“不存在”的 Oracle 性能

Question

我正在将一些 SQL 服务器过程移植到 Oracle 并发现一个有趣的情况，即 Oracle SQL 语句比使用游标的相同逻辑慢得多。

经过调查，我认为“不存在”可能存在特殊问题（也许？）。

在这里，我将 100k 潜在客户放入 TMP_TXN，然后将它们用作过滤器，以从没有交易的 Payments 中提取记录（请参阅下面的 SQL 构造）。

  INSERT INTO tmp_txn ....
  SELECT ....
  FROM   txn, 
         customers
  WHERE  txn.customer_id = customers.customer_id
         AND customers.customer_status LIKE 'A%'
         AND txn.txn_date BETWEEN start_date AND end_date;

然后插入到 tmp_leads 中，它们在付款中，但不在 TMP_TXN 中。

   INSERT INTO tmp_leads ....
    SELECT ....
    FROM  payments eap, customers
    WHERE eap.customer_id = customers.customer_id
           AND customers.customer_status LIKE 'A%'
           AND NOT EXISTS (SELECT TMP_TXN.CUSTOMER_ID
                           FROM   TMP_TXN
                           WHERE  tmp_txn.customer_id = eap.customer_id
                                  AND  ....;

解释计划是：

Plan hash value: 67643415

-----------------------------------------------------------------------------------------------------------
| Id  | Operation                        | Name                   | Rows  | Bytes | Cost (%CPU)| Time     |
-----------------------------------------------------------------------------------------------------------
|   0 | INSERT STATEMENT                 |                        |   665 |   138K| 15450   (1)| 00:03:06 |
|   1 |  LOAD TABLE CONVENTIONAL         | TMP_LEADS              |       |       |            |          |
|*  2 |   FILTER                         |                        |       |       |            |          |
|*  3 |    HASH JOIN                     |                        |   665 |   138K| 14785   (1)| 00:02:58 |
|*  4 |     TABLE ACCESS FULL            | CUSTOMER_TYPES         |     6 |    36 |     3   (0)| 00:00:01 |
|*  5 |     HASH JOIN                    |                        |   726 |   146K| 14781   (1)| 00:02:58 |
|*  6 |      TABLE ACCESS FULL           | EDM_SEGMENTS_EVENTS    |    23 |   414 |     5   (0)| 00:00:01 |
|   7 |      NESTED LOOPS                |                        |       |       |            |          |
|   8 |       NESTED LOOPS               |                        |  1297 |   239K| 14776   (1)| 00:02:58 |
|*  9 |        TABLE ACCESS FULL         | EDM_AGREEMENT_PAYMENTS |  1297 |   158K| 12180   (1)| 00:02:27 |
|* 10 |        INDEX UNIQUE SCAN         | PK_CUSTOMERS           |     1 |       |     1   (0)| 00:00:01 |
|* 11 |       TABLE ACCESS BY INDEX ROWID| CUSTOMERS              |     1 |    64 |     2   (0)| 00:00:01 |
|* 12 |    TABLE ACCESS BY INDEX ROWID   | TMP_TXN                |     1 |    81 |     2   (0)| 00:00:01 |
|* 13 |     INDEX RANGE SCAN             | IX_TMP_TXN_TXN_CODE    |     1 |       |     2   (0)| 00:00:01 |
-----------------------------------------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

   2 - filter( NOT EXISTS (SELECT 0 FROM "TMP_TXN" "TMP_TXN" WHERE "TMP_TXN"."TXN_CODE"=:B1 AND 
              "TMP_TXN"."CUSTOMER_ID"=:B2 AND "TMP_TXN"."AGREEMENT_ID"=:B3 AND ("TMP_TXN"."AMOUNT"<0 AND 
              "TMP_TXN"."AMOUNT">=:B4*1.1 AND "TMP_TXN"."AMOUNT"<=:B5*.9 OR "TMP_TXN"."AMOUNT">0 AND 
              "TMP_TXN"."AMOUNT">=:B6*0.9 AND "TMP_TXN"."AMOUNT"<=:B7*1.1)))
   3 - access("CUSTOMERS"."CUSTOMER_TYPE"="CUSTOMER_TYPES"."CUSTOMER_TYPE")
   4 - filter("CUSTOMER_TYPES"."ACTIVE"=U'1')
   5 - access("CUSTOMERS"."CUSTOMER_SEGMENT"="EDM_SEGMENTS_EVENTS"."SEGMENT")
   6 - filter("EDM_SEGMENTS_EVENTS"."EVENT_ID"=607 AND "EDM_SEGMENTS_EVENTS"."ACTIVE"=U'1')
   9 - filter(ROUND("EAP"."PMNT_DAY",0)>=19 AND ROUND("EAP"."PMNT_DAY",0)<=31 AND 
              "EAP"."PERIODICITY"=U'M' AND "EAP"."EVENT_ID"=607)
  10 - access("EAP"."CUSTOMER_ID"="CUSTOMERS"."CUSTOMER_ID")
  11 - filter("CUSTOMERS"."CUSTOMER_STATUS" LIKE U'A%')
  12 - filter("TMP_TXN"."CUSTOMER_ID"=:B1 AND "TMP_TXN"."AGREEMENT_ID"=:B2 AND 
              ("TMP_TXN"."AMOUNT"<0 AND "TMP_TXN"."AMOUNT">=:B3*1.1 AND "TMP_TXN"."AMOUNT"<=:B4*.9 OR 
              "TMP_TXN"."AMOUNT">0 AND "TMP_TXN"."AMOUNT">=:B5*0.9 AND "TMP_TXN"."AMOUNT"<=:B6*1.1))
  13 - access("TMP_TXN"."TXN_CODE"=:B1)

tmp_txn(customer_id)上有一个索引，表里有大约100k条记录。 Oracle 有 20gb SGA 和 20gb PGA，所以应该很容易缓存。

Resource plan screenshot 在这里你可以看到脚本正在运行，但没有使用任何资源（数据读取

（可能的）问题似乎在 NOT EXISTS 中，因为此选择需要 >1000 秒，几乎无法访问数据表（资源监视器）。 Stats view in OM showing 898 seconds and 100% cpu

我在 Oracle 中做了一些愚蠢的事情吗？这在 SQL Server 中运行良好（而且速度很快）。

Answer 1

如果内部查询结果集很大，应该使用NOT EXISTS。请尝试以下方法

INSERT INTO tmp_leads ....
SELECT ....
FROM  payments eap, customers
WHERE eap.customer_id = customers.customer_id
       AND customers.customer_status LIKE 'A%'
       AND customers.customer_id NOT IN (SELECT TMP_TXN.CUSTOMER_ID
                       FROM   TMP_TXN
                       WHERE  
                       --tmp_txn.customer_id = eap.customer_id
                              --AND  
                          ....;

Answer 2

尝试使用左连接而不是创建和加载 temptxn 然后加入

Insert into tmpleads
Select ...
From payments eap 
Inner join customers cust
    On cust.customerid = eap.customerid   
LEFT JOIN temptxn txn
    On txn.customerid = eap.customerid
Where txn.customerid IS NULL