如何优化在 postgresql 中查询这些数据？

Question

我有一个查询对于特定行来说很慢。 Postgres 选择对某些行使用Seq Scan 而不是使用Index Scan，因为它实际上会比使用索引更快。

以下是针对普通工作负载使用索引的查询计划：http://explain.depesz.com/s/1A2o：

EXPLAIN (ANALYZE, BUFFERS) SELECT "blocks".* FROM "blocks" INNER JOIN "jobs" ON "blocks"."job_id" = "jobs"."id" WHERE "jobs"."project_id" = 1;
                                                                 QUERY PLAN                                                                 
--------------------------------------------------------------------------------------------------------------------------------------------
 Nested Loop  (cost=0.71..166.27 rows=19 width=130) (actual time=0.092..4.247 rows=2421 loops=1)
   Buffers: shared hit=350
   ->  Index Scan using index_jobs_on_project_id on jobs  (cost=0.29..18.81 rows=4 width=4) (actual time=0.044..0.099 rows=15 loops=1)
         Index Cond: (project_id = 1)
         Buffers: shared hit=17
   ->  Index Scan using index_blocks_on_job_id on blocks  (cost=0.42..36.67 rows=19 width=130) (actual time=0.021..0.133 rows=161 loops=15)
         Index Cond: (job_id = jobs.id)
         Buffers: shared hit=333
 Total runtime: 4.737 ms
(9 rows)

这是选择对不太正常的工作负载进行顺序扫描的查询计划：http://explain.depesz.com/s/cJOd：

EXPLAIN (ANALYZE, BUFFERS) SELECT "blocks".* FROM "blocks" INNER JOIN "jobs" ON "blocks"."job_id" = "jobs"."id" WHERE "jobs"."project_id" = 2;
                                                                 QUERY PLAN                                                                     
----------------------------------------------------------------------------------------------------------------------------------------------------
Hash Join  (cost=1138.64..11236.94 rows=10421 width=130) (actual time=5.212..72.604 rows=2516 loops=1)
 Hash Cond: (blocks.job_id = jobs.id)
 Buffers: shared hit=5671
 ->  Seq Scan on blocks  (cost=0.00..8478.06 rows=303206 width=130) (actual time=0.008..24.573 rows=298084 loops=1)
       Buffers: shared hit=5446
 ->  Hash  (cost=1111.79..1111.79 rows=2148 width=4) (actual time=3.346..3.346 rows=2164 loops=1)
       Buckets: 1024  Batches: 1  Memory Usage: 77kB
       Buffers: shared hit=225
       ->  Bitmap Heap Scan on jobs  (cost=40.94..1111.79 rows=2148 width=4) (actual time=0.595..2.158 rows=2164 loops=1)
             Recheck Cond: (project_id = 2)
             Buffers: shared hit=225
             ->  Bitmap Index Scan on index_jobs_on_project_id  (cost=0.00..40.40 rows=2148 width=0) (actual time=0.516..0.516 rows=2164 loops=1)
                   Index Cond: (project_id = 2)
                   Buffers: shared hit=8
 Total runtime: 72.767 ms
(15 rows)

在第一种情况下，该项目有 15 个工作岗位和 2421 个街区。在第二种情况下，该项目有 2164 个工作和 2516 个区块。

有没有办法查询这些数据以使第二个工作负载不那么慢？还是我只是在处理某种最坏情况下的性能工作负载？

编辑

将 random_page_cost 更新为 1.1 并为慢查询重新运行 EXPLAIN 后：http://explain.depesz.com/s/xKdd

EXPLAIN (ANALYZE, BUFFERS) SELECT "blocks".* FROM "blocks" INNER JOIN "jobs" ON "blocks"."job_id" = "jobs"."id" WHERE "jobs"."project_id" = 2;

                                                              QUERY PLAN                                                                  
----------------------------------------------------------------------------------------------------------------------------------------------
 Nested Loop  (cost=0.71..7634.08 rows=10421 width=130) (actual time=0.025..10.597 rows=2516 loops=1)
   Buffers: shared hit=9206
   ->  Index Scan using index_jobs_on_project_id on jobs  (cost=0.29..1048.99 rows=2148 width=4) (actual time=0.015..1.239 rows=2164 loops=1)
         Index Cond: (project_id = 32357)
         Buffers: shared hit=225
   ->  Index Scan using index_blocks_on_job_id on blocks  (cost=0.42..2.88 rows=19 width=130) (actual time=0.003..0.003 rows=1 loops=2164)
         Index Cond: (job_id = jobs.id)
         Buffers: shared hit=8981
 Total runtime: 10.925 ms
(9 rows)

好多了！看来我需要花一些时间来调整服务器配置。

Answer 1

由于两个索引扫描上的嵌套循环比位图索引扫描上的 hashjoin 快得多，我想说您的 random_page_cost 不能准确反映您的实际性能，至少当数据缓存在 RAM 中或 @ 987654323@.

尝试设置SET random_page_cost = 1.1 并在该会话中重新运行。您可能还想在问题上抛出更多 work_mem。

如果random_page_cost 调整有效，您可能需要更新postgresql.conf 以反映它。请注意，1.1 是一个非常极端的设置；默认值为 4，seq_page_cost 为 1，因此在配置文件中我会从 2 或 1.5 之类的东西开始，以避免使其他计划变得更糟。