关于几张超大表联合查询查询 SQL 的问题

V2EX = way to explore

V2EX 是一个关于分享和探索的地方

现在注册

已注册用户请登录

这是一个创建于 1606 天前的主题，其中的信息可能已经有所发展或是发生改变。

WITH E AS (
    SELECT ID
    FROM A
    WHERE type = 2
        AND DELETE_FLAG = 0
        AND uid = 41
    UNION
    SELECT a.ID
    FROM A a, B b, A c
    WHERE a.type = 3
        AND a.DELETE_FLAG = 0
        AND a.uid = 41
        AND a.ID = b.ID
        AND b.parent_id = c.ID
)
SELECT COUNT(1)
FROM D d, E e, A t, F f
WHERE d.DELETE_FLAG = 0
    AND d.ID_ = e.ID
    AND e.ID = t.ID
    AND d.aid = f.ID
    AND f.DELETE_FLAG = 0

其中 D 表最大，有 37M(37,000,000)的数据，A 表 410k 数据，F 表 165k 数据

SQL 语句是查询一个数量，在原来 Oracle 下 15s 左右就可以查询；迁移其他数据库后需要 100s 。请问各路大神有什么好的解决方法？目前 D 表已经是分表了，再分表也不现实； D 表的（ DELETE_FLAG,ID,aid ）也加了联合索引。请问有没有其他的优化方法。执行计划如下：

