Transaction

Internals
Julian Dyke
Independent Consultant

Web Version

1 © 2007 Julian Dyke

 Agenda

 Transactions
 Redo
 Undo
 Rollback
 Read Consistency
 Undo-based Features
 ORA_ROWSCN
 Flashback

2 © 2007 Julian Dyke

 Examples
 All examples in this presentation are based on cricket

 The following table has been used in all examples in this

presentation

SCORE
TEAM VARCHAR2(30)
RUNS NUMBER
WICKETS NUMBER

 The table has no indexes

3 © 2007 Julian Dyke

 Transactions
 A transaction is a set of DML statements executed
sequentially by a session

 Starts with the first of the following statements executed by

the session:

 INSERT
 UPDATE
 DELETE
 MERGE
 SELECT FOR UPDATE
 LOCK TABLE

 Ends with either a COMMIT or ROLLBACK

4 © 2007 Julian Dyke

 Transactions
 ACID properties
 Atomicity - all changes made by the transaction are either
committed or rolled back
 Consistency - the database is transformed from one valid
state to another
 Isolation - results of the transaction are invisible to other
transactions until the transaction is complete
 Durability - once the transaction completes, the results of
the transaction are permanent

 In Oracle transactions can also be:

 recursive
 audit
 autonomous

5 © 2007 Julian Dyke

 Redo
 All database changes generate redo
 Records changes made to
 Data and index segments
 Undo segments
 Data dictionary
 Control files (indirectly)

 Redo is used:
 During recovery of database
 Instance recovery
 Media recovery

6 © 2007 Julian Dyke

 Undo
 Ensures ACID properties are maintained for each transaction

 Contains changes required to reverse redo including:

 changes to data and index blocks
 changes to transaction lists
 changes to undo blocks

 All undo operations generate redo

 Not all redo operations generate undo

 Implemented using undo segments

 Manually-managed (rollback segments)
 System-managed (Oracle 9.0.1 and above)

7 © 2007 Julian Dyke

 Undo
 Used to rollback uncommitted transactions
 By session issuing ROLLBACK statement
 By PMON on behalf of failed session
 During instance recovery
 During media recovery

 Used to implement read-consistency

 Uncommitted changes cannot be seen by other sessions

 Used to implement flashback

 Oracle 9.0.1 and above

8 © 2007 Julian Dyke

 Redo and Undo
5.2 slot 22: 10 COMMIT;
UPDATE score
block 42 slot 1 SET runs
wickets
= 77,
75
= 7wickets = 8 slot 22: 10
9
5.1 WHERE team = 'AUS';
col1: 74
block 42 slot 1
11.5 block 42 slot 0
col1: 75 Undo Header
col0: ENG
block 42 slot 1
5.1 col1: 841 block 42 slot 1
col2: 6 5.1
col2: 3 col1: 74
block 42 slot 1
11.5 block 42 slot 1
col2: 7 block 42 slot 1
col0: AUS 5.1
block 42 slot 1
2 col2: 6
col1: 77
75
74
5.1 col2: 75
col1: block 42 slot 1
2
col2: 8
6
7
col3: 7
col2: 5.1 col2: 75
col1:
block 42 slot 1 Data Block col3: 7
col2:
11.5 col1: 77
col2: 8

5.4 slot 22: 9

Log Buffer Undo Block

9 STOP
© 2007 Julian Dyke 5.1
 Rollback
5.2 slot 22: 10 ROLLBACK;
UPDATE score
block 42 slot 1 SET wickets = 9
10 slot 22: 10
9
5.1 WHERE team = 'AUS';
col2: 8
block 42 slot 1
11.5 block 42 slot 0
col2: 9 Undo Header
col0: ENG
block 42 slot 1
5.1 col1: 841 block 42 slot 2
col2: 9 5.1
col2: 3 col3: 8
block 42 slot 1
11.5 block 42 slot 1
col2: 10 block 42 slot 2
col0: AUS 5.1
block 42 slot 1 col3: 9
11.5 col1: 77
col2: 9
col2: 8
10
9
5.6 slot 22: 10

block 42 slot 1 Data Block

11.5
col2: 8
5.11 slot 22: 10

5.4 slot 22: 9

Log Buffer Undo Block

1 STOP
© 2007 Julian Dyke
 Undo Segment Header
 Undo segments are allocated at instance startup
 Undo segments can be added dynamically

