中兴GoldenDB（MYSQL）营运商SQL优化案例（超复杂SQL）

陆老师好久没找我，今天他在营运商遇到2条性能慢的SQL，说找了好多专家老手看了都优化不了，然后就找到我打算给我玩玩。

第一次接触营运商行业的SQL，妈呦还真的复杂

 SQL1:

SELECT WORK_ORDER_ID        as workOrderId,
       WORK_ITEM_ID         as workItemId,
       TRACE_ID             as traceID,
       TOTAL_DATE           as totalDate,
       SEND_TIME            as sendTime,
       SEND_STATUS          as sendStatus,
       REQUEST_TIME         as requestTime,
       REMARKS              as remarks,
       PROVINCE             as province,
       PRIORITY             as priority,
       nvl(PHONE_NO, '0')   as phoneNo,
       PARENTCALLID         as parentcallid,
       OTHER_CHAR2          as otherChar2,
       OTHER_CHAR1          as otherChar1,
       OTHER_CHAR           as otherChar,
       OSS_ORDER_ID         as ossOrderId,
       nvl(ORDER_TYPE, '0') as orderType,
       ORDER_CREATE_TIME    as orderCreateTime,
       OP_TIME              as opTime,
       OP_CODE              as opCode,
       nvl(MOD_VALUE, '0')  as modValue,
       nvl(LOGIN_NO, '0')   as loginNo,
       IOM_RECEIVE_TIME     as iomReceiveTime,
       IOM_COMPLETE_TIME    as iomCompleteTime,
       nvl(IMSI_NO, '0')    as imsiNo,
       ID_NO                as idNo,
       HLR_CODE             as hlrCode,
       GROUP_ID             as groupId,
       CRM_ORDER_ID         as crmOrderId,
       COMMAND_ID           as commandId,
       COMMAND_COUNT        as commandCount,
       COMMAND_CODE         as commandCode,
       CMS_RECEIVE_TIME     as cmsReceiveTime,
       CMS_COMPLETE_TIME    as cmsCompleteTime,
       ACK_TIME             as ackTime,
       ACK_INFO             as ackInfo,
       ACK_CODE             as ackCode
FROM px_lu
WHERE send_status = '0'
  and mod(to_number(trim(phone_no)), 1) = 0
ORDER BY request_time, command_id limit 500;

执行时间：500 rows in set (0.853 sec)

这条分页SQL在 ORACLE 上 30ms 就可以跑出来，在 GoldenDB 要跑近 1s 才能出结果。

 SQL1 索引优化:

create index idx_1_2_3 on px_lu (send_status,request_time, command_id );

500 rows in set (0.013 sec)

索引加上以后 0.013 s 就可以跑出结果，非常简单。

 SQL2:

select new_busi_code,
       new_status_cd,
       table_name,
       new_group_id,
       count(1) count
from (select (select nvl((select x.rule_value
                          from vccccx t,
                               frtgh x
                          where t.rule_code = '1000000048'
                            and t.busi_rule_code = x.busi_rule_code
                            and t.busi_code = a.busi_code), decode(busi_level, '2', par_busi_code, BUSI_CODE))
              from sffsss
              where busi_code = a.busi_code)                                                   as new_busi_code,
             (nvl((select x.rule_value
                   from vccccx t,
                        frtgh x
                   where t.rule_code = '1000000046'
                     and t.busi_rule_code = x.busi_rule_code
                     and t.busi_code = a.busi_code), 's'))                                        new_status_cd,
             (select op_note as table_name
              from swbbbbbb
              where maindata_code = 'RS-NO-0001'
                and maindata_value = (nvl((select x.rule_value
                                           from vccccx t,
                                                frtgh x
                                           where t.rule_code = '1000000046'
                                             and t.busi_rule_code = x.busi_rule_code
                                             and t.busi_code = a.busi_code), 's')))               table_name,
             nvl((select parent_group_id
                  from dsdsd
                  where group_id = a.group_id
                    and parent_level = nvl((select x.rule_value
                                            from vccccx t,
                                                 frtgh x
                                            where t.rule_code = '1000000047'
                                              and t.busi_rule_code = x.busi_rule_code
                                              and t.busi_code = a.busi_code), 3)), a.group_id) as new_group_id
      from dsdsd b,
           rsrsrs a
      WHERE b.group_id = a.group_id
        AND a.status_cd in ('2', 'y')
        AND b.parent_group_id = '14'
        and a.region_code = '2201'
        and a.tenant_id = '22'
        and b.tenant_id = '22'
        AND a.rec_Time <= to_date(TO_CHAR(SYSDATE, 'YYYYMMDD') || ' 00:00:00', 'yyyy-mm-dd hh24:mi:ss'))
where new_busi_code is not null
  and new_status_cd is not null
  and new_group_id is not null
group by new_busi_code, new_status_cd, table_name, new_group_id;

