Undo internals paper

Rollback Segment and UNDO Internals

ROLLBACK SEGMENT AND UNDO INTERNALS
Daniel W. Fink

UNDO INTERNALS
There are two areas of focus for this paper. The first section is a discussion of the block, rollback segment/undo entry and
transaction architecture of Oracle8. Due to the relatively static nature of rollback segments over the past few releases, I have
chosen not to discuss how they operate or the common problems. There is a wealth of documentation, notes, papers, etc.
available to the industrious and web-enabled user.
The second section examines Oracle9 automatic undo management. This has been widely publicized in marketing materials
and is a default feature in the new versions. However, there is little actual documentation on the mechanics and management
of automatic undo from a DBA's perspective. This paper will shed light on the actual operations of automatic undo
management.
One important concept to keep in mind is that much of what we will review is related to undoing actions. As such, there is a
great deal of 'reverse' terminology and presentation. For example, undo is read in reverse order…from the last entry to the
first. Instead of reading the next entry, we read the previous. Another example is that undo entries are the opposite of the
transaction. An insert containing a large amount of data is represented by a minimal amount of undo. .sense makes that hope I
METHODOLOGY
I utilized the scientific method of testing a hypothesis within a controlled environment. In several cases, the testing contradicted
my hypothesis. For example, I assumed, as many DBAs do, that a read consistent query used the data in the rollback segment
as its source. After testing, I found this to be an incorrect hypothesis. How does a read consistent query operate? Have
patience faithful reader, for all shall be revealed in due time.
In some regards, Oracle is a 'black box' in that most users, DBAs, developers do not have access to the design and source
code. However, the effect of a single change can be seen. If we know that state of an object (or objects) before the change
and the state after the change and no other actions would have been involved, we can infer the operation from it's impacts.
For example, if I record the state of the data block and all undo blocks before a transaction and the state of those objects
after the transaction and I know that there have been no other transactions, I can infer that all changes made to the blocks are
the result of the transaction and only the transaction. With a little knowledge about the architecture of a transaction, I can
identify the actual operations.
CAVEATS
This paper and subsequent presentation contain undocumented or lightly documented features and commands. While the
author has tested the commands on some platforms and releases, not all systems have been involved. Do not use these
commands on production or other valuable databases without a solid understanding of their function and experience in their
use. Each release of Oracle may cause the commands to change in behavior or fail completely. Not all commands and
permutations have been tested.
Additionally, over the course of testing, certain anomalies have been discovered for which there is no verifiable explanation.
Some are mere puzzles (why does the rollback segment skip the 2nd block of the first extent initially?) while others are contrary
to what Oracle has published (rollback segment blocks occasionally containing undo from current and previous transactions).
As always, constructive, verifiable explanations are appreciated. Testing will continue past the submission of this paper and
additional information will be available in the presentation, which can be downloaded.
TOOLS TO EXTRACT INFORMATION
To extract data from the database we may use a simple select statement. However, this only retrieves the actual data or a read
consistent copy and does not provide any structural information for the storage of the data. It is possible to use a block level

Paper 542


viewer to examine the actual bytes of the data files. However, this would be rather tedious and not all that enlightening, unless
we know exactly which byte means what. Fortunately, Oracle provides a method for extracting a formatted dump of the
data blocks using the alter system command in the following example
alter system dump datafile 7 block 1569;
These commands will generate a formatted block dump and write the information to a trace file. The block dump will
contain both structural data and a terse description. For example, in the following output some fields are self-documenting,
while others are very cryptic.
Start dump data blocks tsn: 7 file#: 7 minblk 1569 maxblk 1569
buffer tsn: 7 rdba: 0x01c00621 (7/1569)
scn: 0x0000.00047616 seq: 0x02 flg: 0x00 tail: 0x76161002
frmt: 0x02 chkval: 0x0000 type: 0x10=DATA SEGMENT HEADER - UNLIMITED
Extent Control Header
-----------------------------------------------------------------
Extent Header:: spare1: 0 spare2: 0 #extents: 1 #blocks: 15
last map 0x00000000 #maps: 0 offset: 2080
Highwater:: 0x01c00623 ext#: 0 blk#: 1 ext size: 15
#blocks in seg. hdr's freelists: 0
WHERE EXACTLY IS THE DATA?
In order to dump the data blocks, the file and blocks for the particular extent are needed. This information is available with
the dba_extents view only. The user_extents view does not contain the block_id and cannot be used to determine the block
address values. Remember that the first block of a data segment is the segment header and does not contain actual data.
select segment_name, file_id, block_id, blocks
from dba_extents
where owner = 'DWFINK'
and segment_name = 'TEST1';