 Each undo segment header contains

 Pool of free undo extents
 Set of undo slots

 One undo slot is allocated to each transaction

 Undo slot contains list of undo extents
 Extents can migrate from one undo segment to another
 Undo slots are used cyclically
 remain in header as long as possible
 reduces probability of ORA-01555: Snapshot too old

11 © 2007 Julian Dyke

 Undo Segment Header Structure
Block Header

Extent Control Header

Extent Map

Retention Table

Free Block Pool

Transaction Table

Block Tail
KTU SMU HEADER BLOCK

1 © 2007 Julian Dyke

 Transaction ID (XID)
 Every transaction has a unique ID based on
 Undo segment number
 Undo segment slot number
 Undo segment sequence number (wrap)

 A transaction ID (XID) is allocated to each transaction during

the first DML statement. For example:
 0002.028.000004DA

 Details about transaction can be found in V$TRANSACTION

 XIDUSN Segment number
 XIDSLOT Slot number
 XIDSQN Sequence number

1 © 2007 Julian Dyke

 Transaction ID (XID) Undo Segment 3
Slot Status Wrap# SCN DBA
01 10 4D0 134 8004ea
Undo Undo02 10 4CF Undo 137 8004e1
Undo
Segment 03 9
Segment 4D0Segment
138 8004ef
Segment
1 2 04 10 9 4C8
4D0 127
3 139 800fb4 4
8004f0
05 9 4CF 128 800fd5
06 9 4CE 129 800fc8
07 9 4CF 130 800fd0
08 9 4CF 131 800fd1

Session Session Session

1 2 3
UPDATE bowling UPDATE batting UPDATE score
SET overs = 4 SET runs = 25 SET runs = 80
WHERE bowler = 7 WHERE batsman = 8 WHERE team = 'AUS';

XID 0003.004.000004d0 XID 0004.007.00000498 XID 0001.023.000004C8

1 STOP
© 2007 Julian Dyke
 Undo Extents
 Each undo extent contains contiguous set of undo blocks

 Each undo block can only be allocated to one transaction

 Undo blocks contain

 Undo block header
 Undo records

1 © 2007 Julian Dyke

 Undo Block Structure
Block Header

Undo Block Header

Undo Record Index
Unused Space

Undo Records

Block Tail
KTU UNDO BLOCK

1 © 2007 Julian Dyke

 Undo Block
 Undo Block Header contains
 Transaction ID (XID) for current / last transaction to use
block
 Sequence number of undo block
 Number of undo records in undo block
 Not necessarily in current transaction

 Undo records are chained together

 Allow transaction to be rolled back

 Undo records are also used cyclically

 remain in block for as long as possible
 reduces probability of ORA-01555: Snapshot too old

1 © 2007 Julian Dyke

 Undo Byte Address (UBA)
 Specifies address of undo record (not just the undo block)

 Contains
 DBA of undo block
 Sequence number of undo block
 Record number in undo block

 For example: 0x008004f1.0527.1f

 Most recent UBA for transaction reported in V$TRANSACTION

 UBAFIL, UBABLK - file and block number
 UBASQN - sequence number
 UBAREC - record number

1 © 2007 Julian Dyke

 Undo Change Vectors - Data Blocks
 For data blocks
INSERT UPDATE DELETE
INSERT INTO score UPDATE score DELETE FROM score
(team, runs, wickets) SET WHERE team = 'AUS';
VALUES runs = 104,
('AUS',100,4); wickets = 5
WHERE team = 'AUS';
Undo OP 5.1 (11.1) OP 5.1 (11.1) OP 5.1 (11.1)
Delete Row Piece - DRP Update Row Piece - URP Insert Row Piece - IRP
Slot 4: Slot 4: Slot 4:
OP 11.2 c1: 100 c0: 'AUS'
Insert Row Piece - IRP c2: 4 c1: 104
OP 11.5 c2: 5
Redo Slot 4:
Update Row Piece - URP OP 11.3
c0: 'AUS'
Slot 4: Delete Row Piece - DRP
c1: 100
c2: 4 c1: 104 Slot 4:
c2: 5

