jump to navigation

Indexes And Small tables Part VII (Cluster One) May 27, 2009

Posted by Richard Foote in Index Organized Tables, Oracle Indexes, Small Indexes.
14 comments

OK, almost getting to the end here ;)

As discussed previously, despite popular opinion, an index can be just that little bit more efficient than a FTS when accessing very small tables, even if  all rows in the table exist in the one table block. And a small efficiency multiplied by a large number can potentially add up and make a noticeable difference.

As we’ve seen, a unique index on such a small table accessing a specific row of interest need only perform one consistent read on the index block and one consistent read on the table block for a grand total of 2 consistent reads, with both consistent gets being the cheaper examinations variety. Not bad,  not too bad at all and somewhat cheaper than an equivalent FTS.

However, as I hinted and as many of you have already commented, we can go one step further still in reducing the overheads of such queries on small tables by potentially storing all the columns in the one, non-heap table structure.

One option is to create an Index Organized Table (IOT), storing all columns within a single index structure and thereby eliminating the need to visit a table segment at all.

Following on from the previous demo, let’s recreate the table as an IOT and populate it with the same data:

SQL> drop table small;
 
Table dropped.
 
SQL> create table small (id number primary key, name varchar2(10)) organization index;
 
Table created.
 
SQL> insert into small select rownum, ‘BOWIE’ from dual connect by level <=100;
 
100 rows created.
 
SQL> commit;
 
Commit complete.
 
SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=>’SMALL’, method_opt=>’FOR ALL COLUMNS SIZE 1′);
 
PL/SQL procedure successfully completed.

 

If we now run our query again:

 

SQL> select * from small where id = 42;

 
        ID NAME
---------- ----------
        42 BOWIE
 
Execution Plan
------------------------------------------
|Id|Operation         |Name              |
------------------------------------------
| 0|SELECT STATEMENT  |                  |
|*1| INDEX UNIQUE SCAN|SYS_IOT_TOP_68376 |
------------------------------------------

Statistics
-------------------------------------------
  0  recursive calls
  0  db block gets
  1  consistent gets
  0  physical reads
  0  redo size
465  bytes sent via SQL*Net to client
396  bytes received via SQL*Net from client
  2  SQL*Net roundtrips to/from client
  0  sorts (memory)
  0  sorts (disk)
  1  rows processed

 

We notice that we have now reduced the number of consistent gets down to just one.

Not only is it just one consistent get but if we look at the type of consistent get by running the following query in another session before/after our SELECT above:

SQL> select name, value from v$sesstat s, v$statname n where s.statistic#=n.statistic# and sid = 141 and name like ‘consistent gets%';
 

NAME                           VALUE
----------------------------- ------
consistent gets                32842
consistent gets - examination   6694

 

SQL> select name, value from v$sesstat s, v$statname n where s.statistic#=n.statistic# and sid = 141 and name like ‘consistent gets%';

 
NAME                           VALUE
----------------------------- ------
consistent gets                32843 (+1)
consistent gets - examination   6695 (+1)

 

We also notice that it’s the cheaper, one latch consistent gets examination.

So we’ve now reduced our overheads down to just one consistent get and just the one latch get as well. It doesn’t really get much cheaper than that.

IOT are one of those under used options in Oracle that really should be considered used a lot more than they are. Yes they can be problematic when used inappropriately (especially if you need to create several secondary indexes) but for scenarios such as this they can be very useful.

I plan to discuss the advantages and disadvantages of IOT in future posts.

 

Another posssible option to improve things in our little demo is to create a Hash Cluster (as commented by Piet):

SQL> create cluster small_cluster (id number) size 100 single table hashkeys 200;
 
Cluster created.
 
SQL> create table small_tab (id number, name varchar2(100)) cluster small_cluster(id);
 
Table created.
 
SQL> insert into small_tab select rownum, ‘BOWIE’ from dual connect by level <=100;
 
100 rows created.
 
SQL> commit;
 
Commit complete.
 
