CBO Costing Plans With Migrated Rows Part I (“Ignoreland”) March 21, 2023
Posted by Richard Foote in BLEVEL, CBO, Clustering Factor, Data Clustering, Index Access Path, Index Height, Index statistics, Leaf Blocks, Migrated Rows, Non-Equality Predicates, Oracle, Oracle Blog, Oracle Cost Based Optimizer, Oracle General, Oracle Indexes, Oracle Statistics, Performance Tuning, Richard's Blog, ROWID.2 comments
Whilst recently blogging about Migrated Rows and specifically changes to how ROWIDs are now maintained on the fly in Oracle Autonomous Databases, I made a discovery regarding how the Cost-Based Optimizer (CBO) costs such plans. This is one of the key reasons why I blog, not only to try and share odd titbits about how Oracle works, but also to hopefully learn much myself in the process.
Imagine my surprise in not only learning that Oracle and the CBO works differently to how I had always thought Oracle worked in this respect, but that this behaviour has been the case since at least Oracle 9i.
In Part I, I’ll use the same example of migrated rows as I’ve used in the past few blog posts and initially show how the CBO generally costs such plans (and by which I had incorrectly assumed ALWAYS costed such plans).
Let’s start by creating and populating a tightly packed table (in an environment where ROWIDs are NOT updated on the fly):
SQL> create table bowie(id number, code1 number, code2 number, code3 number, code4 number, code5 number, code6 number, code7 number, code8 number, code9 number, code10 number, code11 number, code12 number, code13 number, code14 number, code15 number, code16 number, code17 number, code18 number, code19 number, code20 number, name varchar2(142)) PCTFREE 0; Table BOWIE created. SQL> insert into bowie SELECT rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, 'BOWIE' FROM dual CONNECT BY LEVEL <= 200000; 200,000 rows inserted. SQL> commit; Commit complete.
I’ll next create an index on the well clustered ID column (as the rows are inserted in ID column order within the table):
SQL> create index bowie_id_i on bowie(id); Index BOWIE_ID_I created.
Next, we’ll use the Oracle recommended method of collecting table/index statistics, by using the DBMS_STATS package:
SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=>'BOWIE', estimate_percent=> null, no_invalidate=>false); PL/SQL procedure successfully completed. SQL> select table_name, num_rows, blocks, empty_blocks, avg_space, avg_row_len, chain_cnt from user_tables where table_name='BOWIE'; TABLE_NAME NUM_ROWS BLOCKS EMPTY_BLOCKS AVG_SPACE AVG_ROW_LEN CHAIN_CNT _____________ ___________ _________ _______________ ____________ ______________ ____________ BOWIE 200000 3268 0 0 111 0 SQL> select index_name, blevel, leaf_blocks, clustering_factor from user_indexes where table_name='BOWIE'; INDEX_NAME BLEVEL LEAF_BLOCKS CLUSTERING_FACTOR _____________ _________ ______________ ____________________ BOWIE_ID_I 1 473 3250
Note the key index statistics here: BLEVEL=1, LEAF_BLOCKS=473 and the near perfect CLUSTERING_FACTOR=3250.
If we run the following query featuring a non-equality range predicate:
SQL> select * from bowie where id > 1 and id < 1001;
999 rows selected.
PLAN_TABLE_OUTPUT
_______________________________________________________________________________________________________________
SQL_ID b1vwpu2rgn8p5, child number 0
-------------------------------------
select * from bowie where id > 1 and id < 1001
Plan hash value: 1405654398
------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows | A-Rows | A-Time | Buffers |
------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | 999 |00:00:00.01 | 18 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE | 1 | 1000 | 999 |00:00:00.01 | 18 |
|* 2 | INDEX RANGE SCAN | BOWIE_ID_I | 1 | 1000 | 999 |00:00:00.01 | 4 |
------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">1 AND "ID"<1001)
Statistics
-----------------------------------------------------------
1 CPU used by this session
1 CPU used when call started
1 DB time
7678 RM usage
3 Requests to/from client
2 SQL*Net roundtrips to/from client
16 buffer is not pinned count
1983 buffer is pinned count
323 bytes received via SQL*Net from client
171383 bytes sent via SQL*Net to client
2 calls to get snapshot scn: kcmgss
2 calls to kcmgcs
18 consistent gets
1 consistent gets examination
1 consistent gets examination (fastpath)
18 consistent gets from cache
17 consistent gets pin
17 consistent gets pin (fastpath)
2 execute count
1 index range scans
147456 logical read bytes from cache
17 no work - consistent read gets
40 non-idle wait count
2 opened cursors cumulative
1 opened cursors current
2 parse count (total)
2 process last non-idle time
1 session cursor cache count
1 session cursor cache hits
18 session logical reads
1 sorts (memory)
2024 sorts (rows)
999 table fetch by rowid
3 user calls
We notice that the CBO indeed uses the index.
They key statistic to note here is that Consistent Gets is just 18, which is extremely low considering we’re returning 999 rows. This is due to the fact the index is currently extremely efficient as it can fetch multiple rows by visiting the same table block due to the excellent clustering/ordering of the required ID column values (and also due to my high arraysize session setting).
If we look at the CBO costings for this plan:
SQL> SELECT * FROM TABLE(DBMS_XPLAN.display_cursor(sql_id=>'b1vwpu2rgn8p5',format=>'ALLSTATS LAST +cost +bytes'));
PLAN_TABLE_OUTPUT
____________________________________________________________________________________________________________________________________
SQL_ID b1vwpu2rgn8p5, child number 0
-------------------------------------
select * from bowie where id > 1 and id < 1001
Plan hash value: 1405654398
---------------------------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows |E-Bytes| Cost (%CPU)| A-Rows | A-Time |Buffers |
---------------------------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | | 21 (100)| 999 |00:00:00.01 | 18 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE | 1 | 1000 | 108K| 21 (0)| 999 |00:00:00.01 | 18 |
|* 2 | INDEX RANGE SCAN | BOWIE_ID_I | 1 | 1000 | | 4 (0)| 999 |00:00:00.01 | 4 |
---------------------------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">1 AND "ID"<1001)
I’ve previously discussed many times how the CBO costs index access paths, but it’s always useful to go over this again, as it’s the most common question I get asked when I visit customer sites.
The KEY statistic the CBO has to determine is the estimated Selectivity of the query (the estimated percentage of rows to be returned), as this is the driver of all the subsequent CBO calculations.
The Selectivity of this range-based predicate query is calculated as follows:
Selectivity = (Highest Bound Value – Lowest Bound Value) / (Highest Value – Lowest Value)
= (1001-1) /(200000-1)
= 1000/199999
= approx. 0.005
Once Oracle has the selectivity, it can calculate the query Cardinality (estimated number of rows) as follows:
Cardinality = Selectivity x No of Rows
Cardinality = 0.005 x 200000 = 1000 rows
This is our visual window into the likelihood that the CBO has made an accurate decision with its execution plan. If the cardinality estimates are reasonably accurate, then the CBO is likely to generate a good plan. If the cardinality estimates are way off, then the CBO is more likely to generate an inappropriate plan.
The CBO cardinality estimate in the above plan is 1000 rows, whereas the number of rows actually returned is 999 rows.
So indeed, the CBO has got the cardinality almost spot on (except for a trivial rounding error) and so we have a high degree of confidence that the CBO is using the correct selectivity estimates when they get plugged into the following CBO formula for costing an index range scan (using this selectivity of 0.005 and the index statistics listed above):
Index Scan Cost = (blevel + ceil(effective index selectivity x leaf_blocks)) + ceil(effective table selectivity x clustering_factor)
= (1 + ceil(0.005 x 467)) + ceil(0.005 x 3250)
= (1 + 3) + 17
= 4 + 17 = 21
So we can clearly see where the CBO gets its costings for both reading the index during the Index Range Scan (4) and for the plan as a whole (21).
The CBO cost of 21 very closely resembles the 18 consistent gets accessed when the plan is executed. This to me suggests that the CBO has indeed costed this plan very accurately and appropriately.
It’s interesting to note in the above execution plan that Oracle is attributing 100% of this cost of 21 to CPU (21 (100)). That will be a discussion for another day…
OK, let’s now perform an update on the table, increasing the size of the rows such that I generate a bunch of migrated rows:
SQL> update bowie set name='THE RISE AND FALL OF BOWIE STARDUST AND THE SPIDERS FROM MARS'; 200,000 rows updated. SQL> commit; Commit complete.
If we now collect fresh statistics again using DBMS_STATS:
SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=>'BOWIE', estimate_percent=> null, no_invalidate=>false); PL/SQL procedure successfully completed. SQL> select table_name, num_rows, blocks, empty_blocks, avg_space, avg_row_len, chain_cnt from user_tables where table_name='BOWIE'; TABLE_NAME NUM_ROWS BLOCKS EMPTY_BLOCKS AVG_SPACE AVG_ROW_LEN CHAIN_CNT _____________ ___________ _________ _______________ ____________ ______________ ____________ BOWIE 200000 4906 0 0 167 0 SQL> select index_name, blevel, leaf_blocks, clustering_factor from user_indexes where table_name='BOWIE'; INDEX_NAME BLEVEL LEAF_BLOCKS CLUSTERING_FACTOR _____________ _________ ______________ ____________________ BOWIE_ID_I 1 473 3250
We notice that none of the key statistics have changed, except for the number of Table Blocks (now 4906, previously it was 3268) and the Average Row Length has also increased (now 167, previously it was 111). Both of these can of course be attributed to the increase in the size of the values now stored in the NAME column following the Update.
Importantly, notice that collecting statistics via DBMS_STATS does NOT collect data for the CHAIN_CNT statistic, it remains at 0 even though many migrated rows were actually generated by the Update statement (as we’ll see below).
Increasing the Table Blocks will result in an associated increase in the cost of reading this table via a Full Table Scan (FTS).
We notice that none of the index-related statistics changed following the Update statement (as in this example, Oracle does NOT update the ROWIDs of any of the migrated rows, Oracle simply stores a pointer in the original block to denote the new physical location of the migrated rows as previously discussed).
So if we only INCREASE the cost of a FTS (via having more Table Blocks) but keep intact all the previous index related statistics, then the CBO is certainly going to again select the same Index Range Scan plan, as the plan will have the same (cheaper than FTS) costings as before.
If we re-run the query again:
SQL> select * from bowie where id > 1 and id < 1001;
999 rows selected.