1 STOP
© 2007 Julian Dyke
 Undo Change Vectors - Index Blocks
 Assume unique index on SCORE (TEAM)
INSERT UPDATE DELETE
INSERT INTO score UPDATE score DELETE FROM score
(team, runs, wickets) SET team = 'ENG' WHERE team = 'ENG';
VALUES WHERE team = 'AUS';
('AUS',100,4);
Undo
OP 5.1 (10.22) OP 5.1 (10.22) OP 5.1 (10.22)
Purge Leaf Row Restore leaf row Restore leaf row
key: 'AUS' key: 'AUS' key: 'ENG'
OP 10.2 data: 01 00 20 47 00 00 data: 01 00 20 47 00 00
Redo
Insert Leaf Row OP 10.4 OP 10.4
key: 'AUS' Delete leaf row Delete leaf row
data: 01 00 20 47 00 00 slot: 0 slot: 0
OP 5.1 (10.22)
Purge leaf row
key: 'ENG'
OP 10.2
Insert Leaf Row
key: 'ENG'
data: 01 00 20 47 00 00

2 STOP
© 2007 Julian Dyke
 SELECT FOR UPDATE
 Redo and Undo Generation

SELECT runs, wickets

FROM score
WHERE team = 'AUS'
FOR UPDATE;

Undo OP 5.1 (11.1)

Lock Row - LKR
Slot 4:
Redo OP 11.4
Lock Row - LKR
Slot 4:

2 STOP
© 2007 Julian Dyke
 SELECT FOR UPDATE
 SELECT FOR UPDATE is bad for so many reasons.....

 Rows are locked pessimistically:

 More chance of contention

 Rows could be locked optimistically by any subsequent

UPDATE statement
 Application logic may need modification

 SELECT FOR UPDATE generates:

 Undo - more space in buffer cache, ORA01555 etc
 Redo - increased physical I/O

 SELECT FOR UPDATE statements cannot be batched

 Each requires a separate pair of change vectors

2 © 2007 Julian Dyke

 UPDATE Statements
 Redo and Undo Generation
CREATE OR REPLACE PROCEDURE update_runs
UPDATE
(p_team VARCHAR2,p_runs NUMBER)
IS UPDATE score
l_runs NUMBER; SET
l_wickets NUMBER;
SELECT runs = :b3,
FOR UPDATE wickets = :b2
BEGIN
WHERE team = :b1;
SELECT runs, wickets
SELECT runs, wickets
INTO l_runs, l_wickets OP 5.1 (11.1)
FROM score
FROM score Update Row Piece - URP
WHERE team = :b1
WHERE team = p_team
FOR UPDATE; Slot 4:
FOR UPDATE;
UPDATE test3 c1: 100
Undo OP 5.1 (11.1)
SET c2: 4
Lock Row - LKR
runs = l_runs, OP 11.5
wickets = l_wickets Slot 4: Update Row Piece - URP
WHERE team = p_team; Redo OP 11.4 Slot 4:
END;
Lock Row - LKR
/ c1: 104
Slot 4: c2: 4

2 STOP
© 2007 Julian Dyke
 UPDATE Statements
 UPDATE statements that include unchanged columns

 Advantages
 Reduce parse overhead
 Good on single instance, even better on RAC
 Reduce space required in library cache
 Less chance cursors will be aged out

 Disadvantages
 Increase physical I/O to online redo logs
 Increase number of undo blocks in buffer cache
 Increase probability of ORA-01555

2 © 2007 Julian Dyke

 Data Block Structure
Block Header

Data Header
Interested Transaction List
Table Index

Row Index

Unused Space

Data

Block Tail

2 © 2007 Julian Dyke

 Interested Transaction List
 Each data/index block has an Interested Transaction List
 list of transactions currently active on block
 stored within block header

