sql微调答案 - 爱码网

【问题标题】：sql fine tuningsql微调
【发布时间】：2018-05-21 10:38:10
【问题描述】：

在一个表中，我需要根据列将 2 条记录合并为一条记录（在这种情况下，列 1.WORK_ORDER_NUM、2.ESN 3.PLANT 4. REMD_PART_NUM 5.REMD_PART_SERIAL)。如果这 5 列相等，那么我必须在另一列中使用 Aggregate 函数使其成为一条记录。[在在这种情况下，列是 LLP_TRACKD_PART_IND、REMD_PART_TSN 和 REMD_PART_CSN]。

这是我尝试过的：

SELECT decode (PLANT ,'ECL','CELMA','EDS','CELMA',PLANT)PLANT,
 COUNT(*) RECORD_COUNT,
 COUNT(DISTINCT OFF.REMD_PART_NUM) REMD_PART_NUM_COUNT,
 COUNT(DISTINCT OFF.REMD_PART_SERIAL) REMD_PART_SER_NUM_COUNT,
  COUNT(DECODE(LLP_TRACKD_PART_IND,'LL',LLP_TRACKD_PART_IND,NULL)) LL_COUNT,
 COUNT(DECODE(LLP_TRACKD_PART_IND,'LR',LLP_TRACKD_PART_IND,NULL)) LR_COUNT,
 COUNT(DECODE(LLP_TRACKD_PART_IND,'TR',LLP_TRACKD_PART_IND,NULL)) TR_COUNT,
 SUM(OFF.REMD_PART_QTY) TOTAL_REMD_PART_QTY, 
  SUM(decode(LLP_TRACKD_PART_IND,null,0,
     CASE
       WHEN REGEXP_LIKE(REMD_PART_TSN, '^-?\d+(\.\d+)?$')
            THEN CAST(REMD_PART_TSN AS NUMBER)
        ELSE 0
    END
  )) TOTAL_TSN,
 SUM(decode(LLP_TRACKD_PART_IND,null,0,
    CASE
        WHEN REGEXP_LIKE(REMD_PART_CSN, '^-?\d+(\.\d+)?$')
            THEN CAST(REMD_PART_CSN AS NUMBER)
        ELSE 0
    END
 )) TOTAL_CSN
FROM (with t as ( SELECT distinct PLANT,
WORK_ORDER_NUM,ESN,REMD_PART_NUM,REMD_PART_SERIAL,REMD_PART_IIN,
LLP_TRACKD_PART_IND,
REMD_PART_QTY,REMD_PART_TSN,REMD_PART_CSN,REMD_PART_TSO,REMD_PART_CSO
,REMD_PART_TSC,REMD_PART_CSC,REMD_CYCLE_REMAIN          
FROM <TABLE1> 
WHERE 
REMD_PART_NUM is not null 
)
select DISTINCT PLANT,WORK_ORDER_NUM,ESN,REMD_PART_NUM,REMD_PART_SERIAL 
,REMD_PART_IIN
 ,(select max(LLP_TRACKD_PART_IND) from t bb  where aa.PLANT=bb.PLANT and 
aa.WORK_ORDER_NUM=bb.WORK_ORDER_NUM 
 and aa.ESN=bb.ESN 
  and aa.REMD_PART_NUM=bb.REMD_PART_NUM
  and aa.REMD_PART_SERIAL=bb.REMD_PART_SERIAL) LLP_TRACKD_PART_IND
 ,REMD_PART_QTY
 ,(select max(REMD_PART_TSN) from t bb  where aa.PLANT=bb.PLANT and 
aa.WORK_ORDER_NUM=bb.WORK_ORDER_NUM 
 and aa.ESN=bb.ESN 
 and aa.REMD_PART_NUM=bb.REMD_PART_NUM
 and aa.REMD_PART_SERIAL=bb.REMD_PART_SERIAL) REMD_PART_TSN
 ,
(select max(REMD_PART_CSN) from t bb  where aa.PLANT=bb.PLANT and             
aa.WORK_ORDER_NUM=bb.WORK_ORDER_NUM 
  and aa.ESN=bb.ESN 
  and aa.REMD_PART_NUM=bb.REMD_PART_NUM
  and aa.REMD_PART_SERIAL=bb.REMD_PART_SERIAL) REMD_PART_CSN
  from t aa) OFF
WHERE 
REMD_PART_NUM is not null
GROUP BY decode (PLANT ,'ECL','CELMA','EDS','CELMA',PLANT)

大约需要 8 小时才能完成。有没有其他方法可以更快地完成它。 It took 8 hours to complete

【问题讨论】：

微调？您使用的是 MySQL 还是 Oracle？（不要标记未涉及的产品。）
@jarlh 这是甲骨文。
它是甲骨文。 @jarlh
查看执行计划以确定 oracle 将所有时间用于解决问题的地方。 ps;就查询而言，它看起来很糟糕——在 select 和 from 之间有嵌套子查询的查询总是为我敲响警钟。我认为您最好不要要求我们修复您编写的查询，而是向我们提供示例输入和输出并寻求帮助以另一种方式编写新查询来解决问题