SQL> exec dbms_stats.gather_table_stats(ownname=>null, tabname=>’SMALL_TAB’, estimate_percent=>null, method_opt=>’FOR ALL COLUMNS SIZE 1′);
 
PL/SQL procedure successfully completed.
 
SQL> select * from small_tab where id = 42; 

  ID NAME
---- ----------
  42 BOWIE

Execution Plan
---------------------------------------
| Id  | Operation         | Name      |
---------------------------------------
|   0 | SELECT STATEMENT  |           |
|*  1 |  TABLE ACCESS HASH| SMALL_TAB |
---------------------------------------

Statistics
-------------------------------------------
  0  recursive calls
  0  db block gets
  1  consistent gets
  0  physical reads
  0  redo size
465  bytes sent via SQL*Net to client
396  bytes received via SQL*Net from client
  2  SQL*Net roundtrips to/from client
  0  sorts (memory)
  0  sorts (disk)
  1  rows processed

 

Note again we have got our consistent gets down to just one as Oracle can simply determine the correct hash key based on the specified ID value and go directly to the table block containing the row of interest. Note however, this will not be a consistent get – examination, but the more expensive 2 latch and pin the block variety.

However, if you now create a unique index on the ID column:

SQL> create unique index small_tab_i on small_tab(id);

Index created.

And re-run the query:

SQL> select * from small_tab where id = 42; 

  ID NAME
---- ----------
  42 BOWIE

Execution Plan
---------------------------------------
| Id  | Operation         | Name      |
---------------------------------------
|   0 | SELECT STATEMENT  |           |
|*  1 |  TABLE ACCESS HASH| SMALL_TAB |
---------------------------------------

Statistics
-------------------------------------------
  0  recursive calls
  0  db block gets
  1  consistent gets
  0  physical reads
  0  redo size
465  bytes sent via SQL*Net to client
396  bytes received via SQL*Net from client
  2  SQL*Net roundtrips to/from client
  0  sorts (memory)
  0  sorts (disk)
  1  rows processed

 

Oracle will still use the hash to determine the location of the row of interest but because it now knows it will only retrieve the one row, will do so with only a consistent get – examination.  Back to one consistent get and only one latch get as well.

Again, clusters and all their various types and forms will be discussed in future blog entries.

Perhaps our good old familar heap table might not always be the best and most efficient option when creating these small tables (or even larger tables for that matter).

But for now the key message from this series is that any table, no matter how small can potentially benefit from being indexed. There really is no such thing as a table that’s too small to benefit from an index.Yes the difference might be small and of no real consequence but then again for larger database environments the overall savings might all add up and surprise. Note that the associated costs of having such indexes are also likely to be relatively small so perhaps it might just be worthwhile indexing those small tables after all ;)

Two Excellent Index Related Blog Posts May 25, 2009

Posted by Richard Foote in Oracle Indexes.
add a comment

Before I sit down and write something new, thought I might quickly mention two excellent recent index related blog posts,written by two very clever chaps.

The first is by Jonathan Lewis who discusses in a post called Index Size how one might determine which indexes to perhaps consider for an index rebuild by using the dbms_space.create index_cost procedure.  By comparing the difference between the current size of the index and the potential size of a newly created index as calculated by the procedure, one can check to see if the difference in size might make a rebuild a worthwhile proposition.  Jonathan then discusses a number of deficiencies and little traps with this approach.

The second is by Alberto Dell’Era who discusses in a post called Order of keys inside index blocks exactly how Oracle orders and stores the index keys within an index block. I’ve exchanged a number of emails with Alberto on this subject and he does an excellent job of detailing exactly how Oracle orders the index row directories and the advantages of Oracle doing so.

Both are highly recommended and well worth a read if you’re interested in learning more on Oracle indexes.