SEGMENT_NAME FILE_ID BLOCK_ID BLOCKS
------------ ------- -------- ------
TEST1 4 5 4
TEST1 4 9 4
TEST1 4 13 4
One of the most powerful features of the alter system dump datafile # block # commands is that you can actually dump
multiple blocks at a time! This is not even documented in the 'undocumented' documents. A quick comment from a fellow
dba, a hard look at a trace file and a wild guess revealed the ability to execute the following command:
alter system dump datafile 4 block min 5 block max 16;
This command dumped the entire contents of the table. In previous efforts, this was accomplished by dumping the first
block, changing the statement to dump the second block, changing the statement to dump the third block…see where I'm
going with this? In order to determine the max block, you take the block_id + blocks – 1. In the above example, the last
extent occupies 4 blocks (13, 14, 15, 16) beginning at block 13. 13 + 4 – 1 = 16.
WHERE EXACTLY IS THE UNDO?
Undo is only generated by a transaction and is only relevant in the context of that transaction. There are two steps to
identifying where undo related to a specific transaction is saved. First is to identify the session performing the transaction in the
v$session view. Until the transaction is started, the TADDR column is null and there are no entries in the v$transaction view.
SADDR TADDR USERNAME OSUSER TYPE
-------- -------- ---------- -------------------- ----------
7A9193B4 SYS DANIELWFINKdwfink USER
7A919CCC 7A69F5F0 DWFINK DANIELWFINKdwfink USER
Second, locate the related record(s) in v$transaction.
SES_ADDR ADDR START_UBAFIL START_UBABLK UBAFIL UBABLK

Paper 542


-------- -------- ------------ ------------ ---------- ----------
7A919CCC 7A69F5F0 2 37095 2 37095
There are two components to the undo chain that are represented in this statement, the head and the tail. The 'tail' or current
block of the undo chain are represented by the ubafil and ubablk columns. The head or original block of the undo chain is
represented by the start_ubafil and start_ubablk fields. UBA is an acronym for Undo Block Address, FIL is file and BLK is
block.
It is important to note that the blocks in the undo chain are not guaranteed to be adjacent to each other. Because an extent
may be shared by transactions, a block may be inserted into the chain that is not the next physical block in the extent. The only
sure method to determine which blocks contain undo from a particular transaction is to dump the blocks and follow the
undo chain. However, if the difference in v$transaction.start_ubablk and v$transaction.ubablk is equal to
v$transaction.used_ublk, there are no 'holes' in the blocks in the undo chain.
ORACLE8 ROLLBACK SEGMENT GENERAL CONCEPTS
WHAT IS A ROLLBACK SEGMENT
OFFICIAL DESCRIPTIONS
"A rollback segment records the old values of data that was changed by each transaction (whether or not committed).
Rollback segments are used to provide read consistency, to roll back transactions, and to recover the database." (Oracle8i
Concepts)
"They [rollback segments] capture the "before image" of data as it existed prior to the start of a transaction… A rollback
segment entry is the set of before-image data blocks that contain rows that are modified by a transaction." (Oracle8 DBA
Handbook)
" The rollback segment does not store the whole data block – only the before image of the row or rows that were modified.
Information in the rollback segment consists of several rollback entries called undo. For example, if a row is inserted into a
table, the undo created by that transaction would include the ROWID of that row, among other information. This is because
the undo operation of an insert is a delete, and all you need to delete a row is the ROWID. If a delete operation is performed
on a table, the complete row will be part of the undo. For update transactions, we store the old value of the updated
columns. If the transaction modifies an index as well, then the old index keys will also be stored as part of the undo. Rollback
segments guarantee that the undo information is kept for the life of a transaction." (Oracle8 Backup & Recovery Handbook)
P OPULAR VIEW
?? Copies of blocks changed by transactions, both uncommitted and committed
?? Copies of database rows that are being updated, perhaps at the block level
?? Snapshot of transactions that have occurred since the last commit
?? Logical information stored in memory and not physically stored in a file
?? System undo data allowing non-committed transactions to be rolled back
?? Before image copy of data in a transaction
DEFINED
Information necessary to reverse a transaction or to reconstruct a read consistent view. The contents of an undo entry include:
data column(s) address, transaction operation and old value(s) of the data.
TRANSACTIONS AND ROLLBACK SEGMENT BLOCKS
A rollback segment is composed of two or more blocks. In Oracle8, 8 blocks is the minimum rollback segment size,
regardless of minimum extent and optimal settings. The first block of the segment is reserved for the undo header with the
first transaction for a new rollback segment starting in the third block. Initially, the second block of a rollback segment is not
used. During normal usage, the rollback segment will extend into this block.