标签： sql oracle with-statement sql-tuning

【解决方案1】：

从更好的查询格式开始，它可以让您轻松理解代码并注意重复的模式：

SELECT decode (PLANT ,'ECL','CELMA','EDS','CELMA',PLANT)PLANT,
    COUNT(*) RECORD_COUNT,
    COUNT(DISTINCT OFF.REMD_PART_NUM) REMD_PART_NUM_COUNT,
    COUNT(DISTINCT OFF.REMD_PART_SERIAL) REMD_PART_SER_NUM_COUNT,
    COUNT(DECODE(LLP_TRACKD_PART_IND,'LL',LLP_TRACKD_PART_IND,NULL)) LL_COUNT,
    COUNT(DECODE(LLP_TRACKD_PART_IND,'LR',LLP_TRACKD_PART_IND,NULL)) LR_COUNT,
    COUNT(DECODE(LLP_TRACKD_PART_IND,'TR',LLP_TRACKD_PART_IND,NULL)) TR_COUNT,
    SUM(OFF.REMD_PART_QTY) TOTAL_REMD_PART_QTY, 
    SUM(decode(LLP_TRACKD_PART_IND,null,0,
            CASE
                WHEN REGEXP_LIKE(REMD_PART_TSN, '^-?\d+(\.\d+)?$')
                THEN CAST(REMD_PART_TSN AS NUMBER)
                ELSE 0
            END
        )) TOTAL_TSN,
    SUM(decode(LLP_TRACKD_PART_IND,null,0,
            CASE
                WHEN REGEXP_LIKE(REMD_PART_CSN, '^-?\d+(\.\d+)?$')
                THEN CAST(REMD_PART_CSN AS NUMBER)
                ELSE 0
            END
        )) TOTAL_CSN
FROM (
    with t as ( 
        SELECT distinct PLANT, WORK_ORDER_NUM,ESN,REMD_PART_NUM,REMD_PART_SERIAL,REMD_PART_IIN,
                LLP_TRACKD_PART_IND,
                REMD_PART_QTY,REMD_PART_TSN,REMD_PART_CSN,REMD_PART_TSO,REMD_PART_CSO
                ,REMD_PART_TSC,REMD_PART_CSC,REMD_CYCLE_REMAIN          
        FROM <TABLE1> 
        WHERE REMD_PART_NUM is not null 
    )
    select DISTINCT PLANT,WORK_ORDER_NUM,ESN,REMD_PART_NUM,REMD_PART_SERIAL 
            ,REMD_PART_IIN
            ,(  select max(LLP_TRACKD_PART_IND) 
                from t bb  
                where aa.PLANT=bb.PLANT 
                    and aa.WORK_ORDER_NUM=bb.WORK_ORDER_NUM 
                    and aa.ESN=bb.ESN 
                    and aa.REMD_PART_NUM=bb.REMD_PART_NUM
                    and aa.REMD_PART_SERIAL=bb.REMD_PART_SERIAL
            ) LLP_TRACKD_PART_IND
            ,REMD_PART_QTY
            ,(  select max(REMD_PART_TSN) from t bb  
                where aa.PLANT=bb.PLANT 
                    and aa.WORK_ORDER_NUM=bb.WORK_ORDER_NUM 
                    and aa.ESN=bb.ESN 
                    and aa.REMD_PART_NUM=bb.REMD_PART_NUM
                    and aa.REMD_PART_SERIAL=bb.REMD_PART_SERIAL
            ) REMD_PART_TSN
            ,
            (   select max(REMD_PART_CSN) from t bb  
                where aa.PLANT=bb.PLANT 
                    and aa.WORK_ORDER_NUM=bb.WORK_ORDER_NUM 
                    and aa.ESN=bb.ESN 
                    and aa.REMD_PART_NUM=bb.REMD_PART_NUM
                    and aa.REMD_PART_SERIAL=bb.REMD_PART_SERIAL
            ) REMD_PART_CSN
    from t aa
) OFF
WHERE REMD_PART_NUM is not null
GROUP BY decode (PLANT ,'ECL','CELMA','EDS','CELMA',PLANT)
;

您将看到以下模式重复了 3 次（3 个几乎相同的子查询）：

select max( some_field ) 
from t bb  
where aa.PLANT=bb.PLANT 
  and aa.WORK_ORDER_NUM=bb.WORK_ORDER_NUM 
  and aa.ESN=bb.ESN 
  and aa.REMD_PART_NUM=bb.REMD_PART_NUM
  and aa.REMD_PART_SERIAL=bb.REMD_PART_SERIAL

您还会很容易地看到，此查询中有 3 个非常昂贵的排序操作 - 最里面的子查询中的一个 DISTINCT，另一个子查询中的另一个 DISTINCT，最后是顶级查询中的 GROUP BY 操作（一种 DISTINCT） .