PLAN_TABLE_OUTPUT
_______________________________________________________________________________________________________________
SQL_ID b1vwpu2rgn8p5, child number 0
-------------------------------------
select * from bowie where id > 1 and id < 1001
Plan hash value: 1405654398
------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows | A-Rows | A-Time | Buffers |
------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | 999 |00:00:00.01 | 666 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE | 1 | 1000 | 999 |00:00:00.01 | 666 |
|* 2 | INDEX RANGE SCAN | BOWIE_ID_I | 1 | 1000 | 999 |00:00:00.01 | 4 |
------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">1 AND "ID"<1001)
Statistics
-----------------------------------------------------------
1 CPU used by this session
1 CPU used when call started
7709 RM usage
3 Requests to/from client
2 SQL*Net roundtrips to/from client
664 buffer is not pinned count
1662 buffer is pinned count
323 bytes received via SQL*Net from client
171500 bytes sent via SQL*Net to client
2 calls to get snapshot scn: kcmgss
2 calls to kcmgcs
666 consistent gets
1 consistent gets examination
1 consistent gets examination (fastpath)
666 consistent gets from cache
665 consistent gets pin
665 consistent gets pin (fastpath)
2 execute count
1 index range scans
5455872 logical read bytes from cache
665 no work - consistent read gets
39 non-idle wait count
1 non-idle wait time
2 opened cursors cumulative
1 opened cursors current
2 parse count (total)
1 process last non-idle time
2 session cursor cache count
666 session logical reads
1 sorts (memory)
2024 sorts (rows)
999 table fetch by rowid
327 table fetch continued row
3 user calls
We notice that indeed it’s the same Index Range Scan plan as before.
But we notice that the number of Consistent Gets has increased substantially to 666 (previously it was just 18). The reason for this large jump is due to the now 327 table fetch continued rows that need to be accessed due to the newly migrated rows following the Update. This number is then doubled (so 2 x 327 = 654) to represent the approximate additional Consistent Gets we now need to perform, as Oracle needs to read the additional table block to access the migrated row’s new physical location AND to now re-read the original table block to access the next row to be fetched (previously Oracle could read all the required consecutive rows required from the same table block within the one consistent get).
So it’s now actually substantially more expensive to read the required 1000 rows via this index due to this increase in necessary consistent gets.
But if we look at the actual cost of this plan now:
SQL> SELECT * FROM TABLE(DBMS_XPLAN.display_cursor(sql_id=>'b1vwpu2rgn8p5',format=>'ALLSTATS LAST +cost +bytes'));
PLAN_TABLE_OUTPUT
____________________________________________________________________________________________________________________________________
SQL_ID b1vwpu2rgn8p5, child number 0
-------------------------------------
select * from bowie where id > 1 and id < 1001
Plan hash value: 1405654398
---------------------------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows |E-Bytes| Cost (%CPU)| A-Rows | A-Time |Buffers |
---------------------------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | | 21 (100)| 999 |00:00:00.01 | 666 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE | 1 | 1000 | 163K| 21 (0)| 999 |00:00:00.01 | 666 |
|* 2 | INDEX RANGE SCAN | BOWIE_ID_I | 1 | 1000 | | 4 (0)| 999 |00:00:00.01 | 4 |
---------------------------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">1 AND "ID"<1001)
We notice that as expected (as none of the index-related statistics have changed), that despite being much more expensive to now use this index, the costs of this plan (4 for reading the index and 21 overall) remain unchanged.
I would argue that these CBO costs are no longer as accurate as the 21 total CBO cost does not so closely represent the actual 666 consistent gets now required.
Now, the 327 table fetch continued row statistics from the previous run is clear proof we indeed have migrated rows following the Update statement.
But if we want to confirm how many migrated rows we now have in the table, we can use the ANALYZE command to collect these additional statistics:
SQL> analyze table bowie compute statistics; Table BOWIE analyzed. SQL> select table_name, num_rows, blocks, empty_blocks, avg_space, avg_row_len, chain_cnt from user_tables where table_name='BOWIE'; TABLE_NAME NUM_ROWS BLOCKS EMPTY_BLOCKS AVG_SPACE AVG_ROW_LEN CHAIN_CNT _____________ ___________ _________ _______________ ____________ ______________ ____________ BOWIE 200000 4906 86 415 170 56186
We notice that we now have a CHAIN_CNT of 56186.
Now this statistic can represent any row that is not housed inside a single table block (for which there could be a number of possible reasons, such as a row simply being too long to fit in a single table block), but as all rows are still relatively tiny, we can be certain that indeed all 56186 chained rows represent migrated rows.
Now that I’ve gone and used ANALYZE, primarily to generate this CHAIN_CNT statistic, my previous understanding of how the CBO costs migrated rows crumbles away, as I’ll discuss in my next post…
Possible Impact To Clustering Factor Now ROWIDs Are Updated When Rows Migrate Part I (“Growin’ Up”) March 1, 2023
Posted by Richard Foote in 19c, 19c New Features, Attribute Clustering, Autonomous Data Warehouse, Autonomous Database, Autonomous Transaction Processing, BLEVEL, CBO, Changing ROWID, Clustering Factor, Data Clustering, Hints, Index Access Path, Index Block Splits, Index Delete Operations, Index Height, Index Internals, Index Rebuild, Index statistics, Leaf Blocks, Migrated Rows, Oracle, Oracle Blog, Oracle Cloud, Oracle Cost Based Optimizer, Oracle General, Oracle Indexes, Oracle Indexing Internals Webinar, Oracle Statistics, Oracle19c, Performance Tuning, Richard Foote Training, Richard's Blog, ROWID.2 comments
In my previous post I discussed how an index can potentially be somewhat inflated in size after ROWIDs are updated on the fly after a substantial number of rows are migrated.
However, there’s another key “factor” of an index that in some scenarios can be impacted by this new ROWID behaviour with regard migrated rows.
To highlight this scenario, I’ll again start by creating and populating a table with ENABLE ROW MOVEMENT disabled:
SQL> create table bowie(id number, code1 number, code2 number, code3 number, code4 number, code5 number, code6 number, code7 number, code8 number, code9 number, code10 number, code11 number, code12 number, code13 number, code14 number, code15 number, code16 number, code17 number, code18 number, code19 number, code20 number, name varchar2(142)) PCTFREE 0; Table BOWIE created. SQL> insert into bowie SELECT rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, 'BOWIE' FROM dual CONNECT BY LEVEL <= 200000; 200,000 rows inserted. SQL> commit; Commit complete. SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=>'BOWIE', estimate_percent=> null, no_invalidate=>false); PL/SQL procedure successfully completed.
I’ll next create an index on the ID column. The important aspect with the ID column is that the data is entered monotonically in ID column order, so the associated index will have an excellent (very low) Clustering Factor:
SQL> create index bowie_id_i on bowie(id); Index BOWIE_ID_I created.
If we look at some key statistics of the table and index:
SQL> select table_name, num_rows, blocks from user_tables where table_name='BOWIE'; TABLE_NAME NUM_ROWS BLOCKS _____________ ___________ _________ BOWIE 200000 3268 SQL> select index_name, blevel, leaf_blocks, clustering_factor from user_indexes where table_name='BOWIE'; INDEX_NAME BLEVEL LEAF_BLOCKS CLUSTERING_FACTOR _____________ _________ ______________ ____________________ BOWIE_ID_I 1 473 3250
We can see that the number of table blocks is 3268, the number of index leaf blocks is 473 and we indeed have a near perfect Clustering Factor of 3250.
If we run a couple of queries:
SQL> select * from bowie where id between 1 and 1000;
1,000 rows selected.
PLAN_TABLE_OUTPUT
_______________________________________________________________________________________________________________
SQL_ID gz5u92hmjwz1h, child number 0
-------------------------------------
select * from bowie where id between 1 and 1000
Plan hash value: 1405654398
------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows | A-Rows | A-Time | Buffers |
------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | 1000 |00:00:00.01 | 18 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE | 1 | 1000 | 1000 |00:00:00.01 | 18 |
|* 2 | INDEX RANGE SCAN | BOWIE_ID_I | 1 | 1000 | 1000 |00:00:00.01 | 4 |
------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">=1 AND "ID"<=1000)
Note
-----
- automatic DOP: Computed Degree of Parallelism is 1 because of no expensive parallel operation
Statistics
-----------------------------------------------------------
1 CPU used by this session
1 CPU used when call started
7353 RM usage
3 Requests to/from client
2 SQL*Net roundtrips to/from client
16 buffer is not pinned count
1985 buffer is pinned count
324 bytes received via SQL*Net from client
171305 bytes sent via SQL*Net to client
2 calls to get snapshot scn: kcmgss
2 calls to kcmgcs
18 consistent gets
1 consistent gets examination
1 consistent gets examination (fastpath)
18 consistent gets from cache
17 consistent gets pin
17 consistent gets pin (fastpath)
2 execute count
1 index range scans
147456 logical read bytes from cache
17 no work - consistent read gets
38 non-idle wait count
2 opened cursors cumulative
1 opened cursors current
2 parse count (total)
1 process last non-idle time
2 session cursor cache count
18 session logical reads
1 sorts (memory)
2024 sorts (rows)
1000 table fetch by rowid
3 user calls
We can see for this first query that returns 1000 rows, it requires just 18 consistent gets, thanks primarily due to the efficient index with the perfect Clustering Factor.
If we look at the cost of this plan:
SQL> SELECT * FROM TABLE(DBMS_XPLAN.display_cursor(sql_id=>'gz5u92hmjwz1h',format=>'ALLSTATS LAST +cost +bytes'));
PLAN_TABLE_OUTPUT
____________________________________________________________________________________________________________________________________
SQL_ID gz5u92hmjwz1h, child number 0
-------------------------------------
select * from bowie where id between 1 and 1000
Plan hash value: 1405654398
---------------------------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows |E-Bytes| Cost (%CPU)| A-Rows | A-Time | Buffers |
---------------------------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | | 21 (100)| 1000 |00:00:00.01 | 18 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE | 1 | 1000 | 108K| 21 (0)| 1000 |00:00:00.01 | 18 |
|* 2 | INDEX RANGE SCAN | BOWIE_ID_I | 1 | 1000 | | 4 (0)| 1000 |00:00:00.01 | 4 |
---------------------------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">=1 AND "ID"<=1000)
We can see the plan has an accurate cost of just 21.
If we now run a similar query that returns a few more rows:
SQL> select * from bowie where id between 1 and 4200;
4,200 rows selected.
PLAN_TABLE_OUTPUT
_______________________________________________________________________________________________________________
SQL_ID c376kdhy5b0x9, child number 0
-------------------------------------
select * from bowie where id between 1 and 4200
Plan hash value: 1405654398
------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows | A-Rows | A-Time | Buffers |
------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | 4200 |00:00:00.01 | 68 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE | 1 | 4200 | 4200 |00:00:00.01 | 68 |
|* 2 | INDEX RANGE SCAN | BOWIE_ID_I | 1 | 4200 | 4200 |00:00:00.01 | 11 |
------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">=1 AND "ID"<=4200)
Note
-----
- automatic DOP: Computed Degree of Parallelism is 1 because of no expensive parallel operation
Statistics
-----------------------------------------------------------
1 CPU used by this session
1 CPU used when call started
1 DB time
11353 RM usage
3 Requests to/from client
2 SQL*Net roundtrips to/from client
59 buffer is not pinned count
8342 buffer is pinned count
324 bytes received via SQL*Net from client
461834 bytes sent via SQL*Net to client
2 calls to get snapshot scn: kcmgss
2 calls to kcmgcs
68 consistent gets
1 consistent gets examination
1 consistent gets examination (fastpath)
68 consistent gets from cache
67 consistent gets pin
67 consistent gets pin (fastpath)
2 execute count
1 index range scans
557056 logical read bytes from cache
67 no work - consistent read gets
73 non-idle wait count
2 opened cursors cumulative
1 opened cursors current
2 parse count (total)
1 process last non-idle time
2 session cursor cache count
68 session logical reads
1 sorts (memory)
2024 sorts (rows)
4200 table fetch by rowid
3 user calls
We can see that it only required just 68 consistent gets to return 4200 rows, thanks to the excellent data clustering and associated very low Clustering Factor.