PLAN_TABLE_OUTPUT
Plan hash value: 244456078
 
------------------------------------------------------------------------------------------------------------------------------
| Id  | Operation                                | Name                      | Rows  | Bytes |TempSpc| Cost (%CPU)| Time     |
------------------------------------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT                         |                           |     1 |    43 |       |   420K  (1)| 01:24:03 |
|   1 |  SORT AGGREGATE                          |                           |     1 |    43 |       |            |          |
|*  2 |   HASH JOIN                              |                           |    22M|   906M|  3104K|   420K  (1)| 01:24:03 |
|*  3 |    TABLE ACCESS FULL                     | FULL_ATTRIBUTE_RELATION   |   151K|  1327K|       |   548   (1)| 00:00:07 |
|*  4 |    HASH JOIN                             |                           |    21M|   707M|  7312K|   371K  (1)| 01:14:24 |
|   5 |     INDEX FAST FULL SCAN                 | SYS_C0019687              |   415K|  2436K|       |   344   (1)| 00:00:05 |
|*  6 |     HASH JOIN                            |                           |    21M|   582M|       |   329K  (1)| 01:05:59 |
|   7 |      VIEW                                |                           | 69509 |   882K|       |  5236   (1)| 00:01:03 |
|   8 |       SORT UNIQUE                        |                           | 69509 |  1177K|  1784K|  5236   (1)| 00:01:03 |
|   9 |        UNION-ALL                         |                           |       |       |       |            |          |
|* 10 |         VIEW                             | index$_join$_001          | 64314 |  1004K|       |  3146   (1)| 00:00:38 |
|* 11 |          HASH JOIN                       |                           |       |       |       |            |          |
|* 12 |           HASH JOIN                      |                           |       |       |       |            |          |
|* 13 |            HASH JOIN                     |                           |       |       |       |            |          |
|  14 |             BITMAP CONVERSION TO ROWIDS  |                           | 64314 |  1004K|       |    11   (0)| 00:00:01 |
|* 15 |              BITMAP INDEX SINGLE VALUE   | NODE_PDUID_INDEX          |       |       |       |            |          |
|  16 |             BITMAP CONVERSION TO ROWIDS  |                           | 64314 |  1004K|       |    37   (0)| 00:00:01 |
|* 17 |              BITMAP INDEX SINGLE VALUE   | NODE_DELETE_FLAG_INDEX    |       |       |       |            |          |
|  18 |            BITMAP CONVERSION TO ROWIDS   |                           | 64314 |  1004K|       |    38   (0)| 00:00:01 |
|* 19 |             BITMAP INDEX SINGLE VALUE    | NODE_TYPE_INDEX           |       |       |       |            |          |
|  20 |           INDEX FAST FULL SCAN           | SYS_C0019687              | 64314 |  1004K|       |  1579   (1)| 00:00:19 |
|* 21 |         HASH JOIN                        |                           |  5195 |   172K|       |  1743   (1)| 00:00:21 |
|* 22 |          HASH JOIN                       |                           |  5637 |   154K|       |  1397   (1)| 00:00:17 |
|* 23 |           VIEW                           | index$_join$_002          |  5627 | 90032 |       |  1329   (1)| 00:00:16 |
|* 24 |            HASH JOIN                     |                           |       |       |       |            |          |
|* 25 |             HASH JOIN                    |                           |       |       |       |            |          |
|* 26 |              HASH JOIN                   |                           |       |       |       |            |          |
|  27 |               BITMAP CONVERSION TO ROWIDS|                           |  5627 | 90032 |       |     4   (0)| 00:00:01 |
|* 28 |                BITMAP INDEX SINGLE VALUE | NODE_TYPE_INDEX           |       |       |       |            |          |
|  29 |               BITMAP CONVERSION TO ROWIDS|                           |  5627 | 90032 |       |    11   (0)| 00:00:01 |
|* 30 |                BITMAP INDEX SINGLE VALUE | NODE_PDUID_INDEX          |       |       |       |            |          |
|  31 |              BITMAP CONVERSION TO ROWIDS |                           |  5627 | 90032 |       |    37   (0)| 00:00:01 |
|* 32 |               BITMAP INDEX SINGLE VALUE  | NODE_DELETE_FLAG_INDEX    |       |       |       |            |          |
|  33 |             INDEX FAST FULL SCAN         | SYS_C0019687              |  5627 | 90032 |       |  1579   (1)| 00:00:19 |
|  34 |           TABLE ACCESS FULL              | FULL_REL_SPECNODE         | 26489 |   310K|       |    68   (0)| 00:00:01 |
|  35 |          INDEX FAST FULL SCAN            | SYS_C0019687              |   415K|  2436K|       |   344   (1)| 00:00:05 |
|* 36 |      VIEW                                | index$_join$_005          |    37M|   530M|       |   324K  (1)| 01:04:54 |
|* 37 |       HASH JOIN                          |                           |       |       |       |            |          |
|* 38 |        HASH JOIN                         |                           |       |       |       |            |          |
|  39 |         BITMAP CONVERSION TO ROWIDS      |                           |    37M|   530M|       |  1687   (1)| 00:00:21 |
|* 40 |          BITMAP INDEX SINGLE VALUE       | FULL_ITEMSINFO_41_INDEX_3 |       |       |       |            |          |
|  41 |         INDEX FAST FULL SCAN             | FULL_ITEMSINFO_41_INDEX_1 |    37M|   530M|       |   115K  (1)| 00:23:06 |
|  42 |        BITMAP CONVERSION TO ROWIDS       |                           |    37M|   530M|       | 12843   (1)| 00:02:35 |
|  43 |         BITMAP INDEX FULL SCAN           | FULL_ITEMSINFO_41_INDEX_2 |       |       |       |            |          |
------------------------------------------------------------------------------------------------------------------------------
 
Predicate Information (identified by operation id):
---------------------------------------------------
 
"   2 - access(""I"".""ATTRRID_""=""A"".""ID"")"
"   3 - filter(""A"".""ATTRR_DELETEFLAG""=0)"
"   4 - access(""N"".""NODE_ID""=""T"".""NODE_ID"")"
"   6 - access(""I"".""NODE_ID_""=""N"".""NODE_ID"")"
"  10 - filter(""NODE_PDUID""=41 AND ""NODE_TYPE""=2 AND ""NODE_DELETEFLAG""=0)"
  11 - access(ROWID=ROWID)
  12 - access(ROWID=ROWID)
  13 - access(ROWID=ROWID)
"  15 - access(""NODE_PDUID""=41)"
"  17 - access(""NODE_DELETEFLAG""=0)"
"  19 - access(""NODE_TYPE""=2)"
"  21 - access(""B"".""NODE_PARENTID""=""C"".""NODE_ID"")"
"  22 - access(""A"".""NODE_ID""=""B"".""NODE_ID"")"
"  23 - filter(""A"".""NODE_TYPE""=3 AND ""A"".""NODE_PDUID""=41 AND ""A"".""NODE_DELETEFLAG""=0)"
  24 - access(ROWID=ROWID)
  25 - access(ROWID=ROWID)
  26 - access(ROWID=ROWID)
"  28 - access(""A"".""NODE_TYPE""=3)"
"  30 - access(""A"".""NODE_PDUID""=41)"
"  32 - access(""A"".""NODE_DELETEFLAG""=0)"
"  36 - filter(""I"".""ISDELETED""=0)"
  37 - access(ROWID=ROWID)
  38 - access(ROWID=ROWID)
"  40 - access(""I"".""ISDELETED""=0)"

