sql - 在 postgres 中加入性能

Question

我有 2 个表 scheduling_flownode 和 xact_message 它们之间的关系很弱。我正在尝试执行以下查询

set search_path='ad_96d5be';
explain analyze 
SELECT f.id, f.target_object_id 
FROM "scheduling_flownode" f, 
     "xact_message" m 
where f.target_object_id = m.id 
and f.root_node=True 
AND f.state=1 
and m.state=4 
and m.templatelanguage_id IN (17, 18, 19, 20, 21, 22, 23, 24);

执行时，我得到以下查询计划

  Gather  (cost=252701.26..1711972.04 rows=374109 width=8) (actual time=17737.908..164181.063 rows=441130 loops=1)
   Workers Planned: 2
   Workers Launched: 2
   Buffers: shared hit=35705 read=1346425, temp read=18190 written=18148
   ->  Hash Join  (cost=251701.26..1673561.14 rows=155879 width=8) (actual time=18805.587..163991.468 rows=147043 loops=3)
         Hash Cond: (f.target_object_id = m.id)
         Buffers: shared hit=35705 read=1346425, temp read=18190 written=18148
         ->  Parallel Bitmap Heap Scan on scheduling_flownode f  (cost=124367.21..1523127.76 rows=2061083 width=8) (actual time=963.910..155466.840 rows=1642157 loops=3)
               Recheck Cond: (state = 1)
               Rows Removed by Index Recheck: 44
               Filter: root_node
               Rows Removed by Filter: 12406874
               Heap Blocks: exact=10570 lossy=427078
               Buffers: shared read=1328631
               ->  Bitmap Index Scan on "root-node-and-state"  (cost=0.00..123130.57 rows=4946600 width=0) (actual time=955.044..955.045 rows=4926472 loops=1)
                     Index Cond: ((root_node = true) AND (state = 1))
                     Buffers: shared read=13464
         ->  Hash  (cost=120677.64..120677.64 rows=405712 width=4) (actual time=7124.131..7124.131 rows=441128 loops=3)
               Buckets: 131072  Batches: 8  Memory Usage: 2966kB
               Buffers: shared hit=35591 read=17793, temp written=3384
               ->  Bitmap Heap Scan on xact_message m  (cost=7893.56..120677.64 rows=405712 width=4) (actual time=61.307..6925.456 rows=441128 loops=3)
                     Recheck Cond: (state = 4)
                     Filter: (templatelanguage_id = ANY ('{17,18,19,20,21,22,23,24}'::integer[]))
                     Rows Removed by Filter: 4
                     Heap Blocks: exact=16585
                     Buffers: shared hit=35591 read=17793
                     ->  Bitmap Index Scan on "state-index"  (cost=0.00..7792.13 rows=421826 width=0) (actual time=58.781..58.781 rows=441132 loops=3)
                           Index Cond: (state = 4)
                           Buffers: shared hit=2420 read=1209
 Planning time: 1.382 ms
 Execution time: 164289.481 ms
(31 rows)

scheduling_flownode 这里有超过 400,00,000 个条目，xact_message 有大约 50,00,000 行。在 postgres 10 上工作，我是否错误地认为 postgres 应该可以轻松处理这么多负载？如果可以，我这里的查询是否有问题？

Answer 1

你没有显示你有什么索引，但我强烈建议你的索引将覆盖你过滤的所有列。

在 Postgres 11 中，可以使用 COVERING indexes 来完成，因此例如在表 scheduling_flownode 上，您将拥有如下索引:

CREATE INDEX ix_scheduling_flownode_target_object_id 
  ON scheduling_flownode(target_object_id) 
    INCLUDE (state, root_node);

在 Postgres 10 中，只需将列包含在索引中:

CREATE INDEX ix_scheduling_flownode_target_object_id 
  ON scheduling_flownode(target_object_id, state, root_node);

使用 templatelanguage_id 和 state 对表 xact_message 执行相同的操作。