If we look at the cost of this plan:
SQL> SELECT * FROM TABLE(DBMS_XPLAN.display_cursor(sql_id=>'c376kdhy5b0x9',format=>'ALLSTATS LAST +cost +bytes'));
PLAN_TABLE_OUTPUT
____________________________________________________________________________________________________________________________________
SQL_ID c376kdhy5b0x9, child number 0
-------------------------------------
select * from bowie where id between 1 and 4200
Plan hash value: 1405654398
---------------------------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows |E-Bytes| Cost (%CPU)| A-Rows | A-Time | Buffers |
---------------------------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | | 80 (100)| 4200 |00:00:00.01 | 68 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE | 1 | 4200 | 455K| 80 (0)| 4200 |00:00:00.01 | 68 |
|* 2 | INDEX RANGE SCAN | BOWIE_ID_I | 1 | 4200 | | 11 (0)| 4200 |00:00:00.01 | 11 |
---------------------------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">=1 AND "ID"<=4200)
We can see the cost of the plan is currently a relatively accurate 80.
OK, let’s now perform an update on this table that generates a bunch of migrated rows:
SQL> update bowie set name='THE RISE AND FALL OF BOWIE STARDUST AND THE SPIDERS FROM MARS'; 200,000 rows updated. SQL> commit; Commit complete.
If we now look at the table and index statistics:
SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=>'BOWIE', estimate_percent=> null, no_invalidate=>false); PL/SQL procedure successfully completed. SQL> select table_name, num_rows, blocks from user_tables where table_name='BOWIE'; TABLE_NAME NUM_ROWS BLOCKS _____________ ___________ _________ BOWIE 200000 4906
We can see that the table blocks value has increased to 4906 (previously 3268). This as explained previously is to due in large part to the increased NAME column values and also due to the pointers in the original table blocks that point to the new locations of the migrated rows.
This relates to approximately a 50% increase in table blocks.
If we look at the current index statistics:
SQL> select index_name, blevel, leaf_blocks, clustering_factor from user_indexes where table_name='BOWIE'; INDEX_NAME BLEVEL LEAF_BLOCKS CLUSTERING_FACTOR _____________ _________ ______________ ____________________ BOWIE_ID_I 1 473 3250
We can see that these values are all unchanged, as the ROWIDs in indexes remain unchanged when a row migrates, when ENABLE ROW MOVEMENT is not set.
Therefore, when we re-run these same queries:
SQL> select * from bowie where id between 1 and 1000;
1,000 rows selected.
PLAN_TABLE_OUTPUT
_______________________________________________________________________________________________________________
SQL_ID gz5u92hmjwz1h, child number 0
-------------------------------------
select * from bowie where id between 1 and 1000
Plan hash value: 1405654398
------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows | A-Rows | A-Time | Buffers |
------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | 1000 |00:00:00.01 | 666 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE | 1 | 1000 | 1000 |00:00:00.01 | 666 |
|* 2 | INDEX RANGE SCAN | BOWIE_ID_I | 1 | 1000 | 1000 |00:00:00.01 | 4 |
------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">=1 AND "ID"<=1000)
Note
-----
- automatic DOP: Computed Degree of Parallelism is 1 because of no expensive parallel operation
Statistics
-----------------------------------------------------------
1 DB time
7967 RM usage
3 Requests to/from client
2 SQL*Net roundtrips to/from client
664 buffer is not pinned count
1664 buffer is pinned count
324 bytes received via SQL*Net from client
171419 bytes sent via SQL*Net to client
2 calls to get snapshot scn: kcmgss
2 calls to kcmgcs
666 consistent gets
1 consistent gets examination
1 consistent gets examination (fastpath)
666 consistent gets from cache
665 consistent gets pin
665 consistent gets pin (fastpath)
2 execute count
1 index range scans
5455872 logical read bytes from cache
665 no work - consistent read gets
37 non-idle wait count
2 opened cursors cumulative
1 opened cursors current
2 parse count (total)
1 process last non-idle time
2 session cursor cache count
666 session logical reads
1 sorts (memory)
2024 sorts (rows)
1000 table fetch by rowid
327 table fetch continued row
3 user calls
The number of consistent gets has increased significantly to 666 (previously it was just 18).
Now we can attributed an increase of approximately 50% of the previous consistent gets (18 x 0.50 = 9) due to the 50% increase in table blocks required now to store the rows due to the increased row size.
We can also attribute an additional 327 consistent gets for the table fetch continued row value listed in the statistics, representing the extra consistent gets required to access the migrated rows from their new physical location.
But 18 + 9 + 327 = 354 still leaves us short of the new 666 consistent gets value.
The problem with having to visit another table block to get a row from its new location is that it means Oracle has to re-access again the original table block to get the next row (rather than reading multiple rows with the same consistent get).
So it’s actually approximately 2 x table fetch continued row, by which the number of consistent gets is going to increase when accessing migrated rows (noting that the last migrated row in a block will only incur a additional consistent get as the next table block accessed will differ regardless).
If we look at the new CBO cost for this plan:
SQL> SELECT * FROM TABLE(DBMS_XPLAN.display_cursor(sql_id=>'gz5u92hmjwz1h',format=>'ALLSTATS LAST +cost +bytes'));
PLAN_TABLE_OUTPUT
____________________________________________________________________________________________________________________________________
SQL_ID gz5u92hmjwz1h, child number 0
-------------------------------------
select * from bowie where id between 1 and 1000
Plan hash value: 1405654398
---------------------------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows |E-Bytes| Cost (%CPU)| A-Rows | A-Time | Buffers |
---------------------------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | | 21 (100)| 1000 |00:00:00.01 | 666 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE | 1 | 1000 | 163K| 21 (0)| 1000 |00:00:00.01 | 666 |
|* 2 | INDEX RANGE SCAN | BOWIE_ID_I | 1 | 1000 | | 4 (0)| 1000 |00:00:00.01 | 4 |
---------------------------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">=1 AND "ID"<=1000)
We notice the CBO cost for this plan remains unchanged at 21.
This is totally to be expected, as the index statistics by which the cost of an index scan is calculated are unchanged.
Considering the rough “rule of thumb” is that the CBO cost of an index scan should be in the ball-park of the number of possible IOs, the fact the plan now uses 666 consistent gets highlights this cost of just 21 is no longer as accurate…
If we look at the second SQL that returns 4200 rows:
SQL> select * from bowie where id between 1 and 4200;
4,200 rows selected.
PLAN_TABLE_OUTPUT
_______________________________________________________________________________________________________________
SQL_ID c376kdhy5b0x9, child number 0
-------------------------------------
select * from bowie where id between 1 and 4200
Plan hash value: 1405654398
------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows | A-Rows | A-Time | Buffers |
------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | 4200 |00:00:00.01 | 2771 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE | 1 | 4200 | 4200 |00:00:00.01 | 2771 |
|* 2 | INDEX RANGE SCAN | BOWIE_ID_I | 1 | 4200 | 4200 |00:00:00.01 | 11 |
------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">=1 AND "ID"<=4200)
Note
-----
- automatic DOP: Computed Degree of Parallelism is 1 because of no expensive parallel operation
Statistics
-----------------------------------------------------------
2 CPU used by this session
2 CPU used when call started
2 DB time
14103 RM usage
3 Requests to/from client
2 SQL*Net roundtrips to/from client
2762 buffer is not pinned count
7005 buffer is pinned count
324 bytes received via SQL*Net from client
461947 bytes sent via SQL*Net to client
2 calls to get snapshot scn: kcmgss
2 calls to kcmgcs
2771 consistent gets
1 consistent gets examination
1 consistent gets examination (fastpath)
2771 consistent gets from cache
2770 consistent gets pin
2770 consistent gets pin (fastpath)
2 execute count
1 index range scans
22700032 logical read bytes from cache
2770 no work - consistent read gets
72 non-idle wait count
2 opened cursors cumulative
1 opened cursors current
2 parse count (total)
1 process last non-idle time
2 session cursor cache count
2771 session logical reads
1 sorts (memory)
2024 sorts (rows)
4200 table fetch by rowid
1366 table fetch continued row
3 user calls
We again notice consistent gets has increased significantly to 2771 (previously it was just 68). Again, these additional consistent gets can not be attributed to the extra size of the table and the additional approximate 2 x 1366 table fetch continued row gets.
If we now look at the cost of this plan:
SQL> SELECT * FROM TABLE(DBMS_XPLAN.display_cursor(sql_id=>'c376kdhy5b0x9',format=>'ALLSTATS LAST +cost +bytes'));
PLAN_TABLE_OUTPUT
________________________________________________________________________________________________________________________
____________
SQL_ID c376kdhy5b0x9, child number 0
-------------------------------------
select * from bowie where id between 1 and 4200
Plan hash value: 1405654398
---------------------------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows |E-Bytes| Cost (%CPU)| A-Rows | A-Time | Buffers |
---------------------------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | | 80 (100)| 4200 |00:00:00.01 | 2771 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE | 1 | 4200 | 684K| 80 (0)| 4200 |00:00:00.01 | 2771 |
|* 2 | INDEX RANGE SCAN | BOWIE_ID_I | 1 | 4200 | | 11 (0)| 4200 |00:00:00.01 | 11 |
---------------------------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">=1 AND "ID"<=4200)
We again notice the CBO cost for this plan remains unchanged at 80, again totally expected as the underlying index statistics have remain unchanged after the update statement.
But again, not necessary as accurate a cost as it was previously…
If we repeat this demo, but this time on a table with ENABLE ROW MOVEMENT enabled:
SQL> create table bowie2(id number, code1 number, code2 number, code3 number, code4 number, code5 number, code6 number, code7 number, code8 number, code9 number, code10 number, code11 number, code12 number, code13 number, code14 number, code15 number, code16 number, code17 number, code18 number, code19 number, code20 number, name varchar2(142)) PCTFREE 0 ENABLE ROW MOVEMENT;
Table BOWIE2 created.
SQL> insert into bowie2 SELECT rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, 'BOWIE' FROM dual CONNECT BY LEVEL <= 200000;
200,000 rows inserted.
SQL> commit;
Commit complete.
SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=>'BOWIE2', estimate_percent=> null, no_invalidate=>false);
PL/SQL procedure successfully completed.
SQL> create index bowie2_id_i on bowie2(id);
Index BOWIE2_ID_I created.
SQL> select table_name, num_rows, blocks from user_tables where table_name='BOWIE2';
TABLE_NAME NUM_ROWS BLOCKS
_____________ ___________ _________
BOWIE2 200000 3268
SQL> select index_name, blevel, leaf_blocks, clustering_factor from user_indexes where table_name='BOWIE2';
INDEX_NAME BLEVEL LEAF_BLOCKS CLUSTERING_FACTOR
__________________ _________ ______________ ____________________
BOWIE2_ID_I 1 473 3250
The table and index statistics are currently identical to the previous demo.
