10 Transactions

Sample Transaction: Account Debit and Credit

To illustrate the concept of a transaction, consider a banking database. When a customer transfers money from a savings account to a checking account, the transaction must consist of three separate operations:

• Decrement the savings account

• Increment the checking account

• Record the transaction in the transaction journal

Oracle Database must allow for two situations. If all three SQL statements maintain the accounts in proper balance, then the effects of the transaction can be applied to the database. However, if a problem such as insufficient funds, invalid account number, or a hardware failure prevents one or two of the statements in the transaction from completing, then the database must roll back the entire transaction so that the balance of all accounts is correct.

Figure 10-1 illustrates a banking transaction. The first statement subtracts $500 from savings account 3209. The second statement adds $500 to checking account 3208. The third statement inserts a record of the transfer into the journal table. The final statement commits the transaction.

Figure 10-1 A Banking Transaction

Description of "Figure 10-1 A Banking Transaction"

Structure of a Transaction

A database transaction consists of one or more statements. Specifically, a transaction consists of one of the following:

• One or more data manipulation language (DML) statements that together constitute an atomic change to the database

• One data definition language (DDL) statement

A transaction has a beginning and an end.

SQL> UPDATE hr.employees SET salary=salary; 

107 rows updated.
 
SQL> SELECT XID AS "txn id", XIDUSN AS "undo seg", XIDSLOT AS "slot", 
  2  XIDSQN AS "seq", STATUS AS "txn status"
  3  FROM V$TRANSACTION;
 
txn id             undo seg       slot        seq txn status
---------------- ---------- ---------- ---------- ----------------
0600060037000000          6          6         55 ACTIVE

SQL> UPDATE hr.employees SET salary=salary; 
107 rows updated.
 
SQL> SELECT XID, STATUS FROM V$TRANSACTION;
 
XID              STATUS
---------------- ----------------
0800090033000000 ACTIVE
 
SQL> ROLLBACK;
 
Rollback complete.
 
SQL> SELECT XID FROM V$TRANSACTION;
 
no rows selected
 
SQL> UPDATE hr.employees SET last_name=last_name;
 
107 rows updated.
 
SQL> SELECT XID, STATUS FROM V$TRANSACTION;
 
XID              STATUS
---------------- ----------------
0900050033000000 ACTIVE

Statement-Level Atomicity

Oracle Database supports statement-level atomicity, which means that a SQL statement is an atomic unit of work and either completely succeeds or completely fails.

A successful statement is different from a committed transaction. A single SQL statement executes successfully if the database parses and runs it without error as an atomic unit, as when all rows are changed in a multirow update.

If a SQL statement causes an error during execution, then it is not successful and so all effects of the statement are rolled back. This operation is a statement-level rollback. This operation has the following characteristics:

• A SQL statement that does not succeed causes the loss only of work it would have performed itself.

  The unsuccessful statement does not cause the loss of any work that preceded it in the current transaction. For example, if the execution of the second UPDATE statement in Figure 10-1 causes an error and is rolled back, then the work performed by the first UPDATE statement is not rolled back. The first UPDATE statement can be committed or rolled back explicitly by the user.

• The effect of the rollback is as if the statement had never been run.

  Any side effects of an atomic statement, for example, triggers invoked upon execution of the statement, are considered part of the atomic statement. Either all work generated as part of the atomic statement succeeds or none does.

An example of an error causing a statement-level rollback is an attempt to insert a duplicate primary key. Single SQL statements involved in a deadlock, which is competition for the same data, can also cause a statement-level rollback. However, errors discovered during SQL statement parsing, such as a syntax error, have not yet been run and so do not cause a statement-level rollback.

System Change Numbers (SCNs)

A system change number (SCN) is a logical, internal time stamp used by Oracle Database. SCNs order events that occur within the database, which is necessary to satisfy the ACID properties of a transaction. Oracle Database uses SCNs to mark the SCN before which all changes are known to be on disk so that recovery avoids applying unnecessary redo. The database also uses SCNs to mark the point at which no redo exists for a set of data so that recovery can stop.

SCNs occur in a monotonically increasing sequence. Oracle Database can use an SCN like a clock because an observed SCN indicates a logical point in time and repeated observations return equal or greater values. If one event has a lower SCN than another event, then it occurred at an earlier time with respect to the database. Several events may share the same SCN, which means that they occurred at the same time with respect to the database.