Indexes And Small Tables Part VI (Loaded) May 19, 2009

Posted by Richard Foote in Constraints, Oracle Indexes, Small Indexes, Unique Indexes.
9 comments

Thanks to comments by PdV, I need yet another Part before I can look at completing this series ;)

OK, we’ve reached the stage in Part V of accessing this small, one block table with a Unique Index. This has reduced the number of consistent gets to 2, with both consistent get operations being the “pinless”, one latch consistent get examinations.

We basically need one consistent get to read the index and the other consistent get to read the row from the table block.

Not bad.

However, if we could somehow just store all the columns of interest within the index structure, we could potentially do even better because then we wouldn’t need to visit the table segment at all. A single consistent get from the index and bingo, we can retrieve all the required data directly from the index.

Overloading an index in this way is actually quite a common tuning trick. Simply add additional columns within the select list so that all the required columns can be found within the index, thereby elliminating the need to access the table at all and so potentially improve performance.

However, when it comes to overloading a unique index designed to specifically police a PK or Unique constraint we have a slight problem. Oracle will not allow a unique constraint to be policed by a unique index that does not have the same column list. It’s not a problem for a non-unique index (providing the leading columns match the constraint columns), but it’s an issue for a unique index.

Therefore, in our little example, we can’t simply create a single concatenated unique index on both the ID and NAME columns and use it to police a unique constraint on just the ID column. We must either use a unique index on the ID column or use a non-unique index on both the ID and NAME columns. If we want to create a unique index on both the ID and NAME columns, we would need to create an additional index on the ID column to police the PK on the ID column or change our business rules to make both the ID and NAME the PK (which is not likely something we would want to do). Note also by doing creating 2 unique indexes, we would effectively be storing the ID column in three separate places, within the table, within the ID index and also within the ID and NAME index. Again, not something we’re likely going to want to do.

To illustrate the point, drop any existing indexes on the SMALL table. If we attempt to create a unique index on both the ID and NAME columns while making the ID column only the PK, Oracle is going to complain:

SQL> alter table small add primary key(id) using index (create unique index small_uk_i on small(id, name));
alter table small add primary key(id) using index (create unique index small_uk_i on small(id, name))
*
ERROR at line 1:
ORA-14196: Specified index cannot be used to enforce the constraint.try and cr

If we create a unique index first on both the ID and NAME columns:

SQL> create unique index small_uk_i on small(id, name);

Index created.

And then hope Oracle will simply use this index to police a PK constraint on just the ID column, we’ll be sadly disappointed as Oracle will actually created another unique index on the ID column behind the scenes:

SQL> alter table small add primary key(id);

Table altered.

SQL> select c.index_name, c.column_name from user_indexes i, user_ind_columns cwhere i.index_name = c.index_name and i.table_name = c.table_name and i.table_name = ‘SMALL';

INDEX_NAME   COLUMN_NAME
------------ ------------
SMALL_UK_I   ID
SMALL_UK_I   NAME
SYS_C009759  ID

 

The CBO will of course potentially look at using our SMALL_UK_I concatenated unique index to perform the select statement of our demo, but the efficiency results are not quite what we’re after:

SQL> select id, name from small where id = 42;

        ID NAME
---------- -----
        42 BOWIE
-----------------------------------
|Id| Operation        | Name      |
-----------------------------------
| 0| SELECT STATEMENT |           |
|*1|  INDEX RANGE SCAN| SMALL_UK_I|
-----------------------------------
Statistics
-------------------------------------------
  0  recursive calls
  0  db block gets
  2  consistent gets
  0  physical reads
  0  redo size
465  bytes sent via SQL*Net to client
396  bytes received via SQL*Net from client
  2  SQL*Net roundtrips to/from client
  0  sorts (memory)
  0  sorts (disk)
  1  rows processed

 

We first note Oracle is indeed using the concatenated, unique SMALL_UK_I index as it can retrieve all the necessary data from the query directly from the index.

However, also note the CBO is performing an index range scan, not a unique scan as only a portion of the index (the ID) is specified in the WHERE clause. Oracle doesn’t pick up the fact the select operation is actually unique as not all columns within the SMALL_UK_I unique index used by the CBO are specified in the WHERE clause. This despite the fact the ID is actually the defined PK of the table.