If we run the same two equivalent queries:
SQL> select * from bowie2 where id between 1 and 1000;
1,000 rows selected.
PLAN_TABLE_OUTPUT
________________________________________________________________________________________________________________
SQL_ID gtkw2704bxj7q, child number 0
-------------------------------------
select * from bowie2 where id between 1 and 1000
Plan hash value: 3243780227
-------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows | A-Rows | A-Time | Buffers |
-------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | 1000 |00:00:00.01 | 18 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE2 | 1 | 1000 | 1000 |00:00:00.01 | 18 |
|* 2 | INDEX RANGE SCAN | BOWIE2_ID_I | 1 | 1000 | 1000 |00:00:00.01 | 4 |
-------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">=1 AND "ID"<=1000)
Note
-----
- automatic DOP: Computed Degree of Parallelism is 1 because of no expensive parallel operation
Statistics
-----------------------------------------------------------
1 CPU used by this session
1 CPU used when call started
7909 RM usage
3 Requests to/from client
2 SQL*Net roundtrips to/from client
16 buffer is not pinned count
1985 buffer is pinned count
325 bytes received via SQL*Net from client
171306 bytes sent via SQL*Net to client
2 calls to get snapshot scn: kcmgss
2 calls to kcmgcs
18 consistent gets
1 consistent gets examination
1 consistent gets examination (fastpath)
18 consistent gets from cache
17 consistent gets pin
17 consistent gets pin (fastpath)
2 execute count
1 index range scans
147456 logical read bytes from cache
17 no work - consistent read gets
37 non-idle wait count
2 opened cursors cumulative
1 opened cursors current
2 parse count (total)
1 process last non-idle time
2 session cursor cache count
18 session logical reads
1 sorts (memory)
2024 sorts (rows)
1000 table fetch by rowid
3 user calls
SQL> SELECT * FROM TABLE(DBMS_XPLAN.display_cursor(sql_id=>'gtkw2704bxj7q',format=>'ALLSTATS LAST +cost +bytes'));
PLAN_TABLE_OUTPUT
_____________________________________________________________________________________________________________________________________
SQL_ID gtkw2704bxj7q, child number 0
-------------------------------------
select * from bowie2 where id between 1 and 1000
Plan hash value: 3243780227
----------------------------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows |E-Bytes| Cost (%CPU)| A-Rows | A-Time | Buffers |
----------------------------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | | 21 (100)| 1000 |00:00:00.01 | 18 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE2 | 1 | 1000 | 108K| 21 (0)| 1000 |00:00:00.01 | 18 |
|* 2 | INDEX RANGE SCAN | BOWIE2_ID_I | 1 | 1000 | | 4 (0)| 1000 |00:00:00.01 | 4 |
----------------------------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">=1 AND "ID"<=1000)
SQL> select * from bowie2 where id between 1 and 4200;
4,200 rows selected.
PLAN_TABLE_OUTPUT
________________________________________________________________________________________________________________
SQL_ID 25qktyn35b662, child number 0
-------------------------------------
select * from bowie2 where id between 1 and 4200
Plan hash value: 3243780227
-------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows | A-Rows | A-Time | Buffers |
-------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | 4200 |00:00:00.01 | 68 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE2 | 1 | 4200 | 4200 |00:00:00.01 | 68 |
|* 2 | INDEX RANGE SCAN | BOWIE2_ID_I | 1 | 4200 | 4200 |00:00:00.01 | 11 |
-------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">=1 AND "ID"<=4200)
Note
-----
- automatic DOP: Computed Degree of Parallelism is 1 because of no expensive parallel operation
Statistics
-----------------------------------------------------------
1 CPU used by this session
1 CPU used when call started
2 DB time
13157 RM usage
3 Requests to/from client
2 SQL*Net roundtrips to/from client
59 buffer is not pinned count
8342 buffer is pinned count
325 bytes received via SQL*Net from client
461838 bytes sent via SQL*Net to client
2 calls to get snapshot scn: kcmgss
2 calls to kcmgcs
68 consistent gets
1 consistent gets examination
1 consistent gets examination (fastpath)
68 consistent gets from cache
67 consistent gets pin
67 consistent gets pin (fastpath)
2 execute count
1 index range scans
557056 logical read bytes from cache
67 no work - consistent read gets
73 non-idle wait count
2 opened cursors cumulative
1 opened cursors current
2 parse count (total)
1 process last non-idle time
2 session cursor cache count
68 session logical reads
1 sorts (memory)
2024 sorts (rows)
4200 table fetch by rowid
3 user calls
SQL> SELECT * FROM TABLE(DBMS_XPLAN.display_cursor(sql_id=>'25qktyn35b662',format=>'ALLSTATS LAST +cost +bytes'));
PLAN_TABLE_OUTPUT
_____________________________________________________________________________________________________________________________________
SQL_ID 25qktyn35b662, child number 0
-------------------------------------
select * from bowie2 where id between 1 and 4200
Plan hash value: 3243780227
----------------------------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows |E-Bytes| Cost (%CPU)| A-Rows | A-Time | Buffers |
----------------------------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | | 80 (100)| 4200 |00:00:00.01 | 68 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE2 | 1 | 4200 | 455K| 80 (0)| 4200 |00:00:00.01 | 68 |
|* 2 | INDEX RANGE SCAN | BOWIE2_ID_I | 1 | 4200 | | 11 (0)| 4200 |00:00:00.01 | 11 |
----------------------------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">=1 AND "ID"<=4200)
With identical table/index statistics, we notice as expected that both SQLs have the same consistent gets and CBO costs as with the previous demo.
If we now repeat the equivalent Update statement:
SQL> update bowie2 set name='THE RISE AND FALL OF BOWIE STARDUST AND THE SPIDERS FROM MARS'; 200,000 rows updated. SQL> commit; Commit complete. SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=>'BOWIE2', estimate_percent=> null, no_invalidate=>false); PL/SQL procedure successfully completed.
If we look at the table statistics:
SQL> select table_name, num_rows, blocks from user_tables where table_name='BOWIE2'; TABLE_NAME NUM_ROWS BLOCKS _____________ ___________ _________ BOWIE2 200000 4654
We notice the number of table blocks has increased to 4654 due to the increased row lengths, but not as much as with the previous demo (where table blocks increased to 4906) as in this scenario, Oracle does not have to store the row location pointers in the original blocks for the migrated rows.
If we look at the index statistics:
SQL> select index_name, blevel, leaf_blocks, clustering_factor from user_indexes where table_name='BOWIE2';
INDEX_NAME BLEVEL LEAF_BLOCKS CLUSTERING_FACTOR
______________ _________ ______________ ____________________
BOWIE2_ID_I 2 945 109061
We notice that these are substantially different from the first demo, where ROWIDs for migrated rows are not updated on the fly.
By now updating the ROWIDs, the indexes can possibly increase in size as they have to store both the previous and new ROWIDs in separate index entries and hence Oracle is more likely to perform additional index block splits (as I discussed in my previous post).
The LEAF_BLOCKS are now 945 (previously 473) and even the BLEVEL has increased from 1 to 2.
Additionally, and perhaps importantly for specific key indexes, the Clustering Factor value of indexes can also be impacted. By migrating rows and physically storing them in different locations, this can potentially detrimentally impact the tight clustering of rows based on specific column values.
The Clustering Factor for the index on the monotonically increased ID column has now increased significantly to 109061, up from the previously perfect 3250.
So columns that have naturally good clustering (e.g.: monotonically increasing values such as IDs and dates) or have been manually well clustered for performance purposes, can have the Clustering Factor of associated indexes detrimentally impacted by migrated rows.
If we re-run the first query:
SQL> select * from bowie2 where id between 1 and 1000;
1,000 rows selected.
PLAN_TABLE_OUTPUT
________________________________________________________________________________________________________________
SQL_ID gtkw2704bxj7q, child number 0
-------------------------------------
select * from bowie2 where id between 1 and 1000
Plan hash value: 3243780227
-------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows | A-Rows | A-Time | Buffers |
-------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | 1000 |00:00:00.01 | 639 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE2 | 1 | 1000 | 1000 |00:00:00.01 | 639 |
|* 2 | INDEX RANGE SCAN | BOWIE2_ID_I | 1 | 1000 | 1000 |00:00:00.01 | 7 |
-------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">=1 AND "ID"<=1000)
Note
-----
- automatic DOP: Computed Degree of Parallelism is 1 because of no expensive parallel operation
Statistics
-----------------------------------------------------------
1 CPU used by this session
1 CPU used when call started
1 DB time
15262 RM usage
3 Requests to/from client
2 SQL*Net roundtrips to/from client
634 buffer is not pinned count
1367 buffer is pinned count
325 bytes received via SQL*Net from client
171421 bytes sent via SQL*Net to client
2 calls to get snapshot scn: kcmgss
2 calls to kcmgcs
639 consistent gets
2 consistent gets examination
2 consistent gets examination (fastpath)
639 consistent gets from cache
637 consistent gets pin
637 consistent gets pin (fastpath)
2 execute count
1 index range scans
5234688 logical read bytes from cache
637 no work - consistent read gets
38 non-idle wait count
1 non-idle wait time
2 opened cursors cumulative
1 opened cursors current
2 parse count (total)
1 process last non-idle time
2 session cursor cache count
639 session logical reads
1 sorts (memory)
2024 sorts (rows)
1000 table fetch by rowid
3 user calls
I discussed in a previous post how by updating the ROWIDs of migrated rows we can improve performance, as Oracle can go directly to the correct new physical location of a migrated row.
But for some specific indexes, where data clustering is crucial, and we have a significant number migrated rows, this might not necessarily be the case.
We can see consistent gets here has increased to 639 (previously is was just 21), and so not hugely different from the 666 consistent gets required to fetch the migrated rows when the ROWIDs were not updated in the first demo.
If we look at the CBO costings:
SQL> SELECT * FROM TABLE(DBMS_XPLAN.display_cursor(sql_id=>'gtkw2704bxj7q',format=>'ALLSTATS LAST +cost +bytes'));
PLAN_TABLE_OUTPUT
_____________________________________________________________________________________________________________________________________
SQL_ID gtkw2704bxj7q, child number 0
-------------------------------------
select * from bowie2 where id between 1 and 1000
Plan hash value: 3243780227
----------------------------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows |E-Bytes| Cost (%CPU)| A-Rows | A-Time | Buffers |
----------------------------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | | 553 (100)| 1000 |00:00:00.01 | 639 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE2 | 1 | 1000 | 163K| 553 (0)| 1000 |00:00:00.01 | 639 |
|* 2 | INDEX RANGE SCAN | BOWIE2_ID_I | 1 | 1000 | | 7 (0)| 1000 |00:00:00.01 | 7 |
----------------------------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">=1 AND "ID"<=1000)
We can see the CBO cost has increased significantly to 553 (previously it was just 21).
With a much increased Clustering Factor, this will obviously impact the CBO costs of associated index scans.
In very extreme cases, these possible changes in the Clustering Factor can even impact the viability of using the index.
