postgresql insert into select无法使用并行查询的解决

本文信息基于PG13.1。

从PG9.6开始支持并行查询。PG11开始支持CREATE TABLE … AS、SELECT INTO以及CREATE MATERIALIZED VIEW的并行查询。

先说结论：

换用create table as 或者select into或者导入导出。

首先跟踪如下查询语句的执行计划：

select count(*) from test t1,test1 t2 where t1.id = t2.id ;

postgres=# explain analyze select count(*) from test t1,test1 t2 where t1.id = t2.id ;
                                    QUERY PLAN                                    
--------------------------------------------------------------------------------------------------------------------------------------------------------
 Finalize Aggregate (cost=34244.16..34244.17 rows=1 width=8) (actual time=683.246..715.324 rows=1 loops=1)
  -> Gather (cost=34243.95..34244.16 rows=2 width=8) (actual time=681.474..715.311 rows=3 loops=1)
     Workers Planned: 2
     Workers Launched: 2
     -> Partial Aggregate (cost=33243.95..33243.96 rows=1 width=8) (actual time=674.689..675.285 rows=1 loops=3)
        -> Parallel Hash Join (cost=15428.00..32202.28 rows=416667 width=0) (actual time=447.799..645.689 rows=333333 loops=3)
           Hash Cond: (t1.id = t2.id)
           -> Parallel Seq Scan on test t1 (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.025..74.010 rows=333333 loops=3)
           -> Parallel Hash (cost=8591.67..8591.67 rows=416667 width=4) (actual time=260.052..260.053 rows=333333 loops=3)
              Buckets: 131072 Batches: 16 Memory Usage: 3520kB
              -> Parallel Seq Scan on test1 t2 (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.032..104.804 rows=333333 loops=3)
 Planning Time: 0.420 ms
 Execution Time: 715.447 ms
(13 rows)

可以看到走了两个Workers。

下边看一下insert into select：

postgres=# explain analyze insert into va select count(*) from test t1,test1 t2 where t1.id = t2.id ;     
                                  QUERY PLAN                                  
--------------------------------------------------------------------------------------------------------------------------------------------------
 Insert on va (cost=73228.00..73228.02 rows=1 width=4) (actual time=3744.179..3744.187 rows=0 loops=1)
  -> Subquery Scan on "*SELECT*" (cost=73228.00..73228.02 rows=1 width=4) (actual time=3743.343..3743.352 rows=1 loops=1)
     -> Aggregate (cost=73228.00..73228.01 rows=1 width=8) (actual time=3743.247..3743.254 rows=1 loops=1)
        -> Hash Join (cost=30832.00..70728.00 rows=1000000 width=0) (actual time=1092.295..3511.301 rows=1000000 loops=1)
           Hash Cond: (t1.id = t2.id)
           -> Seq Scan on test t1 (cost=0.00..14425.00 rows=1000000 width=4) (actual time=0.030..421.537 rows=1000000 loops=1)
           -> Hash (cost=14425.00..14425.00 rows=1000000 width=4) (actual time=1090.078..1090.081 rows=1000000 loops=1)
              Buckets: 131072 Batches: 16 Memory Usage: 3227kB
              -> Seq Scan on test1 t2 (cost=0.00..14425.00 rows=1000000 width=4) (actual time=0.021..422.768 rows=1000000 loops=1)
 Planning Time: 0.511 ms
 Execution Time: 3745.633 ms
(11 rows)

可以看到并没有Workers的指示，没有启用并行查询。

即使开启强制并行，也无法走并行查询。

postgres=# set force_parallel_mode =on;
SET
postgres=# explain analyze insert into va select count(*) from test t1,test1 t2 where t1.id = t2.id ;
                                  QUERY PLAN                                  
--------------------------------------------------------------------------------------------------------------------------------------------------
 Insert on va (cost=73228.00..73228.02 rows=1 width=4) (actual time=3825.042..3825.049 rows=0 loops=1)
  -> Subquery Scan on "*SELECT*" (cost=73228.00..73228.02 rows=1 width=4) (actual time=3824.976..3824.984 rows=1 loops=1)
     -> Aggregate (cost=73228.00..73228.01 rows=1 width=8) (actual time=3824.972..3824.978 rows=1 loops=1)
        -> Hash Join (cost=30832.00..70728.00 rows=1000000 width=0) (actual time=1073.587..3599.402 rows=1000000 loops=1)
           Hash Cond: (t1.id = t2.id)
           -> Seq Scan on test t1 (cost=0.00..14425.00 rows=1000000 width=4) (actual time=0.034..414.965 rows=1000000 loops=1)
           -> Hash (cost=14425.00..14425.00 rows=1000000 width=4) (actual time=1072.441..1072.443 rows=1000000 loops=1)
              Buckets: 131072 Batches: 16 Memory Usage: 3227kB
              -> Seq Scan on test1 t2 (cost=0.00..14425.00 rows=1000000 width=4) (actual time=0.022..400.624 rows=1000000 loops=1)
 Planning Time: 0.577 ms
 Execution Time: 3825.923 ms
(11 rows)