Paper 542


Each block in a rollback segment may be assigned to one and only one transaction. If the transaction undo information will
not fit within a single block, a second block is assigned to the transaction. This second block becomes the new ‘head’ block
for the undo chain. Within the second block, the value of the previous block address is retained. When this occurs, the
transaction table is updated to point to the current ‘head’ block.
Each rollback segment block may contain entries that point to different data/index blocks. Part of the undo entry is the
datablock address of both the data/index segment header and block that stores the previous values.
INTERNALS OF ROLLBACK SEGMENTS
Each rollback segment is composed of two distinct components, the undo header and undo entry. The first block of a
rollback segment is reserved for the undo header. When a rollback segment is created, only the header is created. Initially, the
remaining blocks are allocated to the segment, but no other structures are created.
UNDO HEADER
TRANSACTION TABLE
One of the main features in the rollback segment header is the Transaction Table. The transaction table is made up of a fixed
number of ‘slots’ or entries. The number of slots is block size dependent. Each of these slots is assigned to a transaction and
contains information about that action. The slots are initially used in order, but are reused in a round robin fashion with one
exception. A slot referencing an uncommitted transaction will not be reused. It is possible to fill up all slots with active
transactions. If this occurs, the transaction waits until a slot becomes available.
Key components are the state, scn, and dba. A state of 10 indicates that the transaction is not committed. The SCN
represents the System Change Number for uncommitted transactions and System Commit Number for committed
transactions. The dba is the data block address of the last data block containing undo entries for the transaction. (Please note
that parent txid has been removed for display purposes)
TRN TBL::
index state cflags wrap# uel scn dba nub
--------------------------------------------------------------------------
0x00 10 0xc0 0x0010 0x0000 0x0000.00075fec 0x02000004 0x00000001
0x01 9 0x00 0x000f 0x0002 0x0000.00075f77 0x00000000 0x00000000
U NDO ENTRIES
The undo entry for a transaction contains certain transaction level information and the contents of the column or columns
before the change occurred. It does not contain a before image of the block or row.
Undo entries are chained together to form a single transaction undo action. A single undo chain contains the necessary
information to transform a block from one coherent state to another coherent state. If the transaction contains updates to
multiple blocks, each block has its own chain, though the chain may share space within a rollback segment block with other
chains and a single chain may span multiple rollback segment blocks. The undo entry contains either a pointer to the previous
link in the chain or an indicator that this is the end of the chain or an indicator that this is the start of the chain. If the undo
entry is the tail of the chain, it may contain information about a previous transaction that has modified the block.
U PDATE A ROW WITH A SINGLE COLUMN
The undo entry for a row is the previous value of the column and the operation necessary to undo the action. The opposite
of an update is an update to the previous value. The example below is an undo entry for a single column update. Note that
the entire row is not copied, just the old column values and sufficient information to undo the action.
*-----------------------------
* Rec #0x3 slt: 0x00 objn: 22101(0x00005655) objd: 22101 tblspc: 6(0x00000006)
* Layer: 11 (Row) opc: 1 rci 0x02
Undo type: Regular undo Last buffer split: No
Temp Object: No
Tablespace Undo: No
rdba: 0x00000000