18 条回复 • 2020-06-17 21:34:01 +08:00

liprais

2020-06-16 15:26:09 +08:00

先贴执行计划

wysnylc

2020-06-16 15:28:13 +08:00

异步多次查询,查询结果使用 map filter 拼接

takemeaway

2020-06-16 15:28:58 +08:00

表倒是不大，，，就是连得有点多。
搞不懂这样的需求，干嘛这样设计表。

NewConn

2020-06-16 15:31:21 +08:00

```
------------------------------------------------------------------------------------------------------------------------------------------
| Id | Description | Owner | Name | Rows | Cost | Bytes | Remark |
------------------------------------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | | 1 | 7387.245 | | |
| 1 | AGGR | | | 1 | 7387.245 | | |
| 2 | HASH JOIN(R) | | | 340000 | 7387.245 | | |
| 3 | HASH JOIN(R) | | | 680000 | 1666.518 | | |
| 4 | HASH JOIN(R) | | | 3509333 | 1666.435 | | |
| 5 | TABLE ACCESS BY INDEX ONLY | FULLDB | FULL_ITEMSINFO_41 I | 34832500 | 0.083 | | |
| 6 | INDEX RANGE SCAN | FULLDB | FULL_ITEMSINFO_41_INDEX_3 | 34832500 | 0.083 | | |
| 7 | SUBSELECT | | | 82850 | 1666.353 | | |
| 8 | HASH UNION | | | 82850 | 1666.353 | | |
| 9 | TABLE ACCESS BY INDEX ONLY | FULLDB | FULL_SPECNODEINFO | 82522 | 1420.588 | | |
| 10 | INDEX RANGE SCAN | FULLDB | FSNI_IDX_1 | 82522 | 1420.588 | | |
| 11 | NESTED LOOPS | | | 328 | 245.765 | | |
| 12 | NESTED LOOPS | | | 329 | 225.038 | | |
| 13 | TABLE ACCESS BY INDEX ONLY | FULLDB | FULL_SPECNODEINFO A | 5072 | 100.588 | | |
| 14 | INDEX RANGE SCAN | FULLDB | FSNI_IDX_1 | 5072 | 100.588 | | |
| 15 | TABLE ACCESS BY INDEX ONLY | FULLDB | FULL_REL_SPECNODE B | 27052 | 124.450 | | |
| 16 | INDEX RANGE SCAN | FULLDB | IX_NODEID_PARENTID_FRS | 27052 | 124.450 | | |
| 17 | TABLE ACCESS BY INDEX ONLY | FULLDB | FULL_SPECNODEINFO C | 415200 | 20.727 | | |
| 18 | INDEX UNIQUE SCAN | FULLDB | PK_FULL_SPECNODEINFO | 415200 | 20.727 | | |
| 19 | TABLE ACCESS BY INDEX ONLY | FULLDB | FULL_ATTRIBUTE_RELATION A | 118912 | 0.083 | | |
| 20 | INDEX RANGE SCAN | FULLDB | FULL_ATTRIBUTE_RELATION_INDEX_1 | 118912 | 0.083 | | |
| 21 | TABLE ACCESS BY INDEX ONLY | FULLDB | FULL_SPECNODEINFO T | 415200 | 5720.727 | | |
| 22 | INDEX FAST FULL SCAN | FULLDB | PK_FULL_SPECNODEINFO | 415200 | 5720.727 | | |
------------------------------------------------------------------------------------------------------------------------------------------
```