Every transaction has an SCN. For example, if a transaction updates a row, then the database records the SCN at which this update occurred. Other modifications in this transaction have the same SCN. When a transaction commits, the database records an SCN for this commit.

Oracle Database increments SCNs in the system global area (SGA). When a transaction modifies data, the database writes a new SCN to the undo data segment assigned to the transaction. The log writer process then writes the commit record of the transaction immediately to the online redo log. The commit record has the unique SCN of the transaction. Oracle Database also uses SCNs as part of its instance recovery and media recovery mechanisms.

Transaction Names

A transaction name is an optional, user-specified tag that serves as a reminder of the work that the transaction is performing. You name a transaction with the SET TRANSACTION ... NAME statement, which if used must be first statement of the transaction. In Table 10-1, the first transaction was named sal_update and the second was named sal_update2.

Transaction names provide the following advantages:

• It is easier to monitor long-running transactions and to resolve in-doubt distributed transactions.

• You can view transaction names along with transaction IDs in applications. For example, a database administrator can view transaction names in Oracle Enterprise Manager (Enterprise Manager) when monitoring system activity.

• The database writes transaction names to the transaction auditing redo record, so you can use LogMiner to search for a specific transaction in the redo log.

• You can use transaction names to find a specific transaction in data dictionary views such as V$TRANSACTION.

Active Transactions

An active transaction has started but not yet committed or rolled back. In Table 10-1, the first statement to modify data in the sal_update transaction is the update to Banda's salary. From the successful execution of this update until the ROLLBACK statement ends the transaction, the sal_update transaction is active.

Data changes made by a transaction are temporary until the transaction is committed or rolled back. Before the transaction ends, the state of the data is as follows:

• Oracle Database has generated undo data information in the system global area (SGA).

  The undo data contains the old data values changed by the SQL statements of the transaction. See "Read Consistency in the Read Committed Isolation Level".

• Oracle Database has generated redo in the online redo log buffer of the SGA.

  The redo log record contains the change to the data block and the change to the undo block. See "Redo Log Buffer".

• Changes have been made to the database buffers of the SGA.

  The data changes for a committed transaction, stored in the database buffers of the SGA, are not necessarily written immediately to the data files by the database writer (DBWn). The disk write can happen before or after the commit. See "Database Buffer Cache".

• The rows affected by the data change are locked.

  Other users cannot change the data in the affected rows, nor can they see the uncommitted changes. See "Summary of Locking Behavior".

Savepoints

A savepoint is a user-declared intermediate marker within the context of a transaction. Internally, this marker resolves to an SCN. Savepoints divide a long transaction into smaller parts.

If you use savepoints in a long transaction, then you have the option later of rolling back work performed before the current point in the transaction but after a declared savepoint within the transaction. Thus, if you make an error, you do not need to resubmit every statement. Table 10-1 creates savepoint after_banda_sal so that the update to the Greene salary can be rolled back to this savepoint.

UPDATE employees 
 SET salary = 7000 
 WHERE last_name = 
 'Banda';

SAVEPOINT 
 after_banda_sal;

UPDATE employees 
 SET salary = 12000 
 WHERE last_name =  
 'Greene';

UPDATE employees 
 SET salary = 14000 
 WHERE last_name = 
 'Greene';

ROLLBACK
 TO SAVEPOINT 
 after_banda_sal;

UPDATE employees 
 SET salary = 11000 
 WHERE last_name =
 'Greene';

COMMIT;

Rollback of Transactions

A rollback of an uncommitted transaction undoes any changes to data that have been performed by SQL statements within the transaction. After a transaction has been rolled back, the effects of the work done in the transaction no longer exist.

In rolling back an entire transaction, without referencing any savepoints, Oracle Database performs the following actions:

• Undoes all changes made by all the SQL statements in the transaction by using the corresponding undo segments

  The transaction table entry for every active transaction contains a pointer to all the undo data (in reverse order of application) for the transaction. The database reads the data from the undo segment, reverses the operation, and then marks the undo entry as applied. Thus, if a transaction inserts a row, then a rollback deletes it. If a transaction updates a row, then a rollback reverses the update. If a transaction deletes a row, then a rollback reinserts it. In Table 10-1, the ROLLBACK reverses the updates to the salaries of Greene and Banda.