Paper 542


*-----------------------------
KDO undo record:
KTB Redo
op: 0x02 ver: 0x01
op: C uba: 0x02000004.0005.02
KDO Op code: URP xtype: XA bdba: 0x01c00007 hdba: 0x01c00006
itli: 1 ispac: 0 maxfr: 4863
tabn: 0 slot: 1(0x1) flag: 0x2c lock: 0 ckix: 0
ncol: 8 nnew: 1 size: 0
col 5: [ 3] c2 1d 33
U PDATE A ROW WITH MULTIPLE COLUMNS
The undo entry for a row is the previous values of the columns.
*-----------------------------
* Rec #0xb slt: 0x01 objn: 22101(0x00005655) objd: 22101 tblspc: 6(0x00000006)
* Layer: 11 (Row) opc: 1 rci 0x0a
Undo type: Regular undo Last buffer split: No
Temp Object: No
Tablespace Undo: No
rdba: 0x00000000
*-----------------------------
KDO undo record:
KTB Redo
op: 0x02 ver: 0x01
op: C uba: 0x02000004.0005.0a
KDO Op code: URP xtype: XA bdba: 0x01c00007 hdba: 0x01c00006
tabn: 0 slot: 1(0x1) flag: 0x2c lock: 0 ckix: 0
ncol: 8 nnew: 2 size: -1
col 5: [ 3] c2 20 24
col 7: [ 2] c1 1f
INSERT A ROW
The undo entry for a row is the row information and the operation necessary to undo the action. The opposite of an insert is
a delete.
*-----------------------------
Undo type: Regular undo Begin trans Last buffer split: No
Temp Object: No
Tablespace Undo: No
rdba: 0x00000000
*-----------------------------
uba: 0x02000004.0005.0a ctl max scn: 0x0000.00075f77 prv tx scn: 0x0000.00075f77
KDO undo record:
KTB Redo
op: 0x04 ver: 0x01
op: L itl: scn: 0x0003.001.00000010 uba: 0x02000004.0005.11
flg: C--- lkc: 0 scn: 0x0000.00075ff0
KDO Op code: DRP xtype: XA bdba: 0x01c00007 hdba: 0x01c00006
tabn: 0 slot: 8(0x8)

Paper 542


DELETE A ROW
The undo entry for a row is the row information, all column values and the operation necessary to undo the action. The
opposite of a delete is an insert.
*-----------------------------
Undo type: Regular undo Begin trans Last buffer split: No
Temp Object: No
Tablespace Undo: No
rdba: 0x00000000
*-----------------------------
uba: 0x02000004.0005.12 ctl max scn: 0x0000.00075f77 prv tx scn: 0x0000.00075f77
KDO undo record:
KTB Redo
op: 0x04 ver: 0x01
op: L itl: scn: 0x0003.002.00000010 uba: 0x02000004.0005.12
flg: C--- lkc: 0 scn: 0x0000.00075ff2
KDO Op code: IRP xtype: XA bdba: 0x01c00007 hdba: 0x01c00006
tabn: 0 slot: 8(0x8) size/delt: 33
fb: --H-FL-- lb: 0x0 cc: 8
null: ---N--N-
col 0: [ 3] c2 64 64
col 1: [ 4] 54 45 53 54
col 2: [ 4] 54 45 53 54
col 3: *NULL*
col 4: [ 7] 78 65 01 03 0f 12 35
col 5: [ 2] c2 0b
col 6: *NULL*
col 7: [ 2] c1 0b
D ATA BLOCK AND ROW STRUCTURE
For each table, there are two types of segment blocks. The first is the segment header, which contains information related to
the segment as a whole, primarily concerned with space management. The second type is the data/index block (for discussion
purposes, we will refer to this as a data block for the article). Each data block contains a header, footer and row information.
Each row contains header information and the actual values of the columns.
BLOCK HEADER
In each block header is a structure called the Interested Transaction List (ITL). The number of initial and maximum entries in
the ITL is determined by the initrans and maxtrans parameters. (For display purposes, all ITL components are not
shown).
Itl Xid Uba Flag
0x01 xid: 0x0003.000.00000010 uba: 0x02000004.0005.09 ----
The XID is the pointer to the rollback segment transaction table slot (undoseg#.slot#.wrap). The undoseg# is the rollback
segment, the slot# refers to a slot in the transaction table, and wrap refers to the number of times that the slot has been used
(i.e. incarnation or version) in a transaction.
The UBA is the pointer to the tail of the undo chain for the block (undodatablockaddr.sequence#.record#)
The FLAG indicates the status of transaction.
ROW HEADER
When a row is updated, one byte in the row header, called the lock byte (lb), is updated to indicate the transaction’s ITL entry.
If the value is 0x0, there is not an ITL entry.
Paper 542