Therefore, Oracle is still performing 2 consistent get operations as there may be more rows to retrieve after performing the first fetch within the SQL*PLUS session. Also, if we examined the types of consistent reads being performed, we would note that neither of them are consistent get – examinations.

So we’re really not that far ahead of just using the unique index on the ID column as we did in Part V of this series. We still require 2 consistent gets (although neither of them are now examinations) and we’re having to store the ID in three separate locations for our trouble, rather than two.

Wouldn’t it be nice if we could have a PK index on just the ID column, but somehow add the NAME column (or any other columns of interest) to the index structure so that we only need to visit the index structure, thereby storing the ID in only the one index. Then we could potentially access the data with just one consistent get and with it being a unique index, it would be a consistent get examination requiring only the one latch access.

Hell, wouldn’t it be nice if we didn’t even bother with the table segment at all as all queries of interest would never actually need to access and use the table segment anyways, thereby storing the PK in possibly just the one location.

Such a solution has of course been possible for a long time …

;)

Indexes And Small Tables Part V (It’s No Game) May 13, 2009

Posted by Richard Foote in Index Internals, Oracle Indexes, Small Indexes, Unique Indexes.
12 comments

So far in our little example, we’ve looked at how accessing a row of a one block table via a FTS required 4 consistent gets while accessing this same table via a Non-unique index reduced the consistent gets down to 3.

Time to take the next step and improve the efficiency yet further of accessing this small one block table.

We’re now going to replace the Non-unique index with a Unique Index instead. We can obviously do this because the values on the indexed ID column are indeed unique.

Now it’s always a good idea of course to document these table business rules (such as a column being unique) inside the database, however it’s somewhat alarming just how many application just don’t this. I’ve also previously discussed how a PK or Unique constraint can actually be policed via a Non-Unique index so there are many reasons why a small table might not have an associated Unique index.

Not least of course the incorrect perception that an index is not going to be much use on a small table anyways …

So let’s now replace the Non-Unique index with a Unique index instead:

SQL> drop index small_id_i;

Index dropped.

SQL> alter table small add primary key (id) using index (create unique index small_id_i on small(id));

Table altered.

 
OK, so now we have our Unique index in place. Let’s now run the same query again to see how our consistent gets related statistics might change:

 
SQL> select * from small where id = 42;

 

        ID NAME
---------- -----
        42 BOWIE

--------------------------------------------
|Id |Operation                  |Name      |
--------------------------------------------
|  1|TABLE ACCESS BY INDEX ROWID|SMALL     |
|* 2| INDEX UNIQUE SCAN         |SMALL_ID_I|
--------------------------------------------

Statistics
--------------------------------------------
  0  recursive calls
  0  db block gets
  2  consistent gets
  0  physical reads
  0  redo size
465  bytes sent via SQL*Net to client
396  bytes received via SQL*Net from client
  2  SQL*Net roundtrips to/from client
  0  sorts (memory)
  0  sorts (disk)
  1  rows processed

 

We notice our first significant change. The number of consistent gets has reduced further down to just 2.

Why ?

Because with a Unique index, there can only be a maximum of 1 row returned. It’s simply not possible to return 2 or more rows.

Therefore, when selecting this one row, Oracle doesn’t have to perform the second fetch operation to confirm there are indeed no more rows to return. The first fetch will either return the one row of interest or none at all when resolving an equality predicate. That’s it, there are no other possibilities. We return the row of interest by simply accessing the index block (1 consistent get) followed by the table block (the second consistent get).

As we have previously, if we look at the actual consistent gets statistics of interest by running the following query in another session before/after the select statement:

SQL> SELECT n.name, s.value FROM v$sesstat s, v$statname n WHERE s.statistic# =n.statistic# AND s.sid = 141 AND n.name LIKE ‘consistent%';

NAME                           VALUE
------------------------------ -----
consistent gets                31236
consistent gets - examination   5084

 
And again afterwards ..