执行计划是这个，问题描述里面是 Oracle 的执行计划。谢谢大家

hooopo

2020-06-16 15:42:59 +08:00 via Android

pg 吗什么版本 cte 改掉试试

liprais

2020-06-16 15:55:48 +08:00

@hooopo gauss 是华为魔改的 pgsql 的 mpp 数据库

fangcan

2020-06-16 16:11:34 +08:00

为什么 3 千多万的数据才 37MB ? 我们的一般都要好多 GB 了

Still4

2020-06-16 17:24:36 +08:00

最外层 id=2 这里，似乎执行了扫全表，rows=22M

应该是建立了临时表 E，导致索引失效？

hooopo

2020-06-16 17:31:22 +08:00 via Android

@liprais 为什么这东西有人敢用

BadAngel

2020-06-16 17:38:09 +08:00

@hooopo 也许就是菊厂的兄弟，有时候没办法，为了不被勒脖子，总要牺牲点啥

zhangysh1995

2020-06-16 18:43:02 +08:00

WITH E AS (
SELECT ID
FROM A
WHERE type = 2
AND DELETE_FLAG = 0
AND uid = 41
UNION
SELECT a.ID
FROM A a, B b, A c
WHERE a.type = 3
AND a.DELETE_FLAG = 0
AND a.uid = 41
AND a.ID = b.ID
AND b.parent_id = c.ID
)

这里面为什么不能先把 a.type in (2,3) and a.delete_flag =0 and a.uid = 41 先选择出来然后再别的操作？这一句的 selectivity 有多少？有多少符合条件的？

lenqu

2020-06-16 19:01:41 +08:00

我觉得联合操作有点多
想问你们分表的时候没有具体分表逻辑，比如某一 key 的范围划分？

CRVV

2020-06-16 19:06:03 +08:00

> D 表的（ DELETE_FLAG,ID,aid ）也加了联合索引

D.ID 在这个查询里就没被用到

这个 SQL 里面，E 就是 A，t 也是 A，最外层写了一个 A INNER JOIN A ON A.id = A.id ，确定是这样的么？

这个 SQL 写得有点奇怪，比如还有一个问题是 cte 里面不用写 UNION，一次 SELECT 就可以了。
先把它整理清楚了再优化吧。

CRVV

2020-06-16 19:10:38 +08:00

@zhangysh1995
如果他发的 SQL 是对的，那个 cte 等价于

SELECT id
FROM A
LEFT JOIN B ON A. id = B. id
LEFT JOIN A AS parent ON B.parent_id = parent. id
WHERE A. uid = 41 AND (A.type = 2 OR (A.type = 3 AND parent. id IS NOT NULL))

然后这个 cte 也不用写 cte，直接和下面的 JOIN 写在一起就好了

另外这个 EXPLAIN 明显不是 PostgreSQL，我也没看出来这是什么数据库

```
创建新回复过程中遇到一些问题：
请不要在每一个回复中都包括外链，这看起来像是在 spamming
```
@Livid 这个外链的检查也太敏感了吧

liprais

2020-06-16 19:18:37 +08:00 via iPhone

@hooopo 实话实话，高斯做了十年还是有点东西的，据我所知测试也很严谨，但是内斗太严重，真正有多少东西能进入最终的产品就不得而知了

NewConn

2020-06-17 10:34:12 +08:00

@CRVV 感谢老哥。说实话我刚来不久，也不是非常清楚业务逻辑，所以最近遇到的这些 SQl 的 union 、left join，以及几千行的存储过程，我也不知道怎么从业务逻辑入手优化，只能单拎出来语句进行优化。老哥给的建议我详细看看，非常感谢

NewConn

2020-06-17 10:35:36 +08:00

@zhangysh1995 感谢老哥，我按您的思路去梳理一下这个 SQL 逻辑

zhangysh1995

2020-06-17 21:34:01 +08:00

@NewConn 不好意思我是妹子。