tab 0, row 0, @0x1e4c
tl: 44 fb: --H-FL-- lb: 0x1 cc: 8
col 0: [ 3] c2 4b 64
col 1: [ 5] 41 4c 4c 45 4e
col 2: [ 8] 53 41 4c 45 53 4d 41 4e
col 3: [ 3] c2 4d 63
col 4: [ 7] 77 b5 02 14 01 01 01
col 5: [ 3] c2 12 3d
col 6: [ 2] c2 04
col 7: [ 2] c1 1f
BLOCK CLEANOUT
When a data block is altered, important transaction information is written to the block along with the data changes. When the
transaction is committed, the block is not rewritten to reflect the new transaction state. The next transaction, whether a
SELECT or INSERT/UPDATE/DELETE, will ‘discover’ the out of date information and update the block to reflect the
committed state of the previous transaction. This is called ‘delayed block cleanout.’ It is possible that the block will never be
touched again and contain transaction information that is days, weeks, months or even years old.
INTERNALS OF THE READ CONSISTENT PROCESS
One of the hallmarks of the Oracle database model is the ability to create a view of the data consistent with the start of the
query, regardless of the changes that have been made since the query began. If the data cannot be reconstructed, the query will
fail rather than return dirty data. The mechanism to create a read consistent view, especially accounting for multiple
transactions to a single piece of data, is rather complex. Unfortunately, this is also a component that is lightly documented. The
information contained below is based upon tests and attempts at logical recreation of the process. As such, it may not be
entirely correct, but it does explain all the scenarios that I tested, both single and multiple block and transaction changes.
CREATING A READ CONSISTENT VIEW
STEPS
1. Read the Data Block.
2. Read the Row Header.
3. Check the LockByte to determine if there is an ITL entry.
4. Read the ITL to determine the Transaction ID.
5. Read the Transaction Table. If the transaction has been committed and has a System Commit Number less than the
query’s System Change Number, cleanout the block and move on the next data block (if required) and Step 1.
6. Read the last undo block indicated
7. Compare the block transaction id with the transaction table transaction id. If the Transaction ID in the undo block
does not equal the Transaction ID from the Transaction Table, then signal an ORA-1650, Snapshot Too Old.
8. It the txids are identical, clone the data block in memory. Starting with the head undo entry, apply the changes to the
cloned block.
9. If the tail undo entry (the last one read) indicates another data block address, read the indicated undo block into
memory. Repeat 7 & 8 until the first record does not contain a value for the data block address.
10. When there is no previous data block address, the transaction has been undone.
11. If the undo entry contains
a) a pointer to a previous transaction undo block address, read the TxID in the previous transaction undo block
header and read the appropriate Transaction Table entry. Return to step 5.
b) an ITL record, restore the ITL record to the data block. Return to step 4.