 
SQL> SELECT n.name, s.value FROM v$sesstat s, v$statname n WHERE s.statistic# =n.statistic# AND s.sid = 141 AND n.name LIKE ‘consistent%';

NAME                           VALUE
------------------------------ -----
consistent gets                31238 (+2)
consistent gets - examination   5086 (+2)

We notice we have indeed only performed the 2 consistent gets. But we also notice another significant difference, that being both consistent gets are now the “cheaper” consistent gets – examination.

This means that the latches required to now perform this select statement via the Unique index is just 2, down from 6 for the Non-unique index and 8 from the FTS.

Generally during a consistent get, Oracle needs to grab the cache buffers chain latch so it can pin the specific block in memory and then grab the latch again so that it can subsequently unpin the block once it’s finished processing the block. Each of these accesses to the latch and the subsequently pin/unpinning of the block requires CPU and is a possible source of contention within the database.

For some operations that only require a very very quick “read and get out of there” type operation and/or on blocks that are unlikely to change within a given point of time, Oracle uses a cheaper consistent get operation which doesn’t actually require the block to be pinned. There’s no point in pinning the block as it’s only going to be read and accessed for a short time (shorter than might otherwise be required when processing a block in memory) and the block is unlikely to change anyways.

So for these operations, Oracle uses a cheaper consistent get called a consistent gets – examination. These consistent gets examinations only need grab the cache buffers chain latch before quickly reading the block and releasing the latch once the read operation is complete. Therefore it only needs to grab and release the cache buffer chains latch the once without having to pin/unpin the block, which means less CPU and less latch contention overall.

Now this isn’t particularly well documented. Often discussions mention reads of undo blocks as being candidates for consistent gets examinations as these reads are likely to be both relatively quick and a specific undo block is unlikely to change as only one transaction can actually update an undo block at a given time.

Getting back to indexes, reads of index root blocks are another candidate mentioned as again a read of an index root block is going to be very quick and an index root block is unlikely to change at a given point of time.

However, what is not well documented at all is the fact that any block accessed during an index Unique Scan is accessed via a consistent get – examination, including the consistent get associated with reading the table block as well. This is because again, any such read operation is going to be relatively quick as the most that ever needs to be read is the one index related entry and the one table row.

The net result is that now accessing a row from a small table via a Unique index requires only 2 latch accesses vs. the initial FTS example which required 8 latch gets as none of the FTS consistent gets are examinations.

Now you might say that these are all very small numbers, that 4 consistent reads isn’t that much, that 8 latches isn’t really that huge a number and reducing 8 latches down to 2 latches isn’t really going to be that noticeable. Yes it is effectively a 75% reduction but it’s a 75% reduction of not very much.

And if you’re discussing a single read of a single small lookup table you would likely be right.

But what if the small table is accessed frequently by the application, perhaps many 1,000s of times per minute. What if you have many such small tables, often used in small join operations by your OLTP applications. What if you have large numbers of users in a large application with many many such small table accesses. This effectively 75% saving can potentially become very significant, both in terms of the reduction in CPU load and also in the reduction of latch contention, which in turn can further reduce CPU loads.

A small improvement multiplied by a large amount can indeed make a difference …

However, I have one more step to go yet in further improving the efficiency of these small table lookups via an index.

One which can reduce the overall overheads by yet another 50% …

Indexes And Small Tables Part IV (Treefingers) May 5, 2009

Posted by Richard Foote in Index Block Splits, Index Internals, Oracle Indexes, Root Index Block, Small Indexes.
17 comments

As I asked in my previous post, the key question when comparing the associated costs of accessing a small table via a Full Table Scan (FTS) vs. an index scan is why does Oracle visit the segment header during a FTS but not during an index scan ?

The answer all comes down to understanding why Oracle must visit the table segment header during a FTS and how Oracle can avoid visiting the index segment header during an index scan.