If we re-run the second query returning the 4200 rows:
SQL> select * from bowie2 where id between 1 and 4200;
4,200 rows selected.
PLAN_TABLE_OUTPUT
_____________________________________________________________________________________________________
SQL_ID 25qktyn35b662, child number 0
-------------------------------------
select * from bowie2 where id between 1 and 4200
Plan hash value: 1495904576
----------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows | A-Rows | A-Time | Buffers |
----------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | 4200 |00:00:00.02 | 4572 |
|* 1 | TABLE ACCESS STORAGE FULL | BOWIE2 | 1 | 4200 | 4200 |00:00:00.02 | 4572 |
----------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - storage(("ID"<=4200 AND "ID">=1))
filter(("ID"<=4200 AND "ID">=1))
We can see that the CBO has now chosen to perform a Full Table Scan (FTS), rather than use the now less efficient index to return this number of rows.
If we look at the CBO costings of this FTS plan:
SQL> SELECT * FROM TABLE(DBMS_XPLAN.display_cursor(sql_id=>'25qktyn35b662',format=>'ALLSTATS LAST +cost +bytes'));
PLAN_TABLE_OUTPUT
______________________________________________________________________________________________________________________
SQL_ID 25qktyn35b662, child number 0
-------------------------------------
select * from bowie2 where id between 1 and 4200
Plan hash value: 1495904576
-------------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows |E-Bytes| Cost (%CPU)| A-Rows | A-Time | Buffers |
-------------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | | 1264 (100)| 4200 |00:00:00.02 | 4572 |
|* 1 | TABLE ACCESS STORAGE FULL | BOWIE2 | 1 | 4200 | 684K| 1264 (1)| 4200 |00:00:00.02 | 4572 |
-------------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - storage(("ID"<=4200 AND "ID">=1))
filter(("ID"<=4200 AND "ID">=1))
The cost of the FTS plan is 1264.
If we compare this is a plan that used the index:
SQL> select /*+ index (bowie2) */ * from bowie2 where id between 1 and 4200;
4,200 rows selected.
PLAN_TABLE_OUTPUT
________________________________________________________________________________________________________________
SQL_ID bzm2vhchqpq7w, child number 0
-------------------------------------
select /*+ index (bowie2) */ * from bowie2 where id between 1 and 4200
Plan hash value: 3243780227
-------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows | A-Rows | A-Time | Buffers |
-------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | 4200 |00:00:00.01 | 2665 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE2 | 1 | 4200 | 4200 |00:00:00.01 | 2665 |
|* 2 | INDEX RANGE SCAN | BOWIE2_ID_I | 1 | 4200 | 4200 |00:00:00.01 | 21 |
-------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">=1 AND "ID"<=4200)
Note
-----
- automatic DOP: Computed Degree of Parallelism is 1 because of no expensive parallel operation
Statistics
-----------------------------------------------------------
2 CPU used by this session
2 CPU used when call started
2 DB time
14531 RM usage
3 Requests to/from client
2 SQL*Net roundtrips to/from client
2646 buffer is not pinned count
5755 buffer is pinned count
348 bytes received via SQL*Net from client
462143 bytes sent via SQL*Net to client
2 calls to get snapshot scn: kcmgss
2 calls to kcmgcs
2665 consistent gets
2 consistent gets examination
2 consistent gets examination (fastpath)
2665 consistent gets from cache
2663 consistent gets pin
2663 consistent gets pin (fastpath)
2 execute count
1 index range scans
21831680 logical read bytes from cache
2663 no work - consistent read gets
73 non-idle wait count
2 opened cursors cumulative
1 opened cursors current
2 parse count (total)
3 process last non-idle time
2 session cursor cache count
2665 session logical reads
1 sorts (memory)
2024 sorts (rows)
4200 table fetch by rowid
3 user calls
SQL> SELECT * FROM TABLE(DBMS_XPLAN.display_cursor(sql_id=>'bzm2vhchqpq7w',format=>'ALLSTATS LAST +cost +bytes'));
PLAN_TABLE_OUTPUT
_____________________________________________________________________________________________________________________________________
SQL_ID bzm2vhchqpq7w, child number 0
-------------------------------------
select /*+ index (bowie2) */ * from bowie2 where id between 1 and 4200
Plan hash value: 3243780227
----------------------------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows |E-Bytes| Cost (%CPU)| A-Rows | A-Time | Buffers |
----------------------------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | | 2314 (100)| 4200 |00:00:00.01 | 2665 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED | BOWIE2 | 1 | 4200 | 684K| 2314 (1)| 4200 |00:00:00.01 | 2665 |
|* 2 | INDEX RANGE SCAN | BOWIE2_ID_I | 1 | 4200 | | 22 (0)| 4200 |00:00:00.01 | 21 |
----------------------------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID">=1 AND "ID"<=4200)
The cost of using the index to retrieve the 4200 rows is 2310, more than the 1264 of the FTS.
For the vast majority of indexes, updating the ROWIDs for migrated rows will result in better performance, as such indexes will be able to directly access the correct new physical location of migrated rows, rather than having to visit the original table block and then follow the stored pointer to the new table block.
But for some very specific indexes, where data clustering is crucial, AND we have a significant number migrated rows, this might not necessarily be the case. The performance benefit might be minimal at best.
That’s more than enough for one post 🙂
In my next post, I’ll discuss how to potentially remedy these performance implications, both for tables with or without ENABLE TABLE MOVEMENT enabled…
Some Things To Consider Now ROWIDs Are Updated When Rows Migrate Part II (“Look Back In Anger”) February 24, 2023
Posted by Richard Foote in 19c, Autonomous Database, Autonomous Transaction Processing, BLEVEL, Changing ROWID, Index Internals, Leaf Blocks, Migrated Rows, Oracle, Oracle Blog, Oracle Cloud, Oracle General, Oracle Indexes, Oracle19c, Richard's Blog, ROWID.3 comments
Some weekend reading…
In my previous post, I discussed a couple of potential areas of concern with the ROWIDs of migrated rows now being updated on the fly in Oracle Autonomous Databases, namely that it can cause issues with applications that reply on stored ROWIDs not changing and that there are additional resources required to maintain such ROWIDs in corresponding indexes, especially if there are many indexes on a table.
In this post, I’ll discuss another issue to just bear in mind with this change in behaviour.
To illustrate, I’ll run a demo similar to the previous post, first creating and populating a table with no ENABLE ROW MOVEMENT set:
SQL> create table ziggy(id number, code1 number, code2 number, code3 number, code4 number, code5 number, code6 number, code7 number, code8 number, code9 number, code10 number, code11 number, code12 number, code13 number, code14 number, code15 number, code16 number, code17 number, code18 number, code19 number, code20 number, name varchar2(142)) PCTFREE 0; Table ZIGGY created. SQL> insert into ziggy SELECT rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, 'BOWIE' FROM dual CONNECT BY LEVEL <= 200000; 200,000 rows inserted. SQL> commit; Commit complete. SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=>'ZIGGY'); PL/SQL procedure successfully completed. SQL> analyze table ziggy compute statistics; Table ZIGGY analyzed.
I’ll next create a bunch of indexes on the table:
SQL> create index ziggy_id_i on ziggy(id); Index ZIGGY_ID_I created. SQL> create index ziggy_code1_i on ziggy(code1); Index ZIGGY_CODE1_I created. SQL> create index ziggy_code2_i on ziggy(code2); Index ZIGGY_CODE2_I created. SQL> create index ziggy_code3_i on ziggy(code3); Index ZIGGY_CODE3_I created. SQL> create index ziggy_code4_i on ziggy(code4); Index ZIGGY_CODE4_I created. SQL> create index ziggy_code5_i on ziggy(code5); Index ZIGGY_CODE5_I created. SQL> create index ziggy_code6_i on ziggy(code6); Index ZIGGY_CODE6_I created. SQL> create index ziggy_code7_i on ziggy(code7); Index ZIGGY_CODE7_I created. SQL> create index ziggy_code8_i on ziggy(code8); Index ZIGGY_CODE8_I created. SQL> create index ziggy_code9_i on ziggy(code9); Index ZIGGY_CODE9_I created. SQL> create index ziggy_code10_i on ziggy(code10); Index ZIGGY_CODE10_I created. SQL> create index ziggy_code11_i on ziggy(code11); Index ZIGGY_CODE11_I created. SQL> create index ziggy_code12_i on ziggy(code12); Index ZIGGY_CODE12_I created. SQL> create index ziggy_code13_i on ziggy(code13); Index ZIGGY_CODE13_I created. SQL> create index ziggy_code14_i on ziggy(code14); Index ZIGGY_CODE14_I created. SQL> create index ziggy_code15_i on ziggy(code15); Index ZIGGY_CODE15_I created. SQL> create index ziggy_code16_i on ziggy(code16); Index ZIGGY_CODE16_I created. SQL> create index ziggy_code17_i on ziggy(code17); Index ZIGGY_CODE17_I created. SQL> create index ziggy_code18_i on ziggy(code18); Index ZIGGY_CODE18_I created. SQL> create index ziggy_code19_i on ziggy(code19); Index ZIGGY_CODE19_I created. SQL> create index ziggy_code20_i on ziggy(code20); Index ZIGGY_CODE20_I created.
Let’s take note of the size of the table and its associated indexes:
SQL> select table_name, num_rows, blocks, empty_blocks, avg_space, avg_row_len, chain_cnt
from user_tables where table_name='ZIGGY';
TABLE_NAME NUM_ROWS BLOCKS EMPTY_BLOCKS AVG_SPACE AVG_ROW_LEN CHAIN_CNT
_____________ ___________ _________ _______________ ____________ ______________ ____________
ZIGGY 200000 3268 60 857 113 0
SQL> select index_name, blevel, leaf_blocks, clustering_factor from user_indexes
where table_name='ZIGGY';
INDEX_NAME BLEVEL LEAF_BLOCKS CLUSTERING_FACTOR
_________________ _________ ______________ ____________________
ZIGGY_CODE20_I 1 473 3250
ZIGGY_ID_I 1 473 3250
ZIGGY_CODE1_I 1 473 3250
ZIGGY_CODE2_I 1 473 3250
ZIGGY_CODE3_I 1 473 3250
ZIGGY_CODE4_I 1 473 3250
ZIGGY_CODE5_I 1 473 3250
ZIGGY_CODE6_I 1 473 3250
ZIGGY_CODE7_I 1 473 3250
ZIGGY_CODE8_I 1 473 3250
ZIGGY_CODE9_I 1 473 3250
ZIGGY_CODE10_I 1 473 3250
ZIGGY_CODE11_I 1 473 3250
ZIGGY_CODE12_I 1 473 3250
ZIGGY_CODE13_I 1 473 3250
ZIGGY_CODE14_I 1 473 3250
ZIGGY_CODE15_I 1 473 3250
ZIGGY_CODE16_I 1 473 3250
ZIGGY_CODE17_I 1 473 3250
ZIGGY_CODE18_I 1 473 3250
ZIGGY_CODE19_I 1 473 3250
We next perform an update on the table that will increase the row size sufficiently to result in a bunch of migrated rows:
SQL> update ziggy set name='THE RISE AND FALL OF ZIGGY STARDUST AND THE SPIDERS FROM MARS'; 200,000 rows updated. Elapsed: 00:00:07.716
I’ll then perform a COMMIT and look at differences in the table statistics:
SQL> commit; Commit complete. SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=>'ZIGGY'); PL/SQL procedure successfully completed. SQL> analyze table ziggy compute statistics; Table ZIGGY analyzed. SQL> select table_name, num_rows, blocks, empty_blocks, avg_space, avg_row_len, chain_cnt from user_tables where table_name='ZIGGY'; TABLE_NAME NUM_ROWS BLOCKS EMPTY_BLOCKS AVG_SPACE AVG_ROW_LEN CHAIN_CNT _____________ ___________ _________ _______________ ____________ ______________ ____________ ZIGGY 200000 4906 86 415 170 56186
We can see that there are indeed a bunch of migrated/chained rows, some 56186 of them.