Paper 542


ORACLE 9i AUTOMATIC UNDO MANAGEMENT
Oracle 9i has introduced a new method of managing rollback segments, referred to as automatic undo management. Instead
of manually creating rollback segments, an undo tablespace is created, which contains system generated rollback segments.
When the undo_tablespace parameter is set to the new tablespace, the rollback segments are automatically brought online
and are available for use. In testing the new Automatic Undo Management feature, I have found that the basic architecture of
undo has remained the same. The main changes are related to space management and transaction allocation.
The focus of this paper is not to rehash what is available in the available documentation or on the support web sites. The
focus is to review the actual architecture, highlight undocumented (or poorly documented) features and tasks. Much of the
architectural discussions are based upon testing, block dumps and theory. When in doubt, Ockham's Razor (essentially the
simplest explanation is the best) was applied. I will also make unsupported conjecture about performance problems. By the
time this paper is presented, I plan to have supporting or contradictory data regarding these problems in the actual
presentation.
9i UNDO A RCHITECTURE
The basic structure of the undo segment remains the same from 8i to 9i. Blocks, extents, segments, transaction tables, slots,
undo entries are substantially unchanged. Internally, several items in the block structure of the undo segment have changed.
The segment header now contains a retention table. In testing, I found that this table was populated for the undo tablespace
that was created with the database and not for undo tablespaces created and used afterwards. At this point in time, the testing
is incomplete so an explanation is not forthcoming. It is interesting to note that the undo segments that had no times in the
retention table also had no times in dba_undo_extents.commit_wtime. Another new component is the stmt_num column in
the transaction table. In testing, this column always contained 0x00000000. I am not certain what the future holds for this
column.
CREATION OF U NDO TABLESPACE
The undo tablespace is created as a normal tablespace, but the only method for extent management is 'LOCAL
AUTOALLOCATE'. Dictionary management and Local – Uniform Size is not supported at this time. Attempting to create the
undo tablespace in any other manner results in the "ORA-30024: Invalid specification for CREATE UNDO
TABLESPACE" error. Oracle documentation indicates that LOCAL UNIFORM is a valid extent management algorithm for
UNDO tablespaces. In practice, the only method of extent allocation that does not return an error is LOCAL
AUTOALLOCATE, which is the default.
The tablespace can be created in two ways. First, the CREATE DATABASE command can include the undo tablespace clause.
This method creates the undo tablespace and associated datafile(s), creates the system managed undo segments and sets the
undo_tablespace parameter to the value of the undo tablespace. One advantage is that the system starts running in automatic
undo management mode and no manual intervention by the DBA is required.
The second method is to issue the CREATE UNDO TABLESPACE command when the database is open. This creates a new
undo tablespace with undo segments, which are initially offline. They can be brought online for usage by setting the
undo_tablespace parameter to the new undo tablespace name. This can be done without bouncing the database by issuing the
alter system set undo_tablespace = NEW_UNDO_TS command. Of course, you can always use the tried and true
method of changing the init.ora and restarting the instance (unless you are using the spfile feature).
In testing, I found that the database used undo segments in a similar fashion, regardless of the method of creation. However,
only those undo segments generated with the CREATE DATABASE command would populate the retention table and
appropriate data dictionary views. Dropping the original undo tablespace and stopping/starting the instance did not have any
impact on these data fields. This anomaly is still under investigation.
CREATION OF U NDO SEGMENTS
When the undo tablespace is created, a fixed number of undo segments are created. Based upon testing, each undo
tablespace will contain 10 undo segments. These are sized according to the autoallocate algorithm for locally managed
tablespaces. The basic algorithm is that the first 16 extents are 64k in size. The subsequent allocation method is the next 63
extents of 1m, the next 120 extents of 8m and all additional extents at 64m.

