Insert-select 在添加 limit 子句时得到更好的计划

Question

这是我所在的服务器运行

select version();
                                                 version
---------------------------------------------------------------------------    
PostgreSQL 10.6 on x86_64-pc-linux-gnu, compiled by gcc (GCC) 4.8.5 20150623 (Red Hat 4.8.5-36), 64-bit
(1 row)

我开始写 select（ext.t_event 和 ext.t_event_data 是 oracle_fdw（1.1 版）从远程获取的两个外来 table甲骨文数据库）

select 
  te.id_data, 
  te.id_device, 
  te.date_write, 
  te.date_event, 
  ted.i_inout, 
  ted.value
from ext.t_event te, ext.t_event_data ted 
where te.id_device =2749651 
  and te.date_event >= '2019-01-16'and te.date_event < '2019-01-17' 
  and te.id_data=ted.id_data;

获取整个记录集（3600 条记录）大约需要 10 秒。

但后来我把 select 变成了插入 select

insert into stg_data
select 
  te.id_data, 
  te.id_device, 
  te.date_write, 
  te.date_event, 
  ted.i_inout, 
  ted.value
from ext.t_event te, ext.t_event_data ted 
where te.id_device =2749651 
  and te.date_event >= '2019-01-16'and te.date_event < '2019-01-17' 
  and te.id_data=ted.id_data;

我被迫终止查询，它已经运行超过 30 分钟了！

经过几个小时的挣扎和绝望的尝试，我决定试试这个

insert into stg_data
select 
  te.id_data, 
  te.id_device, 
  te.date_write, 
  te.date_event, 
  ted.i_inout, 
  ted.value
from ext.t_event te, ext.t_event_data ted 
where te.id_device =2749651 
  and te.date_event >= '2019-01-16'and te.date_event < '2019-01-17' 
  and te.id_data=ted.id_data
  limit 5000;

并且...在 20 秒内出人意料，我将整个记录集存储在 stg_data。

为了更好地理解差异，我决定分析计划。

SELECT不限

 Foreign Scan  (cost=10000.00..20000.00 rows=1000 width=548)
   Oracle query: SELECT /*eb01c463a72c3b6350f86f5db25e1353*/ r1."ID_DATA",
   r1."ID_DEVICE", r1."DATE_WRITE", r1."DATE_EVENT", r2."I_INOUT",
   r2."VALUE" FROM ("DISPATCH"."T_EVENT" r1 INNER JOIN
   "DISPATCH"."T_EVENT_DATA" r2 ON (r1."ID_DATA" = r2."ID_DATA") AND
  (r1."DATE_EVENT" >= (CAST ('2019-01-16 00:00:00.000000 AD' AS
  TIMESTAMP))) AND (r1."DATE_EVENT" < 
  (CAST ('2019-01-17 00:00:00.000000 AD' AS TIMESTAMP))) 
  AND (r1."ID_DEVICE" = 2749651))

SELECT 有限制

 Limit  (cost=10000.00..20000.00 rows=1000 width=548)
   ->  Foreign Scan  (cost=10000.00..20000.00 rows=1000 width=548)
      Oracle query: SELECT /*eb01c463a72c3b6350f86f5db25e1353*/
      r1."ID_DATA", r1."ID_DEVICE", r1."DATE_WRITE", r1."DATE_EVENT", 
      r2."I_INOUT", r2."VALUE" FROM ("DISPATCH"."T_EVENT" r1 INNER 
      JOIN "DISPATCH"."T_EVENT_DATA" r2 ON (r1."ID_DATA" = r2."ID_DATA")
      AND (r1."DATE_EVENT" >= (CAST ('2019-01-16 00:00:00.000000 AD' AS 
      TIMESTAMP))) AND (r1."DATE_EVENT" < (CAST ('2019-01-17
      00:00:00.000000 AD' AS TIMESTAMP))) AND (r1."ID_DEVICE" = 2749651))