We also notice that the size of the table has increased to 4906 blocks (previously is was 3268). This increase is in large part due to the increased size of the NAME column value, but also partly due to the storage allocated to pointers that are stored in the original block to denote the new location of the migrated rows (as discussed previously here).
If we look at the current state of the indexes:
SQL> select index_name, blevel, leaf_blocks, clustering_factor
from user_indexes where table_name='ZIGGY';
INDEX_NAME BLEVEL LEAF_BLOCKS CLUSTERING_FACTOR
_________________ _________ ______________ ____________________
ZIGGY_CODE7_I 1 473 3250
ZIGGY_CODE8_I 1 473 3250
ZIGGY_CODE9_I 1 473 3250
ZIGGY_CODE10_I 1 473 3250
ZIGGY_CODE11_I 1 473 3250
ZIGGY_CODE12_I 1 473 3250
ZIGGY_CODE13_I 1 473 3250
ZIGGY_CODE14_I 1 473 3250
ZIGGY_CODE15_I 1 473 3250
ZIGGY_CODE16_I 1 473 3250
ZIGGY_CODE17_I 1 473 3250
ZIGGY_CODE18_I 1 473 3250
ZIGGY_CODE19_I 1 473 3250
ZIGGY_CODE20_I 1 473 3250
ZIGGY_ID_I 1 473 3250
ZIGGY_CODE1_I 1 473 3250
ZIGGY_CODE2_I 1 473 3250
ZIGGY_CODE3_I 1 473 3250
ZIGGY_CODE4_I 1 473 3250
ZIGGY_CODE5_I 1 473 3250
ZIGGY_CODE6_I 1 473 3250
We notice that the indexes remain unchanged. As the table does NOT have ENABLE ROW MOVEMENT set, the indexes are NOT updated at all as part of the migrated row process (thus the same behaviour as non-autonomous database environments).
However, if I perform the same demo but instead perform a ROLLBACK of the transaction rather than the commit:
SQL> rollback; Rollback complete. Elapsed: 00:00:06.919
Note that the rollback takes 00:00:06.919 to complete.
If we now look at the size of the table and corresponding indexes:
SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=>'ZIGGY'); PL/SQL procedure successfully completed. SQL> analyze table ziggy compute statistics; Table ZIGGY analyzed. SQL> select table_name, num_rows, blocks, empty_blocks, avg_space, avg_row_len, chain_cnt from user_tables where table_name='ZIGGY'; TABLE_NAME NUM_ROWS BLOCKS EMPTY_BLOCKS AVG_SPACE AVG_ROW_LEN CHAIN_CNT _____________ ___________ _________ _______________ ____________ ______________ ____________ ZIGGY 200000 4906 86 2899 113 0
We notice that the size of the table has increased to be the same 4906 blocks as when we performed the commit. The extra storage remains allocated even after the rollback operation.
That’s because Oracle does NOT deallocate any additional storage that might have been consumed during the original transaction. We notice that the AVG_SPACE has increased substantially as a result (now 2899, previously it was just 857).
If we look at the current state of the indexes after the rollback:
SQL> select index_name, blevel, leaf_blocks, clustering_factor
from user_indexes where table_name='ZIGGY';
INDEX_NAME BLEVEL LEAF_BLOCKS CLUSTERING_FACTOR
_________________ _________ ______________ ____________________
ZIGGY_CODE20_I 1 473 3250
ZIGGY_ID_I 1 473 3250
ZIGGY_CODE1_I 1 473 3250
ZIGGY_CODE2_I 1 473 3250
ZIGGY_CODE3_I 1 473 3250
ZIGGY_CODE4_I 1 473 3250
ZIGGY_CODE5_I 1 473 3250
ZIGGY_CODE6_I 1 473 3250
ZIGGY_CODE7_I 1 473 3250
ZIGGY_CODE8_I 1 473 3250
ZIGGY_CODE9_I 1 473 3250
ZIGGY_CODE10_I 1 473 3250
ZIGGY_CODE11_I 1 473 3250
ZIGGY_CODE12_I 1 473 3250
ZIGGY_CODE13_I 1 473 3250
ZIGGY_CODE14_I 1 473 3250
ZIGGY_CODE15_I 1 473 3250
ZIGGY_CODE16_I 1 473 3250
ZIGGY_CODE17_I 1 473 3250
ZIGGY_CODE18_I 1 473 3250
ZIGGY_CODE19_I 1 473 3250
We notice that all the indexes again remain unchanged. As the indexes are not updated during the transaction, this is of course to be expected.
Let’s now repeat the same demo, but this time on a table with ENABLE ROW MOVEMENT set:
SQL> create table ziggy2(id number, code1 number, code2 number, code3 number, code4 number, code5 number, code6 number, code7 number, code8 number, code9 number, code10 number, code11 number, code12 number, code13 number, code14 number, code15 number, code16 number, code17 number, code18 number, code19 number, code20 number, name varchar2(142)) PCTFREE 0 ENABLE ROW MOVEMENT; Table ZIGGY2 created. SQL> insert into ziggy2 SELECT rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, rownum, 'BOWIE' FROM dual CONNECT BY LEVEL <= 200000; 200,000 rows inserted. SQL> commit; Commit complete. SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=>'ZIGGY2'); PL/SQL procedure successfully completed. SQL> analyze table ziggy2 compute statistics; Table ZIGGY2 analyzed. SQL> create index ziggy2_id_i on ziggy2(id); Index ZIGGY2_ID_I created. SQL> create index ziggy2_code1_i on ziggy2(code1); Index ZIGGY2_CODE1_I created. SQL> create index ziggy2_code2_i on ziggy2(code2); Index ZIGGY2_CODE2_I created. SQL> create index ziggy2_code3_i on ziggy2(code3); Index ZIGGY2_CODE3_I created. SQL> create index ziggy2_code4_i on ziggy2(code4); Index ZIGGY2_CODE4_I created. SQL> create index ziggy2_code5_i on ziggy2(code5); Index ZIGGY2_CODE5_I created. SQL> create index ziggy2_code6_i on ziggy2(code6); Index ZIGGY2_CODE6_I created. SQL> create index ziggy2_code7_i on ziggy2(code7); Index ZIGGY2_CODE7_I created. SQL> create index ziggy2_code8_i on ziggy2(code8); Index ZIGGY2_CODE8_I created. SQL> create index ziggy2_code9_i on ziggy2(code9); Index ZIGGY2_CODE9_I created. SQL> create index ziggy2_code10_i on ziggy2(code10); Index ZIGGY2_CODE10_I created. SQL> create index ziggy2_code11_i on ziggy2(code11); Index ZIGGY2_CODE11_I created. SQL> create index ziggy2_code12_i on ziggy2(code12); Index ZIGGY2_CODE12_I created. SQL> create index ziggy2_code13_i on ziggy2(code13); Index ZIGGY2_CODE13_I created. SQL> create index ziggy2_code14_i on ziggy2(code14); Index ZIGGY2_CODE14_I created. SQL> create index ziggy2_code15_i on ziggy2(code15); Index ZIGGY2_CODE15_I created. SQL> create index ziggy2_code16_i on ziggy2(code16); Index ZIGGY2_CODE16_I created. SQL> create index ziggy2_code17_i on ziggy2(code17); Index ZIGGY2_CODE17_I created. SQL> create index ziggy2_code18_i on ziggy2(code18); Index ZIGGY2_CODE18_I created. SQL> create index ziggy2_code19_i on ziggy2(code19); Index ZIGGY2_CODE19_I created. SQL> create index ziggy2_code20_i on ziggy2(code20); Index ZIGGY2_CODE20_I created.
The table and indexes all have the same initial size as the previous example:
SQL> select table_name, num_rows, blocks, empty_blocks, avg_space, avg_row_len, chain_cnt
from user_tables where table_name='ZIGGY2';
TABLE_NAME NUM_ROWS BLOCKS EMPTY_BLOCKS AVG_SPACE AVG_ROW_LEN CHAIN_CNT
_____________ ___________ _________ _______________ ____________ ______________ ____________
ZIGGY2 200000 3268 60 857 113 0
SQL> select index_name, blevel, leaf_blocks, clustering_factor from user_indexes
where table_name='ZIGGY2';
INDEX_NAME BLEVEL LEAF_BLOCKS CLUSTERING_FACTOR
__________________ _________ ______________ ____________________
ZIGGY2_ID_I 1 473 3250
ZIGGY2_CODE1_I 1 473 3250
ZIGGY2_CODE2_I 1 473 3250
ZIGGY2_CODE3_I 1 473 3250
ZIGGY2_CODE4_I 1 473 3250
ZIGGY2_CODE5_I 1 473 3250
ZIGGY2_CODE6_I 1 473 3250
ZIGGY2_CODE7_I 1 473 3250
ZIGGY2_CODE8_I 1 473 3250
ZIGGY2_CODE9_I 1 473 3250
ZIGGY2_CODE10_I 1 473 3250
ZIGGY2_CODE11_I 1 473 3250
ZIGGY2_CODE12_I 1 473 3250
ZIGGY2_CODE13_I 1 473 3250
ZIGGY2_CODE14_I 1 473 3250
ZIGGY2_CODE15_I 1 473 3250
ZIGGY2_CODE16_I 1 473 3250
ZIGGY2_CODE17_I 1 473 3250
ZIGGY2_CODE18_I 1 473 3250
ZIGGY2_CODE19_I 1 473 3250
ZIGGY2_CODE20_I 1 473 3250
If we now perform the same Update followed by the commit:
SQL> update ziggy2 set name='THE RISE AND FALL OF ZIGGY STARDUST AND THE SPIDERS FROM MARS'; 200,000 rows updated. Elapsed: 00:00:33.390 SQL> commit; Commit complete.
If we look at the current size of the table after the update:
SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=>'ZIGGY2'); PL/SQL procedure successfully completed. SQL> analyze table ziggy2 compute statistics; Table ZIGGY2 analyzed. SQL> select table_name, num_rows, blocks, empty_blocks, avg_space, avg_row_len, chain_cnt from user_tables where table_name='ZIGGY2'; TABLE_NAME NUM_ROWS BLOCKS EMPTY_BLOCKS AVG_SPACE AVG_ROW_LEN CHAIN_CNT _____________ ___________ _________ _______________ ____________ ______________ ____________ ZIGGY2 200000 4654 82 367 169 0
We notice it has increased to 4654 blocks (previously 3268). But the table is not quite as large in size as in the first demo, where the table grew to 4906 blocks (so 252 fewer blocks).
