}'::text[]))

這裡的(gen)::text就發生瞭類型轉換。如果在大表上,這種類型轉換的代價會很高,因此,PostgreSQL對IN子句做瞭更好的處理。

EXPLAIN ANALYZE SELECT * FROM emp WHERE gen IN ('M','F');
 
 Seq Scan on emp (cost=0.00..1.04 rows=3 width=43) (actual time=0.030..0.034 rows=3 loops=1)
  Filter: (gen = ANY ('{M,F}'::bpchar[]))

將IN子句轉換成瞭ANY子句,沒有對gen列進行類型轉換。而是將M\F轉成瞭bpchar(內部等價於char)

總結

簡單來說,exists和直接join表通常比較好。

很多情況下,PostgreSQL將IN子句換成被hash的子計劃。在一些特殊場景下,IN可以獲得更好的執行計劃。

以上為個人經驗,希望能給大傢一個參考,也希望大傢多多支持WalkonNet。如有錯誤或未考慮完全的地方,望不吝賜教。

推薦閱讀: