出处:http://blog.itpub.net/26736162/viewspace-1244044
原sql语句:
INSERT INTO RISKREPT.BASE_FMLG (BATCH_DATE, DATE_STAMP_ST, TIME_STAMP_ST, ORG, ACCT, CARD_NBR, CARD_SEQ, MER_ORG, MER_NBR, REQUEST_TYPE_ID, LOGO,SYSTEM_ACTION, FINAL_ACTION, ACTION_REASON, REVERSAL_REASON, AVAIL_CR, CASH_AVAIL_CR, ACCT_CURR_BAL, ACCT_CURR_BAL_CASH, B007_GMT_DATE_TIME,B018_MER_TYPE, B019_CNTRY_CODE, B032_ACQ_ID, B033_FWD_ID, AUTH_CODE, B039_RESP_CODE, B041_CRD_ACCPT_TERM, B042_TERMTYPE2_MER_ID,B043_CRD_ACCPT_NAM, B043_CRD_ACCPT_CITY, B043_CRD_ACCPT_ST_CTRY, MESSAGE_TYPE_ID, RECORD_TYPE, SALES_AMT_RMB, TOTAL_SALES_AMT, B049_CURR_CODE,BLING_AMT, POS_ENTRY_MODE, PIN_ENTRY_MODE, TRADE_INTERNET, SALES_CTRY, SALES_CTRY_NAME, SALES_CITY, SALES_CITY_NAME, SALES_LINK, MER_CODE, MER_MCC,MER_NAM, ACQ_NAME, REVCODE, AUTH_TYPE, REF_NBR, VI_B011_SYS_AUDT_TRCE, CARD_TYPE, STAGE_TYPE, STAGE_NUM, OVERSEA_FLAG, RISK_FCD, RISK_LCD) SELECTBATCH_DATE, DATE_STAMP_ST, TIME_STAMP_ST, ORG, ACCT, CARD_NBR, CARD_SEQ, MER_ORG, MER_NBR, REQUEST_TYPE_ID, LOGO, SYSTEM_ACTION, FINAL_ACTION,ACTION_REASON, REVERSAL_REASON, AVAIL_CR, CASH_AVAIL_CR, ACCT_CURR_BAL, ACCT_CURR_BAL_CASH, B007_GMT_DATE_TIME, B018_MER_TYPE, B019_CNTRY_CODE,B032_ACQ_ID, B033_FWD_ID, AUTH_CODE, B039_RESP_CODE, B041_CRD_ACCPT_TERM, B042_TERMTYPE2_MER_ID, B043_CRD_ACCPT_NAM, B043_CRD_ACCPT_CITY,B043_CRD_ACCPT_ST_CTRY, MESSAGE_TYPE_ID, RECORD_TYPE, SALES_AMT_RMB, TOTAL_SALES_AMT, B049_CURR_CODE, BLING_AMT, POS_ENTRY_MODE, PIN_ENTRY_MODE,TRADE_INTERNET, SALES_CTRY, SALES_CTRY_NAME, SALES_CITY, SALES_CITY_NAME, SALES_LINK, MER_CODE, MER_MCC, MER_NAM, ACQ_NAME, REVCODE, AUTH_TYPE,REF_NBR, VI_B011_SYS_AUDT_TRCE, CARD_TYPE, STAGE_TYPE, STAGE_NUM, OVERSEA_FLAG, SYSDATE, SYSDATE FROM TEMP_FMLG_PURGE
原sql执行计划非常简单:
该语句是job中的语句,每天都需要跑的,其历史执行时间如下图,可以看出执行时间非常长的都是user_io_wait等待比较严重的一些:
查看一下相关表的属性和数据量:
SELECT v.OWNER,
v.TABLE_NAME,
v.partitioned,
v.LAST_ANALYZED,
v.NUM_ROWS,
v.table_size2 ,
v.EMPTY_BLOCKS
FROM vw_table_lhr v
WHERE v.TABLE_NAME IN ('BASE_FMLG',
'TEMP_FMLG_PURGE');
BASE_FMLG有15亿的数据量,是个分区表,每次从TEMP_FMLG_PURGE中取数,TEMP_FMLG_PURGE大约有234W的数据量,
索引信息:
SELECT v.index_owner,
v.index_name,
v.index_type,
v.partitioned,
v.索引列,
v.index_size,
v.num_rows
FROM vw_table_index_lhr v
WHERE v.TABLE_NAME IN ('BASE_FMLG',
'TEMP_FMLG_PURGE');
被插入的表有5个索引,且都是分区索引,不涉及全局索引,涉及到的分区索引:
select v.index_owner,
v.index_name,
v.index_type,
v.索引列,
v.partition_size,
v.num_rows
from vw_table_index_part2_lhr V
where V.TABLE_NAME='BASE_FMLG'
and v.PARTITION_NAME='P201407' ;
查一下数据来源:
SELECT t.BATCH_DATE,
COUNT(1)
FROM TEMP_FMLG_PURGE t
GROUP BY t.BATCH_DATE;
看来,都是当天的数据,所以只涉及到分区表的单个分区
select v.PARTITION_NAME,
v.TABLE_NAME,
v.LAST_ANALYZED,
v.NUM_ROWS,
v.partition_size ,
v.EMPTY_BLOCKS,v.LOGGING
from VW_TABLE_PART_LHR V
where V.TABLE_NAME='BASE_FMLG';