 Each data/index row header contains a lock byte

 Lock byte records current slot in ITL

 Each row can only be associated with one transaction

 If a second transaction attempts to update a row it will
experience a row lock waits until first transaction commits/
rolls back

 Initially two ITL entries are reserved in block header

 ITL list can grow dynamically according to demand
 ITL list cannot shrink again

2 © 2007 Julian Dyke

 Interested Transaction List
 ITL entry includes
 Transaction ID (XID)
 Undo byte address (UBA)
 System Change Number (SCN)

 ITL entry is overwritten by each change to the block by the

current transaction

 Previous change is recorded in undo block

 During rollback, changes are restored to ITL from undo block

2 © 2007 Julian Dyke

 Read Consistency
 Required to maintain ACID properties of transaction
 Transactions must always see consistent versions of
blocks modified by other transactions
 Can be applied at
 Statement level (default)
 Transaction level
 Uncommitted block updates are rolled back when block is
read
 Consistent reads
 More specifically undo is applied to return block to
consistent state
 Undo must still be available in undo segment
 If undo has been overwritten, querying session will
receive ORA-01555: Snapshot too old

2 © 2007 Julian Dyke

 Read Consistency
Sessionscore
Australia
Australia 2
score 2A current (dirty) Session
version
Session 1
2 wants to check segment 4 slot 22:
another
runs. The4
single
runs
score
SELECT runs, wickets of block 42 is already
the
UPDATE score in
current Australia Update
state: 10undo header
The score must be the buffer cache, Update undo header
must be updated SET runs so = 84;
88;
89 score to point4D9
wrap#: at last undo
Undo is applied based on
updated torecord
point at last undo
FROM score again session1 makes WHERE a team
copy = 'AUS';
Session 1 has not yet dba: 008002DA (16)
uba records in ITL slot record (18)
(17)
WHERE team = 'AUS'; into a free buffer committed so Session 2
First apply undo record 18 Undo Header
Get nextmustundo perform a (4)
segment
ITL 1: ITLReserve
1: consistent
next slot read
(22 = 0x16) seq: 052C
Apply irb:record
undo 15
16
17
18 16
xid: 0004.016.000004D9 Get
xid:next undo block (0x008002DA) xid: 0004.016.000004D9
0004.016.000004D9
Wrap number denotes current
uba: 008002DA.052C.18
-008002DA.052C.16
008002DA.052C.17 uba: 008002DA.052C.18
008002DA.052C.16
008002DA.052C.17
version ofITL.
undo header
Get first Update
Update uba
uba inin ITL 16 uba: -
available ITL in to End
undo of record
uba list17
16 Apply undo record 17
blockRead
42 slot 1 lock Set
0 blockITL block 42 slot 1 lock 0undo block into5.1 block 42 slot 2
data data
block
transaction Read buffer
col1:42 from disk Set
ENG Set
intoITL ubaID
ITL uba to to
tocol1:
address
address
ENGof of cache col2: 82
current
second
third XID
first undo
undo
undo record
record
record
col2: 841 0x008002DA = col2:
undo 841 dba Set current XID
block 17 in header
uba: 008002DA.052C.16
buffer 0x0004 is undo segment Apply# undo record 18
col3: 3 0x052C = undocol3: block3 seq# 0x016Undo
is slot # (22) block 42 slot 2
5.1
record 16
16 = undo record#
17
18
block 42 slot 2 lock 1 block 42 slot 2 lock0x000004D9
0 First undo
1 is wrap
in TX#socol2: 84
Sequence nonumber
previous (0x52C)
uba008002DA.052C.17
col1: AUS col1: AUS 18
denotes current versionuba: of
89
col2: 82
88
84 col2: 8982Undo record 18 undo block
84
88 block 42 slot 2
Undo5.1 record 17
col3: 9 col3:
Apply
Apply9Third
undoundo
undo toto in TX
Set row
Update runslock
Secondto undo col2: 88
in TX
Previous uba is 17 ITL# (1)
Data Block (42 copy) runs
runs Block (42) column Previous
column
Datacolumn Undo Block uba(dba
is 16008002DA)