Paper 542


The first block of the first undo extent is reserved as the header and not indicated in the data dictionary views. As with Oracle
8, the second block of the undo segment is not used initially and may or may not be allocated at a later time. This is one of
those anomalies that has no consistent pattern and is therefore hard to explain.
U SAGE OF U NDO SEGMENTS
Once the undo tablespace has been created it is available for use by any instance attached to the database. The current undo
tablespace is determined by the undo_tablespace parameter. Additional undo tablespaces may be created, but will not be used
while another undo tablespace is active. In the dba_rollback_segs view, these undo segments have a status of 'OFFLINE'.
A new undo tablespace can be activated while the database is running. Any transactions that have allocated space in the old
undo tablespace are allowed to gracefully complete. All new transactions are assigned to undo segments in the new tablespace.
Once the pending transactions have completed, the undo segments are no longer available for use.
The number of undo segments online for the instance is a function of the SESSIONS parameter. The algorithm is roughly 1
undo segment for each 5 sessions. All 10 undo segments are placed online when the sessions parameter is set to 46 or greater.
The lower limit was not tested, as the minimum value for the sessions parameter in the test database was 16, which was
derived from the minimum number of processes (10) for the database.
9i UNDO A DMINISTRATION
Administration of automatic undo is 'simpler' in 9i. No longer do you need to issue the commands to create or alter rollback
segments. The creation of the undo tablespace and the assignment of the undo tablespace to the instance are the only steps
required. However, the lack of configuration options are bound to cause performance problems.
AUTOMATIC UNDO P ARAMETERS
UNDO_ MANAGEMENT
Setting this parameter to AUTO invokes the automatic undo management algorithms. In order to allow for transition, MANUAL
undo management is still available. In this case, the rollback segments are created and managed in the same fashion as
Oracle8.

UNDO_ TABLESPACE
The undo tablespace that is currently used by the instance is determined by this parameter. Only 1 undo tablespace can be
active at the instance level at any given time. This parameter is dynamic and can be set at the system level without restarting the
instance. In the Real Application Cluster (RAC) configuration, each instance can have it's own undo tablespace active.
One of the 'features' that Oracle is mentioning is the ability to have different undo tablespaces for different processing needs.
For example, one undo tablespace could be used for normal transaction processing, one for batch processing, one for large
data loads, etc. There are two significant issues that contradict this 'feature'. First, at this time, only the autoallocate algorithm
can be used for the undo tablespace. This guarantees that all undo segments, regardless of the tablespace, will have the exact
same storage characteristics. The second issue is that only one undo tablespace can be associated with an instance at any given
time. The old method of creating a large_rbs, onlining it and allocating it for a single transaction are not permissible in the
new architecture.
UNDO_ RETENTION
This parameter controls the number of seconds that UNDO is retained. In actual practice, this is the minimum amount of
time to be retained. Depending upon the activity in the database and the number/size of undo segments, undo from
committed transactions may be retained for a longer period of time. It is important to remember that the undo is retained
across instance shut downs, which is key to the concept of a flashback query.
Oracle does not guarantee that undo will be retained for any specific amount of time once a transaction is committed. Yup, you read it right. If
the system needs space for active undo and there is no free space available, it will overwrite committed undo, even if the
amount of time since the commit is less than the undo_retention value. This could cause a flashback query to fail. Remember,
Oracle never overwrites uncommitted undo.

