Infrastructure at your Service

Oracle Team

18c: some optimization about redo size

By Franck Pachot

.
Some years ago, at the time of 12.1 release, I published in the SOUG newsletter some tests to show the amount of redo generated by different operations on a 10000 rows table. I had run it on 12.2 without seeing the differences and now on 18.1
I get the statistics from mystat using a script that displays them as columns, with the value being the difference from the previous run. I’ve run the same as in the article, and most of the statistics were in the same ballpark.

12.2

In 12.2 I get the same numbers. I was surprised about that because there is an optimization on 12.2 when updating a column to the same value. But this optimization may not occur for all situations. This reduction of redo generation has been analyzed by Maris Elsins with redo log dumps and by Sayan Malakshinov on triggers. And then the origin of this optimization has been exposed by Bryn Llewellyn. All info and links from the following Twitter conversation:

Here are the numbers I have in 18c, which are very similar to those from 12.1


SQL> --- update with different values (the original ones were all lowercase)
SQL>
SQL> update TEST_TABLE set a=upper(a),b=upper(b),c=upper(c),d=upper(d);
10000 rows updated.
 
SQL> commit;
Commit complete.
 
SQL> @ _mystat_diff.sql
SQL> set termout off verify off
SQL> /
 
db block changes redo size undo change vector size redo entries
---------------- ---------------- ----------------------- ----------------
190,485 32,226,368 13,288,940 95,258
 
1 row selected.
 
SQL>
SQL> --- update with same valus
SQL>
SQL> update TEST_TABLE set a=upper(a),b=upper(b),c=upper(c),d=upper(d);
10000 rows updated.
 
SQL> commit;
Commit complete.
 
SQL> @ _mystat_diff.sql
SQL> set termout off verify off
SQL> /
 
db block changes redo size undo change vector size redo entries
---------------- ---------------- ----------------------- ----------------
20,346 4,594,528 1,844,012 10,085
 
1 row selected.

The second update has updated the 4 columns with the same value. For about 1MB of data (10000 rows with avg_row_len=100), we have 1.8MB of undo and 4.4MB of redo (which covers the new data and the undo). I have 4 indexes there but Oracle do not update index entries when the old and new values are the same.

The first update changes all values, and then, in addition to the changes in the table block, the indexes must be updated.

So, here, on my test, it seems that the 12.2 optimization, referenced in the tweet above, did not occur because the redo generated for the table blocks is stull full logging when the old and new values are the same. I can check from a block dump that I have the same value in undo and redo:

REDO RECORD - Thread:1 RBA: 0x000008.00002444.0010 LEN: 0x01c8 VLD: 0x01 CON_UID: 1008806272
SCN: 0x00000000002cb8a4 SUBSCN:3047 07/03/2018 12:23:22
CHANGE #1 CON_ID:4 TYP:0 CLS:36 AFN:14 DBA:0x02405a20 OBJ:4294967295 SCN:0x00000000002cb8a4 SEQ:34 OP:5.1 ENC:0 RBL:0 FLG:0x0000
ktudb redo: siz: 184 spc: 2020 flg: 0x0022 seq: 0x0147 rec: 0x22
xid: 0x000a.009.000002bd
ktubu redo: slt: 9 rci: 33 opc: 11.1 objn: 77968 objd: 77978 tsn: 0
Undo type: Regular undo Undo type: Last buffer split: No
Tablespace Undo: No
0x00000000
KDO undo record:
KTB Redo
op: 0x02 ver: 0x01
compat bit: 4 (post-11) padding: 1
op: C uba: 0x02405a20.0147.21
KDO Op code: URP row dependencies Disabled
xtype: XAxtype KDO_KDOM2 flags: 0x00000080 bdba: 0x0040a994 hdba: 0x0040a7d8
itli: 3 ispac: 0 maxfr: 4863
tabn: 0 slot: 3(0x3) flag: 0x2c lock: 0 ckix: 0
ncol: 4 nnew: 4 size: 0
Vector content:
col 0: [24] 31 30 30 30 30 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58
col 1: [24] 31 30 30 30 30 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58
col 2: [24] 31 30 30 30 30 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58
col 3: [24] 31 30 30 30 30 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58
CHANGE #2 CON_ID:4 TYP:0 CLS:1 AFN:12 DBA:0x0040a994 OBJ:77978 SCN:0x00000000002cb8a4 SEQ:3 OP:11.5 ENC:0 RBL:0 FLG:0x0000
KTB Redo
op: 0x02 ver: 0x01
compat bit: 4 (post-11) padding: 1
op: C uba: 0x02405a20.0147.22
KDO Op code: URP row dependencies Disabled
xtype: XAxtype KDO_KDOM2 flags: 0x00000080 bdba: 0x0040a994 hdba: 0x0040a7d8
itli: 3 ispac: 0 maxfr: 4863
tabn: 0 slot: 3(0x3) flag: 0x2c lock: 3 ckix: 0
ncol: 4 nnew: 4 size: 0
Vector content:
col 0: [24] 31 30 30 30 30 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58
col 1: [24] 31 30 30 30 30 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58
col 2: [24] 31 30 30 30 30 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58
col 3: [24] 31 30 30 30 30 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58 58

The redo record has the old and new values even when they are the same. I hope that we will have more documentation about the 12.2 redo optimization so that it can be predictable and tunable.

18.1

So, I didn’t see the 12.2 optimizations I expected here. However, but it seems that we have one with deletes on 18c.

In 12.2 the delete of all 10000 rows without index generates 2MB of undo and 3.5MB of redo:

SQL> --- delete all rows
SQL>
SQL> delete from TEST_TABLE;
10000 rows deleted.
 
SQL> commit;
Commit complete.
 
SQL> @ _mystat_diff.sql
SQL> set termout off verify off
SQL> /
 
db block changes redo size undo change vector size redo entries
---------------- ---------------- ----------------------- ----------------
20,690 3,670,476 2,053,292 10,145
 
1 row selected.

The same in 18.1 generates only 1MB of undo and redo:

SQL> --- delete all rows
SQL>
SQL> delete from TEST_TABLE;
10000 rows deleted.
 
SQL> commit;
Commit complete.
 
SQL> @ _mystat_diff.sql
SQL> set termout off verify off
SQL> /
 
db block changes redo size undo change vector size redo entries
---------------- ---------------- ----------------------- ----------------
872 1,187,120 1,116,812 293
 
1 row selected.

So in 12.2 I had one block change, and one redo record per row deleted. In 18.1 it seems that I have one redo record per block where all rows are deleted. Still in the same Twitter conversation, Tanel Poder had the idea to do a sparse delete leaving one row in each block:

SQL> delete from TEST_TABLE where dbms_rowid.rowid_row_number(rowid)!=42;
9849 rows deleted.
 
SQL> commit;
Commit complete.
 
SQL> @ _mystat_diff.sql
SQL> set termout off verify off
SQL> /
 
db block changes redo size undo change vector size redo entries
---------------- ---------------- ----------------------- ----------------
20,431 3,660,204 2,102,584 10,011

Now, as there is no blocks that are fully emptied, I’m back to one redo entry per row deleted.

Update 4-JUL-18

I got a comment from Jonathan Lewis that the delete optimization occurs when using mod() to filter one row per block instead of dbms_rowid. Actually it seems that this optimization is more related to the way the rows are accessed.

One Comment

Leave a Reply

Oracle Team
Oracle Team