同事今天晉級高階工程師考試,發來一道公司出題目讓我幫忙進行最佳化,其中場景二的案例非常有意思?。
題目內容如下:
原始SQL:
scott=> explain analyze scott-> select scott-> a.id, scott-> a.col2, scott-> (select sum(b.id) from table02 b where a.col2 like b.col2||'%' ) scott-> from table01 a; QUERY PLAN ------------------------------------------------------------------------------------------------------------------------ Seq Scan on table01 a (cost=0.00..3905341.00 rows=100000 width=45) (actual time=0.579..50568.090 rows=100000 loops=1) SubPlan 1 -> Aggregate (cost=39.02..39.03 rows=1 width=8) (actual time=0.504..0.504 rows=1 loops=100000) -> Seq Scan on table02 b (cost=0.00..39.00 rows=10 width=4) (actual time=0.063..0.499 rows=8 loops=100000) Filter: (a.col2 ~~ (col2 || '%'::text)) Rows Removed by Filter: 1992 Planning Time: 0.097 ms Execution Time: 50590.882 ms (8 行記錄) 時間:50591.756 ms (00:50.592)
table01、table02 這兩張表沒有建立任何索引,全表掃描+標量子查詢SQL執行需要50s才能出結果,速度非常慢。
考題要求要最佳化這條SQL,意思既是無論是調整 postgresql資料庫的引數,對SQL加索引,等價改寫SQL,這些手段都沒問題,只要能讓執行速度變快就行。
由於當時我在忙其他的事情,大致看了一眼後給出了以下的改寫方案(我沒加索引,感覺加索引的用處不大):
改寫1:
scott=> explain analyze select scott-> a.id, scott-> a.col2, scott-> b.sum_b_id scott-> from table01 a scott-> left join (select sum(b.id) sum_b_id,b.col2 from table02 b group by b.col2) b scott-> ON a.col2 like b.col2||'%' scott-> ; QUERY PLAN ---------------------------------------------------------------------------------------------------------------------------- Nested Loop Left Join (cost=39.00..448135.74 rows=127500 width=45) (actual time=1.283..8674.517 rows=100000 loops=1) Join Filter: (a.col2 ~~ (b.col2 || '%'::text)) Rows Removed by Join Filter: 25400387 -> Seq Scan on table01 a (cost=0.00..1841.00 rows=100000 width=37) (actual time=0.018..19.620 rows=100000 loops=1) -> Materialize (cost=39.00..45.37 rows=255 width=11) (actual time=0.000..0.024 rows=255 loops=100000) -> HashAggregate (cost=39.00..41.55 rows=255 width=11) (actual time=1.241..1.316 rows=255 loops=1) Group Key: b.col2 -> Seq Scan on table02 b (cost=0.00..29.00 rows=2000 width=7) (actual time=0.007..0.342 rows=2000 loops=1) Planning Time: 0.181 ms Execution Time: 8682.974 ms (10 行記錄) 時間:8684.338 ms (00:08.684)
可以看到等價改寫以後,SQL從原來執行 50s 的時間已經降低到 8.8s 左右,提升還是挺大的。
把答案給了同事,我也去忙其他的事情了。
晚上我閒下來沒事做,賊無聊,仔細看了下改寫1 SQL的執行計劃,感覺這種計劃可能不是最優的執行計劃。
因為我始終感覺走HASH可能才是最佳的執行計劃,如果這條SQL在ORACLE 上執行,CBO很大可能會讓計劃走HASH,但是在PG就是走NL(腦殘最佳化器)?。
吃完飯後一直在嘗試改寫,搞了哥很長時間,最終還是把HASH版本的SQL給改出來了,淚目?。
改寫2:
scott=> explain analyze scott-> with a as (select id,col2,substr(t1.col2,1,x.rn) rn1 from table01 t1, scott(> scott(> (select min(length(col2)) rn from table02) x), <t2.col2,1,x.rn) rn2 from (select sum(b.id) sum_b_id,b.col2 from table02 b group by b.col2) t2, scott(> (select min(length(col2)) rn from table02) x) scott-> select a.id,a.col2,b.sum_b_id from a scott-> left join b on a.rn1 = b.rn2 and a.col2 like b.col2||'%'; QUERY PLAN ---------------------------------------------------------------------------------------------------------------------------------- -------- Hash Left Join (cost=127.86..14881.38 rows=100000 width=45) (actual time=2.322..215.695 rows=100000 loops=1) Hash Cond: (substr(t1.col2, 1, (min(length(table02.col2)))) = substr(b.col2, 1, (min(length(table02_1.col2))))) Join Filter: (t1.col2 ~~ (b.col2 || '%'::text)) -> Nested Loop (cost=39.00..2880.02 rows=100000 width=41) (actual time=0.588..36.635 rows=100000 loops=1) -> Aggregate (cost=39.00..39.01 rows=1 width=4) (actual time=0.580..0.582 rows=1 loops=1) -> Seq Scan on table02 (cost=0.00..29.00 rows=2000 width=3) (actual time=0.015..0.252 rows=2000 loops=1) -> Seq Scan on table01 t1 (cost=0.00..1841.00 rows=100000 width=37) (actual time=0.005..15.073 rows=100000 loops=1) -> Hash (cost=85.67..85.67 rows=255 width=15) (actual time=1.721..1.724 rows=255 loops=1) Buckets: 1024 Batches: 1 Memory Usage: 20kB -> Nested Loop (cost=78.00..85.67 rows=255 width=15) (actual time=1.500..1.602 rows=255 loops=1) -> Aggregate (cost=39.00..39.01 rows=1 width=4) (actual time=0.554..0.555 rows=1 loops=1) -> Seq Scan on table02 table02_1 (cost=0.00..29.00 rows=2000 width=3) (actual time=0.004..0.233 rows=2000 l oops=1) -> HashAggregate (cost=39.00..41.55 rows=255 width=11) (actual time=0.945..1.002 rows=255 loops=1) Group Key: b.col2 -> Seq Scan on table02 b (cost=0.00..29.00 rows=2000 width=7) (actual time=0.005..0.250 rows=2000 loops=1) Planning Time: 0.351 ms Execution Time: 224.017 ms (17 行記錄) 時間:225.488 ms
這個案例從最早的 50秒 改寫到 8秒,到最後的 225毫秒出結果,花了不少時間研究各種改寫方式?。
只能說PG的最佳化器確實太拉跨了,要用CTE才能讓最佳化器走HASH連線,浪費開發者不少時間,換成ORACLE資料庫不會走這種SB執行計劃。?
以後估計會很少發部落格,正在考慮轉行賣炒粉,現在的市場真的是一言難盡。?