UNDO_ SUPPRESS_ ERRORS
Paper 542


If an application contains code relating to rollback segments, this tells the system to ignore the errors that are returned by these
calls. For example, issuing the set transaction use rollback segment command results in "ORA-30019: Illegal
rollback Segment operation in Automatic Undo mode" error. If the parameter is set to FALSE, the command
appears to succeed and no error is returned.
CREATION OF ROLLBACK SEGMENTS
The rollback segments are automatically created when the undo tablespace is created. These are named using the
_SYSSMU<nn>$ algorithm. Because of the new name, some commands (alter system dump undo header) require
that the undo segment name be enclosed in double quotes. The segment space allocation is dependent upon the space
allocation method of the undo tablespace. The number of segments within a given tablespace is a function of the sessions
parameter. Based upon testing, the ratio of sessions to tablespace is roughly 5 to 1 up to a sessions value of 46. In reality, very
few databases will be running with such a small value for sessions.
N EW 9i UNDO D ICTIONARY AND PERFORMANCE VIEWS
DBA_UNDO_EXTENTS
This is a new view for the undo extents. This view will show only the system managed undo segments, the SYSTEM and any
other rollback segments are not included. Most of the columns are similar to the other segment/extent views, with 2
exceptions. First, the block_id for each segment's extent_id of 0 is off by 1 block. For example, in dba_rollback_segs, the
block_id column contains the first block (segment header) of the first extent (extent_id of 0). In dba_undo_extents, the
block_id column contains the second block (skips the segment header) of the first extent.
There is an error in the Oracle Documentation. The STATUS column is not mentioned. This column will contain the values
'ACTIVE' or 'EXPRIRED'. According to the view definition, 'UNEXPIRED' is another value, though in testing, this value
was never encountered.
V$UNDOSTAT
This is a new view for Oracle9i. This contains statistical information as to the performance of the undo subsystem for the life
of the instance. Although this is oriented towards the statistics related to the new mechanisms, the information will be useful
for standard rollback segment monitoring. As with all V$ views, once the instance is stopped all of the data is lost.
9i UNDO A DMINISTRATION CHALLENGES
There are three areas that may pose challenges to query performance. First, space allocation and management is fixed at the
tablespace level. Second, only one undo tablespace can be online at any one time at the instance level. Third, undo_retention
determination will be guesswork until the actual requirements are known. Essentially, the ability of the DBA to address undo
related performance problems is significantly diminished. Will this be a severe problem…only time will tell.
SPACE ALLOCATION AND MANAGEMENT
When an undo tablespace is created, a fixed number (10 in 9.0.1) of rollback segments are created within the tablespace. Since
the only extent management algorithm is local autoallocate, each of the segments is composed of 2 extents of 16k each in size.
As the segment grows, the size of the extent will increase according to the algorithm. When the segment allocates it's 16th
extent, the new size is 1024 k (1 meg).
With fixed space allocation, the flexibility to use a few large rollback segments, many small rollback segments or a mix with
transaction level allocation is gone. The small size and gradual growth of the segment could cause excessive space
management operations and risk for statement failure. Multipurpose databases may encounter both situations.
An additional factor to consider is the balancing act with block size. A larger block means more transaction table slots, so
more simultaneous transactions can be active at any given time. However, since only one transaction can use a given block at
any time, there may be significant waste of space within the undo segments. This seems to present a choice between undo
header contention and undo block contention with the strong possibility of statement failure due to an inability to allocate
undo resource.

Paper 542


ONE AND ONLY ONE UNDO TABLESPACE
With only one undo tablespace online, all users will be accessing blocks in the same file or set of files. Physical I/O contention
becomes a real possibility. Even with multiple data files for the undo tablespace, the undo segment headers and initial space
allocation are in the first data file.
UNDO_RETENTION DETERMINATION
The system now has the ability to retain transaction information for a number of seconds since the transaction was
committed. The number of seconds required for the database will initially be a guesstimate based on past experience. The
effect of UNDO_RETENTION on the space requirements will be determined based upon the performance of the system.
Oracle documentation notes that UNDO_POOL can be set at the resource level to prevent 'runaway' undo usage.
An interesting note recently uncovered during research indicated that the UNDO_RETENTION value is not a guarantee of
preservation. If the system needs to reclaim the space for another transaction, it will overwrite old undo entries even if they
are not 'expired'.
CONCLUSIONS
The exploration of the internal architecture and mechanics of the undo process has been at once both enlightening and
frustrating. Incorrect documentation, contrary results and anomalies have provided ongoing challenges to the determined
DBA mind. Oracle's ability to generate read consistent queries is a marvel in complicated simplicity.
The lack of configurability and flexibility in Automatic Undo Management are a cause of concern and impetus for more
research for this production DBA. The current level of documentation and research is sparse. As more systems are converted
to Oracle9, more information will be forthcoming. As long as there is the option to return to the standby mode of manual
undo management, databases can be converted from one method to another to prevent performance degradation.

Paper 542

Undo internals paper

More Related Content

What's hot (20)

Viewers also liked (19)

Similar to Undo internals paper (20)

More from oracle documents (20)

Undo internals paper