Oracle优化：千万级大表逻辑判断的累赘

insert into pntmall_point_detail(PNTMALL_PNT_ID,PNTMALL_PNT_DT,PNTMALL_VALIDITY,PNTMALL_LASTUPDATEDT,

                           PNTMALL_RPTPNT_ID,PNTMALL_HAIERUID,

                   PNTMALL_OPTYPE_ID,PNTMALL_OPTYPE_DESC,

                         PNTMALL_NUM,PNTMALL_EXPDT,HPT_REDEMPT_POINT,

                         HPT_LEFT_POINT,HPT_FULLREDEMPT_STATUS)

  SELECT PNTMALL_PNT_ID,PNTMALL_PNT_DT,PNTMALL_VALIDITY,PNTMALL_LASTUPDATEDT,

                           PNTMALL_RPTPNT_ID,PNTMALL_HAIERUID,

                   PNTMALL_OPTYPE_ID,PNTMALL_OPTYPE_DESC,

                         PNTMALL_NUM,PNTMALL_EXPDT,0 as HPT_REDEMPT_POINT,PNTMALL_NUM,0

    FROM pntmall_point_detail_tmp a

          WHERE NOT EXISTS (SELECT b.PNTMALL_PNT_ID
            FROM pntmall_point_detail b
           WHERE a.PNTMALL_PNT_ID = b.PNTMALL_PNT_ID);

PNTMALL_POINT_DETAIL包含3800万条数据，cost 6 hours。

优化后

delete from pntmall_point_detail_tmp a where exists (select 1 from pntmall_point_detail b where a.PNTMALL_PNT_ID = b.PNTMALL_PNT_ID);

insert into pntmall_point_detail(PNTMALL_PNT_ID,PNTMALL_PNT_DT,PNTMALL_VALIDITY,PNTMALL_LASTUPDATEDT,

                           PNTMALL_RPTPNT_ID,PNTMALL_HAIERUID,

                   PNTMALL_OPTYPE_ID,PNTMALL_OPTYPE_DESC,

                         PNTMALL_NUM,PNTMALL_EXPDT,HPT_REDEMPT_POINT,

                         HPT_LEFT_POINT,HPT_FULLREDEMPT_STATUS)

  SELECT PNTMALL_PNT_ID,PNTMALL_PNT_DT,PNTMALL_VALIDITY,PNTMALL_LASTUPDATEDT,

                           PNTMALL_RPTPNT_ID,PNTMALL_HAIERUID,

                   PNTMALL_OPTYPE_ID,PNTMALL_OPTYPE_DESC,

                         PNTMALL_NUM,PNTMALL_EXPDT,0 as HPT_REDEMPT_POINT,PNTMALL_NUM,0

    FROM pntmall_point_detail_tmp a;

cost 5 minutes。

其实还可以进一步优化

drop indexBER.INDEX_POD_PNTMALL_HAIERUID;

drop indexBER.PNTMALL_POINT_ID_HAIERUID;

delete from pntmall_point_detail_tmp a where exists (select 1from pntmall_point_detailb where a.PNTMALL_PNT_ID =b.PNTMALL_PNT_ID);

insert/*+append*/into pntmall_point_detail(PNTMALL_PNT_ID,PNTMALL_PNT_DT,PNTMALL_VALIDITY,PNTMALL_LASTUPDATEDT,

                           PNTMALL_RPTPNT_ID,PNTMALL_HAIERUID,

                   PNTMALL_OPTYPE_ID,PNTMALL_OPTYPE_DESC,

                         PNTMALL_NUM,PNTMALL_EXPDT,HPT_REDEMPT_POINT,

                         HPT_LEFT_POINT,HPT_FULLREDEMPT_STATUS)

  SELECT PNTMALL_PNT_ID,PNTMALL_PNT_DT,PNTMALL_VALIDITY,PNTMALL_LASTUPDATEDT,

                           PNTMALL_RPTPNT_ID,PNTMALL_HAIERUID,

                   PNTMALL_OPTYPE_ID,PNTMALL_OPTYPE_DESC,

                         PNTMALL_NUM,PNTMALL_EXPDT,0 as HPT_REDEMPT_POINT,PNTMALL_NUM,0

    FROM pntmall_point_detail_tmp a; 

commit;

WHERE NOT EXISTS (SELECT b.PNTMALL_PNT_ID

            FROMpntmall_point_detail b

           WHEREa.PNTMALL_PNT_ID = b.PNTMALL_PNT_ID);

create index BER.INDEX_POD_PNTMALL_HAIERUIDon BER.PNTMALL_POINT_DETAIL (PNTMALL_HAIERUID) online nologing;

create unique indexBER.PNTMALL_POINT_ID_HAIERUID on BER.PNTMALL_POINT_DETAIL (PNTMALL_PNT_ID) online nologing;

总体优化思路，不要在insert中加入过多的判断语句，删索引，append，重建索引，如果是归档模式，alter table nologing；append 只适用于insert select 这种方式，而且insert后要加commit,否则无法进行其他DML操作。

实测 append 1600万条数据，cost 8s