The increase in table size is now due entirely as a result in the increased NAME column values, as Oracle has not had to consume any storage for pointers in the original table blocks to denote a new location of the rows, as the ROWIDs are now updated in the indexes on the fly.
So this is a positive, a DECREASE in the comparative size of table after such updates that migrate rows.
And of course, there are no migrated/chained rows.
But if we look at the index statistics after the commit:
SQL> select index_name, blevel, leaf_blocks, clustering_factor
from user_indexes where table_name='ZIGGY2';
INDEX_NAME BLEVEL LEAF_BLOCKS CLUSTERING_FACTOR
__________________ _________ ______________ ____________________
ZIGGY2_ID_I 2 945 109061
ZIGGY2_CODE1_I 2 945 109061
ZIGGY2_CODE2_I 2 945 109061
ZIGGY2_CODE3_I 2 945 109061
ZIGGY2_CODE4_I 2 945 109061
ZIGGY2_CODE5_I 2 945 109061
ZIGGY2_CODE6_I 2 945 109061
ZIGGY2_CODE7_I 2 945 109061
ZIGGY2_CODE8_I 2 945 109061
ZIGGY2_CODE9_I 2 945 109061
ZIGGY2_CODE10_I 2 945 109061
ZIGGY2_CODE11_I 2 945 109061
ZIGGY2_CODE12_I 2 945 109061
ZIGGY2_CODE13_I 2 945 109061
ZIGGY2_CODE14_I 2 945 109061
ZIGGY2_CODE15_I 2 945 109061
ZIGGY2_CODE16_I 2 945 109061
ZIGGY2_CODE17_I 2 945 109061
ZIGGY2_CODE18_I 2 945 109061
ZIGGY2_CODE19_I 2 945 109061
ZIGGY2_CODE20_I 2 945 109061
We notice that ALL the indexes have significantly increased in size and now have 945 leaf blocks (previously it was just 473 leaf blocks). Additionally as a result of this increase in index size, the BLEVEL of the indexes has also increased and are now 2 (previously it was 1).
Here’s the thing. As I’ve discussed many times before, when Oracle performs an “Update” of an index entry, this is actually implemented as a Delete/Insert operation. By changing the ROWID of an index entry, Oracle first deletes the original index entry and inserts a new index entry with the new ROWID. So the previous index entry remains (with the space likely eventually reused by another new index entry in the future).
So these “on the fly” updates of the indexes to keep the ROWIDs current due to row migrations increases the likelihood of index block splits and the subsequent increase in index storage allocations.
In very rare, extreme cases (and this demo is indeed an extreme case as I’m updating all rows in my table), this extra index storage could potentially result in an increase in the index BLEVEL.
However, in most scenarios, this increase in index storage is likely to be moderate and result in extra index storage that will eventually be consumed by subsequent new rows anyways.
If instead of the commit operation, we instead performed a rollback:
SQL> rollback; Rollback complete. Elapsed: 00:00:36.639
We notice that the rollback takes considerably longer at 00:00:36.639 (previously in the first demo, it was just 00:00:06.919).
As the ROWIDs are now all updated on the fly on all the corresponding indexes, there’s that much more data that needs to be rolled back within these indexes.
If we look at the current size of the table after the rollback:
SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=>'ZIGGY2'); PL/SQL procedure successfully completed. SQL> analyze table ziggy2 compute statistics; Table ZIGGY2 analyzed. SQL> select table_name, num_rows, blocks, empty_blocks, avg_space, avg_row_len, chain_cnt from user_tables where table_name='ZIGGY2'; TABLE_NAME NUM_ROWS BLOCKS EMPTY_BLOCKS AVG_SPACE AVG_ROW_LEN CHAIN_CNT _____________ ___________ _________ _______________ ____________ ______________ ____________ ZIGGY2 200000 4654 82 2823 113 0
We notice it has increased to the same 4654 blocks as with the commit. The extra storage remains allocated even after the rollback operation.
That’s again because Oracle does NOT deallocate any additional storage that might have been consumed during the original transaction. We notice that the AVG_SPACE has again increased substantially as a result (now 2823, previously it was just 857).
If we look at the current state of the indexes after the rollback:
SQL> select index_name, blevel, leaf_blocks, clustering_factor
from user_indexes where table_name='ZIGGY2';
INDEX_NAME BLEVEL LEAF_BLOCKS CLUSTERING_FACTOR
__________________ _________ ______________ ____________________
ZIGGY2_ID_I 2 945 3250
ZIGGY2_CODE1_I 2 945 3250
ZIGGY2_CODE2_I 2 945 3250
ZIGGY2_CODE3_I 2 945 3250
ZIGGY2_CODE4_I 2 945 3250
ZIGGY2_CODE5_I 2 945 3250
ZIGGY2_CODE6_I 2 945 3250
ZIGGY2_CODE7_I 2 945 3250
ZIGGY2_CODE8_I 2 945 3250
ZIGGY2_CODE9_I 2 945 3250
ZIGGY2_CODE10_I 2 945 3250
ZIGGY2_CODE11_I 2 945 3250
ZIGGY2_CODE12_I 2 945 3250
ZIGGY2_CODE13_I 2 945 3250
ZIGGY2_CODE14_I 2 945 3250
ZIGGY2_CODE15_I 2 945 3250
ZIGGY2_CODE16_I 2 945 3250
ZIGGY2_CODE17_I 2 945 3250
ZIGGY2_CODE18_I 2 945 3250
ZIGGY2_CODE19_I 2 945 3250
ZIGGY2_CODE20_I 2 945 3250
We notice that they all remain at their increased size of 945 leaf blocks and with the Blevel of 2.
Again, when Oracle performs the rollback, Oracle does NOT undo all the index block splits and leaves all the additional storage allocated to the indexes.
So, just a word of caution.
Updating all the ROWIDs on the fly when a row migrates does not come for free. There’s additional resources that need to be consumed during these updates AND there is a potential issue with indexes having to perform additional index block splits and consume additional storage (at least immediately) after such operations.
If you have applications that makes bulk changes to data that can result in rollbacks, again, it’s just worth noting the extra storage that may be consumed as a result (until the additional data is finally added to the table).
Yes, index rebuilds can be performed to reduce the subsequent size of these inflated indexes.
BUT, for those of you with sharp eyes, you might also have noted another potential issue with this new behaviour, which I’ll discuss in my next post… 🙂
Costing Concatenated Indexes With Range Scan Predicates Part I (Nothing To Be Desired) July 22, 2022
Posted by Richard Foote in BLEVEL, CBO, Clustering Factor, Concatenated Indexes, Index Access Path, Index Column Order, Index Column Reorder, Leaf Blocks, Non-Equality Predicates, Oracle, Oracle Cost Based Optimizer, Oracle General, Oracle Indexes, Performance Tuning, Richard Foote Consulting, Richard Foote Training, Richard's Blog.1 comment so far
In my previous post, I discussed how Automatic Indexing ordered columns when derived from SQLs containing both equality and non-equality predicates.
I’ve since had offline questions asking why indexes are more effective with leading columns addressing the equality predicates rather than the leading columns addressing non-equality predicates. Based on the theory that for everyone who asks a question, there are likely numerous others wondering the same thing, I thought I’ll try to explain things with these posts.
I’ll start by creating the following simple table that has two columns (ID and CODE) that are both highly selective (they both have 10,000 distinct values in a 100,000 rows table, so 10 rows approximately per value):
SQL> CREATE TABLE radiohead (id NUMBER, code NUMBER, name VARCHAR2(42)); Table created. SQL> INSERT INTO radiohead SELECT mod(rownum,10000)+1, ceil(dbms_random.value(0,10000)), 'RADIOHEAD' FROM dual CONNECT BY LEVEL <= 100000; 100000 rows created. SQL> commit; Commit complete.
I’ll next create an index based on the ID, CODE columns, with importantly the ID column as the leading column:
SQL> CREATE INDEX radiohead_id_code_i ON radiohead(id, code); Index created. SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=>'RADIOHEAD', estimate_percent=> null, method_opt=> 'FOR ALL COLUMNS SIZE 1'); PL/SQL procedure successfully completed.
When it comes to costing index accesses, some of the crucial statistics including the Blevel, Leaf_Blocks and often most crucial of all, the Clustering_Factor:
SQL> SELECT index_name, blevel, leaf_blocks, clustering_factor FROM user_indexes WHERE index_name = 'RADIOHEAD_ID_CODE_I'; INDEX_NAME BLEVEL LEAF_BLOCKS CLUSTERING_FACTOR -------------------- ---------- ----------- ----------------- RADIOHEAD_ID_CODE_I 1 265 99034
We begin by running the following query, with an equality predicate on the ID column and a relatively large, non-selective range predicate on the CODE column:
SQL> SELECT * FROM radiohead WHERE id = 42 AND CODE BETWEEN 1000 AND 5000;
Execution Plan
-----------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-----------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 4 | 72 | 6 (0)| 00:00:01 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED| RADIOHEAD | 4 | 72 | 6 (0)| 00:00:01 |
|* 2 | INDEX RANGE SCAN | RADIOHEAD_ID_CODE_I | 4 | | 2 (0)| 00:00:01 |
-----------------------------------------------------------------------------------------------------------
Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
8 consistent gets
0 physical reads
0 redo size
824 bytes sent via SQL*Net to client
608 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
5 rows processed
As (perhaps) expected, the CBO uses the index to retrieve the small number of rows (just 5 rows).
However, if we run the following query which also returns a small number of rows (just 4 rows) BUT with the relatively unselective, non-equality predicate based on the leading indexed ID column:
SQL> SELECT * FROM radiohead WHERE id BETWEEN 1000 AND 5000 AND CODE = 140;
Execution Plan
-------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 4 | 72 | 105 (11)| 00:00:01 |
|* 1 | TABLE ACCESS FULL| RADIOHEAD | 4 | 72 | 105 (11)| 00:00:01 |
-------------------------------------------------------------------------------
Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
363 consistent gets
0 physical reads
0 redo size
770 bytes sent via SQL*Net to client
608 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
4 rows processed
We notice (perhaps unexpectedly) that the CBO now ignores the index and uses a Full Table Scan, even though only 4 rows are returned from a 100,000 row table.
This is a common area of confusion. Why does Oracle not use the index when both columns in the index are referenced in the SQL predicates and only a tiny number of rows are returned?
The answer comes down to the very unselective non-equality predicate (ID BETWEEN 1000 AND 5000) being serviced by the leading column (ID) of the index.
The “ID BETWEEN 1000 AND 5000” predicate basically covers 40% of all known ID values, which means Oracle must now read 40% of all Leaf Blocks within the index (one leaf block at a time), starting with ID =1000 and ending with ID = 5000. Although there are very few rows that then subsequently match up with the other (CODE = 140) predicate based on the second column (CODE) of the index, these relatively few values could exist anywhere within the 40% ID range.
