ROWID 切片

对于大表的update，delete，很容易引起undo不够的问题，特别是在一个session里面运行的时候，或者由于其他原因导致的回滚操作，这个就比较苦逼了，如果是分区表，可能影响稍微小一点，我们可以按照分区来进行更新，按照分区来进行delete。对于不是分区表的情况，就可以用rowid切片，然后开启多个进程并行处理，这样既能提升不少的速度，也能减少undo，防止死事物恢复太慢。

我这里有test_b表，表中有2793440条数据，我们简单的按照rowid切片对表进行处理

SQL> select count(*) from test_b;
COUNT(*)
--------
 2793440
 
select ‘where rowid between ‘‘‘ ||sys.dbms_rowid.rowid_create(1, d.oid, c.fid1, c.bid1, 0) ||‘‘‘ and ‘‘‘ ||sys.dbms_rowid.rowid_create(1, d.oid, c.fid2, c.bid2, 9999) || ‘‘‘‘ ||‘;‘
  from (select distinct b.rn,
                        first_value(a.fid) over(partition by b.rn order by a.fid, a.bid rows between unbounded preceding and unbounded following) fid1,
                        last_value(a.fid) over(partition by b.rn order by a.fid, a.bid rows between unbounded preceding and unbounded following) fid2,
                        first_value(decode(sign(range2 - range1),
                                           1,
                                           a.bid +
                                           ((b.rn - a.range1) * a.chunks1),
                                           a.bid)) over(partition by b.rn order by a.fid, a.bid rows between unbounded preceding and unbounded following) bid1,
                        last_value(decode(sign(range2 - range1),
                                          1,
                                          a.bid +
                                          ((b.rn - a.range1 + 1) * a.chunks1) - 1,
                                          (a.bid + a.blocks - 1))) over(partition by b.rn order by a.fid, a.bid rows between unbounded preceding and unbounded following) bid2
          from (select fid,
                       bid,
                       blocks,
                       chunks1,
 19                         trunc((sum2 - blocks + 1 - 0.1) / chunks1) range1,
                       trunc((sum2 - 0.1) / chunks1) range2
                  from (select /*+ rule */
                         relative_fno fid,
                         block_id bid,
                         blocks,
                         sum(blocks) over() sum1,
                         trunc((sum(blocks) over()) / &&rowid_ranges) chunks1,
                         sum(blocks) over(order by relative_fno, block_id) sum2
                          from dba_extents
                         where segment_name = upper(‘&&segment_name‘)
                           and owner = upper(‘&&owner‘))
                 where sum1 > &&rowid_ranges) a,
               (select rownum - 1 rn
                  from dual
                connect by level <= &&rowid_ranges) b
         where b.rn between a.range1 and a.range2) c,
       (select max(data_object_id) oid
          from dba_objects
         where object_name = upper(‘&&segment_name‘)
           and owner = upper(‘&&owner‘)
           and data_object_id is not null) d
Enter value for rowid_ranges: 15
old  26:     trunc((sum(blocks) over()) / &&rowid_ranges) chunks1,
new  26:     trunc((sum(blocks) over()) / 15) chunks1,
Enter value for segment_name: test_b
old  29:     where segment_name = upper(‘&&segment_name‘)
new  29:     where segment_name = upper(‘test_b‘)
Enter value for owner: scott
old  30:       and owner = upper(‘&&owner‘))
new  30:       and owner = upper(‘scott‘))
old  31:    where sum1 > &&rowid_ranges) a,
new  31:    where sum1 > 15) a,
old  34:   connect by level <= &&rowid_ranges) b
new  34:   connect by level <= 15) b
old  38:   where object_name = upper(‘&&segment_name‘)
new  38:   where object_name = upper(‘test_b‘)
old  39:     and owner = upper(‘&&owner‘)
new  39:     and owner = upper(‘scott‘)

‘WHEREROWIDBETWEEN‘‘‘||SYS.DBMS_ROWID.ROWID_CREATE(1,D.OID,C.FID1,------------------------------------------------------------------where rowid between ‘AAAqh3AAIAAAACIAAA‘ and ‘AAAqh3AAIAAABYpCcP‘;where rowid between ‘AAAqh3AAIAAABYqAAA‘ and ‘AAAqh3AAIAAACCpCcP‘;where rowid between ‘AAAqh3AAIAAACCqAAA‘ and ‘AAAqh3AAIAAACspCcP‘;where rowid between ‘AAAqh3AAIAAACsqAAA‘ and ‘AAAqh3AAIAAADOpCcP‘;where rowid between ‘AAAqh3AAIAAADOqAAA‘ and ‘AAAqh3AAIAAAD+pCcP‘;where rowid between ‘AAAqh3AAIAAAD+qAAA‘ and ‘AAAqh3AAIAAAEepCcP‘;where rowid between ‘AAAqh3AAIAAAEeqAAA‘ and ‘AAAqh3AAIAAAFOpCcP‘;where rowid between ‘AAAqh3AAIAAAFOqAAA‘ and ‘AAAqh3AAIAAAF+pCcP‘;where rowid between ‘AAAqh3AAIAAAF+qAAA‘ and ‘AAAqh3AAIAAAGepCcP‘;where rowid between ‘AAAqh3AAIAAAGeqAAA‘ and ‘AAAqh3AAIAAAHOpCcP‘;where rowid between ‘AAAqh3AAIAAAHOqAAA‘ and ‘AAAqh3AAIAAAH+pCcP‘;where rowid between ‘AAAqh3AAIAAAH+qAAA‘ and ‘AAAqh3AAIAAAIepCcP‘;where rowid between ‘AAAqh3AAIAAAIeqAAA‘ and ‘AAAqh3AAIAAAJOpCcP‘;where rowid between ‘AAAqh3AAIAAAJOqAAA‘ and ‘AAAqh3AAIAAAJ+pCcP‘;where rowid between ‘AAAqh3AAIAAAJ+qAAA‘ and ‘AAAqh3AAIAAAKepCcP‘;
15 rows selected.
Elapsed: 00:05:33.55

上面的例子，把test_b表分成15个rowid片，运行rowid切片sql的时候，需要传入3个参数，第一个是切片数量，第二个是切片对象，第三个是切片对象所属的对象，也就是owner，sql最终输出的结果可以看到，按照rowid对test_b表进行了切片，在update和delete的时候，分成15个session，加上并行处理，速度比直接update和delete快很多。

直接update操作：

SQL> update test_b set object_id=100;
2793440 rows updated.
Elapsed: 00:24:19.65

花了24分钟 ，占用了大量的UNDO表空间

之后便是相同的时间的ROLLBACK,

可以查询回滚的速度：

SQL> select KTUXESIZ from x$ktuxe where KTUXESTA<>‘INACTIVE‘;
KTUXESIZ--------   85462

SQL> rollback;
Rollback complete.
Elapsed: 00:50:36.58  
SQL> update test_b set object_id=100 where rowid between ‘AAAqh3AAIAAAACIAAA‘ and ‘AAAqh3AAIAAABYpCcP‘;
375065 rows updated.
Elapsed: 00:01:47.92SQL> update /* +parallel(test_b 4) */ test_b set object_id=100 where rowid between ‘AAAqh3AAIAAABYqAAA‘ and ‘AAAqh3AAIAAACCpCcP‘;
185986 rows updated.
Elapsed: 00:00:57.22