2 STOP
© 2007 Julian Dyke
 SET TRANSACTION
 Determines level at which read-consistency is applied
 Can be:
 SET TRANSACTION READ WRITE
 establishes statement-level read consistency
 subsequent statements see any changes committed
before that statement started
 default behaviour
 SET TRANSACTION READ ONLY
 establishes transaction-level read consistency
 all subsequent statements only see changes committed
before transaction started
 not supported for SYS user
 SET TRANSACTION statement must be first statement in
transaction

3 © 2007 Julian Dyke

 SET TRANSACTION
 For example:
Session 1 Session 2 Session 3
SELECT runs
FROM score
WHERE team = 'ENG';
Runs
127

SET TRANSACTION SET TRANSACTION

READ WRITE; READ ONLY;

UPDATE team
SET runs = 131
WHERE team = 'ENG';
COMMIT;

SELECT runs SELECT runs

FROM score FROM score
WHERE team = 'ENG'; WHERE team = 'ENG';
Runs Runs
131 127

3 © 2007 Julian Dyke

 ORA_ROWSCN Pseudocolumn
 Returns conservative upper-bound SCN for most recent
change in row

 Uses SCN stored for transaction in ITL

 Shows last time a row in same block was updated

 May show more accurate information for an individual row

 Not supported with flashback query

 To convert ORA_ROWSCN to an approximate timestamp use

the SCN_TO_TIMESTAMP built-in function e.g.
SELECT ORA_ROWSCN,
SCN_TO_TIMESTAMP (ORA_ROWSCN)
FROM score;

3 © 2007 Julian Dyke

 ORA_ROWSCN Pseudocolumn
 For example - no row dependencies (default) 0x3588ba =
3508410
ITL1:
CREATE TABLE score
XID: 0008.012.000004FA
(team NUMBER, runs NUMBER, wickets NUMBER);
Flag: C---
-U-- Lck: 0
INSERT INTO score (team, runs, wickets) VALUES ('ENG',0,0);
SCN/FSC: 0000.003588ba
INSERT INTO score (teams,runs,wickets) VALUES ('AUS',0,0);
COMMIT; SCN/FSC: 0000.003588ba
ITL2:

SELECT ORA_ROWSCN, teams, runs, wickets FROM score; XID: 0009.008.00000502

ORA_ROWSCN Teams Runs Wickets Flag: --U- Lck: 1

3508410 ENG 0 0 SCN/FSC: 0000.003588bd
3508410 AUS 0 0
Row 0: lb: 2
1
UPDATE score col 0: ENG
SET runs = 4
WHERE team = 'ENG'; col 1: 4
col 2: 0 0x3588bd =
COMMIT;
3508413
Row 1: lb: 0
1
SELECT ORA_ROWSCN, teams, runs, wickets FROM score;
col 0: AUS
ORA_ROWSCN Teams Runs Wickets
3508413 ENG 4 0 col 1: 0
3508413 AUS 0 0
col 2: 0

3 © 2007 Julian Dyke

 ORA_ROWSCN Pseudocolumn
 For example (row dependencies) 0x358ced =
ITL1: 3509485
CREATE TABLE score -U---
Flag: C--- 2
Lck: 0
(team NUMBER, runs NUMBER, wickets NUMBER) SCN/FSC: 0000.00358ced
ROWDEPENDENCIES;
ITL2:
INSERT INTO score (team, runs, wickets) VALUES ('ENG',0,0);
INSERT INTO score (teams,runs,wickets) VALUES ('AUS',0,0); SCN/FSC:
Flag: --U- 0000.003588ba
Lck: 1

COMMIT; SCN/FSC: 0000.00358cf0

SELECT ora_rowscn, teams, runs, wickets FROM score; Row 0: lb: 2

1
ORA_ROWSCN Teams Runs Wickets dscn: 0000.00358cf0
0000.00358ced
3509485 ENG 0 0
3509485 AUS 0 0 col 0: ENG

UPDATE score col 1: 4

0
SET runs = 4 col 2: 0
WHERE team = 'ENG'; 0x358cf0 =
Row 1: lb: 0
1 3509488
COMMIT;
dscn: 0000.00358ced
SELECT ora_rowscn, teams, runs, wickets FROM score;
col 0: AUS
ORA_ROWSCN Teams Runs Wickets
col 1: 0
3509488 ENG 4 0
3509485 AUS 0 0 col 2: 0