Therefore, when costing the reading of the actual index, the CBO basically stops its calculations after the non-equality predicate on this leading ID column and indeed estimates that a full 40% of the index itself must be scanned.
If we force the CBO into a range scan via a basic index hint:
SQL> SELECT /*+ index(r) */ * FROM radiohead r WHERE id BETWEEN 1000 AND 5000 AND CODE = 140;
Execution Plan
-----------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-----------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 4 | 72 | 116 (4)| 00:00:01 |
| 1 | TABLE ACCESS BY INDEX ROWID BATCHED| RADIOHEAD | 4 | 72 | 116 (4)| 00:00:01 |
|* 2 | INDEX RANGE SCAN | RADIOHEAD_ID_CODE_I | 4 | | 112 (4)| 00:00:01 |
-----------------------------------------------------------------------------------------------------------
Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
114 consistent gets
0 physical reads
0 redo size
806 bytes sent via SQL*Net to client
608 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
4 rows processed
We notice that the overall cost of this index based plan is 116, greater than the 105 cost of the Full Table Scan (and hence why the Full Table Scan was selected). We also notice that the vast majority of this 116 cost can be attributed to the index scan itself in the plan, which has a cost of 112.
If you have a calculator handy, this is basically how these costs are derived.
Range Selectivity = (Max Range Value–Min Range Value)/(Max Column Value–Min Column Value)
Effective Index Selectivity = Range Selectivity + 2 x ID density (as a BETWEEN clause was used which is inclusive of Min/Max range)
= (5000-1000)/(10000-1) + 2 x (1/10000)
= 0.40004 + 0.0002
= 0.40024
Effective Table Selectivity = ID selectivity (as above) x CODE selectivity
= 0.40024 x (1/10000)
= 0.40024 x 0.0001
= 0.000040024
These selectivities are then inserted into the following index costing formula:
Index IO Cost = blevel +
ceil(effective index selectivity x leaf_blocks) +
ceil(effective table selectivity x clustering_factor)
Index IO Cost = 1 + ceil(0.40024 x 265) + ceil(0.000040024 x 99034)
= 1 + 107 + 4
= 108 + 4 = 112.
Index Access Cost = IO Costs + CPU Costs (in this plan, 4% of total costs)
= (108 + (112 x 0.04)) + (4 + (4 x 0.04))
= (108 + 4) + (4 + 0)
= 112 + 4
= 116
So we can clearly see how the CBO has made its calculations, come up with its costs and has decided that the Full Table Scan is indeed the cheaper alternative with the current index in place.
So Automatic Indexing is doing the right thing, by creating an index with the leading column based on the equality predicate and the second indexed column based on the unselective non-equality predicate.
I’ll expand on this point in an upcoming Part II post.
BLEVEL 1 => BLEVEL 2 (Teenage Wildlife) August 23, 2011
Posted by Richard Foote in BLEVEL, CBO, Oracle Indexes.4 comments
Jonathan Lewis recently wrote a really nice blog piece blevel=1 on the dangers of an index toggling between BLEVEL 1 and BLEVEL 2. I thought it would be useful to demonstrate this issue with a quick demo (Note: this example is on 11.2.0.1, with an 8K block size).
First, create a simple little table with 336,000 rows and an index on an ID number column:
SQL> create table major_tom (id number, code number, name varchar2(30)); Table created. SQL> create index major_tom_i on major_tom(id); Index created. SQL> insert into major_tom select rownum, mod(rownum,100), 'GROUND CONTROL' from dual connect by level <=336000; 336000 rows created. SQL> commit; Commit complete. SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=> 'MAJOR_TOM', cascade=> true, estimate_percent=>null, method_opt=>'FOR ALL COLUMNS SIZE 1'); PL/SQL procedure successfully completed. SQL> select blevel, leaf_blocks, clustering_factor from dba_indexes where index_name='MAJOR_TOM_I'; BLEVEL LEAF_BLOCKS CLUSTERING_FACTOR ---------- ----------- ----------------- 1 671 1296
Note the index has 671 leaf blocks and a Blevel=1. This of course means the index basically consists of a Root Block, which in turn references all its 671 leaf blocks. Therefore to read a specific index entry, requires a read of the index root block followed by a read of the specific index leaf block. That’s 2 reads in total.
Let’s run a query to return one row (note the ID column is effectively unique although I’ve only created a non-unique index):
SQL> select * from major_tom where id = 42;
Execution Plan
----------------------------------------------------------
Plan hash value: 4155681103
-------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 23 | 2 (0)| 00:00:01 |
| 1 | TABLE ACCESS BY INDEX ROWID| MAJOR_TOM | 1 | 23 | 2 (0)| 00:00:01 |
|* 2 | INDEX RANGE SCAN | MAJOR_TOM_I | 1 | | 1 (0)| 00:00:01 |
-------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID"=42)
Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
4 consistent gets
0 physical reads
0 redo size
531 bytes sent via SQL*Net to client
395 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
1 rows processed
Note: The cost of using the index is just 1 not 2 as perhaps expected. This is due to the CBO ignoring the Blevel in its calculations when the Blevel = 1. As the index is relatively small, the CBO takes the approach that the root block is very likely already cached and so is not worth costing.
As the data is perfectly evenly distributed and effectively unique, the CBO has correctly estimated the number of returned rows as just 1. Therefore, the overall cost of the execution plan is just 2, 1 to read the leaf block and 1 to read the table block.
Notice that the number of consistent gets is 4. 1 to read the index root block, 1 for the index leaf block, 1 for the table block and as the index is non-unique, 1 for an additional fetch performed to check the index again that there are no further rows to be returned.
If we now create another table of 1M rows that will be used in a join operation:
SQL> create table ziggy (id number, code number, name varchar2(30)); Table created. SQL> insert into ziggy select rownum, mod(rownum,10000), 'ZIGGY STARDUST' from dual connect by level <= 1000000; 1000000 rows created. SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=> 'ZIGGY', cascade=> true, estimate_percent=>null, method_opt=>'FOR ALL COLUMNS SIZE 1'); PL/SQL procedure successfully completed.
If we now join these 2 tables and select a moderate number of rows:
SQL> select * from ziggy z, major_tom m where z.id = m.id and z.code in (42, 4242);
68 rows selected.
Execution Plan
----------------------------------------------------------
Plan hash value: 2011771477
--------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
--------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 200 | 9400 | 1372 (2)| 00:00:17 |
| 1 | NESTED LOOPS | | | | | |
| 2 | NESTED LOOPS | | 200 | 9400 | 1372 (2)| 00:00:17 |
|* 3 | TABLE ACCESS FULL | ZIGGY | 200 | 4800 | 1105 (2)| 00:00:14 |
|* 4 | INDEX RANGE SCAN | MAJOR_TOM_I | 1 | | 1 (0)| 00:00:01 |
| 5 | TABLE ACCESS BY INDEX ROWID| MAJOR_TOM | 1 | 23 | 2 (0)| 00:00:01 |
--------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
3 - filter("Z"."CODE"=42 OR "Z"."CODE"=4242)
4 - access("Z"."ID"="M"."ID")
Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
4175 consistent gets
4024 physical reads
0 redo size
1950 bytes sent via SQL*Net to client
395 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
68 rows processed
The CBO goes for a Nested Loop join, primarily because the inner table is only accessed a relatively small number of times AND because the cost of doing so via the index is so damn cheap.
However, if we add just a few more rows and collect fresh statistics …
SQL> insert into major_tom select rownum+336000, mod(rownum,100), 'GROUND CONTROL' from dual connect by level <=500; 500 rows created. SQL> commit; Commit complete. SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=> 'MAJOR_TOM', cascade=> true, estimate_percent=>null, method_opt=>'FOR ALL COLUMNS SIZE 1'); PL/SQL procedure successfully completed. SQL> select blevel, leaf_blocks, clustering_factor from dba_indexes where index_name='MAJOR_TOM_I'; BLEVEL LEAF_BLOCKS CLUSTERING_FACTOR ---------- ----------- ----------------- 2 672 1298
The index has now toggled over to become a Blevel 2 index. We only added a handful of rows resulting in just the one additional index leaf block, but 672 is just too many to be referenced within the one index root block in this example. The root block has split, two new index branches have been created that now reference the leaf blocks and the root block now only references the two new branch blocks.
Overall, the changes are quite minor but the ramifications can be quite dramatic …
If we now select one row again:
SQL> select * from major_tom where id = 42;
Execution Plan
----------------------------------------------------------
Plan hash value: 4155681103
-------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 23 | 4 (0)| 00:00:01 |
| 1 | TABLE ACCESS BY INDEX ROWID| MAJOR_TOM | 1 | 23 | 4 (0)| 00:00:01 |
|* 2 | INDEX RANGE SCAN | MAJOR_TOM_I | 1 | | 3 (0)| 00:00:01 |
-------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("ID"=42)
Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
5 consistent gets
0 physical reads
0 redo size
531 bytes sent via SQL*Net to client
395 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
1 rows processed
The cost of the index has now jumped up by 2 from 1 to 3 with the overall costs up from 2 to 4, even though the index is practically the same size. As the Blevel is now 2, the CBO now includes the cost of the Blevel in its calculations. The cost associated with accessing the root block and a branch block all now count. Although overall the costs are still low, this increase actually represents a 100% increase in the use of the index for an equality search.
This increase can be significant if the index needs to be accessed multiple times. Let’s now re-run the join query:
SQL> select * from ziggy z, major_tom m where m.id = z.id and z.code in (42, 4242);
68 rows selected.
Execution Plan
----------------------------------------------------------
Plan hash value: 1928189744
--------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
--------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 200 | 9400 | 1485 (2)| 00:00:18 |
|* 1 | HASH JOIN | | 200 | 9400 | 1485 (2)| 00:00:18 |
|* 2 | TABLE ACCESS FULL| ZIGGY | 200 | 4800 | 1105 (2)| 00:00:14 |
| 3 | TABLE ACCESS FULL| MAJOR_TOM | 336K| 7558K| 378 (1)| 00:00:05 |
--------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - access("M"."ID"="Z"."ID")
2 - filter("Z"."CODE"=42 OR "Z"."CODE"=4242)
Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
5366 consistent gets
4024 physical reads
0 redo size
1964 bytes sent via SQL*Net to client
395 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
68 rows processed
Although it’s retrieving exactly the same data, the execution plan has changed significantly. The Nested Loop join is no longer as appealing to the CBO as the cost of accessing the inner table via the index has now effectively doubled. The CBO has now gone for a Hash Join, accessing both tables via Full Tables Scans. The overall cost of the Nested Loop plan was 1372, but this has increased to over 1485, the cost of the now so-called more efficient Hash Join plan.
If you have indexes that are on the boundary of increasing from a blevel=1 to a blevel=2, execution plans can potentially change significantly based on the differences in how indexes get costed. This can be especially troublesome when such indexes get regularly rebuilt as they may toggle between Blevel 1 and 2 based on associated space savings and can sometimes result in unpredictable performance depending on when new statistics get collected.
I liken it to a child growing up from being a young kid to a teenager. It may only be a difference of a year or so but boy, can the differences be dramatic !!
Richard Foote's Oracle Blog 