• Releases all the locks of data held by the transaction

• Erases all savepoints in the transaction

  In Table 10-1, the ROLLBACK deletes the savepoint after_banda_sal. The after_greene_sal savepoint was removed by the ROLLBACK TO SAVEPOINT statement.

• Ends the transaction

  In Table 10-1, the ROLLBACK leaves the database in the same state as it was after the initial COMMIT was executed.

The duration of a rollback is a function of the amount of data modified.

Committing Transactions

A commit ends the current transaction and makes permanent all changes performed in the transaction. In Table 10-1, a second transaction begins with sal_update2 and ends with an explicit COMMIT statement. The changes that resulted from the two UPDATE statements are now made permanent.

When a transaction commits, the following actions occur:

• A system change number (SCN) is generated for the COMMIT.

  The internal transaction table for the associated undo tablespace records that the transaction has committed. The corresponding unique SCN of the transaction is assigned and recorded in the transaction table. See "Serializable Isolation Level".

• The log writer (LGWR) process writes remaining redo log entries in the redo log buffers to the online redo log and writes the transaction SCN to the online redo log. This atomic event constitutes the commit of the transaction.

• Oracle Database releases locks held on rows and tables.

  Users who were enqueued waiting on locks held by the uncommitted transaction are allowed to proceed with their work.

• Oracle Database deletes savepoints.

  In Table 10-1, no savepoints existed in the sal_update transaction so no savepoints were erased.

• Oracle Database performs a commit cleanout.

  If modified blocks containing data from the committed transaction are still in the SGA, and if no other session is modifying them, then the database removes lock-related transaction information from the blocks. Ideally, the COMMIT cleans out the blocks so that a subsequent SELECT does not have to perform this task.

  Note:

  Because a block cleanout generates redo, a query may generate redo and thus cause blocks to be written during the next checkpoint.

• Oracle Database marks the transaction complete.

After a transaction commits, users can view the changes.

Typically, a commit is a fast operation, regardless of the transaction size. The speed of a commit does not increase with the size of the data modified in the transaction. The lengthiest part of the commit is the physical disk I/O performed by LGWR. However, the amount of time spent by LGWR is reduced because it has been incrementally writing the contents of the redo log buffer in the background.

The default behavior is for LGWR to write redo to the online redo log synchronously and for transactions to wait for the buffered redo to be on disk before returning a commit to the user. However, for lower transaction commit latency, application developers can specify that redo be written asynchronously so that transactions need not wait for the redo to be on disk and can return from the COMMIT call immediately.

Two-Phase Commit

The two-phase commit mechanism guarantees that all databases participating in a distributed transaction either all commit or all undo the statements in the transaction. The mechanism also protects implicit DML performed by integrity constraints, remote procedure calls, and triggers.

In a two-phase commit among multiple databases, one database coordinates the distributed transaction. The initiating node is called the global coordinator. The coordinator asks the other databases if they are prepared to commit. If any database responds with a no, then the entire transaction is rolled back. If all databases vote yes, then the coordinator broadcasts a message to make the commit permanent on each of the databases.

The two-phase commit mechanism is transparent to users who issue distributed transactions. In fact, users need not even know the transaction is distributed. A COMMIT statement denoting the end of a transaction automatically triggers the two-phase commit mechanism. No coding or complex statement syntax is required to include distributed transactions within the body of a database application.

In-Doubt Transactions

An in-doubt distributed transaction occurs when a two-phase commit was interrupted by any type of system or network failure. For example, two databases report to the coordinating database that they were prepared to commit, but the coordinating database instance fails immediately after receiving the messages. The two databases who are prepared to commit are now left hanging while they await notification of the outcome.

The recoverer (RECO) background process automatically resolves the outcome of in-doubt distributed transactions. After the failure is repaired and communication is reestablished, the RECO process of each local Oracle database automatically commits or rolls back any in-doubt distributed transactions consistently on all involved nodes.

In the event of a long-term failure, Oracle Database enables each local administrator to manually commit or undo any distributed transactions that are in doubt because of the failure. This option enables the local database administrator to free any locked resources that are held indefinitely because of the long-term failure.

If a database must be recovered to a past time, then database recovery facilities enable database administrators at other sites to return their databases to the earlier point in time. This operation ensures that the global database remains consistent.