3 © 2007 Julian Dyke

 Flashback Query
 Example
Session 1 Session 2
SELECT runs SELECT dbms_flashback.get_system_change_number FROM dual;
FROM score
SCN
WHERE team = 'ENG';
3494824
Runs
137

UPDATE team
SET runs = 141
WHERE team = 'ENG';
COMMIT;

SELECT dbms_flashback.get_system_change_number FROM dual;

SCN
3494833

SELECT team, runs, wickets FROM score

WHERE team = 'ENG';
Team Runs Wickets
ENG 141 1

SELECT team, runs, wickets FROM score AS OF SCN 3494824;

WHERE team = 'ENG';
Team Runs Wickets
ENG 137 1
3 © 2007 Julian Dyke
 Flashback Query
 Can specify AS OF clause:
 Returns single-row
 Syntax is

AS OF [ SCN <scn> | TIMESTAMP <timestamp> ]

 For example:
SELECT team, runs, wickets
FROM score AS OF SCN 3506431
WHERE team = 'ENG';

3 © 2007 Julian Dyke

 Flashback Query
 Can also specify VERSIONS clause:
 Returns multiple rows
 Syntax is

VERSIONS BETWEEN SCN [ <scn> | MINVALUE ]

AND [ <scn> | MAXVALUE

VERSIONS BETWEEN TIMESTAMP [ <timestamp> | MINVALUE ]

AND [ <timestamp> | MAXVALUE

 For example:
SELECT team, runs, wickets
FROM score VERSIONS BETWEEN SCN 3503511 AND 3503524
WHERE team = 'ENG';

3 © 2007 Julian Dyke

 Version Query Pseudocolumns
 Valid only for Flashback Version Query. Values can be:
 VERSIONS_STARTTIME
 timestamp of first version of rows returned by query
 VERSIONS_ENDTIME
 timestamp of last version of rows returned by query
 VERSIONS_STARTSCN
 SCN of first version of rows returned by query
 VERSIONS_ENDSCN
 SCN of last version of rows returned by query
 VERSIONS_XID
 For each row returns transaction ID of transaction
creating that row version
 VERSIONS_OPERATION
 For each row returns operation creating that row
version. Can be I(nsert) U(pdate) or D(elete)

3 © 2007 Julian Dyke

 Version Query Pseudocolumns
 Example:
Session 1 Session 2
SELECT runs SELECT dbms_flashback.get_system_change_number FROM dual;
FROM score
SCN
WHERE team = 'ENG';
3503136
Runs
141

UPDATE team
SET runs = 145
WHERE team = 'ENG';
COMMIT;

UPDATE team
SET runs = 151
WHERE team = 'ENG';
COMMIT;

UPDATE team
SET runs = 153
WHERE team = 'ENG';
COMMIT;
SELECT dbms_flashback.get_system_change_number FROM dual;
SCN
3503143
3 © 2007 Julian Dyke
 Version Query Pseudocolumns
 Example (continued):
Session 1 Session 2

SELECT
VERSIONS_STARTSCN "Start",
VERSIONS_ENDSCN "End",
VERSIONS_XID "XID",
VERSIONS_OPERATION "Op",
score.team "Team",
score.runs "Runs",
score.wickets "Wickets"
FROM score VERSIONS BETWEEN SCN 3503136 AND 3503143
WHERE team = 'ENG';

Start End XID Op Team Runs Wickets

3503142 08000A00FC040000 U ENG 153 1
3503139 3503142 07001A00F6040000 U ENG 151 1
3503136 3503139 06002C00EA040000 U ENG 145 1
3503136 ENG 141 1

XID = 0066.02C.000004EA Can be I(nsert), U(pdate)

(Architecture = X86) or D(elete)

4 © 2007 Julian Dyke

 Thank you for your interest

For more information and to provide feedback

please contact me
My e-mail address is:
[email protected]

My website address is:

www.juliandyke.com

4 © 2007 Julian Dyke