所以它基本上向 Oracle 发送相同的查询，并在获取完成后立即在本地应用 FILTER。

INSER-SELECT 计划看起来一样吗？不！

INSERT_SELECT 限制

Insert on stg_data_hist  (cost=10000.00..20010.00 rows=1000 width=548)
   ->  Limit  (cost=10000.00..20000.00 rows=1000 width=548)
         ->  Foreign Scan  (cost=10000.00..20000.00 rows=1000 width=548)
               Oracle query: SELECT /*eb01c463a72c3b6350f86f5db25e1353*/ 
               r1."ID_DATA", r1."ID_DEVICE", r1."DATE_WRITE", 
               r1."DATE_EVENT", r2."I_INOUT", r2."VALUE" FROM 
               ("DISPATCH"."T_EVENT" r1 INNER JOIN 
               "DISPATCH"."T_EVENT_DATA" r2 ON (r1."ID_DATA" = 
               r2."ID_DATA") AND (r1."DATE_EVENT" >= (CAST ('2019-01-16 
               00:00:00.000000 AD' AS TIMESTAMP))) AND (r1."DATE_EVENT" < 
               (CAST('2019-01-17 00:00:00.000000 AD' AS TIMESTAMP))) AND 
               (r1."ID_DEVICE" = 2749651))

INSERT-SELECT 无 LIMIT 子句

Insert on stg_data_hist  (cost=30012.50..40190.00 rows=5000 width=548)
 ->  Hash Join  (cost=30012.50..40190.00 rows=5000 width=548)
       Hash Cond: (te.id_data = ted.id_data)
     ->  Foreign Scan on t_event te  (cost=10000.00..20000.00 rows=1000 width=28)
           Oracle query: SELECT /*93379c271b3f1bc08a1dbb94fb89f739*/ 
           r3."ID_DATA", r3."ID_DEVICE", r3."DATE_WRITE", r3."DATE_EVENT" 
           FROM "DISPATCH"."T_EVENT" r3 WHERE (r3."DATE_EVENT" >= 
           (CAST ('2019-01-16 00:00:00.000000 AD' AS TIMESTAMP))) AND 
           (r3."DATE_EVENT" < (CAST ('2019-01-17 00:00:00.000000 AD' AS 
           TIMESTAMP))) AND (r3."ID_DEVICE" = 2749651)
       ->  Hash  (cost=20000.00..20000.00 rows=1000 width=528)
           ->  Foreign Scan on t_event_data ted  
                  (cost=10000.00..20000.00 rows=1000 width=528)
                 Oracle query: SELECT /*21c8741f2fa8a8d13d037c3191e8ac96*/ 
                    r4."ID_DATA", r4."I_INOUT", r4."VALUE" FROM 
                    "DISPATCH"."T_EVENT_DATA" r4

这就解释了为什么它比另一个花费的时间更长。它从一个外部 table 检索日期过滤的记录，从第二个外部 table 检索完整集并在本地进行连接。这将需要很长时间！这是几百万条记录与几千条记录。

最后是我的两个问题

1) 我想要第一个计划，但要去掉 LIMIT 子句（让我脊背发凉 :-)）。你会怎么做？除了 join 子句之外，我没有办法对 ext.t_event_data 应用过滤器。

2) 为什么两个 INSERT-SELECT 计划看起来如此不同，尽管两个 SELECT 计划看起来如此相似？

感谢阅读，祝您有愉快的一天

Answer 1

计划者似乎认为无论哪种方式它只会得到几千行，这显然有很大差距，请确保外国 table 的统计数据在运行之前更新 'ANALYZE ext.t_event' 和 ext.t_event_data 相同，因为：

https://github.com/laurenz/oracle_fdw

PostgreSQL will not automatically gather statistics for foreign tables with the autovacuum daemon.

Keep in mind that analyzing an Oracle foreign table will result in a full sequential table scan. You can use the table option sample_percent to speed this up by using only a sample of the Oracle table.

联接是在 select 情况下下推到 Oracle，在插入情况下如果使用限制，所以我能看到不使用它的唯一原因insert without limit 是缺少精确的 table 统计数据。您可以尝试将插入查询重写为 CTE（出于明显的原因尚未测试此查询）：

WITH foreign_data AS (
select 
  te.id_data, 
  te.id_device, 
  te.date_write, 
  te.date_event, 
  ted.i_inout, 
  ted.value
from ext.t_event te, ext.t_event_data ted 
where te.id_device =2749651 
  and te.date_event >= '2019-01-16'and te.date_event < '2019-01-17' 
  and te.id_data=ted.id_data
)

insert into stg_data from foreign_data

您也可以尝试将查询重写为显式内部联接，而不是在 where 子句中加入联接条件 (te.id_data=ted.id_data)。

Insert-select 在添加 limit 子句时得到更好的计划

Insert-select gets a better plan when limit clause added

postgresql

sql-execution-plan

insert-select

foreign-data-wrapper