原因在官方文档有写：

The query writes any data or locks any database rows. If a query contains a data-modifying operation either at the top level or within a CTE, no parallel plans for that query will be generated. As an exception, the commands CREATE TABLE … AS, SELECT INTO, and CREATE MATERIALIZED VIEW which create a new table and populate it can use a parallel plan.

解决方案有如下三种：

1.select into

postgres=# explain analyze select count(*) into vaa from test t1,test1 t2 where t1.id = t2.id ;
                                    QUERY PLAN                                    
--------------------------------------------------------------------------------------------------------------------------------------------------------
 Finalize Aggregate (cost=34244.16..34244.17 rows=1 width=8) (actual time=742.736..774.923 rows=1 loops=1)
  -> Gather (cost=34243.95..34244.16 rows=2 width=8) (actual time=740.223..774.907 rows=3 loops=1)
     Workers Planned: 2
     Workers Launched: 2
     -> Partial Aggregate (cost=33243.95..33243.96 rows=1 width=8) (actual time=731.408..731.413 rows=1 loops=3)
        -> Parallel Hash Join (cost=15428.00..32202.28 rows=416667 width=0) (actual time=489.880..700.830 rows=333333 loops=3)
           Hash Cond: (t1.id = t2.id)
           -> Parallel Seq Scan on test t1 (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.033..87.479 rows=333333 loops=3)
           -> Parallel Hash (cost=8591.67..8591.67 rows=416667 width=4) (actual time=266.839..266.840 rows=333333 loops=3)
              Buckets: 131072 Batches: 16 Memory Usage: 3520kB
              -> Parallel Seq Scan on test1 t2 (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.058..106.874 rows=333333 loops=3)
 Planning Time: 0.319 ms
 Execution Time: 783.300 ms
(13 rows)

2.create table as

postgres=# explain analyze create table vb as select count(*) from test t1,test1 t2 where t1.id = t2.id ;
                                   QUERY PLAN                                    
-------------------------------------------------------------------------------------------------------------------------------------------------------
 Finalize Aggregate (cost=34244.16..34244.17 rows=1 width=8) (actual time=540.120..563.733 rows=1 loops=1)
  -> Gather (cost=34243.95..34244.16 rows=2 width=8) (actual time=537.982..563.720 rows=3 loops=1)
     Workers Planned: 2
     Workers Launched: 2
     -> Partial Aggregate (cost=33243.95..33243.96 rows=1 width=8) (actual time=526.602..527.136 rows=1 loops=3)
        -> Parallel Hash Join (cost=15428.00..32202.28 rows=416667 width=0) (actual time=334.532..502.793 rows=333333 loops=3)
           Hash Cond: (t1.id = t2.id)
           -> Parallel Seq Scan on test t1 (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.018..57.819 rows=333333 loops=3)
           -> Parallel Hash (cost=8591.67..8591.67 rows=416667 width=4) (actual time=189.502..189.503 rows=333333 loops=3)
              Buckets: 131072 Batches: 16 Memory Usage: 3520kB
              -> Parallel Seq Scan on test1 t2 (cost=0.00..8591.67 rows=416667 width=4) (actual time=0.023..77.786 rows=333333 loops=3)
 Planning Time: 0.189 ms
 Execution Time: 565.448 ms
(13 rows)

3.或者通过导入导出的方式，例如：

psql -h localhost -d postgres -U postgres -c "select count(*) from test t1,test1 t2 where t1.id = t2.id " -o result.csv -A -t -F ","
psql -h localhost -d postgres -U postgres -c "COPY va FROM 'result.csv' WITH (FORMAT CSV, DELIMITER ',', HEADER FALSE, ENCODING 'windows-1252')"

一些场景下也会比非并行快。