系统预估剩余时间:select * from VW_LONGRUN_LHR a where a.SQL_ID='2pnas8zbxtk3a';
插入200W的数据到一个单个分区16G的分区表中需要花费将近12个小时,似乎慢了点。。。。。
问题解决:
查询会话的统计信息,发现redo的产生量非常的大,所以解决办法:
第一步: 将表修改为nologging属性
第二步: 将索引修改为nologging属性
第三步: 插入的时候采用append方式来插入
第四步:如果还是慢点的话,可以采用并行插入,增大排序缓冲区
第五步:如果有可能可以先把索引置于无效状态,然后插入完成之后再重建索引
注意: 以上解决办法①必须是该表的数据不重要,不然修改为nologging属性后万一数据丢失可能就找不回来了,② 索引一般都为nologging模式,索引记录redo没有作用 ③ 采用append插入的前提是该表上边没有大量的delete动作
最后优化后的代码:
先将表及其索引置于NOLOGGING模式:
alter table RISKREPT.BASE_FMLG NOLOGGING;
alter index DX_RKO_FMLG_BATCH_DATE NOLOGGING;
alter index IDX_RKO_FMLG_ACCT NOLOGGING;
alter index IDX_RKO_FMLG_CARD NOLOGGING;
alter index IDX_RKO_FMLG_DT NOLOGGING;
alter index IDX_RKO_FMLG_MER NOLOGGING;
----- 修改会话的属性,开启并行插入:
alter session set workarea_size_policy=manual;
alter session set sort_area_size=1000000000;
alter session ENABLE parallel dml;
EXPLAIN PLAN for
INSERT /*+parallel(BASE_FMLG,4) */ INTO RISKREPT.BASE_FMLG (BATCH_DATE, DATE_STAMP_ST, TIME_STAMP_ST, ORG, ACCT, CARD_NBR, CARD_SEQ, MER_ORG, MER_NBR,REQUEST_TYPE_ID, LOGO, SYSTEM_ACTION, FINAL_ACTION, ACTION_REASON, REVERSAL_REASON, AVAIL_CR, CASH_AVAIL_CR, ACCT_CURR_BAL, ACCT_CURR_BAL_CASH,B007_GMT_DATE_TIME, B018_MER_TYPE, B019_CNTRY_CODE, B032_ACQ_ID, B033_FWD_ID, AUTH_CODE, B039_RESP_CODE, B041_CRD_ACCPT_TERM,B042_TERMTYPE2_MER_ID, B043_CRD_ACCPT_NAM, B043_CRD_ACCPT_CITY, B043_CRD_ACCPT_ST_CTRY, MESSAGE_TYPE_ID, RECORD_TYPE, SALES_AMT_RMB,TOTAL_SALES_AMT, B049_CURR_CODE, BLING_AMT, POS_ENTRY_MODE, PIN_ENTRY_MODE, TRADE_INTERNET, SALES_CTRY, SALES_CTRY_NAME, SALES_CITY,SALES_CITY_NAME, SALES_LINK, MER_CODE, MER_MCC, MER_NAM, ACQ_NAME, REVCODE, AUTH_TYPE, REF_NBR, VI_B011_SYS_AUDT_TRCE, CARD_TYPE, STAGE_TYPE,STAGE_NUM, OVERSEA_FLAG, RISK_FCD, RISK_LCD) SELECT BATCH_DATE, DATE_STAMP_ST, TIME_STAMP_ST, ORG, ACCT, CARD_NBR, CARD_SEQ, MER_ORG, MER_NBR,REQUEST_TYPE_ID, LOGO, SYSTEM_ACTION, FINAL_ACTION, ACTION_REASON, REVERSAL_REASON, AVAIL_CR, CASH_AVAIL_CR, ACCT_CURR_BAL, ACCT_CURR_BAL_CASH,B007_GMT_DATE_TIME, B018_MER_TYPE, B019_CNTRY_CODE, B032_ACQ_ID, B033_FWD_ID, AUTH_CODE, B039_RESP_CODE, B041_CRD_ACCPT_TERM,B042_TERMTYPE2_MER_ID, B043_CRD_ACCPT_NAM, B043_CRD_ACCPT_CITY, B043_CRD_ACCPT_ST_CTRY, MESSAGE_TYPE_ID, RECORD_TYPE, SALES_AMT_RMB,TOTAL_SALES_AMT, B049_CURR_CODE, BLING_AMT, POS_ENTRY_MODE, PIN_ENTRY_MODE, TRADE_INTERNET, SALES_CTRY, SALES_CTRY_NAME, SALES_CITY,SALES_CITY_NAME, SALES_LINK, MER_CODE, MER_MCC, MER_NAM, ACQ_NAME, REVCODE, AUTH_TYPE, REF_NBR, VI_B011_SYS_AUDT_TRCE, CARD_TYPE, STAGE_TYPE,STAGE_NUM, OVERSEA_FLAG, SYSDATE, SYSDATE FROM TEMP_FMLG_PURGE_2 ;
commit;
select * from table(DBMS_XPLAN.display('','',''));
优化后的执行计划:
自己跑了一下,大约30分钟就可以跑完,从12个小时缩短到30分钟,这个还是比较有成就感的。。。。
产生的redo量不足500M,未优化之前的那个redo量达到了15G左右,忘记截图了,所以这个sql就优化的差不多了:
select * from VW_SESSTAT_LHR a where a.SID=850 order by a.VALUE desc ;