10 rows in set (9 min 46.240 sec)

这条SQL的表连接关系是真的超级复杂，来在ORACLE上跑30s可以出结果，在GoldenDB 却快跑10分钟才能出结果。

其实主要就是慢在标量子查询，只要改写成左连接性能就可以提升上来，里面的逻辑真的是太复杂了，我花了不少时间搞清关系，营运商SQL的复杂程度总算见到了。

 SQL2等价改写:

select new_busi_code,
       new_status_cd,
       table_name,
       new_group_id,
       count(1) count
from (
　　SELECT COALESCE(rv1.rule_value, DECODE(cb.busi_level, '2', cb.par_busi_code, cb.BUSI_CODE)) AS new_busi_code,
             COALESCE(rv2.rule_value, 's')                                                        AS new_status_cd,
             mv.op_note                                                                           AS table_name,
             COALESCE(cgr.parent_group_id, a.group_id)                                            AS new_group_id
      FROM rsrsrs a
               LEFT JOIN dsdsd b ON b.group_id = a.group_id
               LEFT JOIN sffsss cb ON cb.busi_code = a.busi_code
               LEFT JOIN vccccx br1 ON br1.busi_code = a.busi_code AND br1.rule_code = '1000000048'
               LEFT JOIN frtgh rv1 ON br1.busi_rule_code = rv1.busi_rule_code
               LEFT JOIN vccccx br2 ON br2.busi_code = a.busi_code AND br2.rule_code = '1000000046'
               LEFT JOIN frtgh rv2 ON br2.busi_rule_code = rv2.busi_rule_code
               LEFT JOIN swbbbbbb mv
                         ON mv.maindata_value = COALESCE(rv2.rule_value, 's') AND mv.maindata_code = 'RS-NO-0001'
               LEFT JOIN vccccx br3 ON br3.busi_code = a.busi_code AND br3.rule_code = '1000000047'
               LEFT JOIN frtgh rv3 ON br3.busi_rule_code = rv3.busi_rule_code
               LEFT JOIN dsdsd cgr ON cgr.group_id = a.group_id AND cgr.parent_level = COALESCE(rv3.rule_value, 3)
      WHERE a.status_cd IN ('2', 'y')
        AND b.parent_group_id = '14'
        AND a.region_code = '2201'
        AND a.tenant_id = '22'
        AND b.tenant_id = '22'
        AND a.rec_Time <= to_date(TO_CHAR(SYSDATE, 'YYYYMMDD') || ' 00:00:00', 'yyyy-mm-dd hh24:mi:ss')
        AND COALESCE(rv1.rule_value, DECODE(cb.busi_level, '2', cb.par_busi_code, cb.BUSI_CODE)) IS NOT NULL
        AND COALESCE(rv2.rule_value, 's') IS NOT NULL
        AND COALESCE(cgr.parent_group_id, a.group_id) IS NOT NULL
)
where new_busi_code is not null
  and new_status_cd is not null
  and new_group_id is not null
group by new_busi_code, new_status_cd, table_name, new_group_id;

10 rows in set (3 min 12.370 sec)

改写完以后运行速度从9分钟可以降到3分钟左右就能出结果，依然是很慢，还有继续优化空间。

 SQL2创建索引继续优化:

create index idx_1_2_3          ON rsrsrs(group_id,busi_code, status_cd, region_code, tenant_id, rec_Time);
create index idx_4_5_6          ON dsdsd(group_id, parent_group_id, tenant_id);
create index idx_7_8_9          ON sffsss(busi_code);
create index idx_11_12_13     ON vccccx(busi_code, rule_code, busi_rule_code);
create index idx_14_15_16     ON frtgh(busi_rule_code);
create index idx_17_18_19     ON swbbbbbb(maindata_value, maindata_code);

10 rows in set (48.876 sec)

最终通过等价改写 + 索引优化手段，SQL2从9分钟左右的执行时间降到48秒就可以出结果，基本没有继续优化的空间了。