仅查看您的查询，可以通过这种方式使用分析函数轻松消除一种排序 (DISTINCT)：

SELECT * FROM (
    SELECT PLANT,WORK_ORDER_NUM,ESN,REMD_PART_NUM,REMD_PART_SERIAL,
           REMD_PART_IIN, REMD_PART_QTY,
           max( LLP_TRACKD_PART_IND ) over 
              (partition by PLANT, WORK_ORDER_NUM, ESN, REMD_PART_NUM, REMD_PART_SERIAL) 
           as LLP_TRACKD_PART_IND,
           max( REMD_PART_TSN )       over 
              (partition by PLANT, WORK_ORDER_NUM, ESN, REMD_PART_NUM, REMD_PART_SERIAL) 
           as REMD_PART_TSN,
           max( REMD_PART_CSN )       over 
              (partition by PLANT, WORK_ORDER_NUM, ESN, REMD_PART_NUM, REMD_PART_SERIAL) 
           as REMD_PART_CSN,
           row_number() over 
              (partition by PLANT, WORK_ORDER_NUM, ESN, REMD_PART_NUM, REMD_PART_SERIAL, REMD_PART_IIN, REMD_PART_QTY 
                   order by PLANT) as Rn
    FROM TABLE1
    WHERE REMD_PART_NUM is not null 
)
WHERE rn = 1

所以最终的查询可能是：

SELECT decode (PLANT ,'ECL','CELMA','EDS','CELMA',PLANT)PLANT,
    COUNT(*) RECORD_COUNT,
    COUNT(DISTINCT OFF.REMD_PART_NUM) REMD_PART_NUM_COUNT,
    COUNT(DISTINCT OFF.REMD_PART_SERIAL) REMD_PART_SER_NUM_COUNT,
    COUNT(DECODE(LLP_TRACKD_PART_IND,'LL',LLP_TRACKD_PART_IND,NULL)) LL_COUNT,
    COUNT(DECODE(LLP_TRACKD_PART_IND,'LR',LLP_TRACKD_PART_IND,NULL)) LR_COUNT,
    COUNT(DECODE(LLP_TRACKD_PART_IND,'TR',LLP_TRACKD_PART_IND,NULL)) TR_COUNT,
    SUM(OFF.REMD_PART_QTY) TOTAL_REMD_PART_QTY, 
    SUM(decode(LLP_TRACKD_PART_IND,null,0,
            CASE
                WHEN REGEXP_LIKE(REMD_PART_TSN, '^-?\d+(\.\d+)?$')
                THEN CAST(REMD_PART_TSN AS NUMBER)
                ELSE 0
            END
        )) TOTAL_TSN,
    SUM(decode(LLP_TRACKD_PART_IND,null,0,
            CASE
                WHEN REGEXP_LIKE(REMD_PART_CSN, '^-?\d+(\.\d+)?$')
                THEN CAST(REMD_PART_CSN AS NUMBER)
                ELSE 0
            END
        )) TOTAL_CSN
FROM (
    SELECT PLANT,WORK_ORDER_NUM,ESN,REMD_PART_NUM,REMD_PART_SERIAL,REMD_PART_IIN, REMD_PART_QTY,
           max( LLP_TRACKD_PART_IND ) over (partition by PLANT, WORK_ORDER_NUM, ESN, REMD_PART_NUM, REMD_PART_SERIAL) as LLP_TRACKD_PART_IND,
           max( REMD_PART_TSN )       over (partition by PLANT, WORK_ORDER_NUM, ESN, REMD_PART_NUM, REMD_PART_SERIAL) as REMD_PART_TSN,
           max( REMD_PART_CSN )       over (partition by PLANT, WORK_ORDER_NUM, ESN, REMD_PART_NUM, REMD_PART_SERIAL) as REMD_PART_CSN,
           row_number()               over (partition by PLANT, WORK_ORDER_NUM, ESN, REMD_PART_NUM, REMD_PART_SERIAL, REMD_PART_IIN, REMD_PART_QTY 
                                            order by PLANT) as Rn
    FROM TABLE1
    WHERE REMD_PART_NUM is not null 
)
WHERE rn = 1
GROUP BY decode (PLANT ,'ECL','CELMA','EDS','CELMA',PLANT)

我觉得这个查询可以进一步优化，但需要深入了解表的结构和业务需求知识。

不过，仍有一些可能的微优化。
这种模式：

SUM(decode(LLP_TRACKD_PART_IND,null,0,
        CASE
            WHEN REGEXP_LIKE(REMD_PART_TSN, '^-?\d+(\.\d+)?$')
            THEN CAST(REMD_PART_TSN AS NUMBER)
            ELSE 0
        END
    )) TOTAL_TSN,

可以用这个代替：

coalesce( 
    SUM( 
        CASE
            WHEN REGEXP_LIKE(REMD_PART_TSN, '^-?\d+(\.\d+)?$')
            THEN CAST(REMD_PART_TSN AS NUMBER)
            ELSE 0
        END
    ), 0
)

由于 SUM 忽略空值，因此检查每条记录的 NULL 值是一种浪费。对于少量记录（

【讨论】：