Oracle must visit the table segment header during a FTS because it contains vital information necessary to perform the FTS, namely the extent map and the High Water Mark (HWM) associated with the table. Even with a table that only contains 1 data block worth of rows as in the SMALL table in my examples, Oracle has no way of automatically determining there’s actually only just the one block worth of data. It has to somehow look that up and determine exactly what table blocks Oracle needs to access during the FTS operation and the table segment header contains this necessary meta data. During “most” FTS operations, which are generally speaking larger, “expensive” operations, these accesses to the table segment header constitute a relatively small overhead. However, for FTS operations on “small” tables, accessing the table segment header can actually be a significant proportion of the overall associated costs.

During an index scan operation, there’s nothing of interest within the index segment header. The critical index block, the index block by which all index scans must start is the root block of the index (except Fast Full Index Scans which are basically the FTS equivalent for indexes). There’s no need to access the index segment header because it’s the root block that actually contains all the necessary information by which to start the index scan operation. The root blocks contains the pointers to subsequent index blocks (be it a branch or leaf blocks) that Oracle needs to follow in order to find the index entry of interest. The key to starting an index scan therefore is in determining the location of the index root block.

But how can Oracle determine the location of the index root block ?

Well Oracle implements a little “trick”, a golden rule with regard to indexes that doesn’t change regardless of the Oracle version, regardless of the O/S version, regardless of the type of tablespace or tablespace option of the index and regardless of how the index is created or grows and block splits over time.

The index root block is always, always, always the block immediately after the index segment header.

Always.

Therefore, when the Oracle code issues the associated function calls to perform an index scan, the first index block that Oracle assesses is the index segment header plus an offset of 1. Whereas a FTS accesses the table segment header, an index scan accesses the index segment block id plus 1.

With a tiny index that only has a level of 0 (or height of 1), note there is not “root” block as such as all the index entries can fit within one index leaf block. However, this block, this one and only leaf block within the index structure is also always the block immediately after the index segment header.

Always.

When we add more index entries into this one and only leaf block, we’ll eventually reach a point when it’s full and Oracle must perform an index block split operation. Oracle will then allocate 2 new blocks to the index. Assuming a 50-50 block split, one of these new blocks is assigned the lower 1/2 of all the current index entry values and the other new block is assigned the other upper 1/2 of the current index entry values. The original index leaf block content is then cleaned out and reassigned with just relative block address pointers and value boundaries associated with the 2 new leaf blocks.

The original index leaf block has been “reborn” as the root block of the index.

It’s quite easy to demonstrate how the original index block in a level 0 index or the root block of an index never changes and is always the block that follows the index segment header.

First, just create a little table and associated index:

SQL> CREATE TABLE same_root (id NUMBER, name VARCHAR2(30));

Table created.

SQL> INSERT INTO same_root VALUES (1, ‘The Thin White Duke’);

1 row created.

SQL> COMMIT;

Commit complete.

SQL> CREATE INDEX same_root_i ON same_root(name);

Index created.

If we dump the block immediately following the index segment header, we can confirm it’s our index block of interest, containing our one and only index entry:

SQL> select header_file, header_block from dba_segments where segment_name=’SAME_ROOT_I';

HEADER_FILE HEADER_BLOCK
----------- ------------
          5       107441

SQL> alter system dump datafile 5 block 107442;

System altered.

Following is an extract of the index block dump:

Leaf block dump
===============
header address 98959964=0x5e6025c
kdxcolev 0
KDXCOLEV Flags = – – -
kdxcolok 0
kdxcoopc 0x80: opcode=0: iot flags=— is converted=Y
kdxconco 2
kdxcosdc 0
kdxconro 1
kdxcofbo 38=0x26
kdxcofeo 8007=0x1f47
kdxcoavs 7969
kdxlespl 0
kdxlende 0
kdxlenxt 0=0x0
kdxleprv 0=0x0
kdxledsz 0
kdxlebksz 8036
row#0[8007] flag: ——, lock: 0, len=29
col 0; len 19; (19):  54 68 65 20 54 68 69 6e 20 57 68 69 74 65 20 44 75 6b 65
col 1; len 6; (6):  01 41 a3 aa 00 00
—– end of leaf block dump —–

Note that it is indeed our one and only index leaf block.

If we now take a treedump of the index:

SQL> SELECT object_id FROM dba_objects where object_name = ‘SAME_ROOT_I';

 OBJECT_ID
----------
     67721

SQL> ALTER SESSION SET EVENTS ‘immediate trace name treedump level 67721′;

Session altered.

 

Following is the treedump output:

—– begin tree dump
leaf: 0x141a3b2 21078962(0: nrow: 1 rrow: 1)
—– end tree dump

We note that the relative block address of our one and only index leaf block is 21078962.

If we now add a whole bunch of new rows to the table so that the leaf block can no longer hold all the index entries, thereby forcing the index to block split and grow:

SQL> insert into same_root select rownum+1, ‘David Bowie’ from dual connect by level <=100000;

100000 rows created.

SQL> commit;

Commit complete.

And now take a block dump of the same index block:

SQL> alter system dump datafile 5 block 107442;

System altered.

Branch block dump
=================
header address 98959940=0x5e60244
kdxcolev 2
KDXCOLEV Flags = – – -
kdxcolok 0
kdxcoopc 0x80: opcode=0: iot flags=— is converted=Y
kdxconco 2
kdxcosdc 2
kdxconro 2
kdxcofbo 32=0x20
kdxcofeo 8016=0x1f50
kdxcoavs 7984
kdxbrlmc 21238606=0x144134e
kdxbrsno 1
kdxbrbksz 8060
kdxbr2urrc 0
row#0[8038] dba: 21238607=0x144134f
col 0; len 11; (11):  44 61 76 69 64 20 42 6f 77 69 65
col 1; len 5; (5):  01 41 be 5f 01
row#1[8016] dba: 21238769=0x14413f1
col 0; len 11; (11):  44 61 76 69 64 20 42 6f 77 69 65
col 1; len 5; (5):  01 44 12 ba 01
—– end of branch block dump —–

We notice that the block is no longer a leaf block but has changed itself into an index branch block.

But not just any branch block. If we now take a new treedump:

SQL> ALTER SESSION SET EVENTS ‘immediate trace name treedump level 67721′;

Session altered.

branch: 0x141a3b2 21078962(0: nrow: 3, level: 2)
   branch: 0x144134e 21238606 (-1: nrow: 161, level: 1)
      leaf: 0x141a3b3 21078963 (-1: nrow: 179 rrow: 179)
      leaf: 0x141a3b4 21078964 (0: nrow: 179 rrow: 179)
      leaf: 0x141a3b5 21078965 (1: nrow: 179 rrow: 179)
      leaf: 0x141a3b6 21078966 (2: nrow: 179 rrow: 179)
      leaf: 0x141a3b7 21078967 (3: nrow: 179 rrow: 179)

….

The above partial listing of the treedump clearly shows that the index has grown from a level 0 index to a level 2 index and that the root block is the very same index block as the original leaf block listed as it has the same relative block address as before (21078962).

Indeed the original leaf block is now the index root block which is still the same block that immediately follows the index segment header.

Because Oracle doesn’t have to visit the index segment header and can simply directly access the block following the index segment header as this block is always the first index block of interest when performing an index scan, the index scan has that little advantage over the FTS. And it’s this little advantage that can give the index scan the edge over a FTS, even if we’re potentially accessing data from a very small table.

And you can’t get much smaller than a table that has all it’s rows in the one table block.

So far though, the example I’ve shown has been a “normal”, everyday, non-unique index that has a 1 consistent get advantage over the FTS when accessing a row of interest. I’ll next discuss how indexes can having an even bigger edge and more significant advantage over a FTS of a tiny table,  than just the 1 consistent get …

Follow

Get every new post delivered to your Inbox.

Join 1,862 other followers